xref: /dragonfly/sys/vfs/hammer2/hammer2_inode.c (revision c3762235)
1 /*
2  * Copyright (c) 2011-2013 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 static void hammer2_inode_move_to_hidden(hammer2_trans_t *trans,
45 					 hammer2_chain_t **chainp,
46 					 hammer2_tid_t inum);
47 
48 RB_GENERATE2(hammer2_inode_tree, hammer2_inode, rbnode, hammer2_inode_cmp,
49 	     hammer2_tid_t, inum);
50 
51 int
52 hammer2_inode_cmp(hammer2_inode_t *ip1, hammer2_inode_t *ip2)
53 {
54 	if (ip1->inum < ip2->inum)
55 		return(-1);
56 	if (ip1->inum > ip2->inum)
57 		return(1);
58 	return(0);
59 }
60 
61 /*
62  * HAMMER2 inode locks
63  *
64  * HAMMER2 offers shared locks and exclusive locks on inodes.
65  *
66  * An inode's ip->chain pointer is resolved and stable while an inode is
67  * locked, and can be cleaned out at any time (become NULL) when an inode
68  * is not locked.
69  *
70  * This function handles duplication races and hardlink replacement races
71  * which can cause ip's cached chain to become stale.
72  *
73  * The underlying chain is also locked and returned.
74  *
75  * NOTE: We don't combine the inode/chain lock because putting away an
76  *       inode would otherwise confuse multiple lock holders of the inode.
77  */
78 hammer2_chain_t *
79 hammer2_inode_lock_ex(hammer2_inode_t *ip)
80 {
81 	hammer2_chain_t *chain;
82 	hammer2_chain_t *ochain;
83 	hammer2_chain_core_t *core;
84 	int error;
85 
86 	hammer2_inode_ref(ip);
87 	ccms_thread_lock(&ip->topo_cst, CCMS_STATE_EXCLUSIVE);
88 
89 	chain = ip->chain;
90 	core = chain->core;
91 	for (;;) {
92 		if (chain->flags & HAMMER2_CHAIN_DUPLICATED) {
93 			spin_lock(&core->cst.spin);
94 			while (chain->flags & HAMMER2_CHAIN_DUPLICATED)
95 				chain = TAILQ_NEXT(chain, core_entry);
96 			hammer2_chain_ref(chain);
97 			spin_unlock(&core->cst.spin);
98 			hammer2_inode_repoint(ip, NULL, chain);
99 			hammer2_chain_drop(chain);
100 		}
101 		hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
102 		if ((chain->flags & HAMMER2_CHAIN_DUPLICATED) == 0)
103 			break;
104 		hammer2_chain_unlock(chain);
105 	}
106 	if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK &&
107 	    (chain->flags & HAMMER2_CHAIN_DELETED) == 0) {
108 		error = hammer2_hardlink_find(ip->pip, &chain, &ochain);
109 		hammer2_chain_drop(ochain);
110 		KKASSERT(error == 0);
111 		/* XXX error handling */
112 	}
113 	return (chain);
114 }
115 
116 void
117 hammer2_inode_unlock_ex(hammer2_inode_t *ip, hammer2_chain_t *chain)
118 {
119 	/*
120 	 * XXX this will catch parent directories too which we don't
121 	 *     really want.
122 	 */
123 	if (chain)
124 		hammer2_chain_unlock(chain);
125 	ccms_thread_unlock(&ip->topo_cst);
126 	hammer2_inode_drop(ip);
127 }
128 
129 /*
130  * NOTE: We don't combine the inode/chain lock because putting away an
131  *       inode would otherwise confuse multiple lock holders of the inode.
132  *
133  *	 Shared locks are especially sensitive to having too many shared
134  *	 lock counts (from the same thread) on certain paths which might
135  *	 need to upgrade them.  Only one count of a shared lock can be
136  *	 upgraded.
137  */
138 hammer2_chain_t *
139 hammer2_inode_lock_sh(hammer2_inode_t *ip)
140 {
141 	hammer2_chain_t *chain;
142 
143 	hammer2_inode_ref(ip);
144 	for (;;) {
145 		ccms_thread_lock(&ip->topo_cst, CCMS_STATE_SHARED);
146 
147 		chain = ip->chain;
148 		KKASSERT(chain != NULL);	/* for now */
149 		hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
150 					  HAMMER2_RESOLVE_SHARED);
151 
152 		/*
153 		 * Resolve duplication races, resolve hardlinks by giving
154 		 * up and cycling an exclusive lock.
155 		 */
156 		if ((chain->flags & HAMMER2_CHAIN_DUPLICATED) == 0 &&
157 		    chain->data->ipdata.type != HAMMER2_OBJTYPE_HARDLINK) {
158 			break;
159 		}
160 		hammer2_chain_unlock(chain);
161 		ccms_thread_unlock(&ip->topo_cst);
162 		chain = hammer2_inode_lock_ex(ip);
163 		hammer2_inode_unlock_ex(ip, chain);
164 	}
165 	return (chain);
166 }
167 
168 void
169 hammer2_inode_unlock_sh(hammer2_inode_t *ip, hammer2_chain_t *chain)
170 {
171 	if (chain)
172 		hammer2_chain_unlock(chain);
173 	ccms_thread_unlock(&ip->topo_cst);
174 	hammer2_inode_drop(ip);
175 }
176 
177 ccms_state_t
178 hammer2_inode_lock_temp_release(hammer2_inode_t *ip)
179 {
180 	return(ccms_thread_lock_temp_release(&ip->topo_cst));
181 }
182 
183 void
184 hammer2_inode_lock_temp_restore(hammer2_inode_t *ip, ccms_state_t ostate)
185 {
186 	ccms_thread_lock_temp_restore(&ip->topo_cst, ostate);
187 }
188 
189 ccms_state_t
190 hammer2_inode_lock_upgrade(hammer2_inode_t *ip)
191 {
192 	return(ccms_thread_lock_upgrade(&ip->topo_cst));
193 }
194 
195 void
196 hammer2_inode_lock_downgrade(hammer2_inode_t *ip, ccms_state_t ostate)
197 {
198 	ccms_thread_lock_downgrade(&ip->topo_cst, ostate);
199 }
200 
201 /*
202  * Lookup an inode by inode number
203  */
204 hammer2_inode_t *
205 hammer2_inode_lookup(hammer2_pfsmount_t *pmp, hammer2_tid_t inum)
206 {
207 	hammer2_inode_t *ip;
208 
209 	if (pmp) {
210 		spin_lock(&pmp->inum_spin);
211 		ip = RB_LOOKUP(hammer2_inode_tree, &pmp->inum_tree, inum);
212 		if (ip)
213 			hammer2_inode_ref(ip);
214 		spin_unlock(&pmp->inum_spin);
215 	} else {
216 		ip = NULL;
217 	}
218 	return(ip);
219 }
220 
221 /*
222  * Adding a ref to an inode is only legal if the inode already has at least
223  * one ref.
224  */
225 void
226 hammer2_inode_ref(hammer2_inode_t *ip)
227 {
228 	atomic_add_int(&ip->refs, 1);
229 }
230 
231 /*
232  * Drop an inode reference, freeing the inode when the last reference goes
233  * away.
234  */
235 void
236 hammer2_inode_drop(hammer2_inode_t *ip)
237 {
238 	hammer2_pfsmount_t *pmp;
239 	hammer2_inode_t *pip;
240 	u_int refs;
241 
242 	while (ip) {
243 		refs = ip->refs;
244 		cpu_ccfence();
245 		if (refs == 1) {
246 			/*
247 			 * Transition to zero, must interlock with
248 			 * the inode inumber lookup tree (if applicable).
249 			 *
250 			 * NOTE: The super-root inode has no pmp.
251 			 */
252 			pmp = ip->pmp;
253 			if (pmp)
254 				spin_lock(&pmp->inum_spin);
255 
256 			if (atomic_cmpset_int(&ip->refs, 1, 0)) {
257 				KKASSERT(ip->topo_cst.count == 0);
258 				if (ip->flags & HAMMER2_INODE_ONRBTREE) {
259 					atomic_clear_int(&ip->flags,
260 						     HAMMER2_INODE_ONRBTREE);
261 					RB_REMOVE(hammer2_inode_tree,
262 						  &pmp->inum_tree, ip);
263 				}
264 				if (pmp)
265 					spin_unlock(&pmp->inum_spin);
266 
267 				pip = ip->pip;
268 				ip->pip = NULL;
269 				ip->pmp = NULL;
270 
271 				/*
272 				 * Cleaning out ip->chain isn't entirely
273 				 * trivial.
274 				 */
275 				hammer2_inode_repoint(ip, NULL, NULL);
276 
277 				/*
278 				 * We have to drop pip (if non-NULL) to
279 				 * dispose of our implied reference from
280 				 * ip->pip.  We can simply loop on it.
281 				 */
282 				if (pmp) {
283 					KKASSERT((ip->flags &
284 						  HAMMER2_INODE_SROOT) == 0);
285 					kfree(ip, pmp->minode);
286 					atomic_add_long(&pmp->inmem_inodes, -1);
287 				} else {
288 					KKASSERT(ip->flags &
289 						 HAMMER2_INODE_SROOT);
290 					kfree(ip, M_HAMMER2);
291 				}
292 				ip = pip;
293 				/* continue with pip (can be NULL) */
294 			} else {
295 				if (pmp)
296 					spin_unlock(&ip->pmp->inum_spin);
297 			}
298 		} else {
299 			/*
300 			 * Non zero transition
301 			 */
302 			if (atomic_cmpset_int(&ip->refs, refs, refs - 1))
303 				break;
304 		}
305 	}
306 }
307 
308 /*
309  * Get the vnode associated with the given inode, allocating the vnode if
310  * necessary.  The vnode will be returned exclusively locked.
311  *
312  * The caller must lock the inode (shared or exclusive).
313  *
314  * Great care must be taken to avoid deadlocks and vnode acquisition/reclaim
315  * races.
316  */
317 struct vnode *
318 hammer2_igetv(hammer2_inode_t *ip, int *errorp)
319 {
320 	hammer2_inode_data_t *ipdata;
321 	hammer2_pfsmount_t *pmp;
322 	struct vnode *vp;
323 	ccms_state_t ostate;
324 
325 	pmp = ip->pmp;
326 	KKASSERT(pmp != NULL);
327 	*errorp = 0;
328 	ipdata = &ip->chain->data->ipdata;
329 
330 	for (;;) {
331 		/*
332 		 * Attempt to reuse an existing vnode assignment.  It is
333 		 * possible to race a reclaim so the vget() may fail.  The
334 		 * inode must be unlocked during the vget() to avoid a
335 		 * deadlock against a reclaim.
336 		 */
337 		vp = ip->vp;
338 		if (vp) {
339 			/*
340 			 * Inode must be unlocked during the vget() to avoid
341 			 * possible deadlocks, but leave the ip ref intact.
342 			 *
343 			 * vnode is held to prevent destruction during the
344 			 * vget().  The vget() can still fail if we lost
345 			 * a reclaim race on the vnode.
346 			 */
347 			vhold(vp);
348 			ostate = hammer2_inode_lock_temp_release(ip);
349 			if (vget(vp, LK_EXCLUSIVE)) {
350 				vdrop(vp);
351 				hammer2_inode_lock_temp_restore(ip, ostate);
352 				continue;
353 			}
354 			hammer2_inode_lock_temp_restore(ip, ostate);
355 			vdrop(vp);
356 			/* vp still locked and ref from vget */
357 			if (ip->vp != vp) {
358 				kprintf("hammer2: igetv race %p/%p\n",
359 					ip->vp, vp);
360 				vput(vp);
361 				continue;
362 			}
363 			*errorp = 0;
364 			break;
365 		}
366 
367 		/*
368 		 * No vnode exists, allocate a new vnode.  Beware of
369 		 * allocation races.  This function will return an
370 		 * exclusively locked and referenced vnode.
371 		 */
372 		*errorp = getnewvnode(VT_HAMMER2, pmp->mp, &vp, 0, 0);
373 		if (*errorp) {
374 			kprintf("hammer2: igetv getnewvnode failed %d\n",
375 				*errorp);
376 			vp = NULL;
377 			break;
378 		}
379 
380 		/*
381 		 * Lock the inode and check for an allocation race.
382 		 */
383 		ostate = hammer2_inode_lock_upgrade(ip);
384 		if (ip->vp != NULL) {
385 			vp->v_type = VBAD;
386 			vx_put(vp);
387 			hammer2_inode_lock_downgrade(ip, ostate);
388 			continue;
389 		}
390 
391 		switch (ipdata->type) {
392 		case HAMMER2_OBJTYPE_DIRECTORY:
393 			vp->v_type = VDIR;
394 			break;
395 		case HAMMER2_OBJTYPE_REGFILE:
396 			vp->v_type = VREG;
397 			vinitvmio(vp, ipdata->size,
398 				  HAMMER2_LBUFSIZE,
399 				  (int)ipdata->size & HAMMER2_LBUFMASK);
400 			break;
401 		case HAMMER2_OBJTYPE_SOFTLINK:
402 			/*
403 			 * XXX for now we are using the generic file_read
404 			 * and file_write code so we need a buffer cache
405 			 * association.
406 			 */
407 			vp->v_type = VLNK;
408 			vinitvmio(vp, ipdata->size,
409 				  HAMMER2_LBUFSIZE,
410 				  (int)ipdata->size & HAMMER2_LBUFMASK);
411 			break;
412 		case HAMMER2_OBJTYPE_CDEV:
413 			vp->v_type = VCHR;
414 			/* fall through */
415 		case HAMMER2_OBJTYPE_BDEV:
416 			vp->v_ops = &pmp->mp->mnt_vn_spec_ops;
417 			if (ipdata->type != HAMMER2_OBJTYPE_CDEV)
418 				vp->v_type = VBLK;
419 			addaliasu(vp, ipdata->rmajor, ipdata->rminor);
420 			break;
421 		case HAMMER2_OBJTYPE_FIFO:
422 			vp->v_type = VFIFO;
423 			vp->v_ops = &pmp->mp->mnt_vn_fifo_ops;
424 			break;
425 		default:
426 			panic("hammer2: unhandled objtype %d", ipdata->type);
427 			break;
428 		}
429 
430 		if (ip == pmp->iroot)
431 			vsetflags(vp, VROOT);
432 
433 		vp->v_data = ip;
434 		ip->vp = vp;
435 		hammer2_inode_ref(ip);		/* vp association */
436 		hammer2_inode_lock_downgrade(ip, ostate);
437 		break;
438 	}
439 
440 	/*
441 	 * Return non-NULL vp and *errorp == 0, or NULL vp and *errorp != 0.
442 	 */
443 	if (hammer2_debug & 0x0002) {
444 		kprintf("igetv vp %p refs 0x%08x aux 0x%08x\n",
445 			vp, vp->v_refcnt, vp->v_auxrefs);
446 	}
447 	return (vp);
448 }
449 
450 /*
451  * The passed-in chain must be locked and the returned inode will also be
452  * locked.  This routine typically locates or allocates the inode, assigns
453  * ip->chain (adding a ref to chain if necessary), and returns the inode.
454  *
455  * The hammer2_inode structure regulates the interface between the high level
456  * kernel VNOPS API and the filesystem backend (the chains).
457  *
458  * WARNING!  This routine sucks up the chain's lock (makes it part of the
459  *	     inode lock from the point of view of the inode lock API),
460  *	     so callers need to be careful.
461  *
462  * WARNING!  The mount code is allowed to pass dip == NULL for iroot and
463  *	     is allowed to pass pmp == NULL and dip == NULL for sroot.
464  */
465 hammer2_inode_t *
466 hammer2_inode_get(hammer2_pfsmount_t *pmp, hammer2_inode_t *dip,
467 		  hammer2_chain_t *chain)
468 {
469 	hammer2_inode_t *nip;
470 
471 	KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
472 
473 	/*
474 	 * Interlocked lookup/ref of the inode.  This code is only needed
475 	 * when looking up inodes with nlinks != 0 (TODO: optimize out
476 	 * otherwise and test for duplicates).
477 	 */
478 again:
479 	for (;;) {
480 		nip = hammer2_inode_lookup(pmp, chain->data->ipdata.inum);
481 		if (nip == NULL)
482 			break;
483 		ccms_thread_lock(&nip->topo_cst, CCMS_STATE_EXCLUSIVE);
484 		if ((nip->flags & HAMMER2_INODE_ONRBTREE) == 0) { /* race */
485 			ccms_thread_unlock(&nip->topo_cst);
486 			hammer2_inode_drop(nip);
487 			continue;
488 		}
489 		if (nip->chain != chain)
490 			hammer2_inode_repoint(nip, NULL, chain);
491 
492 		/*
493 		 * Consolidated nip/nip->chain is locked (chain locked
494 		 * by caller).
495 		 */
496 		return nip;
497 	}
498 
499 	/*
500 	 * We couldn't find the inode number, create a new inode.
501 	 */
502 	if (pmp) {
503 		nip = kmalloc(sizeof(*nip), pmp->minode, M_WAITOK | M_ZERO);
504 		atomic_add_long(&pmp->inmem_inodes, 1);
505 		hammer2_chain_memory_inc(pmp);
506 		hammer2_chain_memory_wakeup(pmp);
507 	} else {
508 		nip = kmalloc(sizeof(*nip), M_HAMMER2, M_WAITOK | M_ZERO);
509 		nip->flags = HAMMER2_INODE_SROOT;
510 	}
511 	nip->inum = chain->data->ipdata.inum;
512 	nip->size = chain->data->ipdata.size;
513 	nip->mtime = chain->data->ipdata.mtime;
514 	hammer2_inode_repoint(nip, NULL, chain);
515 	nip->pip = dip;				/* can be NULL */
516 	if (dip)
517 		hammer2_inode_ref(dip);	/* ref dip for nip->pip */
518 
519 	nip->pmp = pmp;
520 
521 	/*
522 	 * ref and lock on nip gives it state compatible to after a
523 	 * hammer2_inode_lock_ex() call.
524 	 */
525 	nip->refs = 1;
526 	ccms_cst_init(&nip->topo_cst, &nip->chain);
527 	ccms_thread_lock(&nip->topo_cst, CCMS_STATE_EXCLUSIVE);
528 	/* combination of thread lock and chain lock == inode lock */
529 
530 	/*
531 	 * Attempt to add the inode.  If it fails we raced another inode
532 	 * get.  Undo all the work and try again.
533 	 */
534 	if (pmp) {
535 		spin_lock(&pmp->inum_spin);
536 		if (RB_INSERT(hammer2_inode_tree, &pmp->inum_tree, nip)) {
537 			spin_unlock(&pmp->inum_spin);
538 			ccms_thread_unlock(&nip->topo_cst);
539 			hammer2_inode_drop(nip);
540 			goto again;
541 		}
542 		atomic_set_int(&nip->flags, HAMMER2_INODE_ONRBTREE);
543 		spin_unlock(&pmp->inum_spin);
544 	}
545 
546 	return (nip);
547 }
548 
549 /*
550  * Create a new inode in the specified directory using the vattr to
551  * figure out the type of inode.
552  *
553  * If no error occurs the new inode with its chain locked is returned in
554  * *nipp, otherwise an error is returned and *nipp is set to NULL.
555  *
556  * If vap and/or cred are NULL the related fields are not set and the
557  * inode type defaults to a directory.  This is used when creating PFSs
558  * under the super-root, so the inode number is set to 1 in this case.
559  *
560  * dip is not locked on entry.
561  */
562 hammer2_inode_t *
563 hammer2_inode_create(hammer2_trans_t *trans, hammer2_inode_t *dip,
564 		     struct vattr *vap, struct ucred *cred,
565 		     const uint8_t *name, size_t name_len,
566 		     hammer2_chain_t **chainp, int *errorp)
567 {
568 	hammer2_inode_data_t *dipdata;
569 	hammer2_inode_data_t *nipdata;
570 	hammer2_chain_t *chain;
571 	hammer2_chain_t *parent;
572 	hammer2_inode_t *nip;
573 	hammer2_key_t key_dummy;
574 	hammer2_key_t lhc;
575 	int error;
576 	uid_t xuid;
577 	uuid_t dip_uid;
578 	uuid_t dip_gid;
579 	uint32_t dip_mode;
580 	uint8_t dip_algo;
581 	int cache_index = -1;
582 
583 	lhc = hammer2_dirhash(name, name_len);
584 	*errorp = 0;
585 
586 	/*
587 	 * Locate the inode or indirect block to create the new
588 	 * entry in.  At the same time check for key collisions
589 	 * and iterate until we don't get one.
590 	 *
591 	 * NOTE: hidden inodes do not have iterators.
592 	 */
593 retry:
594 	parent = hammer2_inode_lock_ex(dip);
595 	dipdata = &dip->chain->data->ipdata;
596 	dip_uid = dipdata->uid;
597 	dip_gid = dipdata->gid;
598 	dip_mode = dipdata->mode;
599 	dip_algo = dipdata->comp_algo;
600 
601 	error = 0;
602 	while (error == 0) {
603 		chain = hammer2_chain_lookup(&parent, &key_dummy,
604 					     lhc, lhc, &cache_index, 0);
605 		if (chain == NULL)
606 			break;
607 		if ((lhc & HAMMER2_DIRHASH_VISIBLE) == 0)
608 			error = ENOSPC;
609 		if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
610 			error = ENOSPC;
611 		hammer2_chain_unlock(chain);
612 		chain = NULL;
613 		++lhc;
614 	}
615 
616 	if (error == 0) {
617 		error = hammer2_chain_create(trans, &parent, &chain,
618 					     lhc, 0,
619 					     HAMMER2_BREF_TYPE_INODE,
620 					     HAMMER2_INODE_BYTES);
621 	}
622 
623 	/*
624 	 * Cleanup and handle retries.
625 	 */
626 	if (error == EAGAIN) {
627 		hammer2_chain_ref(parent);
628 		hammer2_inode_unlock_ex(dip, parent);
629 		hammer2_chain_wait(parent);
630 		hammer2_chain_drop(parent);
631 		goto retry;
632 	}
633 	hammer2_inode_unlock_ex(dip, parent);
634 
635 	if (error) {
636 		KKASSERT(chain == NULL);
637 		*errorp = error;
638 		return (NULL);
639 	}
640 
641 	/*
642 	 * Set up the new inode.
643 	 *
644 	 * NOTE: *_get() integrates chain's lock into the inode lock.
645 	 *
646 	 * NOTE: Only one new inode can currently be created per
647 	 *	 transaction.  If the need arises we can adjust
648 	 *	 hammer2_trans_init() to allow more.
649 	 *
650 	 * NOTE: nipdata will have chain's blockset data.
651 	 */
652 	chain->data->ipdata.inum = trans->inode_tid;
653 	nip = hammer2_inode_get(dip->pmp, dip, chain);
654 	nipdata = &chain->data->ipdata;
655 
656 	if (vap) {
657 		KKASSERT(trans->inodes_created == 0);
658 		nipdata->type = hammer2_get_obj_type(vap->va_type);
659 		nipdata->inum = trans->inode_tid;
660 		++trans->inodes_created;
661 
662 		switch (nipdata->type) {
663 		case HAMMER2_OBJTYPE_CDEV:
664 		case HAMMER2_OBJTYPE_BDEV:
665 			nipdata->rmajor = vap->va_rmajor;
666 			nipdata->rminor = vap->va_rminor;
667 			break;
668 		default:
669 			break;
670 		}
671 	} else {
672 		nipdata->type = HAMMER2_OBJTYPE_DIRECTORY;
673 		nipdata->inum = 1;
674 	}
675 
676 	/* Inherit parent's inode compression mode. */
677 	nip->comp_heuristic = 0;
678 	nipdata->comp_algo = dip_algo;
679 	nipdata->version = HAMMER2_INODE_VERSION_ONE;
680 	hammer2_update_time(&nipdata->ctime);
681 	nipdata->mtime = nipdata->ctime;
682 	if (vap)
683 		nipdata->mode = vap->va_mode;
684 	nipdata->nlinks = 1;
685 	if (vap) {
686 		if (dip && dip->pmp) {
687 			xuid = hammer2_to_unix_xid(&dip_uid);
688 			xuid = vop_helper_create_uid(dip->pmp->mp,
689 						     dip_mode,
690 						     xuid,
691 						     cred,
692 						     &vap->va_mode);
693 		} else {
694 			/* super-root has no dip and/or pmp */
695 			xuid = 0;
696 		}
697 		if (vap->va_vaflags & VA_UID_UUID_VALID)
698 			nipdata->uid = vap->va_uid_uuid;
699 		else if (vap->va_uid != (uid_t)VNOVAL)
700 			hammer2_guid_to_uuid(&nipdata->uid, vap->va_uid);
701 		else
702 			hammer2_guid_to_uuid(&nipdata->uid, xuid);
703 
704 		if (vap->va_vaflags & VA_GID_UUID_VALID)
705 			nipdata->gid = vap->va_gid_uuid;
706 		else if (vap->va_gid != (gid_t)VNOVAL)
707 			hammer2_guid_to_uuid(&nipdata->gid, vap->va_gid);
708 		else if (dip)
709 			nipdata->gid = dip_gid;
710 	}
711 
712 	/*
713 	 * Regular files and softlinks allow a small amount of data to be
714 	 * directly embedded in the inode.  This flag will be cleared if
715 	 * the size is extended past the embedded limit.
716 	 */
717 	if (nipdata->type == HAMMER2_OBJTYPE_REGFILE ||
718 	    nipdata->type == HAMMER2_OBJTYPE_SOFTLINK) {
719 		nipdata->op_flags |= HAMMER2_OPFLAG_DIRECTDATA;
720 	}
721 
722 	KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
723 	bcopy(name, nipdata->filename, name_len);
724 	nipdata->name_key = lhc;
725 	nipdata->name_len = name_len;
726 	*chainp = chain;
727 
728 	return (nip);
729 }
730 
731 /*
732  * chain may have been moved around by the create.
733  */
734 void
735 hammer2_chain_refactor(hammer2_chain_t **chainp)
736 {
737 	hammer2_chain_t *chain = *chainp;
738 	hammer2_chain_core_t *core;
739 
740 	core = chain->core;
741 	while (chain->flags & HAMMER2_CHAIN_DUPLICATED) {
742 		spin_lock(&core->cst.spin);
743 		chain = TAILQ_NEXT(chain, core_entry);
744 		while (chain->flags & HAMMER2_CHAIN_DUPLICATED)
745 			chain = TAILQ_NEXT(chain, core_entry);
746 		hammer2_chain_ref(chain);
747 		spin_unlock(&core->cst.spin);
748 		KKASSERT(chain->core == core);
749 
750 		hammer2_chain_unlock(*chainp);
751 		hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
752 					  HAMMER2_RESOLVE_NOREF); /* eat ref */
753 		*chainp = chain;
754 	}
755 }
756 
757 /*
758  * Shift *chainp up to the specified directory, change the filename
759  * to "0xINODENUMBER", and adjust the key.  The chain becomes the
760  * invisible hardlink target.
761  *
762  * The original *chainp has already been marked deleted.
763  */
764 static
765 void
766 hammer2_hardlink_shiftup(hammer2_trans_t *trans, hammer2_chain_t **chainp,
767 			hammer2_inode_t *dip, hammer2_chain_t **dchainp,
768 			int nlinks, int *errorp)
769 {
770 	hammer2_inode_data_t *nipdata;
771 	hammer2_chain_t *chain;
772 	hammer2_chain_t *xchain;
773 	hammer2_key_t key_dummy;
774 	hammer2_key_t lhc;
775 	hammer2_blockref_t bref;
776 	int cache_index = -1;
777 
778 	chain = *chainp;
779 	lhc = chain->data->ipdata.inum;
780 	KKASSERT((lhc & HAMMER2_DIRHASH_VISIBLE) == 0);
781 
782 	/*
783 	 * Locate the inode or indirect block to create the new
784 	 * entry in.  lhc represents the inode number so there is
785 	 * no collision iteration.
786 	 *
787 	 * There should be no key collisions with invisible inode keys.
788 	 *
789 	 * WARNING! Must use inode_lock_ex() on dip to handle a stale
790 	 *	    dip->chain cache.
791 	 */
792 retry:
793 	*errorp = 0;
794 	xchain = hammer2_chain_lookup(dchainp, &key_dummy,
795 				      lhc, lhc, &cache_index, 0);
796 	if (xchain) {
797 		kprintf("X3 chain %p dip %p dchain %p dip->chain %p\n",
798 			xchain, dip, *dchainp, dip->chain);
799 		hammer2_chain_unlock(xchain);
800 		xchain = NULL;
801 		*errorp = ENOSPC;
802 #if 0
803 		Debugger("X3");
804 #endif
805 	}
806 
807 	/*
808 	 * Create entry in common parent directory using the seek position
809 	 * calculated above.
810 	 *
811 	 * We must refactor chain because it might have been shifted into
812 	 * an indirect chain by the create.
813 	 */
814 	if (*errorp == 0) {
815 		KKASSERT(xchain == NULL);
816 #if 0
817 		*errorp = hammer2_chain_create(trans, dchainp, &xchain,
818 					       lhc, 0,
819 					       HAMMER2_BREF_TYPE_INODE,/* n/a */
820 					       HAMMER2_INODE_BYTES);   /* n/a */
821 #endif
822 		/*XXX this somehow isn't working on chain XXX*/
823 		/*KKASSERT(xxx)*/
824 	}
825 
826 	/*
827 	 * Cleanup and handle retries.
828 	 */
829 	if (*errorp == EAGAIN) {
830 		kprintf("R");
831 		hammer2_chain_wait(*dchainp);
832 		hammer2_chain_drop(*dchainp);
833 		goto retry;
834 	}
835 
836 	/*
837 	 * Handle the error case
838 	 */
839 	if (*errorp) {
840 		panic("error2");
841 		KKASSERT(xchain == NULL);
842 		return;
843 	}
844 
845 	/*
846 	 * Use xchain as a placeholder for (lhc).  Duplicate chain to the
847 	 * same target bref as xchain and then delete xchain.  The duplication
848 	 * occurs after xchain in flush order even though xchain is deleted
849 	 * after the duplication. XXX
850 	 *
851 	 * WARNING! Duplications (to a different parent) can cause indirect
852 	 *	    blocks to be inserted, refactor xchain.
853 	 */
854 	bref = chain->bref;
855 	bref.key = lhc;			/* invisible dir entry key */
856 	bref.keybits = 0;
857 	hammer2_chain_duplicate(trans, dchainp, &chain, &bref, 0, 2);
858 
859 	/*
860 	 * chain is now 'live' again.. adjust the filename.
861 	 *
862 	 * Directory entries are inodes but this is a hidden hardlink
863 	 * target.  The name isn't used but to ease debugging give it
864 	 * a name after its inode number.
865 	 */
866 	hammer2_chain_modify(trans, &chain, 0);
867 	nipdata = &chain->data->ipdata;
868 	ksnprintf(nipdata->filename, sizeof(nipdata->filename),
869 		  "0x%016jx", (intmax_t)nipdata->inum);
870 	nipdata->name_len = strlen(nipdata->filename);
871 	nipdata->name_key = lhc;
872 	nipdata->nlinks += nlinks;
873 
874 	*chainp = chain;
875 }
876 
877 /*
878  * Connect the target inode represented by (*chainp) to the media topology
879  * at (dip, name, len).  The caller can pass a rough *chainp, this function
880  * will issue lookup()s to position the parent chain properly for the
881  * chain insertion.
882  *
883  * If hlink is TRUE this function creates an OBJTYPE_HARDLINK directory
884  * entry instead of connecting (*chainp).
885  *
886  * If hlink is FALSE this function uses chain_duplicate() to make a copy
887  * if (*chainp) in the directory entry.  (*chainp) is likely to be deleted
888  * by the caller in this case (e.g. rename).
889  */
890 int
891 hammer2_inode_connect(hammer2_trans_t *trans,
892 		      hammer2_chain_t **chainp, int hlink,
893 		      hammer2_inode_t *dip, hammer2_chain_t **dchainp,
894 		      const uint8_t *name, size_t name_len,
895 		      hammer2_key_t lhc)
896 {
897 	hammer2_inode_data_t *ipdata;
898 	hammer2_chain_t *nchain;
899 	hammer2_chain_t *ochain;
900 	hammer2_key_t key_dummy;
901 	int cache_index = -1;
902 	int error;
903 
904 	/*
905 	 * Since ochain is either disconnected from the topology or represents
906 	 * a hardlink terminus which is always a parent of or equal to dip,
907 	 * we should be able to safely lock dip->chain for our setup.
908 	 *
909 	 * WARNING! Must use inode_lock_ex() on dip to handle a stale
910 	 *	    dip->chain cache.
911 	 */
912 	ochain = *chainp;
913 
914 	/*
915 	 * If name is non-NULL we calculate lhc, else we use the passed-in
916 	 * lhc.
917 	 */
918 	if (name) {
919 		lhc = hammer2_dirhash(name, name_len);
920 
921 		/*
922 		 * Locate the inode or indirect block to create the new
923 		 * entry in.  At the same time check for key collisions
924 		 * and iterate until we don't get one.
925 		 */
926 		error = 0;
927 		while (error == 0) {
928 			nchain = hammer2_chain_lookup(dchainp, &key_dummy,
929 						      lhc, lhc,
930 						      &cache_index, 0);
931 			if (nchain == NULL)
932 				break;
933 			if ((lhc & HAMMER2_DIRHASH_LOMASK) ==
934 			    HAMMER2_DIRHASH_LOMASK) {
935 				error = ENOSPC;
936 			}
937 			hammer2_chain_unlock(nchain);
938 			nchain = NULL;
939 			++lhc;
940 		}
941 	} else {
942 		/*
943 		 * Reconnect to specific key (used when moving
944 		 * unlinked-but-open files into the hidden directory).
945 		 */
946 		nchain = hammer2_chain_lookup(dchainp, &key_dummy,
947 					      lhc, lhc, &cache_index, 0);
948 		KKASSERT(nchain == NULL);
949 	}
950 
951 	if (error == 0) {
952 		if (hlink) {
953 			/*
954 			 * Hardlink pointer needed, create totally fresh
955 			 * directory entry.
956 			 *
957 			 * We must refactor ochain because it might have
958 			 * been shifted into an indirect chain by the
959 			 * create.
960 			 */
961 			KKASSERT(nchain == NULL);
962 			error = hammer2_chain_create(trans, dchainp, &nchain,
963 						     lhc, 0,
964 						     HAMMER2_BREF_TYPE_INODE,
965 						     HAMMER2_INODE_BYTES);
966 			hammer2_chain_refactor(&ochain);
967 		} else {
968 			/*
969 			 * Reconnect the original chain and rename.  Use
970 			 * chain_duplicate().  The caller will likely delete
971 			 * or has already deleted the original chain in
972 			 * this case.
973 			 *
974 			 * NOTE: chain_duplicate() generates a new chain
975 			 *	 with CHAIN_DELETED cleared (ochain typically
976 			 *	 has it set from the file unlink).
977 			 *
978 			 * WARNING! Can cause held-over chains to require a
979 			 *	    refactor.  Fortunately we have none (our
980 			 *	    locked chains are passed into and
981 			 *	    modified by the call).
982 			 */
983 			nchain = ochain;
984 			ochain = NULL;
985 			hammer2_chain_duplicate(trans, NULL, &nchain, NULL,
986 						0, 3);
987 			error = hammer2_chain_create(trans, dchainp, &nchain,
988 						     lhc, 0,
989 						     HAMMER2_BREF_TYPE_INODE,
990 						     HAMMER2_INODE_BYTES);
991 		}
992 	}
993 
994 	/*
995 	 * Unlock stuff.
996 	 */
997 	KKASSERT(error != EAGAIN);
998 
999 	/*
1000 	 * nchain should be NULL on error, leave ochain (== *chainp) alone.
1001 	 */
1002 	if (error) {
1003 		KKASSERT(nchain == NULL);
1004 		return (error);
1005 	}
1006 
1007 	/*
1008 	 * Directory entries are inodes so if the name has changed we have
1009 	 * to update the inode.
1010 	 *
1011 	 * When creating an OBJTYPE_HARDLINK entry remember to unlock the
1012 	 * chain, the caller will access the hardlink via the actual hardlink
1013 	 * target file and not the hardlink pointer entry, so we must still
1014 	 * return ochain.
1015 	 */
1016 	if (hlink && hammer2_hardlink_enable >= 0) {
1017 		/*
1018 		 * Create the HARDLINK pointer.  oip represents the hardlink
1019 		 * target in this situation.
1020 		 *
1021 		 * We will return ochain (the hardlink target).
1022 		 */
1023 		hammer2_chain_modify(trans, &nchain, 0);
1024 		KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
1025 		ipdata = &nchain->data->ipdata;
1026 		bcopy(name, ipdata->filename, name_len);
1027 		ipdata->name_key = lhc;
1028 		ipdata->name_len = name_len;
1029 		ipdata->target_type = ochain->data->ipdata.type;
1030 		ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
1031 		ipdata->inum = ochain->data->ipdata.inum;
1032 		ipdata->nlinks = 1;
1033 		hammer2_chain_unlock(nchain);
1034 		nchain = ochain;
1035 		ochain = NULL;
1036 	} else if (hlink && hammer2_hardlink_enable < 0) {
1037 		/*
1038 		 * Create a snapshot (hardlink fake mode for debugging).
1039 		 * (ochain already flushed above so we can just copy the
1040 		 * bref XXX).
1041 		 *
1042 		 * Since this is a snapshot we return nchain in the fake
1043 		 * hardlink case.
1044 		 */
1045 		hammer2_chain_modify(trans, &nchain, 0);
1046 		KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
1047 		ipdata = &nchain->data->ipdata;
1048 		*ipdata = ochain->data->ipdata;
1049 		bcopy(name, ipdata->filename, name_len);
1050 		ipdata->name_key = lhc;
1051 		ipdata->name_len = name_len;
1052 		atomic_clear_int(&nchain->core->flags,
1053 				 HAMMER2_CORE_COUNTEDBREFS);
1054 		kprintf("created fake hardlink %*.*s\n",
1055 			(int)name_len, (int)name_len, name);
1056 	} else {
1057 		/*
1058 		 * nchain is a duplicate of ochain at the new location.
1059 		 * We must fixup the name stored in oip.  The bref key
1060 		 * has already been set up.
1061 		 */
1062 		hammer2_chain_modify(trans, &nchain, 0);
1063 		ipdata = &nchain->data->ipdata;
1064 
1065 		KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
1066 		bcopy(name, ipdata->filename, name_len);
1067 		ipdata->name_key = lhc;
1068 		ipdata->name_len = name_len;
1069 		ipdata->nlinks = 1;
1070 	}
1071 
1072 	/*
1073 	 * We are replacing ochain with nchain, unlock ochain.  In the
1074 	 * case where ochain is left unchanged the code above sets
1075 	 * nchain to ochain and ochain to NULL, resulting in a NOP here.
1076 	 */
1077 	if (ochain)
1078 		hammer2_chain_unlock(ochain);
1079 	*chainp = nchain;
1080 
1081 	return (0);
1082 }
1083 
1084 /*
1085  * Repoint ip->chain to nchain.  Caller must hold the inode exclusively
1086  * locked.
1087  *
1088  * ip->chain is set to nchain.  The prior chain in ip->chain is dropped
1089  * and nchain is ref'd.
1090  */
1091 void
1092 hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
1093 		      hammer2_chain_t *nchain)
1094 {
1095 	hammer2_chain_t *ochain;
1096 	hammer2_inode_t *opip;
1097 
1098 	/*
1099 	 * Repoint ip->chain if requested.
1100 	 */
1101 	ochain = ip->chain;
1102 	ip->chain = nchain;
1103 	if (nchain)
1104 		hammer2_chain_ref(nchain);
1105 	if (ochain)
1106 		hammer2_chain_drop(ochain);
1107 
1108 	/*
1109 	 * Repoint ip->pip if requested (non-NULL pip).
1110 	 */
1111 	if (pip && ip->pip != pip) {
1112 		opip = ip->pip;
1113 		hammer2_inode_ref(pip);
1114 		ip->pip = pip;
1115 		if (opip)
1116 			hammer2_inode_drop(opip);
1117 	}
1118 }
1119 
1120 /*
1121  * Unlink the file from the specified directory inode.  The directory inode
1122  * does not need to be locked.
1123  *
1124  * isdir determines whether a directory/non-directory check should be made.
1125  * No check is made if isdir is set to -1.
1126  *
1127  * isopen specifies whether special unlink-with-open-descriptor handling
1128  * must be performed.  If set to -1 the caller is deleting a PFS and we
1129  * check whether the chain is mounted or not (chain->pmp != NULL).  1 is
1130  * implied if it is mounted.
1131  *
1132  * If isopen is 1 and nlinks drops to 0 this function must move the chain
1133  * to a special hidden directory until last-close occurs on the file.
1134  *
1135  * NOTE!  The underlying file can still be active with open descriptors
1136  *	  or if the chain is being manually held (e.g. for rename).
1137  *
1138  *	  The caller is responsible for fixing up ip->chain if e.g. a
1139  *	  rename occurs (see chain_duplicate()).
1140  */
1141 int
1142 hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
1143 		    const uint8_t *name, size_t name_len,
1144 		    int isdir, int *hlinkp, struct nchandle *nch)
1145 {
1146 	hammer2_inode_data_t *ipdata;
1147 	hammer2_chain_t *parent;
1148 	hammer2_chain_t *ochain;
1149 	hammer2_chain_t *chain;
1150 	hammer2_chain_t *dparent;
1151 	hammer2_chain_t *dchain;
1152 	hammer2_key_t key_dummy;
1153 	hammer2_key_t key_next;
1154 	hammer2_key_t lhc;
1155 	int error;
1156 	int cache_index = -1;
1157 	uint8_t type;
1158 
1159 	error = 0;
1160 	ochain = NULL;
1161 	lhc = hammer2_dirhash(name, name_len);
1162 
1163 	/*
1164 	 * Search for the filename in the directory
1165 	 */
1166 	if (hlinkp)
1167 		*hlinkp = 0;
1168 	parent = hammer2_inode_lock_ex(dip);
1169 	chain = hammer2_chain_lookup(&parent, &key_next,
1170 				     lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1171 				     &cache_index, 0);
1172 	while (chain) {
1173 		if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1174 		    name_len == chain->data->ipdata.name_len &&
1175 		    bcmp(name, chain->data->ipdata.filename, name_len) == 0) {
1176 			break;
1177 		}
1178 		chain = hammer2_chain_next(&parent, chain, &key_next,
1179 					   key_next,
1180 					   lhc + HAMMER2_DIRHASH_LOMASK,
1181 					   &cache_index, 0);
1182 	}
1183 	hammer2_inode_unlock_ex(dip, NULL);	/* retain parent */
1184 
1185 	/*
1186 	 * Not found or wrong type (isdir < 0 disables the type check).
1187 	 * If a hardlink pointer, type checks use the hardlink target.
1188 	 */
1189 	if (chain == NULL) {
1190 		error = ENOENT;
1191 		goto done;
1192 	}
1193 	if ((type = chain->data->ipdata.type) == HAMMER2_OBJTYPE_HARDLINK) {
1194 		if (hlinkp)
1195 			*hlinkp = 1;
1196 		type = chain->data->ipdata.target_type;
1197 	}
1198 
1199 	if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 0) {
1200 		error = ENOTDIR;
1201 		goto done;
1202 	}
1203 	if (type != HAMMER2_OBJTYPE_DIRECTORY && isdir >= 1) {
1204 		error = EISDIR;
1205 		goto done;
1206 	}
1207 
1208 	/*
1209 	 * Hardlink must be resolved.  We can't hold the parent locked
1210 	 * while we do this or we could deadlock.
1211 	 *
1212 	 * On success chain will be adjusted to point at the hardlink target
1213 	 * and ochain will point to the hardlink pointer in the original
1214 	 * directory.  Otherwise chain remains pointing to the original.
1215 	 */
1216 	if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK) {
1217 		hammer2_chain_unlock(parent);
1218 		parent = NULL;
1219 		error = hammer2_hardlink_find(dip, &chain, &ochain);
1220 	}
1221 
1222 	/*
1223 	 * If this is a directory the directory must be empty.  However, if
1224 	 * isdir < 0 we are doing a rename and the directory does not have
1225 	 * to be empty, and if isdir > 1 we are deleting a PFS/snapshot
1226 	 * and the directory does not have to be empty.
1227 	 *
1228 	 * NOTE: We check the full key range here which covers both visible
1229 	 *	 and invisible entries.  Theoretically there should be no
1230 	 *	 invisible (hardlink target) entries if there are no visible
1231 	 *	 entries.
1232 	 */
1233 	if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 1) {
1234 		dparent = hammer2_chain_lookup_init(chain, 0);
1235 		dchain = hammer2_chain_lookup(&dparent, &key_dummy,
1236 					      0, (hammer2_key_t)-1,
1237 					      &cache_index,
1238 					      HAMMER2_LOOKUP_NODATA);
1239 		if (dchain) {
1240 			hammer2_chain_unlock(dchain);
1241 			hammer2_chain_lookup_done(dparent);
1242 			error = ENOTEMPTY;
1243 			goto done;
1244 		}
1245 		hammer2_chain_lookup_done(dparent);
1246 		dparent = NULL;
1247 		/* dchain NULL */
1248 	}
1249 
1250 	/*
1251 	 * Ok, we can now unlink the chain.  We always decrement nlinks even
1252 	 * if the entry can be deleted in case someone has the file open and
1253 	 * does an fstat().
1254 	 *
1255 	 * The chain itself will no longer be in the on-media topology but
1256 	 * can still be flushed to the media (e.g. if an open descriptor
1257 	 * remains).  When the last vnode/ip ref goes away the chain will
1258 	 * be marked unmodified, avoiding any further (now unnecesary) I/O.
1259 	 *
1260 	 * A non-NULL ochain indicates a hardlink.
1261 	 */
1262 	if (ochain) {
1263 		/*
1264 		 * Delete the original hardlink pointer unconditionally.
1265 		 * (any open descriptors will migrate to the hardlink
1266 		 * target and have no affect on this operation).
1267 		 *
1268 		 * NOTE: parent from above is NULL when ochain != NULL
1269 		 *	 so we can reuse it.
1270 		 */
1271 		hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
1272 		hammer2_chain_delete(trans, ochain, 0);
1273 		hammer2_chain_unlock(ochain);
1274 	}
1275 
1276 	/*
1277 	 * Decrement nlinks on the hardlink target (or original file if
1278 	 * there it was not hardlinked).  Delete the target when nlinks
1279 	 * reaches 0 with special handling if (isopen) is set.
1280 	 *
1281 	 * NOTE! In DragonFly the vnops function calls cache_unlink() after
1282 	 *	 calling us here to clean out the namecache association,
1283 	 *	 (which does not represent a ref for the open-test), and to
1284 	 *	 force finalization of the vnode if/when the last ref gets
1285 	 *	 dropped.
1286 	 *
1287 	 * NOTE! Files are unlinked by rename and then relinked.  nch will be
1288 	 *	 passed as NULL in this situation.  hammer2_inode_connect()
1289 	 *	 will bump nlinks.
1290 	 */
1291 	KKASSERT(chain != NULL);
1292 	hammer2_chain_modify(trans, &chain, 0);
1293 	ipdata = &chain->data->ipdata;
1294 	--ipdata->nlinks;
1295 	if ((int64_t)ipdata->nlinks < 0)	/* XXX debugging */
1296 		ipdata->nlinks = 0;
1297 	if (ipdata->nlinks == 0) {
1298 		if ((chain->flags & HAMMER2_CHAIN_PFSROOT) && chain->pmp) {
1299 			error = EINVAL;
1300 			kprintf("hammer2: PFS \"%s\" cannot be deleted "
1301 				"while still mounted\n",
1302 				ipdata->filename);
1303 			goto done;
1304 		}
1305 		if (nch && cache_isopen(nch)) {
1306 			kprintf("WARNING: unlinking open file\n");
1307 			atomic_set_int(&chain->flags, HAMMER2_CHAIN_UNLINKED);
1308 			hammer2_inode_move_to_hidden(trans, &chain,
1309 						     ipdata->inum);
1310 		} else {
1311 			hammer2_chain_delete(trans, chain, 0);
1312 		}
1313 	}
1314 	error = 0;
1315 done:
1316 	if (chain)
1317 		hammer2_chain_unlock(chain);
1318 	if (parent)
1319 		hammer2_chain_lookup_done(parent);
1320 	if (ochain)
1321 		hammer2_chain_drop(ochain);
1322 
1323 	return error;
1324 }
1325 
1326 /*
1327  * This is called from the mount code to initialize pmp->ihidden
1328  */
1329 void
1330 hammer2_inode_install_hidden(hammer2_pfsmount_t *pmp)
1331 {
1332 	hammer2_trans_t trans;
1333 	hammer2_chain_t *parent;
1334 	hammer2_chain_t *chain;
1335 	hammer2_chain_t *scan;
1336 	hammer2_inode_data_t *ipdata;
1337 	hammer2_key_t key_dummy;
1338 	hammer2_key_t key_next;
1339 	int cache_index;
1340 	int error;
1341 	int count;
1342 
1343 	if (pmp->ihidden)
1344 		return;
1345 
1346 	/*
1347 	 * Find the hidden directory
1348 	 */
1349 	bzero(&key_dummy, sizeof(key_dummy));
1350 	hammer2_trans_init(&trans, pmp, NULL, 0);
1351 
1352 	parent = hammer2_inode_lock_ex(pmp->iroot);
1353 	chain = hammer2_chain_lookup(&parent, &key_dummy,
1354 				     HAMMER2_INODE_HIDDENDIR,
1355 				     HAMMER2_INODE_HIDDENDIR,
1356 				     &cache_index, 0);
1357 	if (chain) {
1358 		pmp->ihidden = hammer2_inode_get(pmp, pmp->iroot, chain);
1359 		hammer2_inode_ref(pmp->ihidden);
1360 
1361 		/*
1362 		 * Remove any unlinked files which were left open as-of
1363 		 * any system crash.
1364 		 */
1365 		count = 0;
1366 		scan = hammer2_chain_lookup(&chain, &key_next,
1367 					    0, HAMMER2_MAX_TID,
1368 					    &cache_index,
1369 					    HAMMER2_LOOKUP_NODATA);
1370 		while (scan) {
1371 			if (scan->bref.type == HAMMER2_BREF_TYPE_INODE) {
1372 				hammer2_chain_delete(&trans, scan, 0);
1373 				++count;
1374 			}
1375 			scan = hammer2_chain_next(&chain, scan, &key_next,
1376 						   0, HAMMER2_MAX_TID,
1377 						   &cache_index,
1378 						   HAMMER2_LOOKUP_NODATA);
1379 		}
1380 
1381 		hammer2_inode_unlock_ex(pmp->ihidden, chain);
1382 		hammer2_inode_unlock_ex(pmp->iroot, parent);
1383 		hammer2_trans_done(&trans);
1384 		kprintf("hammer2: PFS loaded hidden dir, "
1385 			"removed %d dead entries\n", count);
1386 		return;
1387 	}
1388 
1389 	/*
1390 	 * Create the hidden directory
1391 	 */
1392 	error = hammer2_chain_create(&trans, &parent, &chain,
1393 				     HAMMER2_INODE_HIDDENDIR, 0,
1394 				     HAMMER2_BREF_TYPE_INODE,
1395 				     HAMMER2_INODE_BYTES);
1396 	hammer2_inode_unlock_ex(pmp->iroot, parent);
1397 	hammer2_chain_modify(&trans, &chain, 0);
1398 	ipdata = &chain->data->ipdata;
1399 	ipdata->type = HAMMER2_OBJTYPE_DIRECTORY;
1400 	ipdata->inum = HAMMER2_INODE_HIDDENDIR;
1401 	ipdata->nlinks = 1;
1402 	kprintf("hammer2: PFS root missing hidden directory, creating\n");
1403 
1404 	pmp->ihidden = hammer2_inode_get(pmp, pmp->iroot, chain);
1405 	hammer2_inode_ref(pmp->ihidden);
1406 	hammer2_inode_unlock_ex(pmp->ihidden, chain);
1407 	hammer2_trans_done(&trans);
1408 }
1409 
1410 /*
1411  * If an open file is unlinked H2 needs to retain the file in the topology
1412  * to ensure that its backing store is not recovered by the bulk free scan.
1413  * This also allows us to avoid having to special-case the CHAIN_DELETED flag.
1414  *
1415  * To do this the file is moved to a hidden directory in the PFS root and
1416  * renamed.  The hidden directory must be created if it does not exist.
1417  */
1418 static
1419 void
1420 hammer2_inode_move_to_hidden(hammer2_trans_t *trans, hammer2_chain_t **chainp,
1421 			     hammer2_tid_t inum)
1422 {
1423 	hammer2_chain_t *chain;
1424 	hammer2_chain_t *dchain;
1425 	hammer2_pfsmount_t *pmp;
1426 	int error;
1427 
1428 	chain = *chainp;
1429 	pmp = chain->pmp;
1430 	KKASSERT(pmp != NULL);
1431 	KKASSERT(pmp->ihidden != NULL);
1432 	hammer2_chain_delete(trans, chain, 0);
1433 
1434 	dchain = hammer2_inode_lock_ex(pmp->ihidden);
1435         error = hammer2_inode_connect(trans, chainp, 0,
1436                                       pmp->ihidden, &dchain,
1437 				      NULL, 0, inum);
1438 	hammer2_inode_unlock_ex(pmp->ihidden, dchain);
1439 	KKASSERT(error == 0);
1440 }
1441 
1442 /*
1443  * Given an exclusively locked inode and chain we consolidate its chain
1444  * for hardlink creation, adding (nlinks) to the file's link count and
1445  * potentially relocating the inode to a directory common to ip->pip and tdip.
1446  *
1447  * Replaces (*chainp) if consolidation occurred, unlocking the old chain
1448  * and returning a new locked chain.
1449  *
1450  * NOTE!  This function will also replace ip->chain.
1451  */
1452 int
1453 hammer2_hardlink_consolidate(hammer2_trans_t *trans,
1454 			     hammer2_inode_t *ip, hammer2_chain_t **chainp,
1455 			     hammer2_inode_t *cdip, hammer2_chain_t **cdchainp,
1456 			     int nlinks)
1457 {
1458 	hammer2_inode_data_t *ipdata;
1459 	hammer2_chain_t *chain;
1460 	hammer2_chain_t *nchain;
1461 	int error;
1462 
1463 	chain = *chainp;
1464 	if (nlinks == 0 &&			/* no hardlink needed */
1465 	    (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE)) {
1466 		return (0);
1467 	}
1468 	if (hammer2_hardlink_enable < 0) {	/* fake hardlinks */
1469 		return (0);
1470 	}
1471 
1472 	if (hammer2_hardlink_enable == 0) {	/* disallow hardlinks */
1473 		hammer2_chain_unlock(chain);
1474 		*chainp = NULL;
1475 		return (ENOTSUP);
1476 	}
1477 
1478 	/*
1479 	 * If no change in the hardlink's target directory is required and
1480 	 * this is already a hardlink target, all we need to do is adjust
1481 	 * the link count.
1482 	 */
1483 	if (cdip == ip->pip &&
1484 	    (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) == 0) {
1485 		if (nlinks) {
1486 			hammer2_chain_modify(trans, &chain, 0);
1487 			chain->data->ipdata.nlinks += nlinks;
1488 		}
1489 		error = 0;
1490 		goto done;
1491 	}
1492 
1493 
1494 	/*
1495 	 * chain is the real inode.  If it's visible we have to convert it
1496 	 * to a hardlink pointer.  If it is not visible then it is already
1497 	 * a hardlink target and only needs to be deleted.
1498 	 */
1499 	KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0);
1500 	KKASSERT(chain->data->ipdata.type != HAMMER2_OBJTYPE_HARDLINK);
1501 	if (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) {
1502 		/*
1503 		 * We are going to duplicate chain later, causing its
1504 		 * media block to be shifted to the duplicate.  Even though
1505 		 * we are delete-duplicating nchain here it might decide not
1506 		 * to reallocate the block.  Set FORCECOW to force it to.
1507 		 */
1508 		nchain = chain;
1509 		hammer2_chain_lock(nchain, HAMMER2_RESOLVE_ALWAYS);
1510 		atomic_set_int(&nchain->flags, HAMMER2_CHAIN_FORCECOW);
1511 		hammer2_chain_delete_duplicate(trans, &nchain,
1512 					       HAMMER2_DELDUP_RECORE);
1513 		KKASSERT((chain->flags & HAMMER2_CHAIN_DUPLICATED) == 0);
1514 
1515 		ipdata = &nchain->data->ipdata;
1516 		ipdata->target_type = ipdata->type;
1517 		ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
1518 		ipdata->uflags = 0;
1519 		ipdata->rmajor = 0;
1520 		ipdata->rminor = 0;
1521 		ipdata->ctime = 0;
1522 		ipdata->mtime = 0;
1523 		ipdata->atime = 0;
1524 		ipdata->btime = 0;
1525 		bzero(&ipdata->uid, sizeof(ipdata->uid));
1526 		bzero(&ipdata->gid, sizeof(ipdata->gid));
1527 		ipdata->op_flags = HAMMER2_OPFLAG_DIRECTDATA;
1528 		ipdata->cap_flags = 0;
1529 		ipdata->mode = 0;
1530 		ipdata->size = 0;
1531 		ipdata->nlinks = 1;
1532 		ipdata->iparent = 0;	/* XXX */
1533 		ipdata->pfs_type = 0;
1534 		ipdata->pfs_inum = 0;
1535 		bzero(&ipdata->pfs_clid, sizeof(ipdata->pfs_clid));
1536 		bzero(&ipdata->pfs_fsid, sizeof(ipdata->pfs_fsid));
1537 		ipdata->data_quota = 0;
1538 		ipdata->data_count = 0;
1539 		ipdata->inode_quota = 0;
1540 		ipdata->inode_count = 0;
1541 		ipdata->attr_tid = 0;
1542 		ipdata->dirent_tid = 0;
1543 		bzero(&ipdata->u, sizeof(ipdata->u));
1544 		/* XXX transaction ids */
1545 	} else {
1546 		hammer2_chain_delete(trans, chain, 0);
1547 		nchain = NULL;
1548 	}
1549 
1550 	/*
1551 	 * chain represents the hardlink target and is now flagged deleted.
1552 	 * duplicate it to the parent directory and adjust nlinks.
1553 	 *
1554 	 * WARNING! The shiftup() call can cause nchain to be moved into
1555 	 *	    an indirect block, and our nchain will wind up pointing
1556 	 *	    to the older/original version.
1557 	 */
1558 	KKASSERT(chain->flags & HAMMER2_CHAIN_DELETED);
1559 	hammer2_hardlink_shiftup(trans, &chain, cdip, cdchainp, nlinks, &error);
1560 
1561 	if (error == 0)
1562 		hammer2_inode_repoint(ip, cdip, chain);
1563 
1564 	/*
1565 	 * Unlock the original chain last as the lock blocked races against
1566 	 * the creation of the new hardlink target.
1567 	 */
1568 	if (nchain)
1569 		hammer2_chain_unlock(nchain);
1570 
1571 done:
1572 	/*
1573 	 * Cleanup, chain/nchain already dealt with.
1574 	 */
1575 	*chainp = chain;
1576 	hammer2_inode_drop(cdip);
1577 
1578 	return (error);
1579 }
1580 
1581 /*
1582  * If (*ochainp) is non-NULL it points to the forward OBJTYPE_HARDLINK
1583  * inode while (*chainp) points to the resolved (hidden hardlink
1584  * target) inode.  In this situation when nlinks is 1 we wish to
1585  * deconsolidate the hardlink, moving it back to the directory that now
1586  * represents the only remaining link.
1587  */
1588 int
1589 hammer2_hardlink_deconsolidate(hammer2_trans_t *trans,
1590 			       hammer2_inode_t *dip,
1591 			       hammer2_chain_t **chainp,
1592 			       hammer2_chain_t **ochainp)
1593 {
1594 	if (*ochainp == NULL)
1595 		return (0);
1596 	/* XXX */
1597 	return (0);
1598 }
1599 
1600 /*
1601  * The caller presents a locked *chainp pointing to a HAMMER2_BREF_TYPE_INODE
1602  * with an obj_type of HAMMER2_OBJTYPE_HARDLINK.  This routine will gobble
1603  * the *chainp and return a new locked *chainp representing the file target
1604  * (the original *chainp will be unlocked).
1605  *
1606  * When a match is found the chain representing the original HARDLINK
1607  * will be returned in *ochainp with a ref, but not locked.
1608  *
1609  * When no match is found *chainp is set to NULL and EIO is returned.
1610  * (*ochainp) will still be set to the original chain with a ref but not
1611  * locked.
1612  */
1613 int
1614 hammer2_hardlink_find(hammer2_inode_t *dip, hammer2_chain_t **chainp,
1615 		      hammer2_chain_t **ochainp)
1616 {
1617 	hammer2_chain_t *chain = *chainp;
1618 	hammer2_chain_t *parent;
1619 	hammer2_inode_t *ip;
1620 	hammer2_inode_t *pip;
1621 	hammer2_key_t key_dummy;
1622 	hammer2_key_t lhc;
1623 	int cache_index = -1;
1624 
1625 	pip = dip;
1626 	hammer2_inode_ref(pip);		/* for loop */
1627 	hammer2_chain_ref(chain);	/* for (*ochainp) */
1628 	*ochainp = chain;
1629 
1630 	/*
1631 	 * Locate the hardlink.  pip is referenced and not locked,
1632 	 * ipp.
1633 	 *
1634 	 * chain is reused.
1635 	 */
1636 	lhc = chain->data->ipdata.inum;
1637 	hammer2_chain_unlock(chain);
1638 	chain = NULL;
1639 
1640 	while ((ip = pip) != NULL) {
1641 		parent = hammer2_inode_lock_ex(ip);
1642 		hammer2_inode_drop(ip);			/* loop */
1643 		KKASSERT(parent->bref.type == HAMMER2_BREF_TYPE_INODE);
1644 		chain = hammer2_chain_lookup(&parent, &key_dummy,
1645 					     lhc, lhc, &cache_index, 0);
1646 		hammer2_chain_lookup_done(parent);	/* discard parent */
1647 		if (chain)
1648 			break;
1649 		pip = ip->pip;		/* safe, ip held locked */
1650 		if (pip)
1651 			hammer2_inode_ref(pip);		/* loop */
1652 		hammer2_inode_unlock_ex(ip, NULL);
1653 	}
1654 
1655 	/*
1656 	 * chain is locked, ip is locked.  Unlock ip, return the locked
1657 	 * chain.  *ipp is already set w/a ref count and not locked.
1658 	 *
1659 	 * (parent is already unlocked).
1660 	 */
1661 	if (ip)
1662 		hammer2_inode_unlock_ex(ip, NULL);
1663 	*chainp = chain;
1664 	if (chain) {
1665 		KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
1666 		/* already locked */
1667 		return (0);
1668 	} else {
1669 		return (EIO);
1670 	}
1671 }
1672 
1673 /*
1674  * Find the directory common to both fdip and tdip, hold and return
1675  * its inode.
1676  */
1677 hammer2_inode_t *
1678 hammer2_inode_common_parent(hammer2_inode_t *fdip, hammer2_inode_t *tdip)
1679 {
1680 	hammer2_inode_t *scan1;
1681 	hammer2_inode_t *scan2;
1682 
1683 	/*
1684 	 * We used to have a depth field but it complicated matters too
1685 	 * much for directory renames.  So now its ugly.  Check for
1686 	 * simple cases before giving up and doing it the expensive way.
1687 	 *
1688 	 * XXX need a bottom-up topology stability lock
1689 	 */
1690 	if (fdip == tdip || fdip == tdip->pip) {
1691 		hammer2_inode_ref(fdip);
1692 		return(fdip);
1693 	}
1694 	if (fdip->pip == tdip) {
1695 		hammer2_inode_ref(tdip);
1696 		return(tdip);
1697 	}
1698 
1699 	/*
1700 	 * XXX not MPSAFE
1701 	 */
1702 	for (scan1 = fdip; scan1->pmp == fdip->pmp; scan1 = scan1->pip) {
1703 		scan2 = tdip;
1704 		while (scan2->pmp == tdip->pmp) {
1705 			if (scan1 == scan2) {
1706 				hammer2_inode_ref(scan1);
1707 				return(scan1);
1708 			}
1709 			scan2 = scan2->pip;
1710 			if (scan2 == NULL)
1711 				break;
1712 		}
1713 	}
1714 	panic("hammer2_inode_common_parent: no common parent %p %p\n",
1715 	      fdip, tdip);
1716 	/* NOT REACHED */
1717 	return(NULL);
1718 }
1719 
1720 /*
1721  * Synchronize the inode's frontend state with the chain state prior
1722  * to any explicit flush of the inode or any strategy write call.
1723  *
1724  * Called with a locked inode.
1725  */
1726 void
1727 hammer2_inode_fsync(hammer2_trans_t *trans, hammer2_inode_t *ip,
1728 		    hammer2_chain_t **chainp)
1729 {
1730 	hammer2_inode_data_t *ipdata;
1731 	hammer2_chain_t *parent;
1732 	hammer2_chain_t *chain;
1733 	hammer2_key_t lbase;
1734 	hammer2_key_t key_next;
1735 	int cache_index;
1736 
1737 	ipdata = &ip->chain->data->ipdata;
1738 
1739 	if (ip->flags & HAMMER2_INODE_MTIME) {
1740 		ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
1741 		atomic_clear_int(&ip->flags, HAMMER2_INODE_MTIME);
1742 		ipdata->mtime = ip->mtime;
1743 	}
1744 	if ((ip->flags & HAMMER2_INODE_RESIZED) && ip->size < ipdata->size) {
1745 		ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
1746 		ipdata->size = ip->size;
1747 		atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED);
1748 
1749 		/*
1750 		 * We must delete any chains beyond the EOF.  The chain
1751 		 * straddling the EOF will be pending in the bioq.
1752 		 */
1753 		lbase = (ipdata->size + HAMMER2_PBUFMASK64) &
1754 			~HAMMER2_PBUFMASK64;
1755 		parent = hammer2_chain_lookup_init(ip->chain, 0);
1756 		chain = hammer2_chain_lookup(&parent, &key_next,
1757 					     lbase, (hammer2_key_t)-1,
1758 					     &cache_index,
1759 					     HAMMER2_LOOKUP_NODATA);
1760 		while (chain) {
1761 			/*
1762 			 * Degenerate embedded case, nothing to loop on
1763 			 */
1764 			if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1765 				hammer2_chain_unlock(chain);
1766 				break;
1767 			}
1768 			if (chain->bref.type == HAMMER2_BREF_TYPE_DATA) {
1769 				hammer2_chain_delete(trans, chain, 0);
1770 			}
1771 			chain = hammer2_chain_next(&parent, chain, &key_next,
1772 						   key_next, (hammer2_key_t)-1,
1773 						   &cache_index,
1774 						   HAMMER2_LOOKUP_NODATA);
1775 		}
1776 		hammer2_chain_lookup_done(parent);
1777 	} else
1778 	if ((ip->flags & HAMMER2_INODE_RESIZED) && ip->size > ipdata->size) {
1779 		ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
1780 		ipdata->size = ip->size;
1781 		atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED);
1782 
1783 		/*
1784 		 * When resizing larger we may not have any direct-data
1785 		 * available.
1786 		 */
1787 		if ((ipdata->op_flags & HAMMER2_OPFLAG_DIRECTDATA) &&
1788 		    ip->size > HAMMER2_EMBEDDED_BYTES) {
1789 			ipdata->op_flags &= ~HAMMER2_OPFLAG_DIRECTDATA;
1790 			bzero(&ipdata->u.blockset, sizeof(ipdata->u.blockset));
1791 		}
1792 	}
1793 }
1794