xref: /dflybsd-src/sys/vfs/hammer/hammer_inode.c (revision 31c068aaf635ad9fa72dbc4c65b32d890ff7544d)
1 /*
2  * Copyright (c) 2007-2008 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  */
34 
35 #include "hammer.h"
36 #include <vm/vm_extern.h>
37 
38 static int	hammer_unload_inode(struct hammer_inode *ip);
39 static void	hammer_free_inode(hammer_inode_t ip);
40 static void	hammer_flush_inode_core(hammer_inode_t ip,
41 					hammer_flush_group_t flg, int flags);
42 static int	hammer_setup_child_callback(hammer_record_t rec, void *data);
43 #if 0
44 static int	hammer_syncgrp_child_callback(hammer_record_t rec, void *data);
45 #endif
46 static int	hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
47 					hammer_flush_group_t flg);
48 static int	hammer_setup_parent_inodes_helper(hammer_record_t record,
49 					int depth, hammer_flush_group_t flg);
50 static void	hammer_inode_wakereclaims(hammer_inode_t ip);
51 static struct hammer_inostats *hammer_inode_inostats(hammer_mount_t hmp,
52 					pid_t pid);
53 
54 #ifdef DEBUG_TRUNCATE
55 extern struct hammer_inode *HammerTruncIp;
56 #endif
57 
58 struct krate hammer_gen_krate = { 1 };
59 
60 /*
61  * RB-Tree support for inode structures
62  */
63 int
64 hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
65 {
66 	if (ip1->obj_localization < ip2->obj_localization)
67 		return(-1);
68 	if (ip1->obj_localization > ip2->obj_localization)
69 		return(1);
70 	if (ip1->obj_id < ip2->obj_id)
71 		return(-1);
72 	if (ip1->obj_id > ip2->obj_id)
73 		return(1);
74 	if (ip1->obj_asof < ip2->obj_asof)
75 		return(-1);
76 	if (ip1->obj_asof > ip2->obj_asof)
77 		return(1);
78 	return(0);
79 }
80 
81 int
82 hammer_redo_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
83 {
84 	if (ip1->redo_fifo_start < ip2->redo_fifo_start)
85 		return(-1);
86 	if (ip1->redo_fifo_start > ip2->redo_fifo_start)
87 		return(1);
88 	return(0);
89 }
90 
91 /*
92  * RB-Tree support for inode structures / special LOOKUP_INFO
93  */
94 static int
95 hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
96 {
97 	if (info->obj_localization < ip->obj_localization)
98 		return(-1);
99 	if (info->obj_localization > ip->obj_localization)
100 		return(1);
101 	if (info->obj_id < ip->obj_id)
102 		return(-1);
103 	if (info->obj_id > ip->obj_id)
104 		return(1);
105 	if (info->obj_asof < ip->obj_asof)
106 		return(-1);
107 	if (info->obj_asof > ip->obj_asof)
108 		return(1);
109 	return(0);
110 }
111 
112 /*
113  * Used by hammer_scan_inode_snapshots() to locate all of an object's
114  * snapshots.  Note that the asof field is not tested, which we can get
115  * away with because it is the lowest-priority field.
116  */
117 static int
118 hammer_inode_info_cmp_all_history(hammer_inode_t ip, void *data)
119 {
120 	hammer_inode_info_t info = data;
121 
122 	if (ip->obj_localization > info->obj_localization)
123 		return(1);
124 	if (ip->obj_localization < info->obj_localization)
125 		return(-1);
126 	if (ip->obj_id > info->obj_id)
127 		return(1);
128 	if (ip->obj_id < info->obj_id)
129 		return(-1);
130 	return(0);
131 }
132 
133 /*
134  * Used by hammer_unload_pseudofs() to locate all inodes associated with
135  * a particular PFS.
136  */
137 static int
138 hammer_inode_pfs_cmp(hammer_inode_t ip, void *data)
139 {
140 	u_int32_t localization = *(u_int32_t *)data;
141 	if (ip->obj_localization > localization)
142 		return(1);
143 	if (ip->obj_localization < localization)
144 		return(-1);
145 	return(0);
146 }
147 
148 /*
149  * RB-Tree support for pseudofs structures
150  */
151 static int
152 hammer_pfs_rb_compare(hammer_pseudofs_inmem_t p1, hammer_pseudofs_inmem_t p2)
153 {
154 	if (p1->localization < p2->localization)
155 		return(-1);
156 	if (p1->localization > p2->localization)
157 		return(1);
158 	return(0);
159 }
160 
161 
162 RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
163 RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
164 		hammer_inode_info_cmp, hammer_inode_info_t);
165 RB_GENERATE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
166              hammer_pfs_rb_compare, u_int32_t, localization);
167 
168 /*
169  * The kernel is not actively referencing this vnode but is still holding
170  * it cached.
171  *
172  * This is called from the frontend.
173  *
174  * MPALMOSTSAFE
175  */
176 int
177 hammer_vop_inactive(struct vop_inactive_args *ap)
178 {
179 	struct hammer_inode *ip = VTOI(ap->a_vp);
180 	hammer_mount_t hmp;
181 
182 	/*
183 	 * Degenerate case
184 	 */
185 	if (ip == NULL) {
186 		vrecycle(ap->a_vp);
187 		return(0);
188 	}
189 
190 	/*
191 	 * If the inode no longer has visibility in the filesystem try to
192 	 * recycle it immediately, even if the inode is dirty.  Recycling
193 	 * it quickly allows the system to reclaim buffer cache and VM
194 	 * resources which can matter a lot in a heavily loaded system.
195 	 *
196 	 * This can deadlock in vfsync() if we aren't careful.
197 	 *
198 	 * Do not queue the inode to the flusher if we still have visibility,
199 	 * otherwise namespace calls such as chmod will unnecessarily generate
200 	 * multiple inode updates.
201 	 */
202 	if (ip->ino_data.nlinks == 0) {
203 		hmp = ip->hmp;
204 		lwkt_gettoken(&hmp->fs_token);
205 		hammer_inode_unloadable_check(ip, 0);
206 		if (ip->flags & HAMMER_INODE_MODMASK)
207 			hammer_flush_inode(ip, 0);
208 		lwkt_reltoken(&hmp->fs_token);
209 		vrecycle(ap->a_vp);
210 	}
211 	return(0);
212 }
213 
214 /*
215  * Release the vnode association.  This is typically (but not always)
216  * the last reference on the inode.
217  *
218  * Once the association is lost we are on our own with regards to
219  * flushing the inode.
220  *
221  * We must interlock ip->vp so hammer_get_vnode() can avoid races.
222  */
223 int
224 hammer_vop_reclaim(struct vop_reclaim_args *ap)
225 {
226 	struct hammer_inode *ip;
227 	hammer_mount_t hmp;
228 	struct vnode *vp;
229 
230 	vp = ap->a_vp;
231 
232 	if ((ip = vp->v_data) != NULL) {
233 		hmp = ip->hmp;
234 		lwkt_gettoken(&hmp->fs_token);
235 		hammer_lock_ex(&ip->lock);
236 		vp->v_data = NULL;
237 		ip->vp = NULL;
238 
239 		if ((ip->flags & HAMMER_INODE_RECLAIM) == 0) {
240 			++hammer_count_reclaims;
241 			++hmp->count_reclaims;
242 			ip->flags |= HAMMER_INODE_RECLAIM;
243 		}
244 		hammer_unlock(&ip->lock);
245 		vclrisdirty(vp);
246 		hammer_rel_inode(ip, 1);
247 		lwkt_reltoken(&hmp->fs_token);
248 	}
249 	return(0);
250 }
251 
252 /*
253  * Inform the kernel that the inode is dirty.  This will be checked
254  * by vn_unlock().
255  *
256  * Theoretically in order to reclaim a vnode the hammer_vop_reclaim()
257  * must be called which will interlock against our inode lock, so
258  * if VRECLAIMED is not set vp->v_mount (as used by vsetisdirty())
259  * should be stable without having to acquire any new locks.
260  */
261 void
262 hammer_inode_dirty(struct hammer_inode *ip)
263 {
264 	struct vnode *vp;
265 
266 	if ((ip->flags & HAMMER_INODE_MODMASK) &&
267 	    (vp = ip->vp) != NULL &&
268 	    (vp->v_flag & (VRECLAIMED | VISDIRTY)) == 0) {
269 		vsetisdirty(vp);
270 	}
271 }
272 
273 /*
274  * Return a locked vnode for the specified inode.  The inode must be
275  * referenced but NOT LOCKED on entry and will remain referenced on
276  * return.
277  *
278  * Called from the frontend.
279  */
280 int
281 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
282 {
283 	hammer_mount_t hmp;
284 	struct vnode *vp;
285 	int error = 0;
286 	u_int8_t obj_type;
287 
288 	hmp = ip->hmp;
289 
290 	for (;;) {
291 		if ((vp = ip->vp) == NULL) {
292 			error = getnewvnode(VT_HAMMER, hmp->mp, vpp, 0, 0);
293 			if (error)
294 				break;
295 			hammer_lock_ex(&ip->lock);
296 			if (ip->vp != NULL) {
297 				hammer_unlock(&ip->lock);
298 				vp = *vpp;
299 				vp->v_type = VBAD;
300 				vx_put(vp);
301 				continue;
302 			}
303 			hammer_ref(&ip->lock);
304 			vp = *vpp;
305 			ip->vp = vp;
306 
307 			obj_type = ip->ino_data.obj_type;
308 			vp->v_type = hammer_get_vnode_type(obj_type);
309 
310 			hammer_inode_wakereclaims(ip);
311 
312 			switch(ip->ino_data.obj_type) {
313 			case HAMMER_OBJTYPE_CDEV:
314 			case HAMMER_OBJTYPE_BDEV:
315 				vp->v_ops = &hmp->mp->mnt_vn_spec_ops;
316 				addaliasu(vp, ip->ino_data.rmajor,
317 					  ip->ino_data.rminor);
318 				break;
319 			case HAMMER_OBJTYPE_FIFO:
320 				vp->v_ops = &hmp->mp->mnt_vn_fifo_ops;
321 				break;
322 			case HAMMER_OBJTYPE_REGFILE:
323 				break;
324 			default:
325 				break;
326 			}
327 
328 			/*
329 			 * Only mark as the root vnode if the ip is not
330 			 * historical, otherwise the VFS cache will get
331 			 * confused.  The other half of the special handling
332 			 * is in hammer_vop_nlookupdotdot().
333 			 *
334 			 * Pseudo-filesystem roots can be accessed via
335 			 * non-root filesystem paths and setting VROOT may
336 			 * confuse the namecache.  Set VPFSROOT instead.
337 			 */
338 			if (ip->obj_id == HAMMER_OBJID_ROOT &&
339 			    ip->obj_asof == hmp->asof) {
340 				if (ip->obj_localization == 0)
341 					vsetflags(vp, VROOT);
342 				else
343 					vsetflags(vp, VPFSROOT);
344 			}
345 
346 			vp->v_data = (void *)ip;
347 			/* vnode locked by getnewvnode() */
348 			/* make related vnode dirty if inode dirty? */
349 			hammer_unlock(&ip->lock);
350 			if (vp->v_type == VREG) {
351 				vinitvmio(vp, ip->ino_data.size,
352 					  hammer_blocksize(ip->ino_data.size),
353 					  hammer_blockoff(ip->ino_data.size));
354 			}
355 			break;
356 		}
357 
358 		/*
359 		 * Interlock vnode clearing.  This does not prevent the
360 		 * vnode from going into a reclaimed state but it does
361 		 * prevent it from being destroyed or reused so the vget()
362 		 * will properly fail.
363 		 */
364 		hammer_lock_ex(&ip->lock);
365 		if ((vp = ip->vp) == NULL) {
366 			hammer_unlock(&ip->lock);
367 			continue;
368 		}
369 		vhold(vp);
370 		hammer_unlock(&ip->lock);
371 
372 		/*
373 		 * loop if the vget fails (aka races), or if the vp
374 		 * no longer matches ip->vp.
375 		 */
376 		if (vget(vp, LK_EXCLUSIVE) == 0) {
377 			if (vp == ip->vp) {
378 				vdrop(vp);
379 				break;
380 			}
381 			vput(vp);
382 		}
383 		vdrop(vp);
384 	}
385 	*vpp = vp;
386 	return(error);
387 }
388 
389 /*
390  * Locate all copies of the inode for obj_id compatible with the specified
391  * asof, reference, and issue the related call-back.  This routine is used
392  * for direct-io invalidation and does not create any new inodes.
393  */
394 void
395 hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo,
396 		            int (*callback)(hammer_inode_t ip, void *data),
397 			    void *data)
398 {
399 	hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root,
400 				   hammer_inode_info_cmp_all_history,
401 				   callback, iinfo);
402 }
403 
404 /*
405  * Acquire a HAMMER inode.  The returned inode is not locked.  These functions
406  * do not attach or detach the related vnode (use hammer_get_vnode() for
407  * that).
408  *
409  * The flags argument is only applied for newly created inodes, and only
410  * certain flags are inherited.
411  *
412  * Called from the frontend.
413  */
414 struct hammer_inode *
415 hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip,
416 		 int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
417 		 int flags, int *errorp)
418 {
419 	hammer_mount_t hmp = trans->hmp;
420 	struct hammer_node_cache *cachep;
421 	struct hammer_inode_info iinfo;
422 	struct hammer_cursor cursor;
423 	struct hammer_inode *ip;
424 
425 
426 	/*
427 	 * Determine if we already have an inode cached.  If we do then
428 	 * we are golden.
429 	 *
430 	 * If we find an inode with no vnode we have to mark the
431 	 * transaction such that hammer_inode_waitreclaims() is
432 	 * called later on to avoid building up an infinite number
433 	 * of inodes.  Otherwise we can continue to * add new inodes
434 	 * faster then they can be disposed of, even with the tsleep
435 	 * delay.
436 	 *
437 	 * If we find a dummy inode we return a failure so dounlink
438 	 * (which does another lookup) doesn't try to mess with the
439 	 * link count.  hammer_vop_nresolve() uses hammer_get_dummy_inode()
440 	 * to ref dummy inodes.
441 	 */
442 	iinfo.obj_id = obj_id;
443 	iinfo.obj_asof = asof;
444 	iinfo.obj_localization = localization;
445 loop:
446 	ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
447 	if (ip) {
448 		if (ip->flags & HAMMER_INODE_DUMMY) {
449 			*errorp = ENOENT;
450 			return(NULL);
451 		}
452 		hammer_ref(&ip->lock);
453 		*errorp = 0;
454 		return(ip);
455 	}
456 
457 	/*
458 	 * Allocate a new inode structure and deal with races later.
459 	 */
460 	ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
461 	++hammer_count_inodes;
462 	++hmp->count_inodes;
463 	ip->obj_id = obj_id;
464 	ip->obj_asof = iinfo.obj_asof;
465 	ip->obj_localization = localization;
466 	ip->hmp = hmp;
467 	ip->flags = flags & HAMMER_INODE_RO;
468 	ip->cache[0].ip = ip;
469 	ip->cache[1].ip = ip;
470 	ip->cache[2].ip = ip;
471 	ip->cache[3].ip = ip;
472 	if (hmp->ronly)
473 		ip->flags |= HAMMER_INODE_RO;
474 	ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
475 		0x7FFFFFFFFFFFFFFFLL;
476 	RB_INIT(&ip->rec_tree);
477 	TAILQ_INIT(&ip->target_list);
478 	hammer_ref(&ip->lock);
479 
480 	/*
481 	 * Locate the on-disk inode.  If this is a PFS root we always
482 	 * access the current version of the root inode and (if it is not
483 	 * a master) always access information under it with a snapshot
484 	 * TID.
485 	 *
486 	 * We cache recent inode lookups in this directory in dip->cache[2].
487 	 * If we can't find it we assume the inode we are looking for is
488 	 * close to the directory inode.
489 	 */
490 retry:
491 	cachep = NULL;
492 	if (dip) {
493 		if (dip->cache[2].node)
494 			cachep = &dip->cache[2];
495 		else
496 			cachep = &dip->cache[0];
497 	}
498 	hammer_init_cursor(trans, &cursor, cachep, NULL);
499 	cursor.key_beg.localization = localization + HAMMER_LOCALIZE_INODE;
500 	cursor.key_beg.obj_id = ip->obj_id;
501 	cursor.key_beg.key = 0;
502 	cursor.key_beg.create_tid = 0;
503 	cursor.key_beg.delete_tid = 0;
504 	cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
505 	cursor.key_beg.obj_type = 0;
506 
507 	cursor.asof = iinfo.obj_asof;
508 	cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
509 		       HAMMER_CURSOR_ASOF;
510 
511 	*errorp = hammer_btree_lookup(&cursor);
512 	if (*errorp == EDEADLK) {
513 		hammer_done_cursor(&cursor);
514 		goto retry;
515 	}
516 
517 	/*
518 	 * On success the B-Tree lookup will hold the appropriate
519 	 * buffer cache buffers and provide a pointer to the requested
520 	 * information.  Copy the information to the in-memory inode
521 	 * and cache the B-Tree node to improve future operations.
522 	 */
523 	if (*errorp == 0) {
524 		ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
525 		ip->ino_data = cursor.data->inode;
526 
527 		/*
528 		 * cache[0] tries to cache the location of the object inode.
529 		 * The assumption is that it is near the directory inode.
530 		 *
531 		 * cache[1] tries to cache the location of the object data.
532 		 * We might have something in the governing directory from
533 		 * scan optimizations (see the strategy code in
534 		 * hammer_vnops.c).
535 		 *
536 		 * We update dip->cache[2], if possible, with the location
537 		 * of the object inode for future directory shortcuts.
538 		 */
539 		hammer_cache_node(&ip->cache[0], cursor.node);
540 		if (dip) {
541 			if (dip->cache[3].node) {
542 				hammer_cache_node(&ip->cache[1],
543 						  dip->cache[3].node);
544 			}
545 			hammer_cache_node(&dip->cache[2], cursor.node);
546 		}
547 
548 		/*
549 		 * The file should not contain any data past the file size
550 		 * stored in the inode.  Setting save_trunc_off to the
551 		 * file size instead of max reduces B-Tree lookup overheads
552 		 * on append by allowing the flusher to avoid checking for
553 		 * record overwrites.
554 		 */
555 		ip->save_trunc_off = ip->ino_data.size;
556 
557 		/*
558 		 * Locate and assign the pseudofs management structure to
559 		 * the inode.
560 		 */
561 		if (dip && dip->obj_localization == ip->obj_localization) {
562 			ip->pfsm = dip->pfsm;
563 			hammer_ref(&ip->pfsm->lock);
564 		} else {
565 			ip->pfsm = hammer_load_pseudofs(trans,
566 							ip->obj_localization,
567 							errorp);
568 			*errorp = 0;	/* ignore ENOENT */
569 		}
570 	}
571 
572 	/*
573 	 * The inode is placed on the red-black tree and will be synced to
574 	 * the media when flushed or by the filesystem sync.  If this races
575 	 * another instantiation/lookup the insertion will fail.
576 	 */
577 	if (*errorp == 0) {
578 		if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
579 			hammer_free_inode(ip);
580 			hammer_done_cursor(&cursor);
581 			goto loop;
582 		}
583 		ip->flags |= HAMMER_INODE_ONDISK;
584 	} else {
585 		if (ip->flags & HAMMER_INODE_RSV_INODES) {
586 			ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
587 			--hmp->rsv_inodes;
588 		}
589 
590 		hammer_free_inode(ip);
591 		ip = NULL;
592 	}
593 	hammer_done_cursor(&cursor);
594 
595 	/*
596 	 * NEWINODE is only set if the inode becomes dirty later,
597 	 * setting it here just leads to unnecessary stalls.
598 	 *
599 	 * trans->flags |= HAMMER_TRANSF_NEWINODE;
600 	 */
601 	return (ip);
602 }
603 
604 /*
605  * Get a dummy inode to placemark a broken directory entry.
606  */
607 struct hammer_inode *
608 hammer_get_dummy_inode(hammer_transaction_t trans, hammer_inode_t dip,
609 		 int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
610 		 int flags, int *errorp)
611 {
612 	hammer_mount_t hmp = trans->hmp;
613 	struct hammer_inode_info iinfo;
614 	struct hammer_inode *ip;
615 
616 	/*
617 	 * Determine if we already have an inode cached.  If we do then
618 	 * we are golden.
619 	 *
620 	 * If we find an inode with no vnode we have to mark the
621 	 * transaction such that hammer_inode_waitreclaims() is
622 	 * called later on to avoid building up an infinite number
623 	 * of inodes.  Otherwise we can continue to * add new inodes
624 	 * faster then they can be disposed of, even with the tsleep
625 	 * delay.
626 	 *
627 	 * If we find a non-fake inode we return an error.  Only fake
628 	 * inodes can be returned by this routine.
629 	 */
630 	iinfo.obj_id = obj_id;
631 	iinfo.obj_asof = asof;
632 	iinfo.obj_localization = localization;
633 loop:
634 	*errorp = 0;
635 	ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
636 	if (ip) {
637 		if ((ip->flags & HAMMER_INODE_DUMMY) == 0) {
638 			*errorp = ENOENT;
639 			return(NULL);
640 		}
641 		hammer_ref(&ip->lock);
642 		return(ip);
643 	}
644 
645 	/*
646 	 * Allocate a new inode structure and deal with races later.
647 	 */
648 	ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
649 	++hammer_count_inodes;
650 	++hmp->count_inodes;
651 	ip->obj_id = obj_id;
652 	ip->obj_asof = iinfo.obj_asof;
653 	ip->obj_localization = localization;
654 	ip->hmp = hmp;
655 	ip->flags = flags | HAMMER_INODE_RO | HAMMER_INODE_DUMMY;
656 	ip->cache[0].ip = ip;
657 	ip->cache[1].ip = ip;
658 	ip->cache[2].ip = ip;
659 	ip->cache[3].ip = ip;
660 	ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
661 		0x7FFFFFFFFFFFFFFFLL;
662 	RB_INIT(&ip->rec_tree);
663 	TAILQ_INIT(&ip->target_list);
664 	hammer_ref(&ip->lock);
665 
666 	/*
667 	 * Populate the dummy inode.  Leave everything zero'd out.
668 	 *
669 	 * (ip->ino_leaf and ip->ino_data)
670 	 *
671 	 * Make the dummy inode a FIFO object which most copy programs
672 	 * will properly ignore.
673 	 */
674 	ip->save_trunc_off = ip->ino_data.size;
675 	ip->ino_data.obj_type = HAMMER_OBJTYPE_FIFO;
676 
677 	/*
678 	 * Locate and assign the pseudofs management structure to
679 	 * the inode.
680 	 */
681 	if (dip && dip->obj_localization == ip->obj_localization) {
682 		ip->pfsm = dip->pfsm;
683 		hammer_ref(&ip->pfsm->lock);
684 	} else {
685 		ip->pfsm = hammer_load_pseudofs(trans, ip->obj_localization,
686 						errorp);
687 		*errorp = 0;	/* ignore ENOENT */
688 	}
689 
690 	/*
691 	 * The inode is placed on the red-black tree and will be synced to
692 	 * the media when flushed or by the filesystem sync.  If this races
693 	 * another instantiation/lookup the insertion will fail.
694 	 *
695 	 * NOTE: Do not set HAMMER_INODE_ONDISK.  The inode is a fake.
696 	 */
697 	if (*errorp == 0) {
698 		if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
699 			hammer_free_inode(ip);
700 			goto loop;
701 		}
702 	} else {
703 		if (ip->flags & HAMMER_INODE_RSV_INODES) {
704 			ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
705 			--hmp->rsv_inodes;
706 		}
707 		hammer_free_inode(ip);
708 		ip = NULL;
709 	}
710 	trans->flags |= HAMMER_TRANSF_NEWINODE;
711 	return (ip);
712 }
713 
714 /*
715  * Return a referenced inode only if it is in our inode cache.
716  *
717  * Dummy inodes do not count.
718  */
719 struct hammer_inode *
720 hammer_find_inode(hammer_transaction_t trans, int64_t obj_id,
721 		  hammer_tid_t asof, u_int32_t localization)
722 {
723 	hammer_mount_t hmp = trans->hmp;
724 	struct hammer_inode_info iinfo;
725 	struct hammer_inode *ip;
726 
727 	iinfo.obj_id = obj_id;
728 	iinfo.obj_asof = asof;
729 	iinfo.obj_localization = localization;
730 
731 	ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
732 	if (ip) {
733 		if (ip->flags & HAMMER_INODE_DUMMY)
734 			ip = NULL;
735 		else
736 			hammer_ref(&ip->lock);
737 	}
738 	return(ip);
739 }
740 
741 /*
742  * Create a new filesystem object, returning the inode in *ipp.  The
743  * returned inode will be referenced.  The inode is created in-memory.
744  *
745  * If pfsm is non-NULL the caller wishes to create the root inode for
746  * a master PFS.
747  */
748 int
749 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
750 		    struct ucred *cred,
751 		    hammer_inode_t dip, const char *name, int namelen,
752 		    hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
753 {
754 	hammer_mount_t hmp;
755 	hammer_inode_t ip;
756 	uid_t xuid;
757 	int error;
758 	int64_t namekey;
759 	u_int32_t dummy;
760 
761 	hmp = trans->hmp;
762 
763 	/*
764 	 * Disallow the creation of new inodes in directories which
765 	 * have been deleted.  In HAMMER, this will cause a record
766 	 * syncing assertion later on in the flush code.
767 	 */
768 	if (dip && dip->ino_data.nlinks == 0) {
769 		*ipp = NULL;
770                 return (EINVAL);
771 	}
772 
773 	/*
774 	 * Allocate inode
775 	 */
776 	ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
777 	++hammer_count_inodes;
778 	++hmp->count_inodes;
779 	trans->flags |= HAMMER_TRANSF_NEWINODE;
780 
781 	if (pfsm) {
782 		KKASSERT(pfsm->localization != 0);
783 		ip->obj_id = HAMMER_OBJID_ROOT;
784 		ip->obj_localization = pfsm->localization;
785 	} else {
786 		KKASSERT(dip != NULL);
787 		namekey = hammer_directory_namekey(dip, name, namelen, &dummy);
788 		ip->obj_id = hammer_alloc_objid(hmp, dip, namekey);
789 		ip->obj_localization = dip->obj_localization;
790 	}
791 
792 	KKASSERT(ip->obj_id != 0);
793 	ip->obj_asof = hmp->asof;
794 	ip->hmp = hmp;
795 	ip->flush_state = HAMMER_FST_IDLE;
796 	ip->flags = HAMMER_INODE_DDIRTY |
797 		    HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
798 	ip->cache[0].ip = ip;
799 	ip->cache[1].ip = ip;
800 	ip->cache[2].ip = ip;
801 	ip->cache[3].ip = ip;
802 
803 	ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
804 	/* ip->save_trunc_off = 0; (already zero) */
805 	RB_INIT(&ip->rec_tree);
806 	TAILQ_INIT(&ip->target_list);
807 
808 	ip->ino_data.atime = trans->time;
809 	ip->ino_data.mtime = trans->time;
810 	ip->ino_data.size = 0;
811 	ip->ino_data.nlinks = 0;
812 
813 	/*
814 	 * A nohistory designator on the parent directory is inherited by
815 	 * the child.  We will do this even for pseudo-fs creation... the
816 	 * sysad can turn it off.
817 	 */
818 	if (dip) {
819 		ip->ino_data.uflags = dip->ino_data.uflags &
820 				      (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
821 	}
822 
823 	ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
824 	ip->ino_leaf.base.localization = ip->obj_localization +
825 					 HAMMER_LOCALIZE_INODE;
826 	ip->ino_leaf.base.obj_id = ip->obj_id;
827 	ip->ino_leaf.base.key = 0;
828 	ip->ino_leaf.base.create_tid = 0;
829 	ip->ino_leaf.base.delete_tid = 0;
830 	ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
831 	ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
832 
833 	ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
834 	ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
835 	ip->ino_data.mode = vap->va_mode;
836 	ip->ino_data.ctime = trans->time;
837 
838 	/*
839 	 * If we are running version 2 or greater directory entries are
840 	 * inode-localized instead of data-localized.
841 	 */
842 	if (trans->hmp->version >= HAMMER_VOL_VERSION_TWO) {
843 		if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
844 			ip->ino_data.cap_flags |=
845 				HAMMER_INODE_CAP_DIR_LOCAL_INO;
846 		}
847 	}
848 	if (trans->hmp->version >= HAMMER_VOL_VERSION_SIX) {
849 		if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
850 			ip->ino_data.cap_flags |=
851 				HAMMER_INODE_CAP_DIRHASH_ALG1;
852 		}
853 	}
854 
855 	/*
856 	 * Setup the ".." pointer.  This only needs to be done for directories
857 	 * but we do it for all objects as a recovery aid.
858 	 */
859 	if (dip)
860 		ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
861 #if 0
862 	/*
863 	 * The parent_obj_localization field only applies to pseudo-fs roots.
864 	 * XXX this is no longer applicable, PFSs are no longer directly
865 	 * tied into the parent's directory structure.
866 	 */
867 	if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY &&
868 	    ip->obj_id == HAMMER_OBJID_ROOT) {
869 		ip->ino_data.ext.obj.parent_obj_localization =
870 						dip->obj_localization;
871 	}
872 #endif
873 
874 	switch(ip->ino_leaf.base.obj_type) {
875 	case HAMMER_OBJTYPE_CDEV:
876 	case HAMMER_OBJTYPE_BDEV:
877 		ip->ino_data.rmajor = vap->va_rmajor;
878 		ip->ino_data.rminor = vap->va_rminor;
879 		break;
880 	default:
881 		break;
882 	}
883 
884 	/*
885 	 * Calculate default uid/gid and overwrite with information from
886 	 * the vap.
887 	 */
888 	if (dip) {
889 		xuid = hammer_to_unix_xid(&dip->ino_data.uid);
890 		xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
891 					     xuid, cred, &vap->va_mode);
892 	} else {
893 		xuid = 0;
894 	}
895 	ip->ino_data.mode = vap->va_mode;
896 
897 	if (vap->va_vaflags & VA_UID_UUID_VALID)
898 		ip->ino_data.uid = vap->va_uid_uuid;
899 	else if (vap->va_uid != (uid_t)VNOVAL)
900 		hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
901 	else
902 		hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
903 
904 	if (vap->va_vaflags & VA_GID_UUID_VALID)
905 		ip->ino_data.gid = vap->va_gid_uuid;
906 	else if (vap->va_gid != (gid_t)VNOVAL)
907 		hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
908 	else if (dip)
909 		ip->ino_data.gid = dip->ino_data.gid;
910 
911 	hammer_ref(&ip->lock);
912 
913 	if (pfsm) {
914 		ip->pfsm = pfsm;
915 		hammer_ref(&pfsm->lock);
916 		error = 0;
917 	} else if (dip->obj_localization == ip->obj_localization) {
918 		ip->pfsm = dip->pfsm;
919 		hammer_ref(&ip->pfsm->lock);
920 		error = 0;
921 	} else {
922 		ip->pfsm = hammer_load_pseudofs(trans,
923 						ip->obj_localization,
924 						&error);
925 		error = 0;	/* ignore ENOENT */
926 	}
927 
928 	if (error) {
929 		hammer_free_inode(ip);
930 		ip = NULL;
931 	} else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
932 		panic("hammer_create_inode: duplicate obj_id %llx",
933 		      (long long)ip->obj_id);
934 		/* not reached */
935 		hammer_free_inode(ip);
936 	}
937 	*ipp = ip;
938 	return(error);
939 }
940 
941 /*
942  * Final cleanup / freeing of an inode structure
943  */
944 static void
945 hammer_free_inode(hammer_inode_t ip)
946 {
947 	struct hammer_mount *hmp;
948 
949 	hmp = ip->hmp;
950 	KKASSERT(hammer_oneref(&ip->lock));
951 	hammer_uncache_node(&ip->cache[0]);
952 	hammer_uncache_node(&ip->cache[1]);
953 	hammer_uncache_node(&ip->cache[2]);
954 	hammer_uncache_node(&ip->cache[3]);
955 	hammer_inode_wakereclaims(ip);
956 	if (ip->objid_cache)
957 		hammer_clear_objid(ip);
958 	--hammer_count_inodes;
959 	--hmp->count_inodes;
960 	if (ip->pfsm) {
961 		hammer_rel_pseudofs(hmp, ip->pfsm);
962 		ip->pfsm = NULL;
963 	}
964 	kfree(ip, hmp->m_inodes);
965 	ip = NULL;
966 }
967 
968 /*
969  * Retrieve pseudo-fs data.  NULL will never be returned.
970  *
971  * If an error occurs *errorp will be set and a default template is returned,
972  * otherwise *errorp is set to 0.  Typically when an error occurs it will
973  * be ENOENT.
974  */
975 hammer_pseudofs_inmem_t
976 hammer_load_pseudofs(hammer_transaction_t trans,
977 		     u_int32_t localization, int *errorp)
978 {
979 	hammer_mount_t hmp = trans->hmp;
980 	hammer_inode_t ip;
981 	hammer_pseudofs_inmem_t pfsm;
982 	struct hammer_cursor cursor;
983 	int bytes;
984 
985 retry:
986 	pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
987 	if (pfsm) {
988 		hammer_ref(&pfsm->lock);
989 		*errorp = 0;
990 		return(pfsm);
991 	}
992 
993 	/*
994 	 * PFS records are stored in the root inode (not the PFS root inode,
995 	 * but the real root).  Avoid an infinite recursion if loading
996 	 * the PFS for the real root.
997 	 */
998 	if (localization) {
999 		ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
1000 				      HAMMER_MAX_TID,
1001 				      HAMMER_DEF_LOCALIZATION, 0, errorp);
1002 	} else {
1003 		ip = NULL;
1004 	}
1005 
1006 	pfsm = kmalloc(sizeof(*pfsm), hmp->m_misc, M_WAITOK | M_ZERO);
1007 	pfsm->localization = localization;
1008 	pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
1009 	pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
1010 
1011 	hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
1012 	cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION +
1013 				      HAMMER_LOCALIZE_MISC;
1014 	cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
1015 	cursor.key_beg.create_tid = 0;
1016 	cursor.key_beg.delete_tid = 0;
1017 	cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
1018 	cursor.key_beg.obj_type = 0;
1019 	cursor.key_beg.key = localization;
1020 	cursor.asof = HAMMER_MAX_TID;
1021 	cursor.flags |= HAMMER_CURSOR_ASOF;
1022 
1023 	if (ip)
1024 		*errorp = hammer_ip_lookup(&cursor);
1025 	else
1026 		*errorp = hammer_btree_lookup(&cursor);
1027 	if (*errorp == 0) {
1028 		*errorp = hammer_ip_resolve_data(&cursor);
1029 		if (*errorp == 0) {
1030 			if (cursor.data->pfsd.mirror_flags &
1031 			    HAMMER_PFSD_DELETED) {
1032 				*errorp = ENOENT;
1033 			} else {
1034 				bytes = cursor.leaf->data_len;
1035 				if (bytes > sizeof(pfsm->pfsd))
1036 					bytes = sizeof(pfsm->pfsd);
1037 				bcopy(cursor.data, &pfsm->pfsd, bytes);
1038 			}
1039 		}
1040 	}
1041 	hammer_done_cursor(&cursor);
1042 
1043 	pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
1044 	hammer_ref(&pfsm->lock);
1045 	if (ip)
1046 		hammer_rel_inode(ip, 0);
1047 	if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
1048 		kfree(pfsm, hmp->m_misc);
1049 		goto retry;
1050 	}
1051 	return(pfsm);
1052 }
1053 
1054 /*
1055  * Store pseudo-fs data.  The backend will automatically delete any prior
1056  * on-disk pseudo-fs data but we have to delete in-memory versions.
1057  */
1058 int
1059 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
1060 {
1061 	struct hammer_cursor cursor;
1062 	hammer_record_t record;
1063 	hammer_inode_t ip;
1064 	int error;
1065 
1066 	ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1067 			      HAMMER_DEF_LOCALIZATION, 0, &error);
1068 retry:
1069 	pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
1070 	hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
1071 	cursor.key_beg.localization = ip->obj_localization +
1072 				      HAMMER_LOCALIZE_MISC;
1073 	cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
1074 	cursor.key_beg.create_tid = 0;
1075 	cursor.key_beg.delete_tid = 0;
1076 	cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
1077 	cursor.key_beg.obj_type = 0;
1078 	cursor.key_beg.key = pfsm->localization;
1079 	cursor.asof = HAMMER_MAX_TID;
1080 	cursor.flags |= HAMMER_CURSOR_ASOF;
1081 
1082 	/*
1083 	 * Replace any in-memory version of the record.
1084 	 */
1085 	error = hammer_ip_lookup(&cursor);
1086 	if (error == 0 && hammer_cursor_inmem(&cursor)) {
1087 		record = cursor.iprec;
1088 		if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
1089 			KKASSERT(cursor.deadlk_rec == NULL);
1090 			hammer_ref(&record->lock);
1091 			cursor.deadlk_rec = record;
1092 			error = EDEADLK;
1093 		} else {
1094 			record->flags |= HAMMER_RECF_DELETED_FE;
1095 			error = 0;
1096 		}
1097 	}
1098 
1099 	/*
1100 	 * Allocate replacement general record.  The backend flush will
1101 	 * delete any on-disk version of the record.
1102 	 */
1103 	if (error == 0 || error == ENOENT) {
1104 		record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
1105 		record->type = HAMMER_MEM_RECORD_GENERAL;
1106 
1107 		record->leaf.base.localization = ip->obj_localization +
1108 						 HAMMER_LOCALIZE_MISC;
1109 		record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
1110 		record->leaf.base.key = pfsm->localization;
1111 		record->leaf.data_len = sizeof(pfsm->pfsd);
1112 		bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
1113 		error = hammer_ip_add_record(trans, record);
1114 	}
1115 	hammer_done_cursor(&cursor);
1116 	if (error == EDEADLK)
1117 		goto retry;
1118 	hammer_rel_inode(ip, 0);
1119 	return(error);
1120 }
1121 
1122 /*
1123  * Create a root directory for a PFS if one does not alredy exist.
1124  *
1125  * The PFS root stands alone so we must also bump the nlinks count
1126  * to prevent it from being destroyed on release.
1127  */
1128 int
1129 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1130 		       hammer_pseudofs_inmem_t pfsm)
1131 {
1132 	hammer_inode_t ip;
1133 	struct vattr vap;
1134 	int error;
1135 
1136 	ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1137 			      pfsm->localization, 0, &error);
1138 	if (ip == NULL) {
1139 		vattr_null(&vap);
1140 		vap.va_mode = 0755;
1141 		vap.va_type = VDIR;
1142 		error = hammer_create_inode(trans, &vap, cred,
1143 					    NULL, NULL, 0,
1144 					    pfsm, &ip);
1145 		if (error == 0) {
1146 			++ip->ino_data.nlinks;
1147 			hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
1148 		}
1149 	}
1150 	if (ip)
1151 		hammer_rel_inode(ip, 0);
1152 	return(error);
1153 }
1154 
1155 /*
1156  * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
1157  * if we are unable to disassociate all the inodes.
1158  */
1159 static
1160 int
1161 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
1162 {
1163 	int res;
1164 
1165 	hammer_ref(&ip->lock);
1166 	if (hammer_isactive(&ip->lock) == 2 && ip->vp)
1167 		vclean_unlocked(ip->vp);
1168 	if (hammer_isactive(&ip->lock) == 1 && ip->vp == NULL)
1169 		res = 0;
1170 	else
1171 		res = -1;	/* stop, someone is using the inode */
1172 	hammer_rel_inode(ip, 0);
1173 	return(res);
1174 }
1175 
1176 int
1177 hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
1178 {
1179 	int res;
1180 	int try;
1181 
1182 	for (try = res = 0; try < 4; ++try) {
1183 		res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
1184 					   hammer_inode_pfs_cmp,
1185 					   hammer_unload_pseudofs_callback,
1186 					   &localization);
1187 		if (res == 0 && try > 1)
1188 			break;
1189 		hammer_flusher_sync(trans->hmp);
1190 	}
1191 	if (res != 0)
1192 		res = ENOTEMPTY;
1193 	return(res);
1194 }
1195 
1196 
1197 /*
1198  * Release a reference on a PFS
1199  */
1200 void
1201 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
1202 {
1203 	hammer_rel(&pfsm->lock);
1204 	if (hammer_norefs(&pfsm->lock)) {
1205 		RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
1206 		kfree(pfsm, hmp->m_misc);
1207 	}
1208 }
1209 
1210 /*
1211  * Called by hammer_sync_inode().
1212  */
1213 static int
1214 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
1215 {
1216 	hammer_transaction_t trans = cursor->trans;
1217 	hammer_record_t record;
1218 	int error;
1219 	int redirty;
1220 
1221 retry:
1222 	error = 0;
1223 
1224 	/*
1225 	 * If the inode has a presence on-disk then locate it and mark
1226 	 * it deleted, setting DELONDISK.
1227 	 *
1228 	 * The record may or may not be physically deleted, depending on
1229 	 * the retention policy.
1230 	 */
1231 	if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
1232 	    HAMMER_INODE_ONDISK) {
1233 		hammer_normalize_cursor(cursor);
1234 		cursor->key_beg.localization = ip->obj_localization +
1235 					       HAMMER_LOCALIZE_INODE;
1236 		cursor->key_beg.obj_id = ip->obj_id;
1237 		cursor->key_beg.key = 0;
1238 		cursor->key_beg.create_tid = 0;
1239 		cursor->key_beg.delete_tid = 0;
1240 		cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1241 		cursor->key_beg.obj_type = 0;
1242 		cursor->asof = ip->obj_asof;
1243 		cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1244 		cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
1245 		cursor->flags |= HAMMER_CURSOR_BACKEND;
1246 
1247 		error = hammer_btree_lookup(cursor);
1248 		if (hammer_debug_inode)
1249 			kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
1250 
1251 		if (error == 0) {
1252 			error = hammer_ip_delete_record(cursor, ip, trans->tid);
1253 			if (hammer_debug_inode)
1254 				kprintf(" error %d\n", error);
1255 			if (error == 0) {
1256 				ip->flags |= HAMMER_INODE_DELONDISK;
1257 			}
1258 			if (cursor->node)
1259 				hammer_cache_node(&ip->cache[0], cursor->node);
1260 		}
1261 		if (error == EDEADLK) {
1262 			hammer_done_cursor(cursor);
1263 			error = hammer_init_cursor(trans, cursor,
1264 						   &ip->cache[0], ip);
1265 			if (hammer_debug_inode)
1266 				kprintf("IPDED %p %d\n", ip, error);
1267 			if (error == 0)
1268 				goto retry;
1269 		}
1270 	}
1271 
1272 	/*
1273 	 * Ok, write out the initial record or a new record (after deleting
1274 	 * the old one), unless the DELETED flag is set.  This routine will
1275 	 * clear DELONDISK if it writes out a record.
1276 	 *
1277 	 * Update our inode statistics if this is the first application of
1278 	 * the inode on-disk.
1279 	 */
1280 	if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
1281 		/*
1282 		 * Generate a record and write it to the media.  We clean-up
1283 		 * the state before releasing so we do not have to set-up
1284 		 * a flush_group.
1285 		 */
1286 		record = hammer_alloc_mem_record(ip, 0);
1287 		record->type = HAMMER_MEM_RECORD_INODE;
1288 		record->flush_state = HAMMER_FST_FLUSH;
1289 		record->leaf = ip->sync_ino_leaf;
1290 		record->leaf.base.create_tid = trans->tid;
1291 		record->leaf.data_len = sizeof(ip->sync_ino_data);
1292 		record->leaf.create_ts = trans->time32;
1293 		record->data = (void *)&ip->sync_ino_data;
1294 		record->flags |= HAMMER_RECF_INTERLOCK_BE;
1295 
1296 		/*
1297 		 * If this flag is set we cannot sync the new file size
1298 		 * because we haven't finished related truncations.  The
1299 		 * inode will be flushed in another flush group to finish
1300 		 * the job.
1301 		 */
1302 		if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1303 		    ip->sync_ino_data.size != ip->ino_data.size) {
1304 			redirty = 1;
1305 			ip->sync_ino_data.size = ip->ino_data.size;
1306 		} else {
1307 			redirty = 0;
1308 		}
1309 
1310 		for (;;) {
1311 			error = hammer_ip_sync_record_cursor(cursor, record);
1312 			if (hammer_debug_inode)
1313 				kprintf("GENREC %p rec %08x %d\n",
1314 					ip, record->flags, error);
1315 			if (error != EDEADLK)
1316 				break;
1317 			hammer_done_cursor(cursor);
1318 			error = hammer_init_cursor(trans, cursor,
1319 						   &ip->cache[0], ip);
1320 			if (hammer_debug_inode)
1321 				kprintf("GENREC reinit %d\n", error);
1322 			if (error)
1323 				break;
1324 		}
1325 
1326 		/*
1327 		 * Note:  The record was never on the inode's record tree
1328 		 * so just wave our hands importantly and destroy it.
1329 		 */
1330 		record->flags |= HAMMER_RECF_COMMITTED;
1331 		record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1332 		record->flush_state = HAMMER_FST_IDLE;
1333 		++ip->rec_generation;
1334 		hammer_rel_mem_record(record);
1335 
1336 		/*
1337 		 * Finish up.
1338 		 */
1339 		if (error == 0) {
1340 			if (hammer_debug_inode)
1341 				kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1342 			ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1343 					    HAMMER_INODE_SDIRTY |
1344 					    HAMMER_INODE_ATIME |
1345 					    HAMMER_INODE_MTIME);
1346 			ip->flags &= ~HAMMER_INODE_DELONDISK;
1347 			if (redirty)
1348 				ip->sync_flags |= HAMMER_INODE_DDIRTY;
1349 
1350 			/*
1351 			 * Root volume count of inodes
1352 			 */
1353 			hammer_sync_lock_sh(trans);
1354 			if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1355 				hammer_modify_volume_field(trans,
1356 							   trans->rootvol,
1357 							   vol0_stat_inodes);
1358 				++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1359 				hammer_modify_volume_done(trans->rootvol);
1360 				ip->flags |= HAMMER_INODE_ONDISK;
1361 				if (hammer_debug_inode)
1362 					kprintf("NOWONDISK %p\n", ip);
1363 			}
1364 			hammer_sync_unlock(trans);
1365 		}
1366 	}
1367 
1368 	/*
1369 	 * If the inode has been destroyed, clean out any left-over flags
1370 	 * that may have been set by the frontend.
1371 	 */
1372 	if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
1373 		ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1374 				    HAMMER_INODE_SDIRTY |
1375 				    HAMMER_INODE_ATIME |
1376 				    HAMMER_INODE_MTIME);
1377 	}
1378 	return(error);
1379 }
1380 
1381 /*
1382  * Update only the itimes fields.
1383  *
1384  * ATIME can be updated without generating any UNDO.  MTIME is updated
1385  * with UNDO so it is guaranteed to be synchronized properly in case of
1386  * a crash.
1387  *
1388  * Neither field is included in the B-Tree leaf element's CRC, which is how
1389  * we can get away with updating ATIME the way we do.
1390  */
1391 static int
1392 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1393 {
1394 	hammer_transaction_t trans = cursor->trans;
1395 	int error;
1396 
1397 retry:
1398 	if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1399 	    HAMMER_INODE_ONDISK) {
1400 		return(0);
1401 	}
1402 
1403 	hammer_normalize_cursor(cursor);
1404 	cursor->key_beg.localization = ip->obj_localization +
1405 				       HAMMER_LOCALIZE_INODE;
1406 	cursor->key_beg.obj_id = ip->obj_id;
1407 	cursor->key_beg.key = 0;
1408 	cursor->key_beg.create_tid = 0;
1409 	cursor->key_beg.delete_tid = 0;
1410 	cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1411 	cursor->key_beg.obj_type = 0;
1412 	cursor->asof = ip->obj_asof;
1413 	cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1414 	cursor->flags |= HAMMER_CURSOR_ASOF;
1415 	cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1416 	cursor->flags |= HAMMER_CURSOR_GET_DATA;
1417 	cursor->flags |= HAMMER_CURSOR_BACKEND;
1418 
1419 	error = hammer_btree_lookup(cursor);
1420 	if (error == 0) {
1421 		hammer_cache_node(&ip->cache[0], cursor->node);
1422 		if (ip->sync_flags & HAMMER_INODE_MTIME) {
1423 			/*
1424 			 * Updating MTIME requires an UNDO.  Just cover
1425 			 * both atime and mtime.
1426 			 */
1427 			hammer_sync_lock_sh(trans);
1428 			hammer_modify_buffer(trans, cursor->data_buffer,
1429 				     HAMMER_ITIMES_BASE(&cursor->data->inode),
1430 				     HAMMER_ITIMES_BYTES);
1431 			cursor->data->inode.atime = ip->sync_ino_data.atime;
1432 			cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1433 			hammer_modify_buffer_done(cursor->data_buffer);
1434 			hammer_sync_unlock(trans);
1435 		} else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1436 			/*
1437 			 * Updating atime only can be done in-place with
1438 			 * no UNDO.
1439 			 */
1440 			hammer_sync_lock_sh(trans);
1441 			hammer_modify_buffer(trans, cursor->data_buffer,
1442 					     NULL, 0);
1443 			cursor->data->inode.atime = ip->sync_ino_data.atime;
1444 			hammer_modify_buffer_done(cursor->data_buffer);
1445 			hammer_sync_unlock(trans);
1446 		}
1447 		ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1448 	}
1449 	if (error == EDEADLK) {
1450 		hammer_done_cursor(cursor);
1451 		error = hammer_init_cursor(trans, cursor,
1452 					   &ip->cache[0], ip);
1453 		if (error == 0)
1454 			goto retry;
1455 	}
1456 	return(error);
1457 }
1458 
1459 /*
1460  * Release a reference on an inode, flush as requested.
1461  *
1462  * On the last reference we queue the inode to the flusher for its final
1463  * disposition.
1464  */
1465 void
1466 hammer_rel_inode(struct hammer_inode *ip, int flush)
1467 {
1468 	/*hammer_mount_t hmp = ip->hmp;*/
1469 
1470 	/*
1471 	 * Handle disposition when dropping the last ref.
1472 	 */
1473 	for (;;) {
1474 		if (hammer_oneref(&ip->lock)) {
1475 			/*
1476 			 * Determine whether on-disk action is needed for
1477 			 * the inode's final disposition.
1478 			 */
1479 			KKASSERT(ip->vp == NULL);
1480 			hammer_inode_unloadable_check(ip, 0);
1481 			if (ip->flags & HAMMER_INODE_MODMASK) {
1482 				hammer_flush_inode(ip, 0);
1483 			} else if (hammer_oneref(&ip->lock)) {
1484 				hammer_unload_inode(ip);
1485 				break;
1486 			}
1487 		} else {
1488 			if (flush)
1489 				hammer_flush_inode(ip, 0);
1490 
1491 			/*
1492 			 * The inode still has multiple refs, try to drop
1493 			 * one ref.
1494 			 */
1495 			KKASSERT(hammer_isactive(&ip->lock) >= 1);
1496 			if (hammer_isactive(&ip->lock) > 1) {
1497 				hammer_rel(&ip->lock);
1498 				break;
1499 			}
1500 		}
1501 	}
1502 }
1503 
1504 /*
1505  * Unload and destroy the specified inode.  Must be called with one remaining
1506  * reference.  The reference is disposed of.
1507  *
1508  * The inode must be completely clean.
1509  */
1510 static int
1511 hammer_unload_inode(struct hammer_inode *ip)
1512 {
1513 	hammer_mount_t hmp = ip->hmp;
1514 
1515 	KASSERT(hammer_oneref(&ip->lock),
1516 		("hammer_unload_inode: %d refs", hammer_isactive(&ip->lock)));
1517 	KKASSERT(ip->vp == NULL);
1518 	KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1519 	KKASSERT(ip->cursor_ip_refs == 0);
1520 	KKASSERT(hammer_notlocked(&ip->lock));
1521 	KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1522 
1523 	KKASSERT(RB_EMPTY(&ip->rec_tree));
1524 	KKASSERT(TAILQ_EMPTY(&ip->target_list));
1525 
1526 	if (ip->flags & HAMMER_INODE_RDIRTY) {
1527 		RB_REMOVE(hammer_redo_rb_tree, &hmp->rb_redo_root, ip);
1528 		ip->flags &= ~HAMMER_INODE_RDIRTY;
1529 	}
1530 	RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1531 
1532 	hammer_free_inode(ip);
1533 	return(0);
1534 }
1535 
1536 /*
1537  * Called during unmounting if a critical error occured.  The in-memory
1538  * inode and all related structures are destroyed.
1539  *
1540  * If a critical error did not occur the unmount code calls the standard
1541  * release and asserts that the inode is gone.
1542  */
1543 int
1544 hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
1545 {
1546 	hammer_record_t rec;
1547 
1548 	/*
1549 	 * Get rid of the inodes in-memory records, regardless of their
1550 	 * state, and clear the mod-mask.
1551 	 */
1552 	while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1553 		TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1554 		rec->target_ip = NULL;
1555 		if (rec->flush_state == HAMMER_FST_SETUP)
1556 			rec->flush_state = HAMMER_FST_IDLE;
1557 	}
1558 	while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1559 		if (rec->flush_state == HAMMER_FST_FLUSH)
1560 			--rec->flush_group->refs;
1561 		else
1562 			hammer_ref(&rec->lock);
1563 		KKASSERT(hammer_oneref(&rec->lock));
1564 		rec->flush_state = HAMMER_FST_IDLE;
1565 		rec->flush_group = NULL;
1566 		rec->flags |= HAMMER_RECF_DELETED_FE; /* wave hands */
1567 		rec->flags |= HAMMER_RECF_DELETED_BE; /* wave hands */
1568 		++ip->rec_generation;
1569 		hammer_rel_mem_record(rec);
1570 	}
1571 	ip->flags &= ~HAMMER_INODE_MODMASK;
1572 	ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1573 	KKASSERT(ip->vp == NULL);
1574 
1575 	/*
1576 	 * Remove the inode from any flush group, force it idle.  FLUSH
1577 	 * and SETUP states have an inode ref.
1578 	 */
1579 	switch(ip->flush_state) {
1580 	case HAMMER_FST_FLUSH:
1581 		RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
1582 		--ip->flush_group->refs;
1583 		ip->flush_group = NULL;
1584 		/* fall through */
1585 	case HAMMER_FST_SETUP:
1586 		hammer_rel(&ip->lock);
1587 		ip->flush_state = HAMMER_FST_IDLE;
1588 		/* fall through */
1589 	case HAMMER_FST_IDLE:
1590 		break;
1591 	}
1592 
1593 	/*
1594 	 * There shouldn't be any associated vnode.  The unload needs at
1595 	 * least one ref, if we do have a vp steal its ip ref.
1596 	 */
1597 	if (ip->vp) {
1598 		kprintf("hammer_destroy_inode_callback: Unexpected "
1599 			"vnode association ip %p vp %p\n", ip, ip->vp);
1600 		ip->vp->v_data = NULL;
1601 		ip->vp = NULL;
1602 	} else {
1603 		hammer_ref(&ip->lock);
1604 	}
1605 	hammer_unload_inode(ip);
1606 	return(0);
1607 }
1608 
1609 /*
1610  * Called on mount -u when switching from RW to RO or vise-versa.  Adjust
1611  * the read-only flag for cached inodes.
1612  *
1613  * This routine is called from a RB_SCAN().
1614  */
1615 int
1616 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1617 {
1618 	hammer_mount_t hmp = ip->hmp;
1619 
1620 	if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1621 		ip->flags |= HAMMER_INODE_RO;
1622 	else
1623 		ip->flags &= ~HAMMER_INODE_RO;
1624 	return(0);
1625 }
1626 
1627 /*
1628  * A transaction has modified an inode, requiring updates as specified by
1629  * the passed flags.
1630  *
1631  * HAMMER_INODE_DDIRTY: Inode data has been updated, not incl mtime/atime,
1632  *			and not including size changes due to write-append
1633  *			(but other size changes are included).
1634  * HAMMER_INODE_SDIRTY: Inode data has been updated, size changes due to
1635  *			write-append.
1636  * HAMMER_INODE_XDIRTY: Dirty in-memory records
1637  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
1638  * HAMMER_INODE_DELETED: Inode record/data must be deleted
1639  * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1640  */
1641 void
1642 hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags)
1643 {
1644 	/*
1645 	 * ronly of 0 or 2 does not trigger assertion.
1646 	 * 2 is a special error state
1647 	 */
1648 	KKASSERT(ip->hmp->ronly != 1 ||
1649 		  (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
1650 			    HAMMER_INODE_SDIRTY |
1651 			    HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1652 			    HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1653 	if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1654 		ip->flags |= HAMMER_INODE_RSV_INODES;
1655 		++ip->hmp->rsv_inodes;
1656 	}
1657 
1658 	/*
1659 	 * Set the NEWINODE flag in the transaction if the inode
1660 	 * transitions to a dirty state.  This is used to track
1661 	 * the load on the inode cache.
1662 	 */
1663 	if (trans &&
1664 	    (ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1665 	    (flags & HAMMER_INODE_MODMASK)) {
1666 		trans->flags |= HAMMER_TRANSF_NEWINODE;
1667 	}
1668 	if (flags & HAMMER_INODE_MODMASK)
1669 		hammer_inode_dirty(ip);
1670 	ip->flags |= flags;
1671 }
1672 
1673 /*
1674  * Attempt to quickly update the atime for a hammer inode.  Return 0 on
1675  * success, -1 on failure.
1676  *
1677  * We attempt to update the atime with only the ip lock and not the
1678  * whole filesystem lock in order to improve concurrency.  We can only
1679  * do this safely if the ATIME flag is already pending on the inode.
1680  *
1681  * This function is called via a vnops path (ip pointer is stable) without
1682  * fs_token held.
1683  */
1684 int
1685 hammer_update_atime_quick(hammer_inode_t ip)
1686 {
1687 	struct timeval tv;
1688 	int res = -1;
1689 
1690 	if ((ip->flags & HAMMER_INODE_RO) ||
1691 	    (ip->hmp->mp->mnt_flag & MNT_NOATIME)) {
1692 		/*
1693 		 * Silently indicate success on read-only mount/snap
1694 		 */
1695 		res = 0;
1696 	} else if (ip->flags & HAMMER_INODE_ATIME) {
1697 		/*
1698 		 * Double check with inode lock held against backend.  This
1699 		 * is only safe if all we need to do is update
1700 		 * ino_data.atime.
1701 		 */
1702 		getmicrotime(&tv);
1703 		hammer_lock_ex(&ip->lock);
1704 		if (ip->flags & HAMMER_INODE_ATIME) {
1705 			ip->ino_data.atime =
1706 			    (unsigned long)tv.tv_sec * 1000000ULL + tv.tv_usec;
1707 			res = 0;
1708 		}
1709 		hammer_unlock(&ip->lock);
1710 	}
1711 	return res;
1712 }
1713 
1714 /*
1715  * Request that an inode be flushed.  This whole mess cannot block and may
1716  * recurse (if not synchronous).  Once requested HAMMER will attempt to
1717  * actively flush the inode until the flush can be done.
1718  *
1719  * The inode may already be flushing, or may be in a setup state.  We can
1720  * place the inode in a flushing state if it is currently idle and flag it
1721  * to reflush if it is currently flushing.
1722  *
1723  * Upon return if the inode could not be flushed due to a setup
1724  * dependancy, then it will be automatically flushed when the dependancy
1725  * is satisfied.
1726  */
1727 void
1728 hammer_flush_inode(hammer_inode_t ip, int flags)
1729 {
1730 	hammer_mount_t hmp;
1731 	hammer_flush_group_t flg;
1732 	int good;
1733 
1734 	/*
1735 	 * fill_flush_group is the first flush group we may be able to
1736 	 * continue filling, it may be open or closed but it will always
1737 	 * be past the currently flushing (running) flg.
1738 	 *
1739 	 * next_flush_group is the next open flush group.
1740 	 */
1741 	hmp = ip->hmp;
1742 	while ((flg = hmp->fill_flush_group) != NULL) {
1743 		KKASSERT(flg->running == 0);
1744 		if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit &&
1745 		    flg->total_count <= hammer_autoflush) {
1746 			break;
1747 		}
1748 		hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
1749 		hammer_flusher_async(ip->hmp, flg);
1750 	}
1751 	if (flg == NULL) {
1752 		flg = kmalloc(sizeof(*flg), hmp->m_misc, M_WAITOK|M_ZERO);
1753 		flg->seq = hmp->flusher.next++;
1754 		if (hmp->next_flush_group == NULL)
1755 			hmp->next_flush_group = flg;
1756 		if (hmp->fill_flush_group == NULL)
1757 			hmp->fill_flush_group = flg;
1758 		RB_INIT(&flg->flush_tree);
1759 		TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1760 	}
1761 
1762 	/*
1763 	 * Trivial 'nothing to flush' case.  If the inode is in a SETUP
1764 	 * state we have to put it back into an IDLE state so we can
1765 	 * drop the extra ref.
1766 	 *
1767 	 * If we have a parent dependancy we must still fall through
1768 	 * so we can run it.
1769 	 */
1770 	if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1771 		if (ip->flush_state == HAMMER_FST_SETUP &&
1772 		    TAILQ_EMPTY(&ip->target_list)) {
1773 			ip->flush_state = HAMMER_FST_IDLE;
1774 			hammer_rel_inode(ip, 0);
1775 		}
1776 		if (ip->flush_state == HAMMER_FST_IDLE)
1777 			return;
1778 	}
1779 
1780 	/*
1781 	 * Our flush action will depend on the current state.
1782 	 */
1783 	switch(ip->flush_state) {
1784 	case HAMMER_FST_IDLE:
1785 		/*
1786 		 * We have no dependancies and can flush immediately.  Some
1787 		 * our children may not be flushable so we have to re-test
1788 		 * with that additional knowledge.
1789 		 */
1790 		hammer_flush_inode_core(ip, flg, flags);
1791 		break;
1792 	case HAMMER_FST_SETUP:
1793 		/*
1794 		 * Recurse upwards through dependancies via target_list
1795 		 * and start their flusher actions going if possible.
1796 		 *
1797 		 * 'good' is our connectivity.  -1 means we have none and
1798 		 * can't flush, 0 means there weren't any dependancies, and
1799 		 * 1 means we have good connectivity.
1800 		 */
1801 		good = hammer_setup_parent_inodes(ip, 0, flg);
1802 
1803 		if (good >= 0) {
1804 			/*
1805 			 * We can continue if good >= 0.  Determine how
1806 			 * many records under our inode can be flushed (and
1807 			 * mark them).
1808 			 */
1809 			hammer_flush_inode_core(ip, flg, flags);
1810 		} else {
1811 			/*
1812 			 * Parent has no connectivity, tell it to flush
1813 			 * us as soon as it does.
1814 			 *
1815 			 * The REFLUSH flag is also needed to trigger
1816 			 * dependancy wakeups.
1817 			 */
1818 			ip->flags |= HAMMER_INODE_CONN_DOWN |
1819 				     HAMMER_INODE_REFLUSH;
1820 			if (flags & HAMMER_FLUSH_SIGNAL) {
1821 				ip->flags |= HAMMER_INODE_RESIGNAL;
1822 				hammer_flusher_async(ip->hmp, flg);
1823 			}
1824 		}
1825 		break;
1826 	case HAMMER_FST_FLUSH:
1827 		/*
1828 		 * We are already flushing, flag the inode to reflush
1829 		 * if needed after it completes its current flush.
1830 		 *
1831 		 * The REFLUSH flag is also needed to trigger
1832 		 * dependancy wakeups.
1833 		 */
1834 		if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1835 			ip->flags |= HAMMER_INODE_REFLUSH;
1836 		if (flags & HAMMER_FLUSH_SIGNAL) {
1837 			ip->flags |= HAMMER_INODE_RESIGNAL;
1838 			hammer_flusher_async(ip->hmp, flg);
1839 		}
1840 		break;
1841 	}
1842 }
1843 
1844 /*
1845  * Scan ip->target_list, which is a list of records owned by PARENTS to our
1846  * ip which reference our ip.
1847  *
1848  * XXX This is a huge mess of recursive code, but not one bit of it blocks
1849  *     so for now do not ref/deref the structures.  Note that if we use the
1850  *     ref/rel code later, the rel CAN block.
1851  */
1852 static int
1853 hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
1854 			   hammer_flush_group_t flg)
1855 {
1856 	hammer_record_t depend;
1857 	int good;
1858 	int r;
1859 
1860 	/*
1861 	 * If we hit our recursion limit and we have parent dependencies
1862 	 * We cannot continue.  Returning < 0 will cause us to be flagged
1863 	 * for reflush.  Returning -2 cuts off additional dependency checks
1864 	 * because they are likely to also hit the depth limit.
1865 	 *
1866 	 * We cannot return < 0 if there are no dependencies or there might
1867 	 * not be anything to wakeup (ip).
1868 	 */
1869 	if (depth == 20 && TAILQ_FIRST(&ip->target_list)) {
1870 		if (hammer_debug_general & 0x10000)
1871 			krateprintf(&hammer_gen_krate,
1872 			    "HAMMER Warning: depth limit reached on "
1873 			    "setup recursion, inode %p %016llx\n",
1874 			    ip, (long long)ip->obj_id);
1875 		return(-2);
1876 	}
1877 
1878 	/*
1879 	 * Scan dependencies
1880 	 */
1881 	good = 0;
1882 	TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1883 		r = hammer_setup_parent_inodes_helper(depend, depth, flg);
1884 		KKASSERT(depend->target_ip == ip);
1885 		if (r < 0 && good == 0)
1886 			good = -1;
1887 		if (r > 0)
1888 			good = 1;
1889 
1890 		/*
1891 		 * If we failed due to the recursion depth limit then stop
1892 		 * now.
1893 		 */
1894 		if (r == -2)
1895 			break;
1896 	}
1897 	return(good);
1898 }
1899 
1900 /*
1901  * This helper function takes a record representing the dependancy between
1902  * the parent inode and child inode.
1903  *
1904  * record->ip		= parent inode
1905  * record->target_ip	= child inode
1906  *
1907  * We are asked to recurse upwards and convert the record from SETUP
1908  * to FLUSH if possible.
1909  *
1910  * Return 1 if the record gives us connectivity
1911  *
1912  * Return 0 if the record is not relevant
1913  *
1914  * Return -1 if we can't resolve the dependancy and there is no connectivity.
1915  */
1916 static int
1917 hammer_setup_parent_inodes_helper(hammer_record_t record, int depth,
1918 				  hammer_flush_group_t flg)
1919 {
1920 	hammer_inode_t pip;
1921 	int good;
1922 
1923 	KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1924 	pip = record->ip;
1925 
1926 	/*
1927 	 * If the record is already flushing, is it in our flush group?
1928 	 *
1929 	 * If it is in our flush group but it is a general record or a
1930 	 * delete-on-disk, it does not improve our connectivity (return 0),
1931 	 * and if the target inode is not trying to destroy itself we can't
1932 	 * allow the operation yet anyway (the second return -1).
1933 	 */
1934 	if (record->flush_state == HAMMER_FST_FLUSH) {
1935 		/*
1936 		 * If not in our flush group ask the parent to reflush
1937 		 * us as soon as possible.
1938 		 */
1939 		if (record->flush_group != flg) {
1940 			pip->flags |= HAMMER_INODE_REFLUSH;
1941 			record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1942 			return(-1);
1943 		}
1944 
1945 		/*
1946 		 * If in our flush group everything is already set up,
1947 		 * just return whether the record will improve our
1948 		 * visibility or not.
1949 		 */
1950 		if (record->type == HAMMER_MEM_RECORD_ADD)
1951 			return(1);
1952 		return(0);
1953 	}
1954 
1955 	/*
1956 	 * It must be a setup record.  Try to resolve the setup dependancies
1957 	 * by recursing upwards so we can place ip on the flush list.
1958 	 *
1959 	 * Limit ourselves to 20 levels of recursion to avoid blowing out
1960 	 * the kernel stack.  If we hit the recursion limit we can't flush
1961 	 * until the parent flushes.  The parent will flush independantly
1962 	 * on its own and ultimately a deep recursion will be resolved.
1963 	 */
1964 	KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1965 
1966 	good = hammer_setup_parent_inodes(pip, depth + 1, flg);
1967 
1968 	/*
1969 	 * If good < 0 the parent has no connectivity and we cannot safely
1970 	 * flush the directory entry, which also means we can't flush our
1971 	 * ip.  Flag us for downward recursion once the parent's
1972 	 * connectivity is resolved.  Flag the parent for [re]flush or it
1973 	 * may not check for downward recursions.
1974 	 */
1975 	if (good < 0) {
1976 		pip->flags |= HAMMER_INODE_REFLUSH;
1977 		record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1978 		return(good);
1979 	}
1980 
1981 	/*
1982 	 * We are go, place the parent inode in a flushing state so we can
1983 	 * place its record in a flushing state.  Note that the parent
1984 	 * may already be flushing.  The record must be in the same flush
1985 	 * group as the parent.
1986 	 */
1987 	if (pip->flush_state != HAMMER_FST_FLUSH)
1988 		hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
1989 	KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
1990 
1991 	/*
1992 	 * It is possible for a rename to create a loop in the recursion
1993 	 * and revisit a record.  This will result in the record being
1994 	 * placed in a flush state unexpectedly.  This check deals with
1995 	 * the case.
1996 	 */
1997 	if (record->flush_state == HAMMER_FST_FLUSH) {
1998 		if (record->type == HAMMER_MEM_RECORD_ADD)
1999 			return(1);
2000 		return(0);
2001 	}
2002 
2003 	KKASSERT(record->flush_state == HAMMER_FST_SETUP);
2004 
2005 #if 0
2006 	if (record->type == HAMMER_MEM_RECORD_DEL &&
2007 	    (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
2008 		/*
2009 		 * Regardless of flushing state we cannot sync this path if the
2010 		 * record represents a delete-on-disk but the target inode
2011 		 * is not ready to sync its own deletion.
2012 		 *
2013 		 * XXX need to count effective nlinks to determine whether
2014 		 * the flush is ok, otherwise removing a hardlink will
2015 		 * just leave the DEL record to rot.
2016 		 */
2017 		record->target_ip->flags |= HAMMER_INODE_REFLUSH;
2018 		return(-1);
2019 	} else
2020 #endif
2021 	if (pip->flush_group == flg) {
2022 		/*
2023 		 * Because we have not calculated nlinks yet we can just
2024 		 * set records to the flush state if the parent is in
2025 		 * the same flush group as we are.
2026 		 */
2027 		record->flush_state = HAMMER_FST_FLUSH;
2028 		record->flush_group = flg;
2029 		++record->flush_group->refs;
2030 		hammer_ref(&record->lock);
2031 
2032 		/*
2033 		 * A general directory-add contributes to our visibility.
2034 		 *
2035 		 * Otherwise it is probably a directory-delete or
2036 		 * delete-on-disk record and does not contribute to our
2037 		 * visbility (but we can still flush it).
2038 		 */
2039 		if (record->type == HAMMER_MEM_RECORD_ADD)
2040 			return(1);
2041 		return(0);
2042 	} else {
2043 		/*
2044 		 * If the parent is not in our flush group we cannot
2045 		 * flush this record yet, there is no visibility.
2046 		 * We tell the parent to reflush and mark ourselves
2047 		 * so the parent knows it should flush us too.
2048 		 */
2049 		pip->flags |= HAMMER_INODE_REFLUSH;
2050 		record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
2051 		return(-1);
2052 	}
2053 }
2054 
2055 /*
2056  * This is the core routine placing an inode into the FST_FLUSH state.
2057  */
2058 static void
2059 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
2060 {
2061 	hammer_mount_t hmp = ip->hmp;
2062 	int go_count;
2063 
2064 	/*
2065 	 * Set flush state and prevent the flusher from cycling into
2066 	 * the next flush group.  Do not place the ip on the list yet.
2067 	 * Inodes not in the idle state get an extra reference.
2068 	 */
2069 	KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
2070 	if (ip->flush_state == HAMMER_FST_IDLE)
2071 		hammer_ref(&ip->lock);
2072 	ip->flush_state = HAMMER_FST_FLUSH;
2073 	ip->flush_group = flg;
2074 	++hmp->flusher.group_lock;
2075 	++hmp->count_iqueued;
2076 	++hammer_count_iqueued;
2077 	++flg->total_count;
2078 	hammer_redo_fifo_start_flush(ip);
2079 
2080 #if 0
2081 	/*
2082 	 * We need to be able to vfsync/truncate from the backend.
2083 	 *
2084 	 * XXX Any truncation from the backend will acquire the vnode
2085 	 *     independently.
2086 	 */
2087 	KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
2088 	if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
2089 		ip->flags |= HAMMER_INODE_VHELD;
2090 		vref(ip->vp);
2091 	}
2092 #endif
2093 
2094 	/*
2095 	 * Figure out how many in-memory records we can actually flush
2096 	 * (not including inode meta-data, buffers, etc).
2097 	 */
2098 	KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
2099 	if (flags & HAMMER_FLUSH_RECURSION) {
2100 		/*
2101 		 * If this is a upwards recursion we do not want to
2102 		 * recurse down again!
2103 		 */
2104 		go_count = 1;
2105 #if 0
2106 	} else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2107 		/*
2108 		 * No new records are added if we must complete a flush
2109 		 * from a previous cycle, but we do have to move the records
2110 		 * from the previous cycle to the current one.
2111 		 */
2112 #if 0
2113 		go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2114 				   hammer_syncgrp_child_callback, NULL);
2115 #endif
2116 		go_count = 1;
2117 #endif
2118 	} else {
2119 		/*
2120 		 * Normal flush, scan records and bring them into the flush.
2121 		 * Directory adds and deletes are usually skipped (they are
2122 		 * grouped with the related inode rather then with the
2123 		 * directory).
2124 		 *
2125 		 * go_count can be negative, which means the scan aborted
2126 		 * due to the flush group being over-full and we should
2127 		 * flush what we have.
2128 		 */
2129 		go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2130 				   hammer_setup_child_callback, NULL);
2131 	}
2132 
2133 	/*
2134 	 * This is a more involved test that includes go_count.  If we
2135 	 * can't flush, flag the inode and return.  If go_count is 0 we
2136 	 * were are unable to flush any records in our rec_tree and
2137 	 * must ignore the XDIRTY flag.
2138 	 */
2139 	if (go_count == 0) {
2140 		if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
2141 			--hmp->count_iqueued;
2142 			--hammer_count_iqueued;
2143 
2144 			--flg->total_count;
2145 			ip->flush_state = HAMMER_FST_SETUP;
2146 			ip->flush_group = NULL;
2147 			if (flags & HAMMER_FLUSH_SIGNAL) {
2148 				ip->flags |= HAMMER_INODE_REFLUSH |
2149 					     HAMMER_INODE_RESIGNAL;
2150 			} else {
2151 				ip->flags |= HAMMER_INODE_REFLUSH;
2152 			}
2153 #if 0
2154 			if (ip->flags & HAMMER_INODE_VHELD) {
2155 				ip->flags &= ~HAMMER_INODE_VHELD;
2156 				vrele(ip->vp);
2157 			}
2158 #endif
2159 
2160 			/*
2161 			 * REFLUSH is needed to trigger dependancy wakeups
2162 			 * when an inode is in SETUP.
2163 			 */
2164 			ip->flags |= HAMMER_INODE_REFLUSH;
2165 			if (--hmp->flusher.group_lock == 0)
2166 				wakeup(&hmp->flusher.group_lock);
2167 			return;
2168 		}
2169 	}
2170 
2171 	/*
2172 	 * Snapshot the state of the inode for the backend flusher.
2173 	 *
2174 	 * We continue to retain save_trunc_off even when all truncations
2175 	 * have been resolved as an optimization to determine if we can
2176 	 * skip the B-Tree lookup for overwrite deletions.
2177 	 *
2178 	 * NOTE: The DELETING flag is a mod flag, but it is also sticky,
2179 	 * and stays in ip->flags.  Once set, it stays set until the
2180 	 * inode is destroyed.
2181 	 */
2182 	if (ip->flags & HAMMER_INODE_TRUNCATED) {
2183 		KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
2184 		ip->sync_trunc_off = ip->trunc_off;
2185 		ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
2186 		ip->flags &= ~HAMMER_INODE_TRUNCATED;
2187 		ip->sync_flags |= HAMMER_INODE_TRUNCATED;
2188 
2189 		/*
2190 		 * The save_trunc_off used to cache whether the B-Tree
2191 		 * holds any records past that point is not used until
2192 		 * after the truncation has succeeded, so we can safely
2193 		 * set it now.
2194 		 */
2195 		if (ip->save_trunc_off > ip->sync_trunc_off)
2196 			ip->save_trunc_off = ip->sync_trunc_off;
2197 	}
2198 	ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
2199 			   ~HAMMER_INODE_TRUNCATED);
2200 	ip->sync_ino_leaf = ip->ino_leaf;
2201 	ip->sync_ino_data = ip->ino_data;
2202 	ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
2203 #ifdef DEBUG_TRUNCATE
2204 	if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
2205 		kprintf("truncateS %016llx\n", ip->sync_trunc_off);
2206 #endif
2207 
2208 	/*
2209 	 * The flusher list inherits our inode and reference.
2210 	 */
2211 	KKASSERT(flg->running == 0);
2212 	RB_INSERT(hammer_fls_rb_tree, &flg->flush_tree, ip);
2213 	if (--hmp->flusher.group_lock == 0)
2214 		wakeup(&hmp->flusher.group_lock);
2215 
2216 	/*
2217 	 * Auto-flush the group if it grows too large.  Make sure the
2218 	 * inode reclaim wait pipeline continues to work.
2219 	 */
2220 	if (flg->total_count >= hammer_autoflush ||
2221 	    flg->total_count >= hammer_limit_reclaims / 4) {
2222 		if (hmp->fill_flush_group == flg)
2223 			hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
2224 		hammer_flusher_async(hmp, flg);
2225 	}
2226 }
2227 
2228 /*
2229  * Callback for scan of ip->rec_tree.  Try to include each record in our
2230  * flush.  ip->flush_group has been set but the inode has not yet been
2231  * moved into a flushing state.
2232  *
2233  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
2234  * both inodes.
2235  *
2236  * We return 1 for any record placed or found in FST_FLUSH, which prevents
2237  * the caller from shortcutting the flush.
2238  */
2239 static int
2240 hammer_setup_child_callback(hammer_record_t rec, void *data)
2241 {
2242 	hammer_flush_group_t flg;
2243 	hammer_inode_t target_ip;
2244 	hammer_inode_t ip;
2245 	int r;
2246 
2247 	/*
2248 	 * Records deleted or committed by the backend are ignored.
2249 	 * Note that the flush detects deleted frontend records at
2250 	 * multiple points to deal with races.  This is just the first
2251 	 * line of defense.  The only time HAMMER_RECF_DELETED_FE cannot
2252 	 * be set is when HAMMER_RECF_INTERLOCK_BE is set, because it
2253 	 * messes up link-count calculations.
2254 	 *
2255 	 * NOTE: Don't get confused between record deletion and, say,
2256 	 * directory entry deletion.  The deletion of a directory entry
2257 	 * which is on-media has nothing to do with the record deletion
2258 	 * flags.
2259 	 */
2260 	if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
2261 			  HAMMER_RECF_COMMITTED)) {
2262 		if (rec->flush_state == HAMMER_FST_FLUSH) {
2263 			KKASSERT(rec->flush_group == rec->ip->flush_group);
2264 			r = 1;
2265 		} else {
2266 			r = 0;
2267 		}
2268 		return(r);
2269 	}
2270 
2271 	/*
2272 	 * If the record is in an idle state it has no dependancies and
2273 	 * can be flushed.
2274 	 */
2275 	ip = rec->ip;
2276 	flg = ip->flush_group;
2277 	r = 0;
2278 
2279 	switch(rec->flush_state) {
2280 	case HAMMER_FST_IDLE:
2281 		/*
2282 		 * The record has no setup dependancy, we can flush it.
2283 		 */
2284 		KKASSERT(rec->target_ip == NULL);
2285 		rec->flush_state = HAMMER_FST_FLUSH;
2286 		rec->flush_group = flg;
2287 		++flg->refs;
2288 		hammer_ref(&rec->lock);
2289 		r = 1;
2290 		break;
2291 	case HAMMER_FST_SETUP:
2292 		/*
2293 		 * The record has a setup dependancy.  These are typically
2294 		 * directory entry adds and deletes.  Such entries will be
2295 		 * flushed when their inodes are flushed so we do not
2296 		 * usually have to add them to the flush here.  However,
2297 		 * if the target_ip has set HAMMER_INODE_CONN_DOWN then
2298 		 * it is asking us to flush this record (and it).
2299 		 */
2300 		target_ip = rec->target_ip;
2301 		KKASSERT(target_ip != NULL);
2302 		KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
2303 
2304 		/*
2305 		 * If the target IP is already flushing in our group
2306 		 * we could associate the record, but target_ip has
2307 		 * already synced ino_data to sync_ino_data and we
2308 		 * would also have to adjust nlinks.   Plus there are
2309 		 * ordering issues for adds and deletes.
2310 		 *
2311 		 * Reflush downward if this is an ADD, and upward if
2312 		 * this is a DEL.
2313 		 */
2314 		if (target_ip->flush_state == HAMMER_FST_FLUSH) {
2315 			if (rec->type == HAMMER_MEM_RECORD_ADD)
2316 				ip->flags |= HAMMER_INODE_REFLUSH;
2317 			else
2318 				target_ip->flags |= HAMMER_INODE_REFLUSH;
2319 			break;
2320 		}
2321 
2322 		/*
2323 		 * Target IP is not yet flushing.  This can get complex
2324 		 * because we have to be careful about the recursion.
2325 		 *
2326 		 * Directories create an issue for us in that if a flush
2327 		 * of a directory is requested the expectation is to flush
2328 		 * any pending directory entries, but this will cause the
2329 		 * related inodes to recursively flush as well.  We can't
2330 		 * really defer the operation so just get as many as we
2331 		 * can and
2332 		 */
2333 #if 0
2334 		if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
2335 		    (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
2336 			/*
2337 			 * We aren't reclaiming and the target ip was not
2338 			 * previously prevented from flushing due to this
2339 			 * record dependancy.  Do not flush this record.
2340 			 */
2341 			/*r = 0;*/
2342 		} else
2343 #endif
2344 		if (flg->total_count + flg->refs >
2345 			   ip->hmp->undo_rec_limit) {
2346 			/*
2347 			 * Our flush group is over-full and we risk blowing
2348 			 * out the UNDO FIFO.  Stop the scan, flush what we
2349 			 * have, then reflush the directory.
2350 			 *
2351 			 * The directory may be forced through multiple
2352 			 * flush groups before it can be completely
2353 			 * flushed.
2354 			 */
2355 			ip->flags |= HAMMER_INODE_RESIGNAL |
2356 				     HAMMER_INODE_REFLUSH;
2357 			r = -1;
2358 		} else if (rec->type == HAMMER_MEM_RECORD_ADD) {
2359 			/*
2360 			 * If the target IP is not flushing we can force
2361 			 * it to flush, even if it is unable to write out
2362 			 * any of its own records we have at least one in
2363 			 * hand that we CAN deal with.
2364 			 */
2365 			rec->flush_state = HAMMER_FST_FLUSH;
2366 			rec->flush_group = flg;
2367 			++flg->refs;
2368 			hammer_ref(&rec->lock);
2369 			hammer_flush_inode_core(target_ip, flg,
2370 						HAMMER_FLUSH_RECURSION);
2371 			r = 1;
2372 		} else {
2373 			/*
2374 			 * General or delete-on-disk record.
2375 			 *
2376 			 * XXX this needs help.  If a delete-on-disk we could
2377 			 * disconnect the target.  If the target has its own
2378 			 * dependancies they really need to be flushed.
2379 			 *
2380 			 * XXX
2381 			 */
2382 			rec->flush_state = HAMMER_FST_FLUSH;
2383 			rec->flush_group = flg;
2384 			++flg->refs;
2385 			hammer_ref(&rec->lock);
2386 			hammer_flush_inode_core(target_ip, flg,
2387 						HAMMER_FLUSH_RECURSION);
2388 			r = 1;
2389 		}
2390 		break;
2391 	case HAMMER_FST_FLUSH:
2392 		/*
2393 		 * The record could be part of a previous flush group if the
2394 		 * inode is a directory (the record being a directory entry).
2395 		 * Once the flush group was closed a hammer_test_inode()
2396 		 * function can cause a new flush group to be setup, placing
2397 		 * the directory inode itself in a new flush group.
2398 		 *
2399 		 * When associated with a previous flush group we count it
2400 		 * as if it were in our current flush group, since it will
2401 		 * effectively be flushed by the time we flush our current
2402 		 * flush group.
2403 		 */
2404 		KKASSERT(
2405 		    rec->ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY ||
2406 		    rec->flush_group == flg);
2407 		r = 1;
2408 		break;
2409 	}
2410 	return(r);
2411 }
2412 
2413 #if 0
2414 /*
2415  * This version just moves records already in a flush state to the new
2416  * flush group and that is it.
2417  */
2418 static int
2419 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
2420 {
2421 	hammer_inode_t ip = rec->ip;
2422 
2423 	switch(rec->flush_state) {
2424 	case HAMMER_FST_FLUSH:
2425 		KKASSERT(rec->flush_group == ip->flush_group);
2426 		break;
2427 	default:
2428 		break;
2429 	}
2430 	return(0);
2431 }
2432 #endif
2433 
2434 /*
2435  * Wait for a previously queued flush to complete.
2436  *
2437  * If a critical error occured we don't try to wait.
2438  */
2439 void
2440 hammer_wait_inode(hammer_inode_t ip)
2441 {
2442 	/*
2443 	 * The inode can be in a SETUP state in which case RESIGNAL
2444 	 * should be set.  If RESIGNAL is not set then the previous
2445 	 * flush completed and a later operation placed the inode
2446 	 * in a passive setup state again, so we're done.
2447 	 *
2448 	 * The inode can be in a FLUSH state in which case we
2449 	 * can just wait for completion.
2450 	 */
2451 	while (ip->flush_state == HAMMER_FST_FLUSH ||
2452 	    (ip->flush_state == HAMMER_FST_SETUP &&
2453 	     (ip->flags & HAMMER_INODE_RESIGNAL))) {
2454 		/*
2455 		 * Don't try to flush on a critical error
2456 		 */
2457 		if (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
2458 			break;
2459 
2460 		/*
2461 		 * If the inode was already being flushed its flg
2462 		 * may not have been queued to the backend.  We
2463 		 * have to make sure it gets queued or we can wind
2464 		 * up blocked or deadlocked (particularly if we are
2465 		 * the vnlru thread).
2466 		 */
2467 		if (ip->flush_state == HAMMER_FST_FLUSH) {
2468 			KKASSERT(ip->flush_group);
2469 			if (ip->flush_group->closed == 0) {
2470 				if (hammer_debug_inode) {
2471 					kprintf("hammer: debug: forcing "
2472 						"async flush ip %016jx\n",
2473 						(intmax_t)ip->obj_id);
2474 				}
2475 				hammer_flusher_async(ip->hmp,
2476 						     ip->flush_group);
2477 				continue; /* retest */
2478 			}
2479 		}
2480 
2481 		/*
2482 		 * In a flush state with the flg queued to the backend
2483 		 * or in a setup state with RESIGNAL set, we can safely
2484 		 * wait.
2485 		 */
2486 		ip->flags |= HAMMER_INODE_FLUSHW;
2487 		tsleep(&ip->flags, 0, "hmrwin", 0);
2488 	}
2489 
2490 #if 0
2491 	/*
2492 	 * The inode may have been in a passive setup state,
2493 	 * call flush to make sure we get signaled.
2494 	 */
2495 	if (ip->flush_state == HAMMER_FST_SETUP)
2496 		hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2497 #endif
2498 
2499 }
2500 
2501 /*
2502  * Called by the backend code when a flush has been completed.
2503  * The inode has already been removed from the flush list.
2504  *
2505  * A pipelined flush can occur, in which case we must re-enter the
2506  * inode on the list and re-copy its fields.
2507  */
2508 void
2509 hammer_flush_inode_done(hammer_inode_t ip, int error)
2510 {
2511 	hammer_mount_t hmp;
2512 	int dorel;
2513 
2514 	KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2515 
2516 	hmp = ip->hmp;
2517 
2518 	/*
2519 	 * Auto-reflush if the backend could not completely flush
2520 	 * the inode.  This fixes a case where a deferred buffer flush
2521 	 * could cause fsync to return early.
2522 	 */
2523 	if (ip->sync_flags & HAMMER_INODE_MODMASK)
2524 		ip->flags |= HAMMER_INODE_REFLUSH;
2525 
2526 	/*
2527 	 * Merge left-over flags back into the frontend and fix the state.
2528 	 * Incomplete truncations are retained by the backend.
2529 	 */
2530 	ip->error = error;
2531 	ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2532 	ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2533 
2534 	/*
2535 	 * The backend may have adjusted nlinks, so if the adjusted nlinks
2536 	 * does not match the fronttend set the frontend's DDIRTY flag again.
2537 	 */
2538 	if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2539 		ip->flags |= HAMMER_INODE_DDIRTY;
2540 
2541 	/*
2542 	 * Fix up the dirty buffer status.
2543 	 */
2544 	if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2545 		ip->flags |= HAMMER_INODE_BUFS;
2546 	}
2547 	hammer_redo_fifo_end_flush(ip);
2548 
2549 	/*
2550 	 * Re-set the XDIRTY flag if some of the inode's in-memory records
2551 	 * could not be flushed.
2552 	 */
2553 	KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2554 		  (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2555 		 (!RB_EMPTY(&ip->rec_tree) &&
2556 		  (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2557 
2558 	/*
2559 	 * Do not lose track of inodes which no longer have vnode
2560 	 * assocations, otherwise they may never get flushed again.
2561 	 *
2562 	 * The reflush flag can be set superfluously, causing extra pain
2563 	 * for no reason.  If the inode is no longer modified it no longer
2564 	 * needs to be flushed.
2565 	 */
2566 	if (ip->flags & HAMMER_INODE_MODMASK) {
2567 		if (ip->vp == NULL)
2568 			ip->flags |= HAMMER_INODE_REFLUSH;
2569 	} else {
2570 		ip->flags &= ~HAMMER_INODE_REFLUSH;
2571 	}
2572 	if (ip->flags & HAMMER_INODE_MODMASK)
2573 		hammer_inode_dirty(ip);
2574 
2575 	/*
2576 	 * Adjust the flush state.
2577 	 */
2578 	if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2579 		/*
2580 		 * We were unable to flush out all our records, leave the
2581 		 * inode in a flush state and in the current flush group.
2582 		 * The flush group will be re-run.
2583 		 *
2584 		 * This occurs if the UNDO block gets too full or there is
2585 		 * too much dirty meta-data and allows the flusher to
2586 		 * finalize the UNDO block and then re-flush.
2587 		 */
2588 		ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2589 		dorel = 0;
2590 	} else {
2591 		/*
2592 		 * Remove from the flush_group
2593 		 */
2594 		RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
2595 		ip->flush_group = NULL;
2596 
2597 #if 0
2598 		/*
2599 		 * Clean up the vnode ref and tracking counts.
2600 		 */
2601 		if (ip->flags & HAMMER_INODE_VHELD) {
2602 			ip->flags &= ~HAMMER_INODE_VHELD;
2603 			vrele(ip->vp);
2604 		}
2605 #endif
2606 		--hmp->count_iqueued;
2607 		--hammer_count_iqueued;
2608 
2609 		/*
2610 		 * And adjust the state.
2611 		 */
2612 		if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2613 			ip->flush_state = HAMMER_FST_IDLE;
2614 			dorel = 1;
2615 		} else {
2616 			ip->flush_state = HAMMER_FST_SETUP;
2617 			dorel = 0;
2618 		}
2619 
2620 		/*
2621 		 * If the frontend is waiting for a flush to complete,
2622 		 * wake it up.
2623 		 */
2624 		if (ip->flags & HAMMER_INODE_FLUSHW) {
2625 			ip->flags &= ~HAMMER_INODE_FLUSHW;
2626 			wakeup(&ip->flags);
2627 		}
2628 
2629 		/*
2630 		 * If the frontend made more changes and requested another
2631 		 * flush, then try to get it running.
2632 		 *
2633 		 * Reflushes are aborted when the inode is errored out.
2634 		 */
2635 		if (ip->flags & HAMMER_INODE_REFLUSH) {
2636 			ip->flags &= ~HAMMER_INODE_REFLUSH;
2637 			if (ip->flags & HAMMER_INODE_RESIGNAL) {
2638 				ip->flags &= ~HAMMER_INODE_RESIGNAL;
2639 				hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2640 			} else {
2641 				hammer_flush_inode(ip, 0);
2642 			}
2643 		}
2644 	}
2645 
2646 	/*
2647 	 * If we have no parent dependancies we can clear CONN_DOWN
2648 	 */
2649 	if (TAILQ_EMPTY(&ip->target_list))
2650 		ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2651 
2652 	/*
2653 	 * If the inode is now clean drop the space reservation.
2654 	 */
2655 	if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2656 	    (ip->flags & HAMMER_INODE_RSV_INODES)) {
2657 		ip->flags &= ~HAMMER_INODE_RSV_INODES;
2658 		--hmp->rsv_inodes;
2659 	}
2660 
2661 	ip->flags &= ~HAMMER_INODE_SLAVEFLUSH;
2662 
2663 	if (dorel)
2664 		hammer_rel_inode(ip, 0);
2665 }
2666 
2667 /*
2668  * Called from hammer_sync_inode() to synchronize in-memory records
2669  * to the media.
2670  */
2671 static int
2672 hammer_sync_record_callback(hammer_record_t record, void *data)
2673 {
2674 	hammer_cursor_t cursor = data;
2675 	hammer_transaction_t trans = cursor->trans;
2676 	hammer_mount_t hmp = trans->hmp;
2677 	int error;
2678 
2679 	/*
2680 	 * Skip records that do not belong to the current flush.
2681 	 */
2682 	++hammer_stats_record_iterations;
2683 	if (record->flush_state != HAMMER_FST_FLUSH)
2684 		return(0);
2685 
2686 #if 1
2687 	if (record->flush_group != record->ip->flush_group) {
2688 		kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2689 		if (hammer_debug_critical)
2690 			Debugger("blah2");
2691 		return(0);
2692 	}
2693 #endif
2694 	KKASSERT(record->flush_group == record->ip->flush_group);
2695 
2696 	/*
2697 	 * Interlock the record using the BE flag.  Once BE is set the
2698 	 * frontend cannot change the state of FE.
2699 	 *
2700 	 * NOTE: If FE is set prior to us setting BE we still sync the
2701 	 * record out, but the flush completion code converts it to
2702 	 * a delete-on-disk record instead of destroying it.
2703 	 */
2704 	KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2705 	record->flags |= HAMMER_RECF_INTERLOCK_BE;
2706 
2707 	/*
2708 	 * The backend has already disposed of the record.
2709 	 */
2710 	if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
2711 		error = 0;
2712 		goto done;
2713 	}
2714 
2715 	/*
2716 	 * If the whole inode is being deleted and all on-disk records will
2717 	 * be deleted very soon, we can't sync any new records to disk
2718 	 * because they will be deleted in the same transaction they were
2719 	 * created in (delete_tid == create_tid), which will assert.
2720 	 *
2721 	 * XXX There may be a case with RECORD_ADD with DELETED_FE set
2722 	 * that we currently panic on.
2723 	 */
2724 	if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2725 		switch(record->type) {
2726 		case HAMMER_MEM_RECORD_DATA:
2727 			/*
2728 			 * We don't have to do anything, if the record was
2729 			 * committed the space will have been accounted for
2730 			 * in the blockmap.
2731 			 */
2732 			/* fall through */
2733 		case HAMMER_MEM_RECORD_GENERAL:
2734 			/*
2735 			 * Set deleted-by-backend flag.  Do not set the
2736 			 * backend committed flag, because we are throwing
2737 			 * the record away.
2738 			 */
2739 			record->flags |= HAMMER_RECF_DELETED_BE;
2740 			++record->ip->rec_generation;
2741 			error = 0;
2742 			goto done;
2743 		case HAMMER_MEM_RECORD_ADD:
2744 			panic("hammer_sync_record_callback: illegal add "
2745 			      "during inode deletion record %p", record);
2746 			break; /* NOT REACHED */
2747 		case HAMMER_MEM_RECORD_INODE:
2748 			panic("hammer_sync_record_callback: attempt to "
2749 			      "sync inode record %p?", record);
2750 			break; /* NOT REACHED */
2751 		case HAMMER_MEM_RECORD_DEL:
2752 			/*
2753 			 * Follow through and issue the on-disk deletion
2754 			 */
2755 			break;
2756 		}
2757 	}
2758 
2759 	/*
2760 	 * If DELETED_FE is set special handling is needed for directory
2761 	 * entries.  Dependant pieces related to the directory entry may
2762 	 * have already been synced to disk.  If this occurs we have to
2763 	 * sync the directory entry and then change the in-memory record
2764 	 * from an ADD to a DELETE to cover the fact that it's been
2765 	 * deleted by the frontend.
2766 	 *
2767 	 * A directory delete covering record (MEM_RECORD_DEL) can never
2768 	 * be deleted by the frontend.
2769 	 *
2770 	 * Any other record type (aka DATA) can be deleted by the frontend.
2771 	 * XXX At the moment the flusher must skip it because there may
2772 	 * be another data record in the flush group for the same block,
2773 	 * meaning that some frontend data changes can leak into the backend's
2774 	 * synchronization point.
2775 	 */
2776 	if (record->flags & HAMMER_RECF_DELETED_FE) {
2777 		if (record->type == HAMMER_MEM_RECORD_ADD) {
2778 			/*
2779 			 * Convert a front-end deleted directory-add to
2780 			 * a directory-delete entry later.
2781 			 */
2782 			record->flags |= HAMMER_RECF_CONVERT_DELETE;
2783 		} else {
2784 			/*
2785 			 * Dispose of the record (race case).  Mark as
2786 			 * deleted by backend (and not committed).
2787 			 */
2788 			KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2789 			record->flags |= HAMMER_RECF_DELETED_BE;
2790 			++record->ip->rec_generation;
2791 			error = 0;
2792 			goto done;
2793 		}
2794 	}
2795 
2796 	/*
2797 	 * Assign the create_tid for new records.  Deletions already
2798 	 * have the record's entire key properly set up.
2799 	 */
2800 	if (record->type != HAMMER_MEM_RECORD_DEL) {
2801 		record->leaf.base.create_tid = trans->tid;
2802 		record->leaf.create_ts = trans->time32;
2803 	}
2804 
2805 	/*
2806 	 * This actually moves the record to the on-media B-Tree.  We
2807 	 * must also generate REDO_TERM entries in the UNDO/REDO FIFO
2808 	 * indicating that the related REDO_WRITE(s) have been committed.
2809 	 *
2810 	 * During recovery any REDO_TERM's within the nominal recovery span
2811 	 * are ignored since the related meta-data is being undone, causing
2812 	 * any matching REDO_WRITEs to execute.  The REDO_TERMs outside
2813 	 * the nominal recovery span will match against REDO_WRITEs and
2814 	 * prevent them from being executed (because the meta-data has
2815 	 * already been synchronized).
2816 	 */
2817 	if (record->flags & HAMMER_RECF_REDO) {
2818 		KKASSERT(record->type == HAMMER_MEM_RECORD_DATA);
2819 		hammer_generate_redo(trans, record->ip,
2820 				     record->leaf.base.key -
2821 					 record->leaf.data_len,
2822 				     HAMMER_REDO_TERM_WRITE,
2823 				     NULL,
2824 				     record->leaf.data_len);
2825 	}
2826 
2827 	for (;;) {
2828 		error = hammer_ip_sync_record_cursor(cursor, record);
2829 		if (error != EDEADLK)
2830 			break;
2831 		hammer_done_cursor(cursor);
2832 		error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2833 					   record->ip);
2834 		if (error)
2835 			break;
2836 	}
2837 	record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2838 
2839 	if (error)
2840 		error = -error;
2841 done:
2842 	hammer_flush_record_done(record, error);
2843 
2844 	/*
2845 	 * Do partial finalization if we have built up too many dirty
2846 	 * buffers.  Otherwise a buffer cache deadlock can occur when
2847 	 * doing things like creating tens of thousands of tiny files.
2848 	 *
2849 	 * We must release our cursor lock to avoid a 3-way deadlock
2850 	 * due to the exclusive sync lock the finalizer must get.
2851 	 *
2852 	 * WARNING: See warnings in hammer_unlock_cursor() function.
2853 	 */
2854         if (hammer_flusher_meta_limit(hmp) ||
2855 	    vm_page_count_severe()) {
2856 		hammer_unlock_cursor(cursor);
2857                 hammer_flusher_finalize(trans, 0);
2858 		hammer_lock_cursor(cursor);
2859 	}
2860 	return(error);
2861 }
2862 
2863 /*
2864  * Backend function called by the flusher to sync an inode to media.
2865  */
2866 int
2867 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2868 {
2869 	struct hammer_cursor cursor;
2870 	hammer_node_t tmp_node;
2871 	hammer_record_t depend;
2872 	hammer_record_t next;
2873 	int error, tmp_error;
2874 	u_int64_t nlinks;
2875 
2876 	if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2877 		return(0);
2878 
2879 	error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2880 	if (error)
2881 		goto done;
2882 
2883 	/*
2884 	 * Any directory records referencing this inode which are not in
2885 	 * our current flush group must adjust our nlink count for the
2886 	 * purposes of synchronizating to disk.
2887 	 *
2888 	 * Records which are in our flush group can be unlinked from our
2889 	 * inode now, potentially allowing the inode to be physically
2890 	 * deleted.
2891 	 *
2892 	 * This cannot block.
2893 	 */
2894 	nlinks = ip->ino_data.nlinks;
2895 	next = TAILQ_FIRST(&ip->target_list);
2896 	while ((depend = next) != NULL) {
2897 		next = TAILQ_NEXT(depend, target_entry);
2898 		if (depend->flush_state == HAMMER_FST_FLUSH &&
2899 		    depend->flush_group == ip->flush_group) {
2900 			/*
2901 			 * If this is an ADD that was deleted by the frontend
2902 			 * the frontend nlinks count will have already been
2903 			 * decremented, but the backend is going to sync its
2904 			 * directory entry and must account for it.  The
2905 			 * record will be converted to a delete-on-disk when
2906 			 * it gets synced.
2907 			 *
2908 			 * If the ADD was not deleted by the frontend we
2909 			 * can remove the dependancy from our target_list.
2910 			 */
2911 			if (depend->flags & HAMMER_RECF_DELETED_FE) {
2912 				++nlinks;
2913 			} else {
2914 				TAILQ_REMOVE(&ip->target_list, depend,
2915 					     target_entry);
2916 				depend->target_ip = NULL;
2917 			}
2918 		} else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2919 			/*
2920 			 * Not part of our flush group and not deleted by
2921 			 * the front-end, adjust the link count synced to
2922 			 * the media (undo what the frontend did when it
2923 			 * queued the record).
2924 			 */
2925 			KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2926 			switch(depend->type) {
2927 			case HAMMER_MEM_RECORD_ADD:
2928 				--nlinks;
2929 				break;
2930 			case HAMMER_MEM_RECORD_DEL:
2931 				++nlinks;
2932 				break;
2933 			default:
2934 				break;
2935 			}
2936 		}
2937 	}
2938 
2939 	/*
2940 	 * Set dirty if we had to modify the link count.
2941 	 */
2942 	if (ip->sync_ino_data.nlinks != nlinks) {
2943 		KKASSERT((int64_t)nlinks >= 0);
2944 		ip->sync_ino_data.nlinks = nlinks;
2945 		ip->sync_flags |= HAMMER_INODE_DDIRTY;
2946 	}
2947 
2948 	/*
2949 	 * If there is a trunction queued destroy any data past the (aligned)
2950 	 * truncation point.  Userland will have dealt with the buffer
2951 	 * containing the truncation point for us.
2952 	 *
2953 	 * We don't flush pending frontend data buffers until after we've
2954 	 * dealt with the truncation.
2955 	 */
2956 	if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2957 		/*
2958 		 * Interlock trunc_off.  The VOP front-end may continue to
2959 		 * make adjustments to it while we are blocked.
2960 		 */
2961 		off_t trunc_off;
2962 		off_t aligned_trunc_off;
2963 		int blkmask;
2964 
2965 		trunc_off = ip->sync_trunc_off;
2966 		blkmask = hammer_blocksize(trunc_off) - 1;
2967 		aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
2968 
2969 		/*
2970 		 * Delete any whole blocks on-media.  The front-end has
2971 		 * already cleaned out any partial block and made it
2972 		 * pending.  The front-end may have updated trunc_off
2973 		 * while we were blocked so we only use sync_trunc_off.
2974 		 *
2975 		 * This operation can blow out the buffer cache, EWOULDBLOCK
2976 		 * means we were unable to complete the deletion.  The
2977 		 * deletion will update sync_trunc_off in that case.
2978 		 */
2979 		error = hammer_ip_delete_range(&cursor, ip,
2980 						aligned_trunc_off,
2981 						0x7FFFFFFFFFFFFFFFLL, 2);
2982 		if (error == EWOULDBLOCK) {
2983 			ip->flags |= HAMMER_INODE_WOULDBLOCK;
2984 			error = 0;
2985 			goto defer_buffer_flush;
2986 		}
2987 
2988 		if (error)
2989 			goto done;
2990 
2991 		/*
2992 		 * Generate a REDO_TERM_TRUNC entry in the UNDO/REDO FIFO.
2993 		 *
2994 		 * XXX we do this even if we did not previously generate
2995 		 * a REDO_TRUNC record.  This operation may enclosed the
2996 		 * range for multiple prior truncation entries in the REDO
2997 		 * log.
2998 		 */
2999 		if (trans->hmp->version >= HAMMER_VOL_VERSION_FOUR &&
3000 		    (ip->flags & HAMMER_INODE_RDIRTY)) {
3001 			hammer_generate_redo(trans, ip, aligned_trunc_off,
3002 					     HAMMER_REDO_TERM_TRUNC,
3003 					     NULL, 0);
3004 		}
3005 
3006 		/*
3007 		 * Clear the truncation flag on the backend after we have
3008 		 * completed the deletions.  Backend data is now good again
3009 		 * (including new records we are about to sync, below).
3010 		 *
3011 		 * Leave sync_trunc_off intact.  As we write additional
3012 		 * records the backend will update sync_trunc_off.  This
3013 		 * tells the backend whether it can skip the overwrite
3014 		 * test.  This should work properly even when the backend
3015 		 * writes full blocks where the truncation point straddles
3016 		 * the block because the comparison is against the base
3017 		 * offset of the record.
3018 		 */
3019 		ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
3020 		/* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
3021 	} else {
3022 		error = 0;
3023 	}
3024 
3025 	/*
3026 	 * Now sync related records.  These will typically be directory
3027 	 * entries, records tracking direct-writes, or delete-on-disk records.
3028 	 */
3029 	if (error == 0) {
3030 		tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
3031 				    hammer_sync_record_callback, &cursor);
3032 		if (tmp_error < 0)
3033 			tmp_error = -error;
3034 		if (tmp_error)
3035 			error = tmp_error;
3036 	}
3037 	hammer_cache_node(&ip->cache[1], cursor.node);
3038 
3039 	/*
3040 	 * Re-seek for inode update, assuming our cache hasn't been ripped
3041 	 * out from under us.
3042 	 */
3043 	if (error == 0) {
3044 		tmp_node = hammer_ref_node_safe(trans, &ip->cache[0], &error);
3045 		if (tmp_node) {
3046 			hammer_cursor_downgrade(&cursor);
3047 			hammer_lock_sh(&tmp_node->lock);
3048 			if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
3049 				hammer_cursor_seek(&cursor, tmp_node, 0);
3050 			hammer_unlock(&tmp_node->lock);
3051 			hammer_rel_node(tmp_node);
3052 		}
3053 		error = 0;
3054 	}
3055 
3056 	/*
3057 	 * If we are deleting the inode the frontend had better not have
3058 	 * any active references on elements making up the inode.
3059 	 *
3060 	 * The call to hammer_ip_delete_clean() cleans up auxillary records
3061 	 * but not DB or DATA records.  Those must have already been deleted
3062 	 * by the normal truncation mechanic.
3063 	 */
3064 	if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
3065 		RB_EMPTY(&ip->rec_tree)  &&
3066 	    (ip->sync_flags & HAMMER_INODE_DELETING) &&
3067 	    (ip->flags & HAMMER_INODE_DELETED) == 0) {
3068 		int count1 = 0;
3069 
3070 		error = hammer_ip_delete_clean(&cursor, ip, &count1);
3071 		if (error == 0) {
3072 			ip->flags |= HAMMER_INODE_DELETED;
3073 			ip->sync_flags &= ~HAMMER_INODE_DELETING;
3074 			ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
3075 			KKASSERT(RB_EMPTY(&ip->rec_tree));
3076 
3077 			/*
3078 			 * Set delete_tid in both the frontend and backend
3079 			 * copy of the inode record.  The DELETED flag handles
3080 			 * this, do not set DDIRTY.
3081 			 */
3082 			ip->ino_leaf.base.delete_tid = trans->tid;
3083 			ip->sync_ino_leaf.base.delete_tid = trans->tid;
3084 			ip->ino_leaf.delete_ts = trans->time32;
3085 			ip->sync_ino_leaf.delete_ts = trans->time32;
3086 
3087 
3088 			/*
3089 			 * Adjust the inode count in the volume header
3090 			 */
3091 			hammer_sync_lock_sh(trans);
3092 			if (ip->flags & HAMMER_INODE_ONDISK) {
3093 				hammer_modify_volume_field(trans,
3094 							   trans->rootvol,
3095 							   vol0_stat_inodes);
3096 				--ip->hmp->rootvol->ondisk->vol0_stat_inodes;
3097 				hammer_modify_volume_done(trans->rootvol);
3098 			}
3099 			hammer_sync_unlock(trans);
3100 		}
3101 	}
3102 
3103 	if (error)
3104 		goto done;
3105 	ip->sync_flags &= ~HAMMER_INODE_BUFS;
3106 
3107 defer_buffer_flush:
3108 	/*
3109 	 * Now update the inode's on-disk inode-data and/or on-disk record.
3110 	 * DELETED and ONDISK are managed only in ip->flags.
3111 	 *
3112 	 * In the case of a defered buffer flush we still update the on-disk
3113 	 * inode to satisfy visibility requirements if there happen to be
3114 	 * directory dependancies.
3115 	 */
3116 	switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
3117 	case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
3118 		/*
3119 		 * If deleted and on-disk, don't set any additional flags.
3120 		 * the delete flag takes care of things.
3121 		 *
3122 		 * Clear flags which may have been set by the frontend.
3123 		 */
3124 		ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
3125 				    HAMMER_INODE_SDIRTY |
3126 				    HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
3127 				    HAMMER_INODE_DELETING);
3128 		break;
3129 	case HAMMER_INODE_DELETED:
3130 		/*
3131 		 * Take care of the case where a deleted inode was never
3132 		 * flushed to the disk in the first place.
3133 		 *
3134 		 * Clear flags which may have been set by the frontend.
3135 		 */
3136 		ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
3137 				    HAMMER_INODE_SDIRTY |
3138 				    HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
3139 				    HAMMER_INODE_DELETING);
3140 		while (RB_ROOT(&ip->rec_tree)) {
3141 			hammer_record_t record = RB_ROOT(&ip->rec_tree);
3142 			hammer_ref(&record->lock);
3143 			KKASSERT(hammer_oneref(&record->lock));
3144 			record->flags |= HAMMER_RECF_DELETED_BE;
3145 			++record->ip->rec_generation;
3146 			hammer_rel_mem_record(record);
3147 		}
3148 		break;
3149 	case HAMMER_INODE_ONDISK:
3150 		/*
3151 		 * If already on-disk, do not set any additional flags.
3152 		 */
3153 		break;
3154 	default:
3155 		/*
3156 		 * If not on-disk and not deleted, set DDIRTY to force
3157 		 * an initial record to be written.
3158 		 *
3159 		 * Also set the create_tid in both the frontend and backend
3160 		 * copy of the inode record.
3161 		 */
3162 		ip->ino_leaf.base.create_tid = trans->tid;
3163 		ip->ino_leaf.create_ts = trans->time32;
3164 		ip->sync_ino_leaf.base.create_tid = trans->tid;
3165 		ip->sync_ino_leaf.create_ts = trans->time32;
3166 		ip->sync_flags |= HAMMER_INODE_DDIRTY;
3167 		break;
3168 	}
3169 
3170 	/*
3171 	 * If DDIRTY or SDIRTY is set, write out a new record.
3172 	 * If the inode is already on-disk the old record is marked as
3173 	 * deleted.
3174 	 *
3175 	 * If DELETED is set hammer_update_inode() will delete the existing
3176 	 * record without writing out a new one.
3177 	 *
3178 	 * If *ONLY* the ITIMES flag is set we can update the record in-place.
3179 	 */
3180 	if (ip->flags & HAMMER_INODE_DELETED) {
3181 		error = hammer_update_inode(&cursor, ip);
3182 	} else
3183 	if (!(ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY)) &&
3184 	    (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
3185 		error = hammer_update_itimes(&cursor, ip);
3186 	} else
3187 	if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY |
3188 			      HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
3189 		error = hammer_update_inode(&cursor, ip);
3190 	}
3191 done:
3192 	if (ip->flags & HAMMER_INODE_MODMASK)
3193 		hammer_inode_dirty(ip);
3194 	if (error) {
3195 		hammer_critical_error(ip->hmp, ip, error,
3196 				      "while syncing inode");
3197 	}
3198 	hammer_done_cursor(&cursor);
3199 	return(error);
3200 }
3201 
3202 /*
3203  * This routine is called when the OS is no longer actively referencing
3204  * the inode (but might still be keeping it cached), or when releasing
3205  * the last reference to an inode.
3206  *
3207  * At this point if the inode's nlinks count is zero we want to destroy
3208  * it, which may mean destroying it on-media too.
3209  */
3210 void
3211 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
3212 {
3213 	struct vnode *vp;
3214 
3215 	/*
3216 	 * Set the DELETING flag when the link count drops to 0 and the
3217 	 * OS no longer has any opens on the inode.
3218 	 *
3219 	 * The backend will clear DELETING (a mod flag) and set DELETED
3220 	 * (a state flag) when it is actually able to perform the
3221 	 * operation.
3222 	 *
3223 	 * Don't reflag the deletion if the flusher is currently syncing
3224 	 * one that was already flagged.  A previously set DELETING flag
3225 	 * may bounce around flags and sync_flags until the operation is
3226 	 * completely done.
3227 	 *
3228 	 * Do not attempt to modify a snapshot inode (one set to read-only).
3229 	 */
3230 	if (ip->ino_data.nlinks == 0 &&
3231 	    ((ip->flags | ip->sync_flags) & (HAMMER_INODE_RO|HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
3232 		ip->flags |= HAMMER_INODE_DELETING;
3233 		ip->flags |= HAMMER_INODE_TRUNCATED;
3234 		ip->trunc_off = 0;
3235 		vp = NULL;
3236 		if (getvp) {
3237 			if (hammer_get_vnode(ip, &vp) != 0)
3238 				return;
3239 		}
3240 
3241 		/*
3242 		 * Final cleanup
3243 		 */
3244 		if (ip->vp)
3245 			nvtruncbuf(ip->vp, 0, HAMMER_BUFSIZE, 0, 0);
3246 		if (ip->flags & HAMMER_INODE_MODMASK)
3247 			hammer_inode_dirty(ip);
3248 		if (getvp)
3249 			vput(vp);
3250 	}
3251 }
3252 
3253 /*
3254  * After potentially resolving a dependancy the inode is tested
3255  * to determine whether it needs to be reflushed.
3256  */
3257 void
3258 hammer_test_inode(hammer_inode_t ip)
3259 {
3260 	if (ip->flags & HAMMER_INODE_REFLUSH) {
3261 		ip->flags &= ~HAMMER_INODE_REFLUSH;
3262 		hammer_ref(&ip->lock);
3263 		if (ip->flags & HAMMER_INODE_RESIGNAL) {
3264 			ip->flags &= ~HAMMER_INODE_RESIGNAL;
3265 			hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
3266 		} else {
3267 			hammer_flush_inode(ip, 0);
3268 		}
3269 		hammer_rel_inode(ip, 0);
3270 	}
3271 }
3272 
3273 /*
3274  * Clear the RECLAIM flag on an inode.  This occurs when the inode is
3275  * reassociated with a vp or just before it gets freed.
3276  *
3277  * Pipeline wakeups to threads blocked due to an excessive number of
3278  * detached inodes.  This typically occurs when atime updates accumulate
3279  * while scanning a directory tree.
3280  */
3281 static void
3282 hammer_inode_wakereclaims(hammer_inode_t ip)
3283 {
3284 	struct hammer_reclaim *reclaim;
3285 	hammer_mount_t hmp = ip->hmp;
3286 
3287 	if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
3288 		return;
3289 
3290 	--hammer_count_reclaims;
3291 	--hmp->count_reclaims;
3292 	ip->flags &= ~HAMMER_INODE_RECLAIM;
3293 
3294 	if ((reclaim = TAILQ_FIRST(&hmp->reclaim_list)) != NULL) {
3295 		KKASSERT(reclaim->count > 0);
3296 		if (--reclaim->count == 0) {
3297 			TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
3298 			wakeup(reclaim);
3299 		}
3300 	}
3301 }
3302 
3303 /*
3304  * Setup our reclaim pipeline.  We only let so many detached (and dirty)
3305  * inodes build up before we start blocking.  This routine is called
3306  * if a new inode is created or an inode is loaded from media.
3307  *
3308  * When we block we don't care *which* inode has finished reclaiming,
3309  * as long as one does.
3310  *
3311  * The reclaim pipeline is primarily governed by the auto-flush which is
3312  * 1/4 hammer_limit_reclaims.  We don't want to block if the count is
3313  * less than 1/2 hammer_limit_reclaims.  From 1/2 to full count is
3314  * dynamically governed.
3315  */
3316 void
3317 hammer_inode_waitreclaims(hammer_transaction_t trans)
3318 {
3319 	hammer_mount_t hmp = trans->hmp;
3320 	struct hammer_reclaim reclaim;
3321 	int lower_limit;
3322 
3323 	/*
3324 	 * Track inode load, delay if the number of reclaiming inodes is
3325 	 * between 2/4 and 4/4 hammer_limit_reclaims, depending.
3326 	 */
3327 	if (curthread->td_proc) {
3328 		struct hammer_inostats *stats;
3329 
3330 		stats = hammer_inode_inostats(hmp, curthread->td_proc->p_pid);
3331 		++stats->count;
3332 
3333 		if (stats->count > hammer_limit_reclaims / 2)
3334 			stats->count = hammer_limit_reclaims / 2;
3335 		lower_limit = hammer_limit_reclaims - stats->count;
3336 		if (hammer_debug_general & 0x10000) {
3337 			kprintf("pid %5d limit %d\n",
3338 				(int)curthread->td_proc->p_pid, lower_limit);
3339 		}
3340 	} else {
3341 		lower_limit = hammer_limit_reclaims * 3 / 4;
3342 	}
3343 	if (hmp->count_reclaims >= lower_limit) {
3344 		reclaim.count = 1;
3345 		TAILQ_INSERT_TAIL(&hmp->reclaim_list, &reclaim, entry);
3346 		tsleep(&reclaim, 0, "hmrrcm", hz);
3347 		if (reclaim.count > 0)
3348 			TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
3349 	}
3350 }
3351 
3352 /*
3353  * Keep track of reclaim statistics on a per-pid basis using a loose
3354  * 4-way set associative hash table.  Collisions inherit the count of
3355  * the previous entry.
3356  *
3357  * NOTE: We want to be careful here to limit the chain size.  If the chain
3358  *	 size is too large a pid will spread its stats out over too many
3359  *	 entries under certain types of heavy filesystem activity and
3360  *	 wind up not delaying long enough.
3361  */
3362 static
3363 struct hammer_inostats *
3364 hammer_inode_inostats(hammer_mount_t hmp, pid_t pid)
3365 {
3366 	struct hammer_inostats *stats;
3367 	int delta;
3368 	int chain;
3369 	static volatile int iterator;	/* we don't care about MP races */
3370 
3371 	/*
3372 	 * Chain up to 4 times to find our entry.
3373 	 */
3374 	for (chain = 0; chain < 4; ++chain) {
3375 		stats = &hmp->inostats[(pid + chain) & HAMMER_INOSTATS_HMASK];
3376 		if (stats->pid == pid)
3377 			break;
3378 	}
3379 
3380 	/*
3381 	 * Replace one of the four chaining entries with our new entry.
3382 	 */
3383 	if (chain == 4) {
3384 		stats = &hmp->inostats[(pid + (iterator++ & 3)) &
3385 				       HAMMER_INOSTATS_HMASK];
3386 		stats->pid = pid;
3387 	}
3388 
3389 	/*
3390 	 * Decay the entry
3391 	 */
3392 	if (stats->count && stats->ltick != ticks) {
3393 		delta = ticks - stats->ltick;
3394 		stats->ltick = ticks;
3395 		if (delta <= 0 || delta > hz * 60)
3396 			stats->count = 0;
3397 		else
3398 			stats->count = stats->count * hz / (hz + delta);
3399 	}
3400 	if (hammer_debug_general & 0x10000)
3401 		kprintf("pid %5d stats %d\n", (int)pid, stats->count);
3402 	return (stats);
3403 }
3404 
3405 #if 0
3406 
3407 /*
3408  * XXX not used, doesn't work very well due to the large batching nature
3409  * of flushes.
3410  *
3411  * A larger then normal backlog of inodes is sitting in the flusher,
3412  * enforce a general slowdown to let it catch up.  This routine is only
3413  * called on completion of a non-flusher-related transaction which
3414  * performed B-Tree node I/O.
3415  *
3416  * It is possible for the flusher to stall in a continuous load.
3417  * blogbench -i1000 -o seems to do a good job generating this sort of load.
3418  * If the flusher is unable to catch up the inode count can bloat until
3419  * we run out of kvm.
3420  *
3421  * This is a bit of a hack.
3422  */
3423 void
3424 hammer_inode_waithard(hammer_mount_t hmp)
3425 {
3426 	/*
3427 	 * Hysteresis.
3428 	 */
3429 	if (hmp->flags & HAMMER_MOUNT_FLUSH_RECOVERY) {
3430 		if (hmp->count_reclaims < hammer_limit_reclaims / 2 &&
3431 		    hmp->count_iqueued < hmp->count_inodes / 20) {
3432 			hmp->flags &= ~HAMMER_MOUNT_FLUSH_RECOVERY;
3433 			return;
3434 		}
3435 	} else {
3436 		if (hmp->count_reclaims < hammer_limit_reclaims ||
3437 		    hmp->count_iqueued < hmp->count_inodes / 10) {
3438 			return;
3439 		}
3440 		hmp->flags |= HAMMER_MOUNT_FLUSH_RECOVERY;
3441 	}
3442 
3443 	/*
3444 	 * Block for one flush cycle.
3445 	 */
3446 	hammer_flusher_wait_next(hmp);
3447 }
3448 
3449 #endif
3450