xref: /dflybsd-src/sys/vfs/hammer2/hammer2_inode.c (revision 25ca8c794df06033ca37cb3fdf9b2992a7b5a9e6)
1 /*
2  * Copyright (c) 2011-2014 The DragonFly Project.  All rights reserved.
3  *
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>
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  *
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
17  *    distribution.
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.
21  *
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
33  * SUCH DAMAGE.
34  */
35 #include <sys/cdefs.h>
36 #include <sys/param.h>
37 #include <sys/systm.h>
38 #include <sys/types.h>
39 #include <sys/lock.h>
40 #include <sys/uuid.h>
41 
42 #include "hammer2.h"
43 
44 #define INODE_DEBUG	0
45 
46 static void hammer2_inode_move_to_hidden(hammer2_trans_t *trans,
47 					 hammer2_cluster_t **cparentp,
48 					 hammer2_cluster_t **clusterp,
49 					 hammer2_tid_t inum);
50 
51 RB_GENERATE2(hammer2_inode_tree, hammer2_inode, rbnode, hammer2_inode_cmp,
52 	     hammer2_tid_t, meta.inum);
53 
54 int
55 hammer2_inode_cmp(hammer2_inode_t *ip1, hammer2_inode_t *ip2)
56 {
57 	if (ip1->meta.inum < ip2->meta.inum)
58 		return(-1);
59 	if (ip1->meta.inum > ip2->meta.inum)
60 		return(1);
61 	return(0);
62 }
63 
64 /*
65  * HAMMER2 inode locks
66  *
67  * HAMMER2 offers shared and exclusive locks on inodes.  Pass a mask of
68  * flags for options:
69  *
70  *	- pass HAMMER2_RESOLVE_SHARED if a shared lock is desired.  The
71  *	  inode locking function will automatically set the RDONLY flag.
72  *
73  *	- pass HAMMER2_RESOLVE_ALWAYS if you need the inode's meta-data.
74  *	  Most front-end inode locks do.
75  *
76  *	- pass HAMMER2_RESOLVE_NEVER if you do not want to require that
77  *	  the inode data be resolved.  This is used by the syncthr because
78  *	  it can run on an unresolved/out-of-sync cluster, and also by the
79  *	  vnode reclamation code to avoid unnecessary I/O (particularly when
80  *	  disposing of hundreds of thousands of cached vnodes).
81  *
82  * The inode locking function locks the inode itself, resolves any stale
83  * chains in the inode's cluster, and allocates a fresh copy of the
84  * cluster with 1 ref and all the underlying chains locked.
85  *
86  * ip->cluster will be stable while the inode is locked.
87  *
88  * NOTE: We don't combine the inode/chain lock because putting away an
89  *       inode would otherwise confuse multiple lock holders of the inode.
90  *
91  * NOTE: In-memory inodes always point to hardlink targets (the actual file),
92  *	 and never point to a hardlink pointer.
93  *
94  * NOTE: If caller passes HAMMER2_RESOLVE_RDONLY the exclusive locking code
95  *	 will feel free to reduce the chain set in the cluster as an
96  *	 optimization.  It will still be validated against the quorum if
97  *	 appropriate, but the optimization might be able to reduce data
98  *	 accesses to one node.  This flag is automatically set if the inode
99  *	 is locked with HAMMER2_RESOLVE_SHARED.
100  */
101 void
102 hammer2_inode_lock(hammer2_inode_t *ip, int how)
103 {
104 	hammer2_inode_ref(ip);
105 
106 	/*
107 	 * Inode structure mutex
108 	 */
109 	if (how & HAMMER2_RESOLVE_SHARED) {
110 		/*how |= HAMMER2_RESOLVE_RDONLY; not used */
111 		hammer2_mtx_sh(&ip->lock);
112 	} else {
113 		hammer2_mtx_ex(&ip->lock);
114 	}
115 }
116 
117 /*
118  * Create a locked copy of ip->cluster.  Note that the copy will have a
119  * ref on the cluster AND its chains and we don't want a second ref to
120  * either when we lock it.
121  *
122  * Exclusive inode locks set the template focus chain in (ip)
123  * as a hint.  Cluster locks can ALWAYS replace the focus in the
124  * working copy if the hint does not work out, so beware.
125  */
126 hammer2_cluster_t *
127 hammer2_inode_cluster(hammer2_inode_t *ip, int how)
128 {
129 	hammer2_cluster_t *cluster;
130 
131 	cluster = hammer2_cluster_copy(&ip->cluster);
132 	hammer2_cluster_lock(cluster, how);
133 	hammer2_cluster_resolve(cluster);
134 
135 	/*
136 	 * cluster->focus will be set if resolving RESOLVE_ALWAYS, but
137 	 * only update the cached focus in the inode structure when taking
138 	 * out an exclusive lock.
139 	 */
140 	if ((how & HAMMER2_RESOLVE_SHARED) == 0)
141 		ip->cluster.focus = cluster->focus;
142 
143 	return cluster;
144 }
145 
146 void
147 hammer2_inode_unlock(hammer2_inode_t *ip, hammer2_cluster_t *cluster)
148 {
149 	if (cluster) {
150 		hammer2_cluster_unlock(cluster);
151 		hammer2_cluster_drop(cluster);
152 	}
153 	hammer2_mtx_unlock(&ip->lock);
154 	hammer2_inode_drop(ip);
155 }
156 
157 /*
158  * Temporarily release a lock held shared or exclusive.  Caller must
159  * hold the lock shared or exclusive on call and lock will be released
160  * on return.
161  *
162  * Restore a lock that was temporarily released.
163  */
164 hammer2_mtx_state_t
165 hammer2_inode_lock_temp_release(hammer2_inode_t *ip)
166 {
167 	return hammer2_mtx_temp_release(&ip->lock);
168 }
169 
170 void
171 hammer2_inode_lock_temp_restore(hammer2_inode_t *ip, hammer2_mtx_state_t ostate)
172 {
173 	hammer2_mtx_temp_restore(&ip->lock, ostate);
174 }
175 
176 /*
177  * Upgrade a shared inode lock to exclusive and return.  If the inode lock
178  * is already held exclusively this is a NOP.
179  *
180  * The caller MUST hold the inode lock either shared or exclusive on call
181  * and will own the lock exclusively on return.
182  *
183  * Returns non-zero if the lock was already exclusive prior to the upgrade.
184  */
185 int
186 hammer2_inode_lock_upgrade(hammer2_inode_t *ip)
187 {
188 	int wasexclusive;
189 
190 	if (mtx_islocked_ex(&ip->lock)) {
191 		wasexclusive = 1;
192 	} else {
193 		hammer2_mtx_unlock(&ip->lock);
194 		hammer2_mtx_ex(&ip->lock);
195 		wasexclusive = 0;
196 	}
197 	return wasexclusive;
198 }
199 
200 /*
201  * Downgrade an inode lock from exclusive to shared only if the inode
202  * lock was previously shared.  If the inode lock was previously exclusive,
203  * this is a NOP.
204  */
205 void
206 hammer2_inode_lock_downgrade(hammer2_inode_t *ip, int wasexclusive)
207 {
208 	if (wasexclusive == 0)
209 		mtx_downgrade(&ip->lock);
210 }
211 
212 /*
213  * Lookup an inode by inode number
214  */
215 hammer2_inode_t *
216 hammer2_inode_lookup(hammer2_pfs_t *pmp, hammer2_tid_t inum)
217 {
218 	hammer2_inode_t *ip;
219 
220 	KKASSERT(pmp);
221 	if (pmp->spmp_hmp) {
222 		ip = NULL;
223 	} else {
224 		hammer2_spin_ex(&pmp->inum_spin);
225 		ip = RB_LOOKUP(hammer2_inode_tree, &pmp->inum_tree, inum);
226 		if (ip)
227 			hammer2_inode_ref(ip);
228 		hammer2_spin_unex(&pmp->inum_spin);
229 	}
230 	return(ip);
231 }
232 
233 /*
234  * Adding a ref to an inode is only legal if the inode already has at least
235  * one ref.
236  *
237  * (can be called with spinlock held)
238  */
239 void
240 hammer2_inode_ref(hammer2_inode_t *ip)
241 {
242 	atomic_add_int(&ip->refs, 1);
243 }
244 
245 /*
246  * Drop an inode reference, freeing the inode when the last reference goes
247  * away.
248  */
249 void
250 hammer2_inode_drop(hammer2_inode_t *ip)
251 {
252 	hammer2_pfs_t *pmp;
253 	hammer2_inode_t *pip;
254 	u_int refs;
255 
256 	while (ip) {
257 		refs = ip->refs;
258 		cpu_ccfence();
259 		if (refs == 1) {
260 			/*
261 			 * Transition to zero, must interlock with
262 			 * the inode inumber lookup tree (if applicable).
263 			 * It should not be possible for anyone to race
264 			 * the transition to 0.
265 			 *
266 			 */
267 			pmp = ip->pmp;
268 			KKASSERT(pmp);
269 			hammer2_spin_ex(&pmp->inum_spin);
270 
271 			if (atomic_cmpset_int(&ip->refs, 1, 0)) {
272 				KKASSERT(hammer2_mtx_refs(&ip->lock) == 0);
273 				if (ip->flags & HAMMER2_INODE_ONRBTREE) {
274 					atomic_clear_int(&ip->flags,
275 						     HAMMER2_INODE_ONRBTREE);
276 					RB_REMOVE(hammer2_inode_tree,
277 						  &pmp->inum_tree, ip);
278 				}
279 				hammer2_spin_unex(&pmp->inum_spin);
280 
281 				pip = ip->pip;
282 				ip->pip = NULL;
283 				ip->pmp = NULL;
284 
285 				/*
286 				 * Cleaning out ip->cluster isn't entirely
287 				 * trivial.
288 				 */
289 				hammer2_inode_repoint(ip, NULL, NULL);
290 
291 				/*
292 				 * We have to drop pip (if non-NULL) to
293 				 * dispose of our implied reference from
294 				 * ip->pip.  We can simply loop on it.
295 				 */
296 				kfree(ip, pmp->minode);
297 				atomic_add_long(&pmp->inmem_inodes, -1);
298 				ip = pip;
299 				/* continue with pip (can be NULL) */
300 			} else {
301 				hammer2_spin_unex(&ip->pmp->inum_spin);
302 			}
303 		} else {
304 			/*
305 			 * Non zero transition
306 			 */
307 			if (atomic_cmpset_int(&ip->refs, refs, refs - 1))
308 				break;
309 		}
310 	}
311 }
312 
313 /*
314  * Get the vnode associated with the given inode, allocating the vnode if
315  * necessary.  The vnode will be returned exclusively locked.
316  *
317  * The caller must lock the inode (shared or exclusive).
318  *
319  * Great care must be taken to avoid deadlocks and vnode acquisition/reclaim
320  * races.
321  */
322 struct vnode *
323 hammer2_igetv(hammer2_inode_t *ip, hammer2_cluster_t *cparent, int *errorp)
324 {
325 	const hammer2_inode_data_t *ripdata;
326 	hammer2_pfs_t *pmp;
327 	struct vnode *vp;
328 
329 	pmp = ip->pmp;
330 	KKASSERT(pmp != NULL);
331 	*errorp = 0;
332 
333 	ripdata = &hammer2_cluster_rdata(cparent)->ipdata;
334 
335 	for (;;) {
336 		/*
337 		 * Attempt to reuse an existing vnode assignment.  It is
338 		 * possible to race a reclaim so the vget() may fail.  The
339 		 * inode must be unlocked during the vget() to avoid a
340 		 * deadlock against a reclaim.
341 		 */
342 		int wasexclusive;
343 
344 		vp = ip->vp;
345 		if (vp) {
346 			/*
347 			 * Inode must be unlocked during the vget() to avoid
348 			 * possible deadlocks, but leave the ip ref intact.
349 			 *
350 			 * vnode is held to prevent destruction during the
351 			 * vget().  The vget() can still fail if we lost
352 			 * a reclaim race on the vnode.
353 			 */
354 			hammer2_mtx_state_t ostate;
355 
356 			vhold(vp);
357 			ostate = hammer2_inode_lock_temp_release(ip);
358 			if (vget(vp, LK_EXCLUSIVE)) {
359 				vdrop(vp);
360 				hammer2_inode_lock_temp_restore(ip, ostate);
361 				continue;
362 			}
363 			hammer2_inode_lock_temp_restore(ip, ostate);
364 			vdrop(vp);
365 			/* vp still locked and ref from vget */
366 			if (ip->vp != vp) {
367 				kprintf("hammer2: igetv race %p/%p\n",
368 					ip->vp, vp);
369 				vput(vp);
370 				continue;
371 			}
372 			*errorp = 0;
373 			break;
374 		}
375 
376 		/*
377 		 * No vnode exists, allocate a new vnode.  Beware of
378 		 * allocation races.  This function will return an
379 		 * exclusively locked and referenced vnode.
380 		 */
381 		*errorp = getnewvnode(VT_HAMMER2, pmp->mp, &vp, 0, 0);
382 		if (*errorp) {
383 			kprintf("hammer2: igetv getnewvnode failed %d\n",
384 				*errorp);
385 			vp = NULL;
386 			break;
387 		}
388 
389 		/*
390 		 * Lock the inode and check for an allocation race.
391 		 */
392 		wasexclusive = hammer2_inode_lock_upgrade(ip);
393 		if (ip->vp != NULL) {
394 			vp->v_type = VBAD;
395 			vx_put(vp);
396 			hammer2_inode_lock_downgrade(ip, wasexclusive);
397 			continue;
398 		}
399 
400 		switch (ripdata->meta.type) {
401 		case HAMMER2_OBJTYPE_DIRECTORY:
402 			vp->v_type = VDIR;
403 			break;
404 		case HAMMER2_OBJTYPE_REGFILE:
405 			vp->v_type = VREG;
406 			vinitvmio(vp, ripdata->meta.size,
407 				  HAMMER2_LBUFSIZE,
408 				  (int)ripdata->meta.size & HAMMER2_LBUFMASK);
409 			break;
410 		case HAMMER2_OBJTYPE_SOFTLINK:
411 			/*
412 			 * XXX for now we are using the generic file_read
413 			 * and file_write code so we need a buffer cache
414 			 * association.
415 			 */
416 			vp->v_type = VLNK;
417 			vinitvmio(vp, ripdata->meta.size,
418 				  HAMMER2_LBUFSIZE,
419 				  (int)ripdata->meta.size & HAMMER2_LBUFMASK);
420 			break;
421 		case HAMMER2_OBJTYPE_CDEV:
422 			vp->v_type = VCHR;
423 			/* fall through */
424 		case HAMMER2_OBJTYPE_BDEV:
425 			vp->v_ops = &pmp->mp->mnt_vn_spec_ops;
426 			if (ripdata->meta.type != HAMMER2_OBJTYPE_CDEV)
427 				vp->v_type = VBLK;
428 			addaliasu(vp,
429 				  ripdata->meta.rmajor,
430 				  ripdata->meta.rminor);
431 			break;
432 		case HAMMER2_OBJTYPE_FIFO:
433 			vp->v_type = VFIFO;
434 			vp->v_ops = &pmp->mp->mnt_vn_fifo_ops;
435 			break;
436 		default:
437 			panic("hammer2: unhandled objtype %d",
438 			      ripdata->meta.type);
439 			break;
440 		}
441 
442 		if (ip == pmp->iroot)
443 			vsetflags(vp, VROOT);
444 
445 		vp->v_data = ip;
446 		ip->vp = vp;
447 		hammer2_inode_ref(ip);		/* vp association */
448 		hammer2_inode_lock_downgrade(ip, wasexclusive);
449 		break;
450 	}
451 
452 	/*
453 	 * Return non-NULL vp and *errorp == 0, or NULL vp and *errorp != 0.
454 	 */
455 	if (hammer2_debug & 0x0002) {
456 		kprintf("igetv vp %p refs 0x%08x aux 0x%08x\n",
457 			vp, vp->v_refcnt, vp->v_auxrefs);
458 	}
459 	return (vp);
460 }
461 
462 /*
463  * Returns the inode associated with the passed-in cluster, creating the
464  * inode if necessary and synchronizing it to the passed-in cluster otherwise.
465  *
466  * The passed-in cluster must be locked and will remain locked on return.
467  * The returned inode will be locked and the caller may dispose of both
468  * via hammer2_inode_unlock() + hammer2_inode_drop().  However, if the caller
469  * needs to resolve a hardlink it must ref/unlock/relock/drop the inode.
470  *
471  * The hammer2_inode structure regulates the interface between the high level
472  * kernel VNOPS API and the filesystem backend (the chains).
473  *
474  * On return the inode is locked with the supplied cluster.
475  */
476 hammer2_inode_t *
477 hammer2_inode_get(hammer2_pfs_t *pmp, hammer2_inode_t *dip,
478 		  hammer2_cluster_t *cluster)
479 {
480 	hammer2_inode_t *nip;
481 	const hammer2_inode_data_t *iptmp;
482 	const hammer2_inode_data_t *nipdata;
483 
484 	KKASSERT(cluster == NULL ||
485 		 hammer2_cluster_type(cluster) == HAMMER2_BREF_TYPE_INODE);
486 	KKASSERT(pmp);
487 
488 	/*
489 	 * Interlocked lookup/ref of the inode.  This code is only needed
490 	 * when looking up inodes with nlinks != 0 (TODO: optimize out
491 	 * otherwise and test for duplicates).
492 	 *
493 	 * Cluster can be NULL during the initial pfs allocation.
494 	 */
495 again:
496 	while (cluster) {
497 		iptmp = &hammer2_cluster_rdata(cluster)->ipdata;
498 		nip = hammer2_inode_lookup(pmp, iptmp->meta.inum);
499 		if (nip == NULL)
500 			break;
501 
502 		hammer2_mtx_ex(&nip->lock);
503 
504 		/*
505 		 * Handle SMP race (not applicable to the super-root spmp
506 		 * which can't index inodes due to duplicative inode numbers).
507 		 */
508 		if (pmp->spmp_hmp == NULL &&
509 		    (nip->flags & HAMMER2_INODE_ONRBTREE) == 0) {
510 			hammer2_mtx_unlock(&nip->lock);
511 			hammer2_inode_drop(nip);
512 			continue;
513 		}
514 		hammer2_inode_repoint(nip, NULL, cluster);
515 
516 		return nip;
517 	}
518 
519 	/*
520 	 * We couldn't find the inode number, create a new inode.
521 	 */
522 	nip = kmalloc(sizeof(*nip), pmp->minode, M_WAITOK | M_ZERO);
523 	spin_init(&nip->cluster_spin, "h2clspin");
524 	atomic_add_long(&pmp->inmem_inodes, 1);
525 	hammer2_pfs_memory_inc(pmp);
526 	hammer2_pfs_memory_wakeup(pmp);
527 	if (pmp->spmp_hmp)
528 		nip->flags = HAMMER2_INODE_SROOT;
529 
530 	/*
531 	 * Initialize nip's cluster.  A cluster is provided for normal
532 	 * inodes but typically not for the super-root or PFS inodes.
533 	 */
534 	nip->cluster.refs = 1;
535 	nip->cluster.pmp = pmp;
536 	nip->cluster.flags |= HAMMER2_CLUSTER_INODE;
537 	if (cluster) {
538 		nipdata = &hammer2_cluster_rdata(cluster)->ipdata;
539 		nip->meta = nipdata->meta;
540 		hammer2_cluster_bref(cluster, &nip->bref);
541 		atomic_set_int(&nip->flags, HAMMER2_INODE_METAGOOD);
542 		hammer2_inode_repoint(nip, NULL, cluster);
543 	} else {
544 		nip->meta.inum = 1;		/* PFS inum is always 1 XXX */
545 		/* mtime will be updated when a cluster is available */
546 		atomic_set_int(&nip->flags, HAMMER2_INODE_METAGOOD);/*XXX*/
547 	}
548 
549 	nip->pip = dip;				/* can be NULL */
550 	if (dip)
551 		hammer2_inode_ref(dip);	/* ref dip for nip->pip */
552 
553 	nip->pmp = pmp;
554 
555 	/*
556 	 * ref and lock on nip gives it state compatible to after a
557 	 * hammer2_inode_lock() call.
558 	 */
559 	nip->refs = 1;
560 	hammer2_mtx_init(&nip->lock, "h2inode");
561 	hammer2_mtx_ex(&nip->lock);
562 	/* combination of thread lock and chain lock == inode lock */
563 
564 	/*
565 	 * Attempt to add the inode.  If it fails we raced another inode
566 	 * get.  Undo all the work and try again.
567 	 */
568 	if (pmp->spmp_hmp == NULL) {
569 		hammer2_spin_ex(&pmp->inum_spin);
570 		if (RB_INSERT(hammer2_inode_tree, &pmp->inum_tree, nip)) {
571 			hammer2_spin_unex(&pmp->inum_spin);
572 			hammer2_mtx_unlock(&nip->lock);
573 			hammer2_inode_drop(nip);
574 			goto again;
575 		}
576 		atomic_set_int(&nip->flags, HAMMER2_INODE_ONRBTREE);
577 		hammer2_spin_unex(&pmp->inum_spin);
578 	}
579 
580 	return (nip);
581 }
582 
583 /*
584  * Create a new inode in the specified directory using the vattr to
585  * figure out the type of inode.
586  *
587  * If no error occurs the new inode with its cluster locked is returned in
588  * *nipp, otherwise an error is returned and *nipp is set to NULL.
589  *
590  * If vap and/or cred are NULL the related fields are not set and the
591  * inode type defaults to a directory.  This is used when creating PFSs
592  * under the super-root, so the inode number is set to 1 in this case.
593  *
594  * dip is not locked on entry.
595  *
596  * NOTE: When used to create a snapshot, the inode is temporarily associated
597  *	 with the super-root spmp. XXX should pass new pmp for snapshot.
598  */
599 hammer2_inode_t *
600 hammer2_inode_create(hammer2_trans_t *trans, hammer2_inode_t *dip,
601 		     struct vattr *vap, struct ucred *cred,
602 		     const uint8_t *name, size_t name_len,
603 		     hammer2_cluster_t **clusterp,
604 		     int flags, int *errorp)
605 {
606 	const hammer2_inode_data_t *dipdata;
607 	hammer2_inode_data_t *nipdata;
608 	hammer2_cluster_t *cluster;
609 	hammer2_cluster_t *cparent;
610 	hammer2_inode_t *nip;
611 	hammer2_key_t key_dummy;
612 	hammer2_key_t lhc;
613 	int error;
614 	uid_t xuid;
615 	uuid_t dip_uid;
616 	uuid_t dip_gid;
617 	uint32_t dip_mode;
618 	uint8_t dip_comp_algo;
619 	uint8_t dip_check_algo;
620 
621 	lhc = hammer2_dirhash(name, name_len);
622 	*errorp = 0;
623 
624 	/*
625 	 * Locate the inode or indirect block to create the new
626 	 * entry in.  At the same time check for key collisions
627 	 * and iterate until we don't get one.
628 	 *
629 	 * NOTE: hidden inodes do not have iterators.
630 	 */
631 retry:
632 	hammer2_inode_lock(dip, HAMMER2_RESOLVE_ALWAYS);
633 	cparent = hammer2_inode_cluster(dip, HAMMER2_RESOLVE_ALWAYS);
634 	dipdata = &hammer2_cluster_rdata(cparent)->ipdata;
635 	dip_uid = dipdata->meta.uid;
636 	dip_gid = dipdata->meta.gid;
637 	dip_mode = dipdata->meta.mode;
638 	dip_comp_algo = dipdata->meta.comp_algo;
639 	dip_check_algo = dipdata->meta.check_algo;
640 
641 	error = 0;
642 	while (error == 0) {
643 		cluster = hammer2_cluster_lookup(cparent, &key_dummy,
644 						 lhc, lhc, 0);
645 		if (cluster == NULL)
646 			break;
647 		if ((lhc & HAMMER2_DIRHASH_VISIBLE) == 0)
648 			error = ENOSPC;
649 		if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
650 			error = ENOSPC;
651 		hammer2_cluster_unlock(cluster);
652 		hammer2_cluster_drop(cluster);
653 		cluster = NULL;
654 		++lhc;
655 	}
656 
657 	if (error == 0) {
658 		error = hammer2_cluster_create(trans, cparent, &cluster,
659 					     lhc, 0,
660 					     HAMMER2_BREF_TYPE_INODE,
661 					     HAMMER2_INODE_BYTES,
662 					     flags);
663 	}
664 #if INODE_DEBUG
665 	kprintf("CREATE INODE %*.*s chain=%p\n",
666 		(int)name_len, (int)name_len, name,
667 		(cluster ? cluster->focus : NULL));
668 #endif
669 
670 	/*
671 	 * Cleanup and handle retries.
672 	 */
673 	if (error == EAGAIN) {
674 		hammer2_cluster_ref(cparent);
675 		hammer2_inode_unlock(dip, cparent);
676 		hammer2_cluster_wait(cparent);
677 		hammer2_cluster_drop(cparent);
678 		goto retry;
679 	}
680 	hammer2_inode_unlock(dip, cparent);
681 	cparent = NULL;
682 
683 	if (error) {
684 		KKASSERT(cluster == NULL);
685 		*errorp = error;
686 		return (NULL);
687 	}
688 
689 	/*
690 	 * Set up the new inode.
691 	 *
692 	 * NOTE: *_get() integrates chain's lock into the inode lock.
693 	 *
694 	 * NOTE: Only one new inode can currently be created per
695 	 *	 transaction.  If the need arises we can adjust
696 	 *	 hammer2_trans_init() to allow more.
697 	 *
698 	 * NOTE: nipdata will have chain's blockset data.
699 	 */
700 	KKASSERT(cluster->focus->flags & HAMMER2_CHAIN_MODIFIED);
701 	nipdata = &hammer2_cluster_wdata(cluster)->ipdata;
702 	nipdata->meta.inum = trans->inode_tid;
703 	hammer2_cluster_modsync(cluster);
704 	nip = hammer2_inode_get(dip->pmp, dip, cluster);
705 	nipdata = &hammer2_cluster_wdata(cluster)->ipdata;
706 
707 	if (vap) {
708 		KKASSERT(trans->inodes_created == 0);
709 		nipdata->meta.type = hammer2_get_obj_type(vap->va_type);
710 		nipdata->meta.inum = trans->inode_tid;
711 		++trans->inodes_created;
712 
713 		switch (nipdata->meta.type) {
714 		case HAMMER2_OBJTYPE_CDEV:
715 		case HAMMER2_OBJTYPE_BDEV:
716 			nipdata->meta.rmajor = vap->va_rmajor;
717 			nipdata->meta.rminor = vap->va_rminor;
718 			break;
719 		default:
720 			break;
721 		}
722 	} else {
723 		nipdata->meta.type = HAMMER2_OBJTYPE_DIRECTORY;
724 		nipdata->meta.inum = 1;
725 	}
726 
727 	/* Inherit parent's inode compression mode. */
728 	nip->comp_heuristic = 0;
729 	nipdata->meta.comp_algo = dip_comp_algo;
730 	nipdata->meta.check_algo = dip_check_algo;
731 	nipdata->meta.version = HAMMER2_INODE_VERSION_ONE;
732 	hammer2_update_time(&nipdata->meta.ctime);
733 	nipdata->meta.mtime = nipdata->meta.ctime;
734 	if (vap)
735 		nipdata->meta.mode = vap->va_mode;
736 	nipdata->meta.nlinks = 1;
737 	if (vap) {
738 		if (dip && dip->pmp) {
739 			xuid = hammer2_to_unix_xid(&dip_uid);
740 			xuid = vop_helper_create_uid(dip->pmp->mp,
741 						     dip_mode,
742 						     xuid,
743 						     cred,
744 						     &vap->va_mode);
745 		} else {
746 			/* super-root has no dip and/or pmp */
747 			xuid = 0;
748 		}
749 		if (vap->va_vaflags & VA_UID_UUID_VALID)
750 			nipdata->meta.uid = vap->va_uid_uuid;
751 		else if (vap->va_uid != (uid_t)VNOVAL)
752 			hammer2_guid_to_uuid(&nipdata->meta.uid, vap->va_uid);
753 		else
754 			hammer2_guid_to_uuid(&nipdata->meta.uid, xuid);
755 
756 		if (vap->va_vaflags & VA_GID_UUID_VALID)
757 			nipdata->meta.gid = vap->va_gid_uuid;
758 		else if (vap->va_gid != (gid_t)VNOVAL)
759 			hammer2_guid_to_uuid(&nipdata->meta.gid, vap->va_gid);
760 		else if (dip)
761 			nipdata->meta.gid = dip_gid;
762 	}
763 
764 	/*
765 	 * Regular files and softlinks allow a small amount of data to be
766 	 * directly embedded in the inode.  This flag will be cleared if
767 	 * the size is extended past the embedded limit.
768 	 */
769 	if (nipdata->meta.type == HAMMER2_OBJTYPE_REGFILE ||
770 	    nipdata->meta.type == HAMMER2_OBJTYPE_SOFTLINK) {
771 		nipdata->meta.op_flags |= HAMMER2_OPFLAG_DIRECTDATA;
772 	}
773 
774 	KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
775 	bcopy(name, nipdata->filename, name_len);
776 	nipdata->meta.name_key = lhc;
777 	nipdata->meta.name_len = name_len;
778 	nip->meta = nipdata->meta;
779 	hammer2_cluster_modsync(cluster);
780 	*clusterp = cluster;
781 
782 	return (nip);
783 }
784 
785 /*
786  * The cluster has been removed from the original directory and replaced
787  * with a hardlink pointer.  Move the cluster to the specified parent
788  * directory, change the filename to "0xINODENUMBER", and adjust the key.
789  * The cluster becomes our invisible hardlink target.
790  *
791  * The original cluster must be deleted on entry.
792  */
793 static
794 void
795 hammer2_hardlink_shiftup(hammer2_trans_t *trans, hammer2_cluster_t *cluster,
796 			hammer2_inode_t *ip, hammer2_inode_t *dip,
797 			hammer2_cluster_t *dcluster,
798 			int nlinks, int *errorp)
799 {
800 	const hammer2_inode_data_t *iptmp;
801 	hammer2_inode_data_t *nipdata;
802 	hammer2_cluster_t *xcluster;
803 	hammer2_key_t key_dummy;
804 	hammer2_key_t lhc;
805 	hammer2_blockref_t bref;
806 
807 	iptmp = &hammer2_cluster_rdata(cluster)->ipdata;
808 	lhc = iptmp->meta.inum;
809 	KKASSERT((lhc & HAMMER2_DIRHASH_VISIBLE) == 0);
810 
811 	/*
812 	 * Locate the inode or indirect block to create the new
813 	 * entry in.  lhc represents the inode number so there is
814 	 * no collision iteration.
815 	 *
816 	 * There should be no key collisions with invisible inode keys.
817 	 *
818 	 * WARNING! Must use inode_lock_ex() on dip to handle a stale
819 	 *	    dip->cluster cache.
820 	 */
821 	*errorp = 0;
822 	xcluster = hammer2_cluster_lookup(dcluster, &key_dummy,
823 				      lhc, lhc, 0);
824 	if (xcluster) {
825 		kprintf("X3 chain %p dip %p dchain %p dip->chain %p\n",
826 			xcluster->focus, dip, dcluster->focus,
827 			dip->cluster.focus);
828 		hammer2_cluster_unlock(xcluster);
829 		hammer2_cluster_drop(xcluster);
830 		xcluster = NULL;
831 		*errorp = ENOSPC;
832 #if 0
833 		Debugger("X3");
834 #endif
835 	}
836 
837 	/*
838 	 * Handle the error case
839 	 */
840 	if (*errorp) {
841 		panic("error2");
842 		KKASSERT(xcluster == NULL);
843 		return;
844 	}
845 
846 	/*
847 	 * Use xcluster as a placeholder for (lhc).  Duplicate cluster to the
848 	 * same target bref as xcluster and then delete xcluster.  The
849 	 * duplication occurs after xcluster in flush order even though
850 	 * xcluster is deleted after the duplication. XXX
851 	 *
852 	 * WARNING! Duplications (to a different parent) can cause indirect
853 	 *	    blocks to be inserted, refactor xcluster.
854 	 *
855 	 * WARNING! Only key and keybits is extracted from a passed-in bref.
856 	 */
857 	hammer2_cluster_bref(cluster, &bref);
858 	bref.key = lhc;			/* invisible dir entry key */
859 	bref.keybits = 0;
860 	hammer2_cluster_rename(trans, &bref, dcluster, cluster, 0);
861 
862 	/*
863 	 * cluster is now 'live' again.. adjust the filename.
864 	 *
865 	 * Directory entries are inodes but this is a hidden hardlink
866 	 * target.  The name isn't used but to ease debugging give it
867 	 * a name after its inode number.
868 	 */
869 	hammer2_cluster_modify(trans, cluster, 0);
870 	nipdata = &hammer2_cluster_wdata(cluster)->ipdata;
871 	ksnprintf(nipdata->filename, sizeof(nipdata->filename),
872 		  "0x%016jx", (intmax_t)nipdata->meta.inum);
873 	nipdata->meta.name_len = strlen(nipdata->filename);
874 	nipdata->meta.name_key = lhc;
875 	nipdata->meta.nlinks += nlinks;
876 
877 	/*
878 	 * Resync ip->meta.  Some fields have to be retained.
879 	 */
880 	nipdata->meta.size = ip->meta.size;
881 	nipdata->meta.mtime = ip->meta.mtime;
882 	ip->meta = nipdata->meta;
883 
884 	hammer2_cluster_modsync(cluster);
885 }
886 
887 /*
888  * Connect the target inode represented by (cluster) to the media topology
889  * at (dip, name, len).  The caller can pass a rough *chainp, this function
890  * will issue lookup()s to position the parent chain properly for the
891  * chain insertion.
892  *
893  * If hlink is TRUE this function creates an OBJTYPE_HARDLINK directory
894  * entry instead of connecting (cluster).
895  *
896  * If hlink is FALSE this function expects (cluster) to be unparented.
897  */
898 int
899 hammer2_inode_connect(hammer2_trans_t *trans,
900 		      hammer2_inode_t *ip, hammer2_cluster_t **clusterp,
901 		      int hlink,
902 		      hammer2_inode_t *dip, hammer2_cluster_t *dcluster,
903 		      const uint8_t *name, size_t name_len,
904 		      hammer2_key_t lhc)
905 {
906 	hammer2_inode_data_t *wipdata;
907 	hammer2_cluster_t *ocluster;
908 	hammer2_cluster_t *ncluster;
909 	hammer2_key_t key_dummy;
910 	int error;
911 
912 	/*
913 	 * Since ocluster is either disconnected from the topology or
914 	 * represents a hardlink terminus which is always a parent of or
915 	 * equal to dip, we should be able to safely lock dip->chain for
916 	 * our setup.
917 	 *
918 	 * WARNING! Must use inode_lock_ex() on dip to handle a stale
919 	 *	    dip->cluster.
920 	 *
921 	 * If name is non-NULL we calculate lhc, else we use the passed-in
922 	 * lhc.
923 	 */
924 	ocluster = *clusterp;
925 
926 	if (name) {
927 		lhc = hammer2_dirhash(name, name_len);
928 
929 		/*
930 		 * Locate the inode or indirect block to create the new
931 		 * entry in.  At the same time check for key collisions
932 		 * and iterate until we don't get one.
933 		 */
934 		error = 0;
935 		while (error == 0) {
936 			ncluster = hammer2_cluster_lookup(dcluster, &key_dummy,
937 						      lhc, lhc, 0);
938 			if (ncluster == NULL)
939 				break;
940 			if ((lhc & HAMMER2_DIRHASH_LOMASK) ==
941 			    HAMMER2_DIRHASH_LOMASK) {
942 				error = ENOSPC;
943 			}
944 			hammer2_cluster_unlock(ncluster);
945 			hammer2_cluster_drop(ncluster);
946 			ncluster = NULL;
947 			++lhc;
948 		}
949 	} else {
950 		/*
951 		 * Reconnect to specific key (used when moving
952 		 * unlinked-but-open files into the hidden directory).
953 		 */
954 		ncluster = hammer2_cluster_lookup(dcluster, &key_dummy,
955 						  lhc, lhc, 0);
956 		KKASSERT(ncluster == NULL);
957 		error = 0;
958 	}
959 
960 	if (error == 0) {
961 		if (hlink) {
962 			/*
963 			 * Hardlink pointer needed, create totally fresh
964 			 * directory entry.
965 			 *
966 			 * We must refactor ocluster because it might have
967 			 * been shifted into an indirect cluster by the
968 			 * create.
969 			 */
970 			KKASSERT(ncluster == NULL);
971 			error = hammer2_cluster_create(trans,
972 						       dcluster, &ncluster,
973 						       lhc, 0,
974 						       HAMMER2_BREF_TYPE_INODE,
975 						       HAMMER2_INODE_BYTES,
976 						       0);
977 		} else {
978 			/*
979 			 * Reconnect the original cluster under the new name.
980 			 * Original cluster must have already been deleted by
981 			 * teh caller.
982 			 *
983 			 * WARNING! Can cause held-over clusters to require a
984 			 *	    refactor.  Fortunately we have none (our
985 			 *	    locked clusters are passed into and
986 			 *	    modified by the call).
987 			 */
988 			ncluster = ocluster;
989 			ocluster = NULL;
990 			error = hammer2_cluster_create(trans,
991 						       dcluster, &ncluster,
992 						       lhc, 0,
993 						       HAMMER2_BREF_TYPE_INODE,
994 						       HAMMER2_INODE_BYTES,
995 						       0);
996 		}
997 	}
998 
999 	/*
1000 	 * Unlock stuff.
1001 	 */
1002 	KKASSERT(error != EAGAIN);
1003 
1004 	/*
1005 	 * ncluster should be NULL on error, leave ocluster
1006 	 * (ocluster == *clusterp) alone.
1007 	 */
1008 	if (error) {
1009 		KKASSERT(ncluster == NULL);
1010 		return (error);
1011 	}
1012 
1013 	/*
1014 	 * Directory entries are inodes so if the name has changed we have
1015 	 * to update the inode.
1016 	 *
1017 	 * When creating an OBJTYPE_HARDLINK entry remember to unlock the
1018 	 * cluster, the caller will access the hardlink via the actual hardlink
1019 	 * target file and not the hardlink pointer entry, so we must still
1020 	 * return ocluster.
1021 	 */
1022 	if (hlink && hammer2_hardlink_enable >= 0) {
1023 		/*
1024 		 * Create the HARDLINK pointer.  oip represents the hardlink
1025 		 * target in this situation.
1026 		 *
1027 		 * We will return ocluster (the hardlink target).
1028 		 */
1029 		hammer2_cluster_modify(trans, ncluster, 0);
1030 		KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
1031 		wipdata = &hammer2_cluster_wdata(ncluster)->ipdata;
1032 		bcopy(name, wipdata->filename, name_len);
1033 		wipdata->meta.name_key = lhc;
1034 		wipdata->meta.name_len = name_len;
1035 		wipdata->meta.target_type =
1036 			    hammer2_cluster_rdata(ocluster)->ipdata.meta.type;
1037 		wipdata->meta.type = HAMMER2_OBJTYPE_HARDLINK;
1038 		wipdata->meta.inum =
1039 			    hammer2_cluster_rdata(ocluster)->ipdata.meta.inum;
1040 		wipdata->meta.version = HAMMER2_INODE_VERSION_ONE;
1041 		wipdata->meta.nlinks = 1;
1042 		wipdata->meta.op_flags = HAMMER2_OPFLAG_DIRECTDATA;
1043 		hammer2_cluster_modsync(ncluster);
1044 		hammer2_cluster_unlock(ncluster);
1045 		hammer2_cluster_drop(ncluster);
1046 		ncluster = ocluster;
1047 		ocluster = NULL;
1048 	} else {
1049 		/*
1050 		 * ncluster is a duplicate of ocluster at the new location.
1051 		 * We must fixup the name stored in the inode data.
1052 		 * The bref key has already been adjusted by inode_connect().
1053 		 */
1054 		hammer2_cluster_modify(trans, ncluster, 0);
1055 		wipdata = &hammer2_cluster_wdata(ncluster)->ipdata;
1056 
1057 		KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
1058 		bcopy(name, wipdata->filename, name_len);
1059 		wipdata->meta.name_key = lhc;
1060 		wipdata->meta.name_len = name_len;
1061 		wipdata->meta.nlinks = 1;
1062 		hammer2_cluster_modsync(ncluster);
1063 
1064 		/*
1065 		 * Resync the in-memory inode, some fields must be retained.
1066 		 */
1067 		if (ip) {	/* XXX move_to_hidden passes NULL */
1068 			wipdata->meta.size = ip->meta.size;
1069 			wipdata->meta.mtime = ip->meta.mtime;
1070 			ip->meta = wipdata->meta;
1071 		}
1072 	}
1073 
1074 	/*
1075 	 * We are replacing ocluster with ncluster, unlock ocluster.  In the
1076 	 * case where ocluster is left unchanged the code above sets
1077 	 * ncluster to ocluster and ocluster to NULL, resulting in a NOP here.
1078 	 */
1079 	if (ocluster) {
1080 		hammer2_cluster_unlock(ocluster);
1081 		hammer2_cluster_drop(ocluster);
1082 	}
1083 	*clusterp = ncluster;
1084 
1085 	return (0);
1086 }
1087 
1088 /*
1089  * Repoint ip->cluster's chains to cluster's chains and fixup the default
1090  * focus.  Only valid elements are repointed.  Invalid elements have to be
1091  * adjusted by the appropriate slave sync threads.
1092  *
1093  * Caller must hold the inode and cluster exclusive locked, if not NULL,
1094  * must also be locked.
1095  *
1096  * Cluster may be NULL to clean out any chains in ip->cluster.
1097  */
1098 void
1099 hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
1100 		      hammer2_cluster_t *cluster)
1101 {
1102 	hammer2_chain_t *dropch[HAMMER2_MAXCLUSTER];
1103 	hammer2_chain_t *ochain;
1104 	hammer2_chain_t *nchain;
1105 	hammer2_inode_t *opip;
1106 	int i;
1107 
1108 	bzero(dropch, sizeof(dropch));
1109 
1110 	/*
1111 	 * Replace chains in ip->cluster with chains from cluster and
1112 	 * adjust the focus if necessary.
1113 	 *
1114 	 * NOTE: nchain and/or ochain can be NULL due to gaps
1115 	 *	 in the cluster arrays.
1116 	 */
1117 	hammer2_spin_ex(&ip->cluster_spin);
1118 	for (i = 0; cluster && i < cluster->nchains; ++i) {
1119 		/*
1120 		 * Do not replace invalid elements as this might race
1121 		 * syncthr replacements.
1122 		 */
1123 		if (cluster->array[i].flags & HAMMER2_CITEM_INVALID)
1124 			continue;
1125 
1126 		/*
1127 		 * Do not replace elements which are the same.  Also handle
1128 		 * element count discrepancies.
1129 		 */
1130 		nchain = cluster->array[i].chain;
1131 		if (i < ip->cluster.nchains) {
1132 			ochain = ip->cluster.array[i].chain;
1133 			if (ochain == nchain)
1134 				continue;
1135 		} else {
1136 			ochain = NULL;
1137 		}
1138 
1139 		/*
1140 		 * Make adjustments
1141 		 */
1142 		ip->cluster.array[i].chain = nchain;
1143 		ip->cluster.array[i].flags &= ~HAMMER2_CITEM_INVALID;
1144 		ip->cluster.array[i].flags |= cluster->array[i].flags &
1145 					      HAMMER2_CITEM_INVALID;
1146 		if (nchain)
1147 			hammer2_chain_ref(nchain);
1148 		dropch[i] = ochain;
1149 	}
1150 
1151 	/*
1152 	 * Release any left-over chains in ip->cluster.
1153 	 */
1154 	while (i < ip->cluster.nchains) {
1155 		nchain = ip->cluster.array[i].chain;
1156 		if (nchain) {
1157 			ip->cluster.array[i].chain = NULL;
1158 			ip->cluster.array[i].flags |= HAMMER2_CITEM_INVALID;
1159 		}
1160 		dropch[i] = nchain;
1161 		++i;
1162 	}
1163 
1164 	/*
1165 	 * Fixup fields.  Note that the inode-embedded cluster is never
1166 	 * directly locked.
1167 	 */
1168 	if (cluster) {
1169 		ip->cluster.nchains = cluster->nchains;
1170 		ip->cluster.focus = cluster->focus;
1171 		ip->cluster.flags = cluster->flags & ~HAMMER2_CLUSTER_LOCKED;
1172 	} else {
1173 		ip->cluster.nchains = 0;
1174 		ip->cluster.focus = NULL;
1175 		ip->cluster.flags &= ~HAMMER2_CLUSTER_ZFLAGS;
1176 	}
1177 
1178 	/*
1179 	 * Repoint ip->pip if requested (non-NULL pip).
1180 	 */
1181 	if (pip && ip->pip != pip) {
1182 		opip = ip->pip;
1183 		hammer2_inode_ref(pip);
1184 		ip->pip = pip;
1185 	} else {
1186 		opip = NULL;
1187 	}
1188 	hammer2_spin_unex(&ip->cluster_spin);
1189 
1190 	/*
1191 	 * Cleanup outside of spinlock
1192 	 */
1193 	while (--i >= 0) {
1194 		if (dropch[i])
1195 			hammer2_chain_drop(dropch[i]);
1196 	}
1197 	if (opip)
1198 		hammer2_inode_drop(opip);
1199 }
1200 
1201 /*
1202  * Repoint a single element from the cluster to the ip.  Used by the
1203  * synchronization threads to piecemeal update inodes.  Does not change
1204  * focus and requires inode to be re-locked to clean-up flags (XXX).
1205  */
1206 void
1207 hammer2_inode_repoint_one(hammer2_inode_t *ip, hammer2_cluster_t *cluster,
1208 			  int idx)
1209 {
1210 	hammer2_chain_t *ochain;
1211 	hammer2_chain_t *nchain;
1212 	int i;
1213 
1214 	hammer2_spin_ex(&ip->cluster_spin);
1215 	KKASSERT(idx < cluster->nchains);
1216 	if (idx < ip->cluster.nchains) {
1217 		ochain = ip->cluster.array[idx].chain;
1218 		nchain = cluster->array[idx].chain;
1219 	} else {
1220 		ochain = NULL;
1221 		nchain = cluster->array[idx].chain;
1222 		ip->cluster.nchains = idx + 1;
1223 		for (i = ip->cluster.nchains; i <= idx; ++i) {
1224 			bzero(&ip->cluster.array[i],
1225 			      sizeof(ip->cluster.array[i]));
1226 			ip->cluster.array[i].flags |= HAMMER2_CITEM_INVALID;
1227 		}
1228 	}
1229 	if (ochain != nchain) {
1230 		/*
1231 		 * Make adjustments.
1232 		 */
1233 		ip->cluster.array[idx].chain = nchain;
1234 		ip->cluster.array[idx].flags &= ~HAMMER2_CITEM_INVALID;
1235 		ip->cluster.array[idx].flags |= cluster->array[idx].flags &
1236 						HAMMER2_CITEM_INVALID;
1237 	}
1238 	hammer2_spin_unex(&ip->cluster_spin);
1239 	if (ochain != nchain) {
1240 		if (nchain)
1241 			hammer2_chain_ref(nchain);
1242 		if (ochain)
1243 			hammer2_chain_drop(ochain);
1244 	}
1245 }
1246 
1247 /*
1248  * Unlink the file (dip, name, name_len), from the specified directory inode.
1249  * If the caller also has the hammer2_inode the caller must pass it locked as
1250  * (ip), but may pass NULL if it does not have the inode in hand.
1251  *
1252  * The directory inode does not need to be locked.
1253  *
1254  * isdir determines whether a directory/non-directory check should be made.
1255  * No check is made if isdir is set to -1.
1256  *
1257  * isopen specifies whether special unlink-with-open-descriptor handling
1258  * must be performed.  If set to -1 the caller is deleting a PFS and we
1259  * check whether the chain is mounted or not (chain->pmp != NULL).  1 is
1260  * implied if it is mounted.
1261  *
1262  * If isopen is 1 and nlinks drops to 0 this function must move the chain
1263  * to a special hidden directory until last-close occurs on the file.
1264  *
1265  * NOTE!  The underlying file can still be active with open descriptors
1266  *	  or if the inode is being manually held (e.g. for rename).
1267  *
1268  * NOTE!  When unlinking an open file the inode will be temporarily moved to
1269  *	  a hidden directory, otherwise the inode will be deleted.
1270  */
1271 int
1272 hammer2_unlink_file(hammer2_trans_t *trans,
1273 		    hammer2_inode_t *dip, hammer2_inode_t *ip,
1274 		    const uint8_t *name, size_t name_len,
1275 		    int isdir, int *hlinkp, struct nchandle *nch,
1276 		    int nlinks)
1277 {
1278 	const hammer2_inode_data_t *ripdata;
1279 	hammer2_cluster_t *cparent;
1280 	hammer2_cluster_t *hcluster;
1281 	hammer2_cluster_t *hparent;
1282 	hammer2_cluster_t *cluster;
1283 	hammer2_cluster_t *dparent;
1284 	hammer2_cluster_t *dcluster;
1285 	hammer2_key_t key_dummy;
1286 	hammer2_key_t key_next;
1287 	hammer2_key_t lhc;
1288 	int last_link;
1289 	int error;
1290 	int hlink;
1291 	int myip;
1292 	uint8_t type;
1293 
1294 	error = 0;
1295 	hlink = 0;
1296 	myip = 0;
1297 	hcluster = NULL;
1298 	hparent = NULL;
1299 	lhc = hammer2_dirhash(name, name_len);
1300 
1301 again:
1302 	/*
1303 	 * Locate the filename in the directory and instantiate the ip
1304 	 * if necessary.  If the ip is already known we must still locate
1305 	 * the filename to adjust cparent for possible deletion.
1306 	 */
1307 	hammer2_inode_lock(dip, HAMMER2_RESOLVE_ALWAYS);
1308 	cparent = hammer2_inode_cluster(dip, HAMMER2_RESOLVE_ALWAYS);
1309 	cluster = hammer2_cluster_lookup(cparent, &key_next,
1310 				     lhc, lhc + HAMMER2_DIRHASH_LOMASK, 0);
1311 	while (cluster) {
1312 		if (hammer2_cluster_type(cluster) == HAMMER2_BREF_TYPE_INODE) {
1313 			ripdata = &hammer2_cluster_rdata(cluster)->ipdata;
1314 			if (ripdata->meta.name_len == name_len &&
1315 			    bcmp(ripdata->filename, name, name_len) == 0) {
1316 				break;
1317 			}
1318 		}
1319 		cluster = hammer2_cluster_next(cparent, cluster, &key_next,
1320 					       key_next,
1321 					       lhc + HAMMER2_DIRHASH_LOMASK,
1322 					       0);
1323 	}
1324 	hammer2_inode_unlock(dip, NULL);	/* retain cparent */
1325 
1326 	/*
1327 	 * Not found or wrong type (isdir < 0 disables the type check).
1328 	 * If a hardlink pointer, type checks use the hardlink target.
1329 	 */
1330 	if (cluster == NULL) {
1331 		error = ENOENT;
1332 		goto done;
1333 	}
1334 
1335 	ripdata = &hammer2_cluster_rdata(cluster)->ipdata;
1336 	type = ripdata->meta.type;
1337 	if (type == HAMMER2_OBJTYPE_HARDLINK) {
1338 		hlink = 1;
1339 		type = ripdata->meta.target_type;
1340 	}
1341 
1342 	if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 0) {
1343 		error = ENOTDIR;
1344 		goto done;
1345 	}
1346 	if (type != HAMMER2_OBJTYPE_DIRECTORY && isdir >= 1) {
1347 		error = EISDIR;
1348 		goto done;
1349 	}
1350 
1351 	/*
1352 	 * Hardlink must be resolved.  We can't hold the parent locked
1353 	 * while we do this or we could deadlock.  The physical file will
1354 	 * be located at or above the current directory.
1355 	 *
1356 	 * We loop to reacquire the hardlink origination.
1357 	 *
1358 	 * NOTE: hammer2_hardlink_find() will locate the hardlink target,
1359 	 *	 returning a modified hparent and hcluster.
1360 	 */
1361 	if (ripdata->meta.type == HAMMER2_OBJTYPE_HARDLINK) {
1362 		if (hcluster == NULL) {
1363 			hcluster = cluster;
1364 			cluster = NULL;	/* safety */
1365 			hammer2_cluster_unlock(cparent);
1366 			hammer2_cluster_drop(cparent);
1367 			cparent = NULL; /* safety */
1368 			ripdata = NULL;	/* safety (associated w/cparent) */
1369 			error = hammer2_hardlink_find(dip, &hparent, &hcluster);
1370 
1371 			/*
1372 			 * If we couldn't find the hardlink target then some
1373 			 * parent directory containing the hardlink pointer
1374 			 * probably got renamed to above the original target,
1375 			 * a case not yet handled by H2.
1376 			 */
1377 			if (error) {
1378 				kprintf("H2 unlink_file: hardlink target for "
1379 					"\"%s\" not found\n",
1380 					name);
1381 				kprintf("(likely due to known directory "
1382 					"rename bug)\n");
1383 				goto done;
1384 			}
1385 			goto again;
1386 		}
1387 	}
1388 
1389 	/*
1390 	 * If this is a directory the directory must be empty.  However, if
1391 	 * isdir < 0 we are doing a rename and the directory does not have
1392 	 * to be empty, and if isdir > 1 we are deleting a PFS/snapshot
1393 	 * and the directory does not have to be empty.
1394 	 *
1395 	 * NOTE: We check the full key range here which covers both visible
1396 	 *	 and invisible entries.  Theoretically there should be no
1397 	 *	 invisible (hardlink target) entries if there are no visible
1398 	 *	 entries.
1399 	 */
1400 	if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 1) {
1401 		dparent = hammer2_cluster_lookup_init(cluster, 0);
1402 		dcluster = hammer2_cluster_lookup(dparent, &key_dummy,
1403 					          0, (hammer2_key_t)-1,
1404 					          HAMMER2_LOOKUP_NODATA);
1405 		if (dcluster) {
1406 			hammer2_cluster_unlock(dcluster);
1407 			hammer2_cluster_drop(dcluster);
1408 			hammer2_cluster_lookup_done(dparent);
1409 			error = ENOTEMPTY;
1410 			goto done;
1411 		}
1412 		hammer2_cluster_lookup_done(dparent);
1413 		dparent = NULL;
1414 		/* dcluster NULL */
1415 	}
1416 
1417 	/*
1418 	 * If this was a hardlink then (cparent, cluster) is the hardlink
1419 	 * pointer, which we can simply destroy outright.  Discard the
1420 	 * clusters and replace with the hardlink target.
1421 	 */
1422 	if (hcluster) {
1423 		hammer2_cluster_delete(trans, cparent, cluster,
1424 				       HAMMER2_DELETE_PERMANENT);
1425 		hammer2_cluster_unlock(cparent);
1426 		hammer2_cluster_drop(cparent);
1427 		hammer2_cluster_unlock(cluster);
1428 		hammer2_cluster_drop(cluster);
1429 		cparent = hparent;
1430 		cluster = hcluster;
1431 		hparent = NULL;
1432 		hcluster = NULL;
1433 	}
1434 
1435 	/*
1436 	 * This leaves us with the hardlink target or non-hardlinked file
1437 	 * or directory in (cparent, cluster).
1438 	 *
1439 	 * Delete the target when nlinks reaches 0 with special handling
1440 	 * to avoid I/O (to avoid actually updating the inode) for the 1->0
1441 	 * transition, if possible.  This optimization makes rm -rf very
1442 	 * fast.
1443 	 *
1444 	 * NOTE! In DragonFly the vnops function calls cache_unlink() after
1445 	 *	 calling us here to clean out the namecache association,
1446 	 *	 (which does not represent a ref for the open-test), and to
1447 	 *	 force finalization of the vnode if/when the last ref gets
1448 	 *	 dropped.
1449 	 *
1450 	 * NOTE! Files are unlinked by rename and then relinked.  nch will be
1451 	 *	 passed as NULL in this situation.  hammer2_inode_connect()
1452 	 *	 will bump nlinks.
1453 	 */
1454 	KKASSERT(cluster != NULL);
1455 
1456 	/*
1457 	 * Instantiate ip if necessary for ip->meta data consolidation.
1458 	 */
1459 	if (ip == NULL) {
1460 		ip = hammer2_inode_get(dip->pmp, dip, cluster);
1461 		myip = 1;
1462 	}
1463 
1464 
1465 	/*
1466 	 * Note: nlinks is negative when decrementing, positive when
1467 	 *	 incrementing.
1468 	 */
1469 	last_link = (ip->meta.nlinks + nlinks == 0);
1470 
1471 	if (last_link) {
1472 		/*
1473 		 * Target nlinks has reached 0, file now unlinked (but may
1474 		 * still be open).
1475 		 *
1476 		 * nlinks will be -1 for a normal remove().  If this is the
1477 		 * last link we must flag the inode so we can optimally
1478 		 * throw away buffer data and destroy the file on reclaim.
1479 		 */
1480 		if (nlinks == -1)
1481 			atomic_set_int(&ip->flags, HAMMER2_INODE_ISUNLINKED);
1482 
1483 		if (nch && cache_isopen(nch)) {
1484 			/*
1485 			 * If an unlinked file is still open we must update
1486 			 * the inodes link count.
1487 			 */
1488 			/*hammer2_cluster_modify(trans, cluster, 0);*/
1489 			hammer2_inode_modify(trans, ip);
1490 			ip->meta.nlinks += nlinks;
1491 			if ((int64_t)ip->meta.nlinks < 0)	/* safety */
1492 				ip->meta.nlinks = 0;
1493 			hammer2_inode_move_to_hidden(trans, &cparent, &cluster,
1494 						     ip->meta.inum);
1495 			/* hammer2_cluster_modsync(cluster); */
1496 		} else {
1497 			/*
1498 			 * This won't get everything if a vnode is still
1499 			 * present, but the cache_unlink() call the caller
1500 			 * makes will.
1501 			 */
1502 			hammer2_cluster_delete(trans, cparent, cluster,
1503 					       HAMMER2_DELETE_PERMANENT);
1504 		}
1505 	} else if (hlink == 0) {
1506 		/*
1507 		 * In this situation a normal non-hardlinked file (which can
1508 		 * only have nlinks == 1) still has a non-zero nlinks, the
1509 		 * caller must be doing a RENAME operation and so is passing
1510 		 * a nlinks adjustment of 0, and only wishes to remove file
1511 		 * in order to be able to reconnect it under a different name.
1512 		 *
1513 		 * In this situation we do a temporary deletion of the
1514 		 * chain in order to allow the file to be reconnected in
1515 		 * a different location.
1516 		 */
1517 		KKASSERT(nlinks == 0);
1518 		hammer2_cluster_delete(trans, cparent, cluster, 0);
1519 	} else {
1520 		/*
1521 		 * Links remain, must update the inode link count.
1522 		 */
1523 		/*hammer2_cluster_modify(trans, cluster, 0);*/
1524 		hammer2_inode_modify(trans, ip);
1525 		ip->meta.nlinks += nlinks;
1526 		if ((int64_t)ip->meta.nlinks < 0)
1527 			ip->meta.nlinks = 0;
1528 		/* hammer2_cluster_modsync(cluster); */
1529 	}
1530 
1531 	if (myip) {
1532 		hammer2_inode_unlock(ip, NULL);
1533 	}
1534 
1535 	error = 0;
1536 done:
1537 	if (cparent) {
1538 		hammer2_cluster_unlock(cparent);
1539 		hammer2_cluster_drop(cparent);
1540 	}
1541 	if (cluster) {
1542 		hammer2_cluster_unlock(cluster);
1543 		hammer2_cluster_drop(cluster);
1544 	}
1545 	if (hparent) {
1546 		hammer2_cluster_unlock(hparent);
1547 		hammer2_cluster_drop(hparent);
1548 	}
1549 	if (hcluster) {
1550 		hammer2_cluster_unlock(hcluster);
1551 		hammer2_cluster_drop(hcluster);
1552 	}
1553 	if (hlinkp)
1554 		*hlinkp = hlink;
1555 
1556 	return error;
1557 }
1558 
1559 /*
1560  * This is called from the mount code to initialize pmp->ihidden
1561  */
1562 void
1563 hammer2_inode_install_hidden(hammer2_pfs_t *pmp)
1564 {
1565 	hammer2_trans_t trans;
1566 	hammer2_cluster_t *cparent;
1567 	hammer2_cluster_t *cluster;
1568 	hammer2_cluster_t *scan;
1569 	const hammer2_inode_data_t *ripdata;
1570 	hammer2_inode_data_t *wipdata;
1571 	hammer2_key_t key_dummy;
1572 	hammer2_key_t key_next;
1573 	int error;
1574 	int count;
1575 	int dip_check_algo;
1576 	int dip_comp_algo;
1577 
1578 	if (pmp->ihidden)
1579 		return;
1580 
1581 	/*
1582 	 * Find the hidden directory
1583 	 */
1584 	bzero(&key_dummy, sizeof(key_dummy));
1585 	hammer2_trans_init(&trans, pmp, 0);
1586 
1587 	/*
1588 	 * Setup for lookup, retrieve iroot's check and compression
1589 	 * algorithm request which was likely generated by newfs_hammer2.
1590 	 *
1591 	 * The check/comp fields will probably never be used since inodes
1592 	 * are renamed into the hidden directory and not created relative to
1593 	 * the hidden directory, chain creation inherits from bref.methods,
1594 	 * and data chains inherit from their respective file inode *_algo
1595 	 * fields.
1596 	 */
1597 	hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_ALWAYS);
1598 	cparent = hammer2_inode_cluster(pmp->iroot, HAMMER2_RESOLVE_ALWAYS);
1599 	ripdata = &hammer2_cluster_rdata(cparent)->ipdata;
1600 	dip_check_algo = ripdata->meta.check_algo;
1601 	dip_comp_algo = ripdata->meta.comp_algo;
1602 	ripdata = NULL;
1603 
1604 	cluster = hammer2_cluster_lookup(cparent, &key_dummy,
1605 					 HAMMER2_INODE_HIDDENDIR,
1606 					 HAMMER2_INODE_HIDDENDIR,
1607 					 0);
1608 	if (cluster) {
1609 		pmp->ihidden = hammer2_inode_get(pmp, pmp->iroot, cluster);
1610 		hammer2_inode_ref(pmp->ihidden);
1611 
1612 		/*
1613 		 * Remove any unlinked files which were left open as-of
1614 		 * any system crash.
1615 		 *
1616 		 * Don't pass NODATA, we need the inode data so the delete
1617 		 * can do proper statistics updates.
1618 		 */
1619 		count = 0;
1620 		scan = hammer2_cluster_lookup(cluster, &key_next,
1621 					      0, HAMMER2_TID_MAX, 0);
1622 		while (scan) {
1623 			if (hammer2_cluster_type(scan) ==
1624 			    HAMMER2_BREF_TYPE_INODE) {
1625 				hammer2_cluster_delete(&trans, cluster, scan,
1626 						   HAMMER2_DELETE_PERMANENT);
1627 				++count;
1628 			}
1629 			scan = hammer2_cluster_next(cluster, scan, &key_next,
1630 						    0, HAMMER2_TID_MAX, 0);
1631 		}
1632 
1633 		hammer2_inode_unlock(pmp->ihidden, cluster);
1634 		hammer2_inode_unlock(pmp->iroot, cparent);
1635 		hammer2_trans_done(&trans);
1636 		kprintf("hammer2: PFS loaded hidden dir, "
1637 			"removed %d dead entries\n", count);
1638 		return;
1639 	}
1640 
1641 	/*
1642 	 * Create the hidden directory
1643 	 */
1644 	error = hammer2_cluster_create(&trans, cparent, &cluster,
1645 				       HAMMER2_INODE_HIDDENDIR, 0,
1646 				       HAMMER2_BREF_TYPE_INODE,
1647 				       HAMMER2_INODE_BYTES,
1648 				       0);
1649 	hammer2_inode_unlock(pmp->iroot, cparent);
1650 
1651 	hammer2_cluster_modify(&trans, cluster, 0);
1652 	wipdata = &hammer2_cluster_wdata(cluster)->ipdata;
1653 	wipdata->meta.type = HAMMER2_OBJTYPE_DIRECTORY;
1654 	wipdata->meta.inum = HAMMER2_INODE_HIDDENDIR;
1655 	wipdata->meta.nlinks = 1;
1656 	wipdata->meta.comp_algo = dip_comp_algo;
1657 	wipdata->meta.check_algo = dip_check_algo;
1658 	hammer2_cluster_modsync(cluster);
1659 	kprintf("hammer2: PFS root missing hidden directory, creating\n");
1660 
1661 	pmp->ihidden = hammer2_inode_get(pmp, pmp->iroot, cluster);
1662 	hammer2_inode_ref(pmp->ihidden);
1663 	hammer2_inode_unlock(pmp->ihidden, cluster);
1664 	hammer2_trans_done(&trans);
1665 }
1666 
1667 /*
1668  * If an open file is unlinked H2 needs to retain the file in the topology
1669  * to ensure that its backing store is not recovered by the bulk free scan.
1670  * This also allows us to avoid having to special-case the CHAIN_DELETED flag.
1671  *
1672  * To do this the file is moved to a hidden directory in the PFS root and
1673  * renamed.  The hidden directory must be created if it does not exist.
1674  */
1675 static
1676 void
1677 hammer2_inode_move_to_hidden(hammer2_trans_t *trans,
1678 			     hammer2_cluster_t **cparentp,
1679 			     hammer2_cluster_t **clusterp,
1680 			     hammer2_tid_t inum)
1681 {
1682 	hammer2_cluster_t *dcluster;
1683 	hammer2_pfs_t *pmp;
1684 	int error;
1685 
1686 	pmp = (*clusterp)->pmp;
1687 	KKASSERT(pmp != NULL);
1688 	KKASSERT(pmp->ihidden != NULL);
1689 
1690 	hammer2_cluster_delete(trans, *cparentp, *clusterp, 0);
1691 	hammer2_inode_lock(pmp->ihidden, HAMMER2_RESOLVE_ALWAYS);
1692 	dcluster = hammer2_inode_cluster(pmp->ihidden, HAMMER2_RESOLVE_ALWAYS);
1693 	error = hammer2_inode_connect(trans,
1694 				      NULL/*XXX*/, clusterp, 0,
1695 				      pmp->ihidden, dcluster,
1696 				      NULL, 0, inum);
1697 	hammer2_inode_unlock(pmp->ihidden, dcluster);
1698 	KKASSERT(error == 0);
1699 }
1700 
1701 /*
1702  * Given an exclusively locked inode and cluster we consolidate the cluster
1703  * for hardlink creation, adding (nlinks) to the file's link count and
1704  * potentially relocating the inode to (cdip) which is a parent directory
1705  * common to both the current location of the inode and the intended new
1706  * hardlink.
1707  *
1708  * Replaces (*clusterp) if consolidation occurred, unlocking the old cluster
1709  * and returning a new locked cluster.
1710  *
1711  * NOTE!  This function will also replace ip->cluster.
1712  */
1713 int
1714 hammer2_hardlink_consolidate(hammer2_trans_t *trans,
1715 			     hammer2_inode_t *ip,
1716 			     hammer2_cluster_t **clusterp,
1717 			     hammer2_inode_t *cdip,
1718 			     hammer2_cluster_t *cdcluster,
1719 			     int nlinks)
1720 {
1721 	hammer2_cluster_t *cluster;
1722 	hammer2_cluster_t *cparent;
1723 	int error;
1724 
1725 	cluster = *clusterp;
1726 	if (nlinks == 0 &&			/* no hardlink needed */
1727 	    (ip->meta.name_key & HAMMER2_DIRHASH_VISIBLE)) {
1728 		return (0);
1729 	}
1730 
1731 	if (hammer2_hardlink_enable == 0) {	/* disallow hardlinks */
1732 		hammer2_cluster_unlock(cluster);
1733 		hammer2_cluster_drop(cluster);
1734 		*clusterp = NULL;
1735 		return (ENOTSUP);
1736 	}
1737 
1738 	cparent = NULL;
1739 
1740 	/*
1741 	 * If no change in the hardlink's target directory is required and
1742 	 * this is already a hardlink target, all we need to do is adjust
1743 	 * the link count.
1744 	 */
1745 	if (cdip == ip->pip &&
1746 	    (ip->meta.name_key & HAMMER2_DIRHASH_VISIBLE) == 0) {
1747 		if (nlinks) {
1748 			hammer2_inode_modify(trans, ip);
1749 			ip->meta.nlinks += nlinks;
1750 #if 0
1751 			hammer2_cluster_modify(trans, cluster, 0);
1752 			wipdata = &hammer2_cluster_wdata(cluster)->ipdata;
1753 			wipdata->meta.nlinks += nlinks;
1754 			hammer2_cluster_modsync(cluster);
1755 			ripdata = wipdata;
1756 #endif
1757 		}
1758 		error = 0;
1759 		goto done;
1760 	}
1761 
1762 	/*
1763 	 * Cluster is the real inode.  The originating directory is locked
1764 	 * by the caller so we can manipulate it without worrying about races
1765 	 * against other lookups.
1766 	 *
1767 	 * If cluster is visible we need to delete it from the current
1768 	 * location and create a hardlink pointer in its place.  If it is
1769 	 * not visible we need only delete it.  Then later cluster will be
1770 	 * renamed to a parent directory and converted (if necessary) to
1771 	 * a hidden inode (via shiftup).
1772 	 *
1773 	 * NOTE! We must hold cparent locked through the delete/create/rename
1774 	 *	 operation to ensure that other threads block resolving to
1775 	 *	 the same hardlink, otherwise the other threads may not see
1776 	 *	 the hardlink.
1777 	 */
1778 	KKASSERT((cluster->focus->flags & HAMMER2_CHAIN_DELETED) == 0);
1779 	cparent = hammer2_cluster_parent(cluster);
1780 
1781 	hammer2_cluster_delete(trans, cparent, cluster, 0);
1782 
1783 	KKASSERT(ip->meta.type != HAMMER2_OBJTYPE_HARDLINK);
1784 	if (ip->meta.name_key & HAMMER2_DIRHASH_VISIBLE) {
1785 		const hammer2_inode_data_t *ripdata;
1786 		hammer2_inode_data_t *wipdata;
1787 		hammer2_cluster_t *ncluster;
1788 		hammer2_key_t lhc;
1789 
1790 		ncluster = NULL;
1791 		lhc = cluster->focus->bref.key;
1792 		error = hammer2_cluster_create(trans, cparent, &ncluster,
1793 					     lhc, 0,
1794 					     HAMMER2_BREF_TYPE_INODE,
1795 					     HAMMER2_INODE_BYTES,
1796 					     0);
1797 		hammer2_cluster_modify(trans, ncluster, 0);
1798 		wipdata = &hammer2_cluster_wdata(ncluster)->ipdata;
1799 
1800 		/* wipdata->meta.comp_algo = ip->meta.comp_algo; */
1801 		wipdata->meta.comp_algo = 0;
1802 		wipdata->meta.check_algo = 0;
1803 		wipdata->meta.version = HAMMER2_INODE_VERSION_ONE;
1804 		wipdata->meta.inum = ip->meta.inum;
1805 		wipdata->meta.target_type = ip->meta.type;
1806 		wipdata->meta.type = HAMMER2_OBJTYPE_HARDLINK;
1807 		wipdata->meta.uflags = 0;
1808 		wipdata->meta.rmajor = 0;
1809 		wipdata->meta.rminor = 0;
1810 		wipdata->meta.ctime = 0;
1811 		wipdata->meta.mtime = 0;
1812 		wipdata->meta.atime = 0;
1813 		wipdata->meta.btime = 0;
1814 		bzero(&wipdata->meta.uid, sizeof(wipdata->meta.uid));
1815 		bzero(&wipdata->meta.gid, sizeof(wipdata->meta.gid));
1816 		wipdata->meta.op_flags = HAMMER2_OPFLAG_DIRECTDATA;
1817 		wipdata->meta.cap_flags = 0;
1818 		wipdata->meta.mode = 0;
1819 		wipdata->meta.size = 0;
1820 		wipdata->meta.nlinks = 1;
1821 		wipdata->meta.iparent = 0;	/* XXX */
1822 		wipdata->meta.pfs_type = 0;
1823 		wipdata->meta.pfs_inum = 0;
1824 		bzero(&wipdata->meta.pfs_clid, sizeof(wipdata->meta.pfs_clid));
1825 		bzero(&wipdata->meta.pfs_fsid, sizeof(wipdata->meta.pfs_fsid));
1826 		wipdata->meta.data_quota = 0;
1827 		/* wipdata->data_count = 0; */
1828 		wipdata->meta.inode_quota = 0;
1829 		/* wipdata->inode_count = 0; */
1830 		wipdata->meta.attr_tid = 0;
1831 		wipdata->meta.dirent_tid = 0;
1832 		bzero(&wipdata->u, sizeof(wipdata->u));
1833 		ripdata = &hammer2_cluster_rdata(cluster)->ipdata;
1834 		KKASSERT(ip->meta.name_len <= sizeof(wipdata->filename));
1835 		bcopy(ripdata->filename, wipdata->filename,
1836 		      ip->meta.name_len);
1837 		wipdata->meta.name_key = ncluster->focus->bref.key;
1838 		wipdata->meta.name_len = ip->meta.name_len;
1839 		/* XXX transaction ids */
1840 		hammer2_cluster_modsync(ncluster);
1841 		hammer2_cluster_unlock(ncluster);
1842 		hammer2_cluster_drop(ncluster);
1843 	}
1844 
1845 	/*
1846 	 * cluster represents the hardlink target and is now flagged deleted.
1847 	 * duplicate it to the parent directory and adjust nlinks.
1848 	 *
1849 	 * WARNING! The shiftup() call can cause ncluster to be moved into
1850 	 *	    an indirect block, and our ncluster will wind up pointing
1851 	 *	    to the older/original version.
1852 	 */
1853 	KKASSERT(cluster->focus->flags & HAMMER2_CHAIN_DELETED);
1854 	hammer2_hardlink_shiftup(trans, cluster, ip, cdip, cdcluster,
1855 				 nlinks, &error);
1856 
1857 	if (error == 0)
1858 		hammer2_inode_repoint(ip, cdip, cluster);
1859 
1860 done:
1861 	/*
1862 	 * Cleanup, cluster/ncluster already dealt with.
1863 	 *
1864 	 * Return the shifted cluster in *clusterp.
1865 	 */
1866 	if (cparent) {
1867 		hammer2_cluster_unlock(cparent);
1868 		hammer2_cluster_drop(cparent);
1869 	}
1870 	*clusterp = cluster;
1871 
1872 	return (error);
1873 }
1874 
1875 /*
1876  * If (*ochainp) is non-NULL it points to the forward OBJTYPE_HARDLINK
1877  * inode while (*chainp) points to the resolved (hidden hardlink
1878  * target) inode.  In this situation when nlinks is 1 we wish to
1879  * deconsolidate the hardlink, moving it back to the directory that now
1880  * represents the only remaining link.
1881  */
1882 int
1883 hammer2_hardlink_deconsolidate(hammer2_trans_t *trans,
1884 			       hammer2_inode_t *dip,
1885 			       hammer2_chain_t **chainp,
1886 			       hammer2_chain_t **ochainp)
1887 {
1888 	if (*ochainp == NULL)
1889 		return (0);
1890 	/* XXX */
1891 	return (0);
1892 }
1893 
1894 /*
1895  * The caller presents a locked cluster with an obj_type of
1896  * HAMMER2_OBJTYPE_HARDLINK in (*clusterp).  This routine will locate
1897  * the inode and replace (*clusterp) with a new locked cluster containing
1898  * the target hardlink, also locked.  The original cluster will be
1899  * unlocked and released.
1900  *
1901  * If cparentp is not NULL a locked cluster representing the hardlink's
1902  * parent is also returned.
1903  *
1904  * If we are unable to locate the hardlink target EIO is returned,
1905  * (*cparentp) is set to NULL, the original passed-in (*clusterp)
1906  * will be unlocked and released and (*clusterp) will be set to NULL
1907  * as well.
1908  */
1909 int
1910 hammer2_hardlink_find(hammer2_inode_t *dip,
1911 		      hammer2_cluster_t **cparentp,
1912 		      hammer2_cluster_t **clusterp)
1913 {
1914 	const hammer2_inode_data_t *ipdata;
1915 	hammer2_cluster_t *cluster;
1916 	hammer2_cluster_t *cparent;
1917 	hammer2_cluster_t *rcluster;
1918 	hammer2_inode_t *ip;
1919 	hammer2_inode_t *pip;
1920 	hammer2_key_t key_dummy;
1921 	hammer2_key_t lhc;
1922 
1923 	cluster = *clusterp;
1924 	pip = dip;
1925 	hammer2_inode_ref(pip);		/* for loop */
1926 
1927 	/*
1928 	 * Locate the hardlink.  pip is referenced and not locked.
1929 	 * Unlock and release (*clusterp) after extracting the needed
1930 	 * data.
1931 	 */
1932 	ipdata = &hammer2_cluster_rdata(cluster)->ipdata;
1933 	lhc = ipdata->meta.inum;
1934 	ipdata = NULL;			/* safety */
1935 	hammer2_cluster_unlock(cluster);
1936 	hammer2_cluster_drop(cluster);
1937 	*clusterp = NULL;		/* safety */
1938 
1939 	rcluster = NULL;
1940 	cparent = NULL;
1941 
1942 	while ((ip = pip) != NULL) {
1943 		hammer2_inode_lock(ip, HAMMER2_RESOLVE_ALWAYS);
1944 		cparent = hammer2_inode_cluster(ip, HAMMER2_RESOLVE_ALWAYS);
1945 		hammer2_inode_drop(ip);			/* loop */
1946 		KKASSERT(hammer2_cluster_type(cparent) ==
1947 			 HAMMER2_BREF_TYPE_INODE);
1948 		rcluster = hammer2_cluster_lookup(cparent, &key_dummy,
1949 					     lhc, lhc, 0);
1950 		if (rcluster)
1951 			break;
1952 		hammer2_cluster_lookup_done(cparent);	/* discard parent */
1953 		cparent = NULL;				/* safety */
1954 		pip = ip->pip;		/* safe, ip held locked */
1955 		if (pip)
1956 			hammer2_inode_ref(pip);		/* loop */
1957 		hammer2_inode_unlock(ip, NULL);
1958 	}
1959 
1960 	/*
1961 	 * chain is locked, ip is locked.  Unlock ip, return the locked
1962 	 * chain.  *ipp is already set w/a ref count and not locked.
1963 	 *
1964 	 * (cparent is already unlocked).
1965 	 */
1966 	*clusterp = rcluster;
1967 	if (rcluster) {
1968 		if (cparentp) {
1969 			*cparentp = cparent;
1970 			hammer2_inode_unlock(ip, NULL);
1971 		} else {
1972 			hammer2_inode_unlock(ip, cparent);
1973 		}
1974 		return (0);
1975 	} else {
1976 		if (cparentp)
1977 			*cparentp = NULL;
1978 		if (ip)
1979 			hammer2_inode_unlock(ip, cparent);
1980 		return (EIO);
1981 	}
1982 }
1983 
1984 /*
1985  * Find the directory common to both fdip and tdip.
1986  *
1987  * Returns a held but not locked inode.  Caller typically locks the inode,
1988  * and when through unlocks AND drops it.
1989  */
1990 hammer2_inode_t *
1991 hammer2_inode_common_parent(hammer2_inode_t *fdip, hammer2_inode_t *tdip)
1992 {
1993 	hammer2_inode_t *scan1;
1994 	hammer2_inode_t *scan2;
1995 
1996 	/*
1997 	 * We used to have a depth field but it complicated matters too
1998 	 * much for directory renames.  So now its ugly.  Check for
1999 	 * simple cases before giving up and doing it the expensive way.
2000 	 *
2001 	 * XXX need a bottom-up topology stability lock
2002 	 */
2003 	if (fdip == tdip || fdip == tdip->pip) {
2004 		hammer2_inode_ref(fdip);
2005 		return(fdip);
2006 	}
2007 	if (fdip->pip == tdip) {
2008 		hammer2_inode_ref(tdip);
2009 		return(tdip);
2010 	}
2011 
2012 	/*
2013 	 * XXX not MPSAFE
2014 	 */
2015 	for (scan1 = fdip; scan1->pmp == fdip->pmp; scan1 = scan1->pip) {
2016 		scan2 = tdip;
2017 		while (scan2->pmp == tdip->pmp) {
2018 			if (scan1 == scan2) {
2019 				hammer2_inode_ref(scan1);
2020 				return(scan1);
2021 			}
2022 			scan2 = scan2->pip;
2023 			if (scan2 == NULL)
2024 				break;
2025 		}
2026 	}
2027 	panic("hammer2_inode_common_parent: no common parent %p %p\n",
2028 	      fdip, tdip);
2029 	/* NOT REACHED */
2030 	return(NULL);
2031 }
2032 
2033 /*
2034  * Set an inode's cluster modified, marking the related chains RW and
2035  * duplicating them if necessary.
2036  *
2037  * The passed-in chain is a localized copy of the chain previously acquired
2038  * when the inode was locked (and possilby replaced in the mean time), and
2039  * must also be updated.  In fact, we update it first and then synchronize
2040  * the inode's cluster cache.
2041  */
2042 void
2043 hammer2_inode_modify(hammer2_trans_t *trans, hammer2_inode_t *ip)
2044 {
2045 	atomic_set_int(&ip->flags, HAMMER2_INODE_MODIFIED);
2046 	if (ip->vp)
2047 		vsetisdirty(ip->vp);
2048 }
2049 
2050 /*
2051  * Synchronize the inode's frontend state with the chain state prior
2052  * to any explicit flush of the inode or any strategy write call.
2053  *
2054  * Called with a locked inode.
2055  */
2056 void
2057 hammer2_inode_fsync(hammer2_trans_t *trans, hammer2_inode_t *ip,
2058 		    hammer2_cluster_t *cparent)
2059 {
2060 	int clear_directdata = 0;
2061 
2062 	/* temporary hack, allow cparent to be NULL */
2063 	if (cparent == NULL) {
2064 		cparent = hammer2_inode_cluster(ip, HAMMER2_RESOLVE_ALWAYS);
2065 		hammer2_inode_fsync(trans, ip, cparent);
2066 		hammer2_cluster_unlock(cparent);
2067 		hammer2_cluster_drop(cparent);
2068 		return;
2069 	}
2070 
2071 	if ((ip->flags & HAMMER2_INODE_RESIZED) == 0) {
2072 		/* do nothing */
2073 	} else if (ip->meta.size < ip->osize) {
2074 		/*
2075 		 * We must delete any chains beyond the EOF.  The chain
2076 		 * straddling the EOF will be pending in the bioq.
2077 		 */
2078 		hammer2_cluster_t *dparent;
2079 		hammer2_cluster_t *cluster;
2080 		hammer2_key_t lbase;
2081 		hammer2_key_t key_next;
2082 
2083 		lbase = (ip->meta.size + HAMMER2_PBUFMASK64) &
2084 			~HAMMER2_PBUFMASK64;
2085 		dparent = hammer2_cluster_lookup_init(&ip->cluster, 0);
2086 		cluster = hammer2_cluster_lookup(dparent, &key_next,
2087 					         lbase, (hammer2_key_t)-1,
2088 						 HAMMER2_LOOKUP_NODATA);
2089 		while (cluster) {
2090 			/*
2091 			 * Degenerate embedded case, nothing to loop on
2092 			 */
2093 			switch (hammer2_cluster_type(cluster)) {
2094 			case HAMMER2_BREF_TYPE_INODE:
2095 				hammer2_cluster_unlock(cluster);
2096 				hammer2_cluster_drop(cluster);
2097 				cluster = NULL;
2098 				break;
2099 			case HAMMER2_BREF_TYPE_DATA:
2100 				hammer2_cluster_delete(trans, dparent, cluster,
2101 						   HAMMER2_DELETE_PERMANENT);
2102 				/* fall through */
2103 			default:
2104 				cluster = hammer2_cluster_next(dparent, cluster,
2105 						   &key_next,
2106 						   key_next, (hammer2_key_t)-1,
2107 						   HAMMER2_LOOKUP_NODATA);
2108 				break;
2109 			}
2110 		}
2111 		hammer2_cluster_lookup_done(dparent);
2112 		atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED);
2113 		KKASSERT(ip->flags & HAMMER2_INODE_MODIFIED);
2114 	} else if (ip->meta.size > ip->osize) {
2115 		/*
2116 		 * When resizing larger we may not have any direct-data
2117 		 * available.
2118 		 */
2119 		if ((ip->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) &&
2120 		    ip->meta.size > HAMMER2_EMBEDDED_BYTES) {
2121 			ip->meta.op_flags &= ~HAMMER2_OPFLAG_DIRECTDATA;
2122 			clear_directdata = 1;
2123 		}
2124 		atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED);
2125 		KKASSERT(ip->flags & HAMMER2_INODE_MODIFIED);
2126 	} else {
2127 		/*
2128 		 * RESIZED was set but size didn't change.
2129 		 */
2130 		atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED);
2131 		KKASSERT(ip->flags & HAMMER2_INODE_MODIFIED);
2132 	}
2133 
2134 	/*
2135 	 * Sync inode meta-data
2136 	 */
2137 	if (ip->flags & HAMMER2_INODE_MODIFIED) {
2138 		hammer2_inode_data_t *wipdata;
2139 
2140 		atomic_clear_int(&ip->flags, HAMMER2_INODE_MODIFIED);
2141 		hammer2_cluster_modify(trans, cparent, 0);
2142 		hammer2_inode_repoint(ip, NULL, cparent);
2143 
2144 		wipdata = &hammer2_cluster_wdata(cparent)->ipdata;
2145 		wipdata->meta = ip->meta;
2146 		if (clear_directdata) {
2147 			bzero(&wipdata->u.blockset,
2148 			      sizeof(wipdata->u.blockset));
2149 		}
2150 		hammer2_cluster_modsync(cparent);
2151 	}
2152 }
2153