xref: /dflybsd-src/sys/vfs/hammer/hammer_inode.c (revision 5aa42fef418118d7414e8b76fbb5ae50738ffea0)
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 associated with the root inode (not the PFS root
995 	 * inode, 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 	/*
1067 	 * PFS records are associated with the root inode (not the PFS root
1068 	 * inode, but the real root).
1069 	 */
1070 	ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1071 			      HAMMER_DEF_LOCALIZATION, 0, &error);
1072 retry:
1073 	pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
1074 	hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
1075 	cursor.key_beg.localization = ip->obj_localization +
1076 				      HAMMER_LOCALIZE_MISC;
1077 	cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
1078 	cursor.key_beg.create_tid = 0;
1079 	cursor.key_beg.delete_tid = 0;
1080 	cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
1081 	cursor.key_beg.obj_type = 0;
1082 	cursor.key_beg.key = pfsm->localization;
1083 	cursor.asof = HAMMER_MAX_TID;
1084 	cursor.flags |= HAMMER_CURSOR_ASOF;
1085 
1086 	/*
1087 	 * Replace any in-memory version of the record.
1088 	 */
1089 	error = hammer_ip_lookup(&cursor);
1090 	if (error == 0 && hammer_cursor_inmem(&cursor)) {
1091 		record = cursor.iprec;
1092 		if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
1093 			KKASSERT(cursor.deadlk_rec == NULL);
1094 			hammer_ref(&record->lock);
1095 			cursor.deadlk_rec = record;
1096 			error = EDEADLK;
1097 		} else {
1098 			record->flags |= HAMMER_RECF_DELETED_FE;
1099 			error = 0;
1100 		}
1101 	}
1102 
1103 	/*
1104 	 * Allocate replacement general record.  The backend flush will
1105 	 * delete any on-disk version of the record.
1106 	 */
1107 	if (error == 0 || error == ENOENT) {
1108 		record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
1109 		record->type = HAMMER_MEM_RECORD_GENERAL;
1110 
1111 		record->leaf.base.localization = ip->obj_localization +
1112 						 HAMMER_LOCALIZE_MISC;
1113 		record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
1114 		record->leaf.base.key = pfsm->localization;
1115 		record->leaf.data_len = sizeof(pfsm->pfsd);
1116 		bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
1117 		error = hammer_ip_add_record(trans, record);
1118 	}
1119 	hammer_done_cursor(&cursor);
1120 	if (error == EDEADLK)
1121 		goto retry;
1122 	hammer_rel_inode(ip, 0);
1123 	return(error);
1124 }
1125 
1126 /*
1127  * Create a root directory for a PFS if one does not alredy exist.
1128  *
1129  * The PFS root stands alone so we must also bump the nlinks count
1130  * to prevent it from being destroyed on release.
1131  */
1132 int
1133 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1134 		       hammer_pseudofs_inmem_t pfsm)
1135 {
1136 	hammer_inode_t ip;
1137 	struct vattr vap;
1138 	int error;
1139 
1140 	ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1141 			      pfsm->localization, 0, &error);
1142 	if (ip == NULL) {
1143 		vattr_null(&vap);
1144 		vap.va_mode = 0755;
1145 		vap.va_type = VDIR;
1146 		error = hammer_create_inode(trans, &vap, cred,
1147 					    NULL, NULL, 0,
1148 					    pfsm, &ip);
1149 		if (error == 0) {
1150 			++ip->ino_data.nlinks;
1151 			hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
1152 		}
1153 	}
1154 	if (ip)
1155 		hammer_rel_inode(ip, 0);
1156 	return(error);
1157 }
1158 
1159 /*
1160  * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
1161  * if we are unable to disassociate all the inodes.
1162  */
1163 static
1164 int
1165 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
1166 {
1167 	int res;
1168 
1169 	hammer_ref(&ip->lock);
1170 	if (hammer_isactive(&ip->lock) == 2 && ip->vp)
1171 		vclean_unlocked(ip->vp);
1172 	if (hammer_isactive(&ip->lock) == 1 && ip->vp == NULL)
1173 		res = 0;
1174 	else
1175 		res = -1;	/* stop, someone is using the inode */
1176 	hammer_rel_inode(ip, 0);
1177 	return(res);
1178 }
1179 
1180 int
1181 hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
1182 {
1183 	int res;
1184 	int try;
1185 
1186 	for (try = res = 0; try < 4; ++try) {
1187 		res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
1188 					   hammer_inode_pfs_cmp,
1189 					   hammer_unload_pseudofs_callback,
1190 					   &localization);
1191 		if (res == 0 && try > 1)
1192 			break;
1193 		hammer_flusher_sync(trans->hmp);
1194 	}
1195 	if (res != 0)
1196 		res = ENOTEMPTY;
1197 	return(res);
1198 }
1199 
1200 
1201 /*
1202  * Release a reference on a PFS
1203  */
1204 void
1205 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
1206 {
1207 	hammer_rel(&pfsm->lock);
1208 	if (hammer_norefs(&pfsm->lock)) {
1209 		RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
1210 		kfree(pfsm, hmp->m_misc);
1211 	}
1212 }
1213 
1214 /*
1215  * Called by hammer_sync_inode().
1216  */
1217 static int
1218 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
1219 {
1220 	hammer_transaction_t trans = cursor->trans;
1221 	hammer_record_t record;
1222 	int error;
1223 	int redirty;
1224 
1225 retry:
1226 	error = 0;
1227 
1228 	/*
1229 	 * If the inode has a presence on-disk then locate it and mark
1230 	 * it deleted, setting DELONDISK.
1231 	 *
1232 	 * The record may or may not be physically deleted, depending on
1233 	 * the retention policy.
1234 	 */
1235 	if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
1236 	    HAMMER_INODE_ONDISK) {
1237 		hammer_normalize_cursor(cursor);
1238 		cursor->key_beg.localization = ip->obj_localization +
1239 					       HAMMER_LOCALIZE_INODE;
1240 		cursor->key_beg.obj_id = ip->obj_id;
1241 		cursor->key_beg.key = 0;
1242 		cursor->key_beg.create_tid = 0;
1243 		cursor->key_beg.delete_tid = 0;
1244 		cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1245 		cursor->key_beg.obj_type = 0;
1246 		cursor->asof = ip->obj_asof;
1247 		cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1248 		cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
1249 		cursor->flags |= HAMMER_CURSOR_BACKEND;
1250 
1251 		error = hammer_btree_lookup(cursor);
1252 		if (hammer_debug_inode)
1253 			kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
1254 
1255 		if (error == 0) {
1256 			error = hammer_ip_delete_record(cursor, ip, trans->tid);
1257 			if (hammer_debug_inode)
1258 				kprintf(" error %d\n", error);
1259 			if (error == 0) {
1260 				ip->flags |= HAMMER_INODE_DELONDISK;
1261 			}
1262 			if (cursor->node)
1263 				hammer_cache_node(&ip->cache[0], cursor->node);
1264 		}
1265 		if (error == EDEADLK) {
1266 			hammer_done_cursor(cursor);
1267 			error = hammer_init_cursor(trans, cursor,
1268 						   &ip->cache[0], ip);
1269 			if (hammer_debug_inode)
1270 				kprintf("IPDED %p %d\n", ip, error);
1271 			if (error == 0)
1272 				goto retry;
1273 		}
1274 	}
1275 
1276 	/*
1277 	 * Ok, write out the initial record or a new record (after deleting
1278 	 * the old one), unless the DELETED flag is set.  This routine will
1279 	 * clear DELONDISK if it writes out a record.
1280 	 *
1281 	 * Update our inode statistics if this is the first application of
1282 	 * the inode on-disk.
1283 	 */
1284 	if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
1285 		/*
1286 		 * Generate a record and write it to the media.  We clean-up
1287 		 * the state before releasing so we do not have to set-up
1288 		 * a flush_group.
1289 		 */
1290 		record = hammer_alloc_mem_record(ip, 0);
1291 		record->type = HAMMER_MEM_RECORD_INODE;
1292 		record->flush_state = HAMMER_FST_FLUSH;
1293 		record->leaf = ip->sync_ino_leaf;
1294 		record->leaf.base.create_tid = trans->tid;
1295 		record->leaf.data_len = sizeof(ip->sync_ino_data);
1296 		record->leaf.create_ts = trans->time32;
1297 		record->data = (void *)&ip->sync_ino_data;
1298 		record->flags |= HAMMER_RECF_INTERLOCK_BE;
1299 
1300 		/*
1301 		 * If this flag is set we cannot sync the new file size
1302 		 * because we haven't finished related truncations.  The
1303 		 * inode will be flushed in another flush group to finish
1304 		 * the job.
1305 		 */
1306 		if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1307 		    ip->sync_ino_data.size != ip->ino_data.size) {
1308 			redirty = 1;
1309 			ip->sync_ino_data.size = ip->ino_data.size;
1310 		} else {
1311 			redirty = 0;
1312 		}
1313 
1314 		for (;;) {
1315 			error = hammer_ip_sync_record_cursor(cursor, record);
1316 			if (hammer_debug_inode)
1317 				kprintf("GENREC %p rec %08x %d\n",
1318 					ip, record->flags, error);
1319 			if (error != EDEADLK)
1320 				break;
1321 			hammer_done_cursor(cursor);
1322 			error = hammer_init_cursor(trans, cursor,
1323 						   &ip->cache[0], ip);
1324 			if (hammer_debug_inode)
1325 				kprintf("GENREC reinit %d\n", error);
1326 			if (error)
1327 				break;
1328 		}
1329 
1330 		/*
1331 		 * Note:  The record was never on the inode's record tree
1332 		 * so just wave our hands importantly and destroy it.
1333 		 */
1334 		record->flags |= HAMMER_RECF_COMMITTED;
1335 		record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1336 		record->flush_state = HAMMER_FST_IDLE;
1337 		++ip->rec_generation;
1338 		hammer_rel_mem_record(record);
1339 
1340 		/*
1341 		 * Finish up.
1342 		 */
1343 		if (error == 0) {
1344 			if (hammer_debug_inode)
1345 				kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1346 			ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1347 					    HAMMER_INODE_SDIRTY |
1348 					    HAMMER_INODE_ATIME |
1349 					    HAMMER_INODE_MTIME);
1350 			ip->flags &= ~HAMMER_INODE_DELONDISK;
1351 			if (redirty)
1352 				ip->sync_flags |= HAMMER_INODE_DDIRTY;
1353 
1354 			/*
1355 			 * Root volume count of inodes
1356 			 */
1357 			hammer_sync_lock_sh(trans);
1358 			if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1359 				hammer_modify_volume_field(trans,
1360 							   trans->rootvol,
1361 							   vol0_stat_inodes);
1362 				++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1363 				hammer_modify_volume_done(trans->rootvol);
1364 				ip->flags |= HAMMER_INODE_ONDISK;
1365 				if (hammer_debug_inode)
1366 					kprintf("NOWONDISK %p\n", ip);
1367 			}
1368 			hammer_sync_unlock(trans);
1369 		}
1370 	}
1371 
1372 	/*
1373 	 * If the inode has been destroyed, clean out any left-over flags
1374 	 * that may have been set by the frontend.
1375 	 */
1376 	if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
1377 		ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1378 				    HAMMER_INODE_SDIRTY |
1379 				    HAMMER_INODE_ATIME |
1380 				    HAMMER_INODE_MTIME);
1381 	}
1382 	return(error);
1383 }
1384 
1385 /*
1386  * Update only the itimes fields.
1387  *
1388  * ATIME can be updated without generating any UNDO.  MTIME is updated
1389  * with UNDO so it is guaranteed to be synchronized properly in case of
1390  * a crash.
1391  *
1392  * Neither field is included in the B-Tree leaf element's CRC, which is how
1393  * we can get away with updating ATIME the way we do.
1394  */
1395 static int
1396 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1397 {
1398 	hammer_transaction_t trans = cursor->trans;
1399 	int error;
1400 
1401 retry:
1402 	if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1403 	    HAMMER_INODE_ONDISK) {
1404 		return(0);
1405 	}
1406 
1407 	hammer_normalize_cursor(cursor);
1408 	cursor->key_beg.localization = ip->obj_localization +
1409 				       HAMMER_LOCALIZE_INODE;
1410 	cursor->key_beg.obj_id = ip->obj_id;
1411 	cursor->key_beg.key = 0;
1412 	cursor->key_beg.create_tid = 0;
1413 	cursor->key_beg.delete_tid = 0;
1414 	cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1415 	cursor->key_beg.obj_type = 0;
1416 	cursor->asof = ip->obj_asof;
1417 	cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1418 	cursor->flags |= HAMMER_CURSOR_ASOF;
1419 	cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1420 	cursor->flags |= HAMMER_CURSOR_GET_DATA;
1421 	cursor->flags |= HAMMER_CURSOR_BACKEND;
1422 
1423 	error = hammer_btree_lookup(cursor);
1424 	if (error == 0) {
1425 		hammer_cache_node(&ip->cache[0], cursor->node);
1426 		if (ip->sync_flags & HAMMER_INODE_MTIME) {
1427 			/*
1428 			 * Updating MTIME requires an UNDO.  Just cover
1429 			 * both atime and mtime.
1430 			 */
1431 			hammer_sync_lock_sh(trans);
1432 			hammer_modify_buffer(trans, cursor->data_buffer,
1433 				     HAMMER_ITIMES_BASE(&cursor->data->inode),
1434 				     HAMMER_ITIMES_BYTES);
1435 			cursor->data->inode.atime = ip->sync_ino_data.atime;
1436 			cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1437 			hammer_modify_buffer_done(cursor->data_buffer);
1438 			hammer_sync_unlock(trans);
1439 		} else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1440 			/*
1441 			 * Updating atime only can be done in-place with
1442 			 * no UNDO.
1443 			 */
1444 			hammer_sync_lock_sh(trans);
1445 			hammer_modify_buffer(trans, cursor->data_buffer,
1446 					     NULL, 0);
1447 			cursor->data->inode.atime = ip->sync_ino_data.atime;
1448 			hammer_modify_buffer_done(cursor->data_buffer);
1449 			hammer_sync_unlock(trans);
1450 		}
1451 		ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1452 	}
1453 	if (error == EDEADLK) {
1454 		hammer_done_cursor(cursor);
1455 		error = hammer_init_cursor(trans, cursor,
1456 					   &ip->cache[0], ip);
1457 		if (error == 0)
1458 			goto retry;
1459 	}
1460 	return(error);
1461 }
1462 
1463 /*
1464  * Release a reference on an inode, flush as requested.
1465  *
1466  * On the last reference we queue the inode to the flusher for its final
1467  * disposition.
1468  */
1469 void
1470 hammer_rel_inode(struct hammer_inode *ip, int flush)
1471 {
1472 	/*hammer_mount_t hmp = ip->hmp;*/
1473 
1474 	/*
1475 	 * Handle disposition when dropping the last ref.
1476 	 */
1477 	for (;;) {
1478 		if (hammer_oneref(&ip->lock)) {
1479 			/*
1480 			 * Determine whether on-disk action is needed for
1481 			 * the inode's final disposition.
1482 			 */
1483 			KKASSERT(ip->vp == NULL);
1484 			hammer_inode_unloadable_check(ip, 0);
1485 			if (ip->flags & HAMMER_INODE_MODMASK) {
1486 				hammer_flush_inode(ip, 0);
1487 			} else if (hammer_oneref(&ip->lock)) {
1488 				hammer_unload_inode(ip);
1489 				break;
1490 			}
1491 		} else {
1492 			if (flush)
1493 				hammer_flush_inode(ip, 0);
1494 
1495 			/*
1496 			 * The inode still has multiple refs, try to drop
1497 			 * one ref.
1498 			 */
1499 			KKASSERT(hammer_isactive(&ip->lock) >= 1);
1500 			if (hammer_isactive(&ip->lock) > 1) {
1501 				hammer_rel(&ip->lock);
1502 				break;
1503 			}
1504 		}
1505 	}
1506 }
1507 
1508 /*
1509  * Unload and destroy the specified inode.  Must be called with one remaining
1510  * reference.  The reference is disposed of.
1511  *
1512  * The inode must be completely clean.
1513  */
1514 static int
1515 hammer_unload_inode(struct hammer_inode *ip)
1516 {
1517 	hammer_mount_t hmp = ip->hmp;
1518 
1519 	KASSERT(hammer_oneref(&ip->lock),
1520 		("hammer_unload_inode: %d refs", hammer_isactive(&ip->lock)));
1521 	KKASSERT(ip->vp == NULL);
1522 	KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1523 	KKASSERT(ip->cursor_ip_refs == 0);
1524 	KKASSERT(hammer_notlocked(&ip->lock));
1525 	KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1526 
1527 	KKASSERT(RB_EMPTY(&ip->rec_tree));
1528 	KKASSERT(TAILQ_EMPTY(&ip->target_list));
1529 
1530 	if (ip->flags & HAMMER_INODE_RDIRTY) {
1531 		RB_REMOVE(hammer_redo_rb_tree, &hmp->rb_redo_root, ip);
1532 		ip->flags &= ~HAMMER_INODE_RDIRTY;
1533 	}
1534 	RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1535 
1536 	hammer_free_inode(ip);
1537 	return(0);
1538 }
1539 
1540 /*
1541  * Called during unmounting if a critical error occured.  The in-memory
1542  * inode and all related structures are destroyed.
1543  *
1544  * If a critical error did not occur the unmount code calls the standard
1545  * release and asserts that the inode is gone.
1546  */
1547 int
1548 hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
1549 {
1550 	hammer_record_t rec;
1551 
1552 	/*
1553 	 * Get rid of the inodes in-memory records, regardless of their
1554 	 * state, and clear the mod-mask.
1555 	 */
1556 	while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1557 		TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1558 		rec->target_ip = NULL;
1559 		if (rec->flush_state == HAMMER_FST_SETUP)
1560 			rec->flush_state = HAMMER_FST_IDLE;
1561 	}
1562 	while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1563 		if (rec->flush_state == HAMMER_FST_FLUSH)
1564 			--rec->flush_group->refs;
1565 		else
1566 			hammer_ref(&rec->lock);
1567 		KKASSERT(hammer_oneref(&rec->lock));
1568 		rec->flush_state = HAMMER_FST_IDLE;
1569 		rec->flush_group = NULL;
1570 		rec->flags |= HAMMER_RECF_DELETED_FE; /* wave hands */
1571 		rec->flags |= HAMMER_RECF_DELETED_BE; /* wave hands */
1572 		++ip->rec_generation;
1573 		hammer_rel_mem_record(rec);
1574 	}
1575 	ip->flags &= ~HAMMER_INODE_MODMASK;
1576 	ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1577 	KKASSERT(ip->vp == NULL);
1578 
1579 	/*
1580 	 * Remove the inode from any flush group, force it idle.  FLUSH
1581 	 * and SETUP states have an inode ref.
1582 	 */
1583 	switch(ip->flush_state) {
1584 	case HAMMER_FST_FLUSH:
1585 		RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
1586 		--ip->flush_group->refs;
1587 		ip->flush_group = NULL;
1588 		/* fall through */
1589 	case HAMMER_FST_SETUP:
1590 		hammer_rel(&ip->lock);
1591 		ip->flush_state = HAMMER_FST_IDLE;
1592 		/* fall through */
1593 	case HAMMER_FST_IDLE:
1594 		break;
1595 	}
1596 
1597 	/*
1598 	 * There shouldn't be any associated vnode.  The unload needs at
1599 	 * least one ref, if we do have a vp steal its ip ref.
1600 	 */
1601 	if (ip->vp) {
1602 		kprintf("hammer_destroy_inode_callback: Unexpected "
1603 			"vnode association ip %p vp %p\n", ip, ip->vp);
1604 		ip->vp->v_data = NULL;
1605 		ip->vp = NULL;
1606 	} else {
1607 		hammer_ref(&ip->lock);
1608 	}
1609 	hammer_unload_inode(ip);
1610 	return(0);
1611 }
1612 
1613 /*
1614  * Called on mount -u when switching from RW to RO or vise-versa.  Adjust
1615  * the read-only flag for cached inodes.
1616  *
1617  * This routine is called from a RB_SCAN().
1618  */
1619 int
1620 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1621 {
1622 	hammer_mount_t hmp = ip->hmp;
1623 
1624 	if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1625 		ip->flags |= HAMMER_INODE_RO;
1626 	else
1627 		ip->flags &= ~HAMMER_INODE_RO;
1628 	return(0);
1629 }
1630 
1631 /*
1632  * A transaction has modified an inode, requiring updates as specified by
1633  * the passed flags.
1634  *
1635  * HAMMER_INODE_DDIRTY: Inode data has been updated, not incl mtime/atime,
1636  *			and not including size changes due to write-append
1637  *			(but other size changes are included).
1638  * HAMMER_INODE_SDIRTY: Inode data has been updated, size changes due to
1639  *			write-append.
1640  * HAMMER_INODE_XDIRTY: Dirty in-memory records
1641  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
1642  * HAMMER_INODE_DELETED: Inode record/data must be deleted
1643  * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1644  */
1645 void
1646 hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags)
1647 {
1648 	/*
1649 	 * ronly of 0 or 2 does not trigger assertion.
1650 	 * 2 is a special error state
1651 	 */
1652 	KKASSERT(ip->hmp->ronly != 1 ||
1653 		  (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
1654 			    HAMMER_INODE_SDIRTY |
1655 			    HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1656 			    HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1657 	if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1658 		ip->flags |= HAMMER_INODE_RSV_INODES;
1659 		++ip->hmp->rsv_inodes;
1660 	}
1661 
1662 	/*
1663 	 * Set the NEWINODE flag in the transaction if the inode
1664 	 * transitions to a dirty state.  This is used to track
1665 	 * the load on the inode cache.
1666 	 */
1667 	if (trans &&
1668 	    (ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1669 	    (flags & HAMMER_INODE_MODMASK)) {
1670 		trans->flags |= HAMMER_TRANSF_NEWINODE;
1671 	}
1672 	if (flags & HAMMER_INODE_MODMASK)
1673 		hammer_inode_dirty(ip);
1674 	ip->flags |= flags;
1675 }
1676 
1677 /*
1678  * Attempt to quickly update the atime for a hammer inode.  Return 0 on
1679  * success, -1 on failure.
1680  *
1681  * We attempt to update the atime with only the ip lock and not the
1682  * whole filesystem lock in order to improve concurrency.  We can only
1683  * do this safely if the ATIME flag is already pending on the inode.
1684  *
1685  * This function is called via a vnops path (ip pointer is stable) without
1686  * fs_token held.
1687  */
1688 int
1689 hammer_update_atime_quick(hammer_inode_t ip)
1690 {
1691 	struct timeval tv;
1692 	int res = -1;
1693 
1694 	if ((ip->flags & HAMMER_INODE_RO) ||
1695 	    (ip->hmp->mp->mnt_flag & MNT_NOATIME)) {
1696 		/*
1697 		 * Silently indicate success on read-only mount/snap
1698 		 */
1699 		res = 0;
1700 	} else if (ip->flags & HAMMER_INODE_ATIME) {
1701 		/*
1702 		 * Double check with inode lock held against backend.  This
1703 		 * is only safe if all we need to do is update
1704 		 * ino_data.atime.
1705 		 */
1706 		getmicrotime(&tv);
1707 		hammer_lock_ex(&ip->lock);
1708 		if (ip->flags & HAMMER_INODE_ATIME) {
1709 			ip->ino_data.atime =
1710 			    (unsigned long)tv.tv_sec * 1000000ULL + tv.tv_usec;
1711 			res = 0;
1712 		}
1713 		hammer_unlock(&ip->lock);
1714 	}
1715 	return res;
1716 }
1717 
1718 /*
1719  * Request that an inode be flushed.  This whole mess cannot block and may
1720  * recurse (if not synchronous).  Once requested HAMMER will attempt to
1721  * actively flush the inode until the flush can be done.
1722  *
1723  * The inode may already be flushing, or may be in a setup state.  We can
1724  * place the inode in a flushing state if it is currently idle and flag it
1725  * to reflush if it is currently flushing.
1726  *
1727  * Upon return if the inode could not be flushed due to a setup
1728  * dependancy, then it will be automatically flushed when the dependancy
1729  * is satisfied.
1730  */
1731 void
1732 hammer_flush_inode(hammer_inode_t ip, int flags)
1733 {
1734 	hammer_mount_t hmp;
1735 	hammer_flush_group_t flg;
1736 	int good;
1737 
1738 	/*
1739 	 * fill_flush_group is the first flush group we may be able to
1740 	 * continue filling, it may be open or closed but it will always
1741 	 * be past the currently flushing (running) flg.
1742 	 *
1743 	 * next_flush_group is the next open flush group.
1744 	 */
1745 	hmp = ip->hmp;
1746 	while ((flg = hmp->fill_flush_group) != NULL) {
1747 		KKASSERT(flg->running == 0);
1748 		if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit &&
1749 		    flg->total_count <= hammer_autoflush) {
1750 			break;
1751 		}
1752 		hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
1753 		hammer_flusher_async(ip->hmp, flg);
1754 	}
1755 	if (flg == NULL) {
1756 		flg = kmalloc(sizeof(*flg), hmp->m_misc, M_WAITOK|M_ZERO);
1757 		flg->seq = hmp->flusher.next++;
1758 		if (hmp->next_flush_group == NULL)
1759 			hmp->next_flush_group = flg;
1760 		if (hmp->fill_flush_group == NULL)
1761 			hmp->fill_flush_group = flg;
1762 		RB_INIT(&flg->flush_tree);
1763 		TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1764 	}
1765 
1766 	/*
1767 	 * Trivial 'nothing to flush' case.  If the inode is in a SETUP
1768 	 * state we have to put it back into an IDLE state so we can
1769 	 * drop the extra ref.
1770 	 *
1771 	 * If we have a parent dependancy we must still fall through
1772 	 * so we can run it.
1773 	 */
1774 	if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1775 		if (ip->flush_state == HAMMER_FST_SETUP &&
1776 		    TAILQ_EMPTY(&ip->target_list)) {
1777 			ip->flush_state = HAMMER_FST_IDLE;
1778 			hammer_rel_inode(ip, 0);
1779 		}
1780 		if (ip->flush_state == HAMMER_FST_IDLE)
1781 			return;
1782 	}
1783 
1784 	/*
1785 	 * Our flush action will depend on the current state.
1786 	 */
1787 	switch(ip->flush_state) {
1788 	case HAMMER_FST_IDLE:
1789 		/*
1790 		 * We have no dependancies and can flush immediately.  Some
1791 		 * our children may not be flushable so we have to re-test
1792 		 * with that additional knowledge.
1793 		 */
1794 		hammer_flush_inode_core(ip, flg, flags);
1795 		break;
1796 	case HAMMER_FST_SETUP:
1797 		/*
1798 		 * Recurse upwards through dependancies via target_list
1799 		 * and start their flusher actions going if possible.
1800 		 *
1801 		 * 'good' is our connectivity.  -1 means we have none and
1802 		 * can't flush, 0 means there weren't any dependancies, and
1803 		 * 1 means we have good connectivity.
1804 		 */
1805 		good = hammer_setup_parent_inodes(ip, 0, flg);
1806 
1807 		if (good >= 0) {
1808 			/*
1809 			 * We can continue if good >= 0.  Determine how
1810 			 * many records under our inode can be flushed (and
1811 			 * mark them).
1812 			 */
1813 			hammer_flush_inode_core(ip, flg, flags);
1814 		} else {
1815 			/*
1816 			 * Parent has no connectivity, tell it to flush
1817 			 * us as soon as it does.
1818 			 *
1819 			 * The REFLUSH flag is also needed to trigger
1820 			 * dependancy wakeups.
1821 			 */
1822 			ip->flags |= HAMMER_INODE_CONN_DOWN |
1823 				     HAMMER_INODE_REFLUSH;
1824 			if (flags & HAMMER_FLUSH_SIGNAL) {
1825 				ip->flags |= HAMMER_INODE_RESIGNAL;
1826 				hammer_flusher_async(ip->hmp, flg);
1827 			}
1828 		}
1829 		break;
1830 	case HAMMER_FST_FLUSH:
1831 		/*
1832 		 * We are already flushing, flag the inode to reflush
1833 		 * if needed after it completes its current flush.
1834 		 *
1835 		 * The REFLUSH flag is also needed to trigger
1836 		 * dependancy wakeups.
1837 		 */
1838 		if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1839 			ip->flags |= HAMMER_INODE_REFLUSH;
1840 		if (flags & HAMMER_FLUSH_SIGNAL) {
1841 			ip->flags |= HAMMER_INODE_RESIGNAL;
1842 			hammer_flusher_async(ip->hmp, flg);
1843 		}
1844 		break;
1845 	}
1846 }
1847 
1848 /*
1849  * Scan ip->target_list, which is a list of records owned by PARENTS to our
1850  * ip which reference our ip.
1851  *
1852  * XXX This is a huge mess of recursive code, but not one bit of it blocks
1853  *     so for now do not ref/deref the structures.  Note that if we use the
1854  *     ref/rel code later, the rel CAN block.
1855  */
1856 static int
1857 hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
1858 			   hammer_flush_group_t flg)
1859 {
1860 	hammer_record_t depend;
1861 	int good;
1862 	int r;
1863 
1864 	/*
1865 	 * If we hit our recursion limit and we have parent dependencies
1866 	 * We cannot continue.  Returning < 0 will cause us to be flagged
1867 	 * for reflush.  Returning -2 cuts off additional dependency checks
1868 	 * because they are likely to also hit the depth limit.
1869 	 *
1870 	 * We cannot return < 0 if there are no dependencies or there might
1871 	 * not be anything to wakeup (ip).
1872 	 */
1873 	if (depth == 20 && TAILQ_FIRST(&ip->target_list)) {
1874 		if (hammer_debug_general & 0x10000)
1875 			krateprintf(&hammer_gen_krate,
1876 			    "HAMMER Warning: depth limit reached on "
1877 			    "setup recursion, inode %p %016llx\n",
1878 			    ip, (long long)ip->obj_id);
1879 		return(-2);
1880 	}
1881 
1882 	/*
1883 	 * Scan dependencies
1884 	 */
1885 	good = 0;
1886 	TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1887 		r = hammer_setup_parent_inodes_helper(depend, depth, flg);
1888 		KKASSERT(depend->target_ip == ip);
1889 		if (r < 0 && good == 0)
1890 			good = -1;
1891 		if (r > 0)
1892 			good = 1;
1893 
1894 		/*
1895 		 * If we failed due to the recursion depth limit then stop
1896 		 * now.
1897 		 */
1898 		if (r == -2)
1899 			break;
1900 	}
1901 	return(good);
1902 }
1903 
1904 /*
1905  * This helper function takes a record representing the dependancy between
1906  * the parent inode and child inode.
1907  *
1908  * record->ip		= parent inode
1909  * record->target_ip	= child inode
1910  *
1911  * We are asked to recurse upwards and convert the record from SETUP
1912  * to FLUSH if possible.
1913  *
1914  * Return 1 if the record gives us connectivity
1915  *
1916  * Return 0 if the record is not relevant
1917  *
1918  * Return -1 if we can't resolve the dependancy and there is no connectivity.
1919  */
1920 static int
1921 hammer_setup_parent_inodes_helper(hammer_record_t record, int depth,
1922 				  hammer_flush_group_t flg)
1923 {
1924 	hammer_inode_t pip;
1925 	int good;
1926 
1927 	KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1928 	pip = record->ip;
1929 
1930 	/*
1931 	 * If the record is already flushing, is it in our flush group?
1932 	 *
1933 	 * If it is in our flush group but it is a general record or a
1934 	 * delete-on-disk, it does not improve our connectivity (return 0),
1935 	 * and if the target inode is not trying to destroy itself we can't
1936 	 * allow the operation yet anyway (the second return -1).
1937 	 */
1938 	if (record->flush_state == HAMMER_FST_FLUSH) {
1939 		/*
1940 		 * If not in our flush group ask the parent to reflush
1941 		 * us as soon as possible.
1942 		 */
1943 		if (record->flush_group != flg) {
1944 			pip->flags |= HAMMER_INODE_REFLUSH;
1945 			record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1946 			return(-1);
1947 		}
1948 
1949 		/*
1950 		 * If in our flush group everything is already set up,
1951 		 * just return whether the record will improve our
1952 		 * visibility or not.
1953 		 */
1954 		if (record->type == HAMMER_MEM_RECORD_ADD)
1955 			return(1);
1956 		return(0);
1957 	}
1958 
1959 	/*
1960 	 * It must be a setup record.  Try to resolve the setup dependancies
1961 	 * by recursing upwards so we can place ip on the flush list.
1962 	 *
1963 	 * Limit ourselves to 20 levels of recursion to avoid blowing out
1964 	 * the kernel stack.  If we hit the recursion limit we can't flush
1965 	 * until the parent flushes.  The parent will flush independantly
1966 	 * on its own and ultimately a deep recursion will be resolved.
1967 	 */
1968 	KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1969 
1970 	good = hammer_setup_parent_inodes(pip, depth + 1, flg);
1971 
1972 	/*
1973 	 * If good < 0 the parent has no connectivity and we cannot safely
1974 	 * flush the directory entry, which also means we can't flush our
1975 	 * ip.  Flag us for downward recursion once the parent's
1976 	 * connectivity is resolved.  Flag the parent for [re]flush or it
1977 	 * may not check for downward recursions.
1978 	 */
1979 	if (good < 0) {
1980 		pip->flags |= HAMMER_INODE_REFLUSH;
1981 		record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1982 		return(good);
1983 	}
1984 
1985 	/*
1986 	 * We are go, place the parent inode in a flushing state so we can
1987 	 * place its record in a flushing state.  Note that the parent
1988 	 * may already be flushing.  The record must be in the same flush
1989 	 * group as the parent.
1990 	 */
1991 	if (pip->flush_state != HAMMER_FST_FLUSH)
1992 		hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
1993 	KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
1994 
1995 	/*
1996 	 * It is possible for a rename to create a loop in the recursion
1997 	 * and revisit a record.  This will result in the record being
1998 	 * placed in a flush state unexpectedly.  This check deals with
1999 	 * the case.
2000 	 */
2001 	if (record->flush_state == HAMMER_FST_FLUSH) {
2002 		if (record->type == HAMMER_MEM_RECORD_ADD)
2003 			return(1);
2004 		return(0);
2005 	}
2006 
2007 	KKASSERT(record->flush_state == HAMMER_FST_SETUP);
2008 
2009 #if 0
2010 	if (record->type == HAMMER_MEM_RECORD_DEL &&
2011 	    (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
2012 		/*
2013 		 * Regardless of flushing state we cannot sync this path if the
2014 		 * record represents a delete-on-disk but the target inode
2015 		 * is not ready to sync its own deletion.
2016 		 *
2017 		 * XXX need to count effective nlinks to determine whether
2018 		 * the flush is ok, otherwise removing a hardlink will
2019 		 * just leave the DEL record to rot.
2020 		 */
2021 		record->target_ip->flags |= HAMMER_INODE_REFLUSH;
2022 		return(-1);
2023 	} else
2024 #endif
2025 	if (pip->flush_group == flg) {
2026 		/*
2027 		 * Because we have not calculated nlinks yet we can just
2028 		 * set records to the flush state if the parent is in
2029 		 * the same flush group as we are.
2030 		 */
2031 		record->flush_state = HAMMER_FST_FLUSH;
2032 		record->flush_group = flg;
2033 		++record->flush_group->refs;
2034 		hammer_ref(&record->lock);
2035 
2036 		/*
2037 		 * A general directory-add contributes to our visibility.
2038 		 *
2039 		 * Otherwise it is probably a directory-delete or
2040 		 * delete-on-disk record and does not contribute to our
2041 		 * visbility (but we can still flush it).
2042 		 */
2043 		if (record->type == HAMMER_MEM_RECORD_ADD)
2044 			return(1);
2045 		return(0);
2046 	} else {
2047 		/*
2048 		 * If the parent is not in our flush group we cannot
2049 		 * flush this record yet, there is no visibility.
2050 		 * We tell the parent to reflush and mark ourselves
2051 		 * so the parent knows it should flush us too.
2052 		 */
2053 		pip->flags |= HAMMER_INODE_REFLUSH;
2054 		record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
2055 		return(-1);
2056 	}
2057 }
2058 
2059 /*
2060  * This is the core routine placing an inode into the FST_FLUSH state.
2061  */
2062 static void
2063 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
2064 {
2065 	hammer_mount_t hmp = ip->hmp;
2066 	int go_count;
2067 
2068 	/*
2069 	 * Set flush state and prevent the flusher from cycling into
2070 	 * the next flush group.  Do not place the ip on the list yet.
2071 	 * Inodes not in the idle state get an extra reference.
2072 	 */
2073 	KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
2074 	if (ip->flush_state == HAMMER_FST_IDLE)
2075 		hammer_ref(&ip->lock);
2076 	ip->flush_state = HAMMER_FST_FLUSH;
2077 	ip->flush_group = flg;
2078 	++hmp->flusher.group_lock;
2079 	++hmp->count_iqueued;
2080 	++hammer_count_iqueued;
2081 	++flg->total_count;
2082 	hammer_redo_fifo_start_flush(ip);
2083 
2084 #if 0
2085 	/*
2086 	 * We need to be able to vfsync/truncate from the backend.
2087 	 *
2088 	 * XXX Any truncation from the backend will acquire the vnode
2089 	 *     independently.
2090 	 */
2091 	KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
2092 	if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
2093 		ip->flags |= HAMMER_INODE_VHELD;
2094 		vref(ip->vp);
2095 	}
2096 #endif
2097 
2098 	/*
2099 	 * Figure out how many in-memory records we can actually flush
2100 	 * (not including inode meta-data, buffers, etc).
2101 	 */
2102 	KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
2103 	if (flags & HAMMER_FLUSH_RECURSION) {
2104 		/*
2105 		 * If this is a upwards recursion we do not want to
2106 		 * recurse down again!
2107 		 */
2108 		go_count = 1;
2109 #if 0
2110 	} else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2111 		/*
2112 		 * No new records are added if we must complete a flush
2113 		 * from a previous cycle, but we do have to move the records
2114 		 * from the previous cycle to the current one.
2115 		 */
2116 #if 0
2117 		go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2118 				   hammer_syncgrp_child_callback, NULL);
2119 #endif
2120 		go_count = 1;
2121 #endif
2122 	} else {
2123 		/*
2124 		 * Normal flush, scan records and bring them into the flush.
2125 		 * Directory adds and deletes are usually skipped (they are
2126 		 * grouped with the related inode rather then with the
2127 		 * directory).
2128 		 *
2129 		 * go_count can be negative, which means the scan aborted
2130 		 * due to the flush group being over-full and we should
2131 		 * flush what we have.
2132 		 */
2133 		go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2134 				   hammer_setup_child_callback, NULL);
2135 	}
2136 
2137 	/*
2138 	 * This is a more involved test that includes go_count.  If we
2139 	 * can't flush, flag the inode and return.  If go_count is 0 we
2140 	 * were are unable to flush any records in our rec_tree and
2141 	 * must ignore the XDIRTY flag.
2142 	 */
2143 	if (go_count == 0) {
2144 		if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
2145 			--hmp->count_iqueued;
2146 			--hammer_count_iqueued;
2147 
2148 			--flg->total_count;
2149 			ip->flush_state = HAMMER_FST_SETUP;
2150 			ip->flush_group = NULL;
2151 			if (flags & HAMMER_FLUSH_SIGNAL) {
2152 				ip->flags |= HAMMER_INODE_REFLUSH |
2153 					     HAMMER_INODE_RESIGNAL;
2154 			} else {
2155 				ip->flags |= HAMMER_INODE_REFLUSH;
2156 			}
2157 #if 0
2158 			if (ip->flags & HAMMER_INODE_VHELD) {
2159 				ip->flags &= ~HAMMER_INODE_VHELD;
2160 				vrele(ip->vp);
2161 			}
2162 #endif
2163 
2164 			/*
2165 			 * REFLUSH is needed to trigger dependancy wakeups
2166 			 * when an inode is in SETUP.
2167 			 */
2168 			ip->flags |= HAMMER_INODE_REFLUSH;
2169 			if (--hmp->flusher.group_lock == 0)
2170 				wakeup(&hmp->flusher.group_lock);
2171 			return;
2172 		}
2173 	}
2174 
2175 	/*
2176 	 * Snapshot the state of the inode for the backend flusher.
2177 	 *
2178 	 * We continue to retain save_trunc_off even when all truncations
2179 	 * have been resolved as an optimization to determine if we can
2180 	 * skip the B-Tree lookup for overwrite deletions.
2181 	 *
2182 	 * NOTE: The DELETING flag is a mod flag, but it is also sticky,
2183 	 * and stays in ip->flags.  Once set, it stays set until the
2184 	 * inode is destroyed.
2185 	 */
2186 	if (ip->flags & HAMMER_INODE_TRUNCATED) {
2187 		KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
2188 		ip->sync_trunc_off = ip->trunc_off;
2189 		ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
2190 		ip->flags &= ~HAMMER_INODE_TRUNCATED;
2191 		ip->sync_flags |= HAMMER_INODE_TRUNCATED;
2192 
2193 		/*
2194 		 * The save_trunc_off used to cache whether the B-Tree
2195 		 * holds any records past that point is not used until
2196 		 * after the truncation has succeeded, so we can safely
2197 		 * set it now.
2198 		 */
2199 		if (ip->save_trunc_off > ip->sync_trunc_off)
2200 			ip->save_trunc_off = ip->sync_trunc_off;
2201 	}
2202 	ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
2203 			   ~HAMMER_INODE_TRUNCATED);
2204 	ip->sync_ino_leaf = ip->ino_leaf;
2205 	ip->sync_ino_data = ip->ino_data;
2206 	ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
2207 #ifdef DEBUG_TRUNCATE
2208 	if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
2209 		kprintf("truncateS %016llx\n", ip->sync_trunc_off);
2210 #endif
2211 
2212 	/*
2213 	 * The flusher list inherits our inode and reference.
2214 	 */
2215 	KKASSERT(flg->running == 0);
2216 	RB_INSERT(hammer_fls_rb_tree, &flg->flush_tree, ip);
2217 	if (--hmp->flusher.group_lock == 0)
2218 		wakeup(&hmp->flusher.group_lock);
2219 
2220 	/*
2221 	 * Auto-flush the group if it grows too large.  Make sure the
2222 	 * inode reclaim wait pipeline continues to work.
2223 	 */
2224 	if (flg->total_count >= hammer_autoflush ||
2225 	    flg->total_count >= hammer_limit_reclaims / 4) {
2226 		if (hmp->fill_flush_group == flg)
2227 			hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
2228 		hammer_flusher_async(hmp, flg);
2229 	}
2230 }
2231 
2232 /*
2233  * Callback for scan of ip->rec_tree.  Try to include each record in our
2234  * flush.  ip->flush_group has been set but the inode has not yet been
2235  * moved into a flushing state.
2236  *
2237  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
2238  * both inodes.
2239  *
2240  * We return 1 for any record placed or found in FST_FLUSH, which prevents
2241  * the caller from shortcutting the flush.
2242  */
2243 static int
2244 hammer_setup_child_callback(hammer_record_t rec, void *data)
2245 {
2246 	hammer_flush_group_t flg;
2247 	hammer_inode_t target_ip;
2248 	hammer_inode_t ip;
2249 	int r;
2250 
2251 	/*
2252 	 * Records deleted or committed by the backend are ignored.
2253 	 * Note that the flush detects deleted frontend records at
2254 	 * multiple points to deal with races.  This is just the first
2255 	 * line of defense.  The only time HAMMER_RECF_DELETED_FE cannot
2256 	 * be set is when HAMMER_RECF_INTERLOCK_BE is set, because it
2257 	 * messes up link-count calculations.
2258 	 *
2259 	 * NOTE: Don't get confused between record deletion and, say,
2260 	 * directory entry deletion.  The deletion of a directory entry
2261 	 * which is on-media has nothing to do with the record deletion
2262 	 * flags.
2263 	 */
2264 	if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
2265 			  HAMMER_RECF_COMMITTED)) {
2266 		if (rec->flush_state == HAMMER_FST_FLUSH) {
2267 			KKASSERT(rec->flush_group == rec->ip->flush_group);
2268 			r = 1;
2269 		} else {
2270 			r = 0;
2271 		}
2272 		return(r);
2273 	}
2274 
2275 	/*
2276 	 * If the record is in an idle state it has no dependancies and
2277 	 * can be flushed.
2278 	 */
2279 	ip = rec->ip;
2280 	flg = ip->flush_group;
2281 	r = 0;
2282 
2283 	switch(rec->flush_state) {
2284 	case HAMMER_FST_IDLE:
2285 		/*
2286 		 * The record has no setup dependancy, we can flush it.
2287 		 */
2288 		KKASSERT(rec->target_ip == NULL);
2289 		rec->flush_state = HAMMER_FST_FLUSH;
2290 		rec->flush_group = flg;
2291 		++flg->refs;
2292 		hammer_ref(&rec->lock);
2293 		r = 1;
2294 		break;
2295 	case HAMMER_FST_SETUP:
2296 		/*
2297 		 * The record has a setup dependancy.  These are typically
2298 		 * directory entry adds and deletes.  Such entries will be
2299 		 * flushed when their inodes are flushed so we do not
2300 		 * usually have to add them to the flush here.  However,
2301 		 * if the target_ip has set HAMMER_INODE_CONN_DOWN then
2302 		 * it is asking us to flush this record (and it).
2303 		 */
2304 		target_ip = rec->target_ip;
2305 		KKASSERT(target_ip != NULL);
2306 		KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
2307 
2308 		/*
2309 		 * If the target IP is already flushing in our group
2310 		 * we could associate the record, but target_ip has
2311 		 * already synced ino_data to sync_ino_data and we
2312 		 * would also have to adjust nlinks.   Plus there are
2313 		 * ordering issues for adds and deletes.
2314 		 *
2315 		 * Reflush downward if this is an ADD, and upward if
2316 		 * this is a DEL.
2317 		 */
2318 		if (target_ip->flush_state == HAMMER_FST_FLUSH) {
2319 			if (rec->type == HAMMER_MEM_RECORD_ADD)
2320 				ip->flags |= HAMMER_INODE_REFLUSH;
2321 			else
2322 				target_ip->flags |= HAMMER_INODE_REFLUSH;
2323 			break;
2324 		}
2325 
2326 		/*
2327 		 * Target IP is not yet flushing.  This can get complex
2328 		 * because we have to be careful about the recursion.
2329 		 *
2330 		 * Directories create an issue for us in that if a flush
2331 		 * of a directory is requested the expectation is to flush
2332 		 * any pending directory entries, but this will cause the
2333 		 * related inodes to recursively flush as well.  We can't
2334 		 * really defer the operation so just get as many as we
2335 		 * can and
2336 		 */
2337 #if 0
2338 		if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
2339 		    (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
2340 			/*
2341 			 * We aren't reclaiming and the target ip was not
2342 			 * previously prevented from flushing due to this
2343 			 * record dependancy.  Do not flush this record.
2344 			 */
2345 			/*r = 0;*/
2346 		} else
2347 #endif
2348 		if (flg->total_count + flg->refs >
2349 			   ip->hmp->undo_rec_limit) {
2350 			/*
2351 			 * Our flush group is over-full and we risk blowing
2352 			 * out the UNDO FIFO.  Stop the scan, flush what we
2353 			 * have, then reflush the directory.
2354 			 *
2355 			 * The directory may be forced through multiple
2356 			 * flush groups before it can be completely
2357 			 * flushed.
2358 			 */
2359 			ip->flags |= HAMMER_INODE_RESIGNAL |
2360 				     HAMMER_INODE_REFLUSH;
2361 			r = -1;
2362 		} else if (rec->type == HAMMER_MEM_RECORD_ADD) {
2363 			/*
2364 			 * If the target IP is not flushing we can force
2365 			 * it to flush, even if it is unable to write out
2366 			 * any of its own records we have at least one in
2367 			 * hand that we CAN deal with.
2368 			 */
2369 			rec->flush_state = HAMMER_FST_FLUSH;
2370 			rec->flush_group = flg;
2371 			++flg->refs;
2372 			hammer_ref(&rec->lock);
2373 			hammer_flush_inode_core(target_ip, flg,
2374 						HAMMER_FLUSH_RECURSION);
2375 			r = 1;
2376 		} else {
2377 			/*
2378 			 * General or delete-on-disk record.
2379 			 *
2380 			 * XXX this needs help.  If a delete-on-disk we could
2381 			 * disconnect the target.  If the target has its own
2382 			 * dependancies they really need to be flushed.
2383 			 *
2384 			 * XXX
2385 			 */
2386 			rec->flush_state = HAMMER_FST_FLUSH;
2387 			rec->flush_group = flg;
2388 			++flg->refs;
2389 			hammer_ref(&rec->lock);
2390 			hammer_flush_inode_core(target_ip, flg,
2391 						HAMMER_FLUSH_RECURSION);
2392 			r = 1;
2393 		}
2394 		break;
2395 	case HAMMER_FST_FLUSH:
2396 		/*
2397 		 * The record could be part of a previous flush group if the
2398 		 * inode is a directory (the record being a directory entry).
2399 		 * Once the flush group was closed a hammer_test_inode()
2400 		 * function can cause a new flush group to be setup, placing
2401 		 * the directory inode itself in a new flush group.
2402 		 *
2403 		 * When associated with a previous flush group we count it
2404 		 * as if it were in our current flush group, since it will
2405 		 * effectively be flushed by the time we flush our current
2406 		 * flush group.
2407 		 */
2408 		KKASSERT(
2409 		    rec->ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY ||
2410 		    rec->flush_group == flg);
2411 		r = 1;
2412 		break;
2413 	}
2414 	return(r);
2415 }
2416 
2417 #if 0
2418 /*
2419  * This version just moves records already in a flush state to the new
2420  * flush group and that is it.
2421  */
2422 static int
2423 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
2424 {
2425 	hammer_inode_t ip = rec->ip;
2426 
2427 	switch(rec->flush_state) {
2428 	case HAMMER_FST_FLUSH:
2429 		KKASSERT(rec->flush_group == ip->flush_group);
2430 		break;
2431 	default:
2432 		break;
2433 	}
2434 	return(0);
2435 }
2436 #endif
2437 
2438 /*
2439  * Wait for a previously queued flush to complete.
2440  *
2441  * If a critical error occured we don't try to wait.
2442  */
2443 void
2444 hammer_wait_inode(hammer_inode_t ip)
2445 {
2446 	/*
2447 	 * The inode can be in a SETUP state in which case RESIGNAL
2448 	 * should be set.  If RESIGNAL is not set then the previous
2449 	 * flush completed and a later operation placed the inode
2450 	 * in a passive setup state again, so we're done.
2451 	 *
2452 	 * The inode can be in a FLUSH state in which case we
2453 	 * can just wait for completion.
2454 	 */
2455 	while (ip->flush_state == HAMMER_FST_FLUSH ||
2456 	    (ip->flush_state == HAMMER_FST_SETUP &&
2457 	     (ip->flags & HAMMER_INODE_RESIGNAL))) {
2458 		/*
2459 		 * Don't try to flush on a critical error
2460 		 */
2461 		if (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
2462 			break;
2463 
2464 		/*
2465 		 * If the inode was already being flushed its flg
2466 		 * may not have been queued to the backend.  We
2467 		 * have to make sure it gets queued or we can wind
2468 		 * up blocked or deadlocked (particularly if we are
2469 		 * the vnlru thread).
2470 		 */
2471 		if (ip->flush_state == HAMMER_FST_FLUSH) {
2472 			KKASSERT(ip->flush_group);
2473 			if (ip->flush_group->closed == 0) {
2474 				if (hammer_debug_inode) {
2475 					kprintf("hammer: debug: forcing "
2476 						"async flush ip %016jx\n",
2477 						(intmax_t)ip->obj_id);
2478 				}
2479 				hammer_flusher_async(ip->hmp,
2480 						     ip->flush_group);
2481 				continue; /* retest */
2482 			}
2483 		}
2484 
2485 		/*
2486 		 * In a flush state with the flg queued to the backend
2487 		 * or in a setup state with RESIGNAL set, we can safely
2488 		 * wait.
2489 		 */
2490 		ip->flags |= HAMMER_INODE_FLUSHW;
2491 		tsleep(&ip->flags, 0, "hmrwin", 0);
2492 	}
2493 
2494 #if 0
2495 	/*
2496 	 * The inode may have been in a passive setup state,
2497 	 * call flush to make sure we get signaled.
2498 	 */
2499 	if (ip->flush_state == HAMMER_FST_SETUP)
2500 		hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2501 #endif
2502 
2503 }
2504 
2505 /*
2506  * Called by the backend code when a flush has been completed.
2507  * The inode has already been removed from the flush list.
2508  *
2509  * A pipelined flush can occur, in which case we must re-enter the
2510  * inode on the list and re-copy its fields.
2511  */
2512 void
2513 hammer_flush_inode_done(hammer_inode_t ip, int error)
2514 {
2515 	hammer_mount_t hmp;
2516 	int dorel;
2517 
2518 	KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2519 
2520 	hmp = ip->hmp;
2521 
2522 	/*
2523 	 * Auto-reflush if the backend could not completely flush
2524 	 * the inode.  This fixes a case where a deferred buffer flush
2525 	 * could cause fsync to return early.
2526 	 */
2527 	if (ip->sync_flags & HAMMER_INODE_MODMASK)
2528 		ip->flags |= HAMMER_INODE_REFLUSH;
2529 
2530 	/*
2531 	 * Merge left-over flags back into the frontend and fix the state.
2532 	 * Incomplete truncations are retained by the backend.
2533 	 */
2534 	ip->error = error;
2535 	ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2536 	ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2537 
2538 	/*
2539 	 * The backend may have adjusted nlinks, so if the adjusted nlinks
2540 	 * does not match the fronttend set the frontend's DDIRTY flag again.
2541 	 */
2542 	if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2543 		ip->flags |= HAMMER_INODE_DDIRTY;
2544 
2545 	/*
2546 	 * Fix up the dirty buffer status.
2547 	 */
2548 	if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2549 		ip->flags |= HAMMER_INODE_BUFS;
2550 	}
2551 	hammer_redo_fifo_end_flush(ip);
2552 
2553 	/*
2554 	 * Re-set the XDIRTY flag if some of the inode's in-memory records
2555 	 * could not be flushed.
2556 	 */
2557 	KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2558 		  (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2559 		 (!RB_EMPTY(&ip->rec_tree) &&
2560 		  (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2561 
2562 	/*
2563 	 * Do not lose track of inodes which no longer have vnode
2564 	 * assocations, otherwise they may never get flushed again.
2565 	 *
2566 	 * The reflush flag can be set superfluously, causing extra pain
2567 	 * for no reason.  If the inode is no longer modified it no longer
2568 	 * needs to be flushed.
2569 	 */
2570 	if (ip->flags & HAMMER_INODE_MODMASK) {
2571 		if (ip->vp == NULL)
2572 			ip->flags |= HAMMER_INODE_REFLUSH;
2573 	} else {
2574 		ip->flags &= ~HAMMER_INODE_REFLUSH;
2575 	}
2576 
2577 	/*
2578 	 * The fs token is held but the inode lock is not held.  Because this
2579 	 * is a backend flush it is possible that the vnode has no references
2580 	 * and cause a reclaim race inside vsetisdirty() if/when it blocks.
2581 	 *
2582 	 * Therefore, we must lock the inode around this particular dirtying
2583 	 * operation.  We don't have to around other dirtying operations
2584 	 * where the vnode is implicitly or explicitly held.
2585 	 */
2586 	if (ip->flags & HAMMER_INODE_MODMASK) {
2587 		hammer_lock_ex(&ip->lock);
2588 		hammer_inode_dirty(ip);
2589 		hammer_unlock(&ip->lock);
2590 	}
2591 
2592 	/*
2593 	 * Adjust the flush state.
2594 	 */
2595 	if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2596 		/*
2597 		 * We were unable to flush out all our records, leave the
2598 		 * inode in a flush state and in the current flush group.
2599 		 * The flush group will be re-run.
2600 		 *
2601 		 * This occurs if the UNDO block gets too full or there is
2602 		 * too much dirty meta-data and allows the flusher to
2603 		 * finalize the UNDO block and then re-flush.
2604 		 */
2605 		ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2606 		dorel = 0;
2607 	} else {
2608 		/*
2609 		 * Remove from the flush_group
2610 		 */
2611 		RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
2612 		ip->flush_group = NULL;
2613 
2614 #if 0
2615 		/*
2616 		 * Clean up the vnode ref and tracking counts.
2617 		 */
2618 		if (ip->flags & HAMMER_INODE_VHELD) {
2619 			ip->flags &= ~HAMMER_INODE_VHELD;
2620 			vrele(ip->vp);
2621 		}
2622 #endif
2623 		--hmp->count_iqueued;
2624 		--hammer_count_iqueued;
2625 
2626 		/*
2627 		 * And adjust the state.
2628 		 */
2629 		if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2630 			ip->flush_state = HAMMER_FST_IDLE;
2631 			dorel = 1;
2632 		} else {
2633 			ip->flush_state = HAMMER_FST_SETUP;
2634 			dorel = 0;
2635 		}
2636 
2637 		/*
2638 		 * If the frontend is waiting for a flush to complete,
2639 		 * wake it up.
2640 		 */
2641 		if (ip->flags & HAMMER_INODE_FLUSHW) {
2642 			ip->flags &= ~HAMMER_INODE_FLUSHW;
2643 			wakeup(&ip->flags);
2644 		}
2645 
2646 		/*
2647 		 * If the frontend made more changes and requested another
2648 		 * flush, then try to get it running.
2649 		 *
2650 		 * Reflushes are aborted when the inode is errored out.
2651 		 */
2652 		if (ip->flags & HAMMER_INODE_REFLUSH) {
2653 			ip->flags &= ~HAMMER_INODE_REFLUSH;
2654 			if (ip->flags & HAMMER_INODE_RESIGNAL) {
2655 				ip->flags &= ~HAMMER_INODE_RESIGNAL;
2656 				hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2657 			} else {
2658 				hammer_flush_inode(ip, 0);
2659 			}
2660 		}
2661 	}
2662 
2663 	/*
2664 	 * If we have no parent dependancies we can clear CONN_DOWN
2665 	 */
2666 	if (TAILQ_EMPTY(&ip->target_list))
2667 		ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2668 
2669 	/*
2670 	 * If the inode is now clean drop the space reservation.
2671 	 */
2672 	if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2673 	    (ip->flags & HAMMER_INODE_RSV_INODES)) {
2674 		ip->flags &= ~HAMMER_INODE_RSV_INODES;
2675 		--hmp->rsv_inodes;
2676 	}
2677 
2678 	ip->flags &= ~HAMMER_INODE_SLAVEFLUSH;
2679 
2680 	if (dorel)
2681 		hammer_rel_inode(ip, 0);
2682 }
2683 
2684 /*
2685  * Called from hammer_sync_inode() to synchronize in-memory records
2686  * to the media.
2687  */
2688 static int
2689 hammer_sync_record_callback(hammer_record_t record, void *data)
2690 {
2691 	hammer_cursor_t cursor = data;
2692 	hammer_transaction_t trans = cursor->trans;
2693 	hammer_mount_t hmp = trans->hmp;
2694 	int error;
2695 
2696 	/*
2697 	 * Skip records that do not belong to the current flush.
2698 	 */
2699 	++hammer_stats_record_iterations;
2700 	if (record->flush_state != HAMMER_FST_FLUSH)
2701 		return(0);
2702 
2703 #if 1
2704 	if (record->flush_group != record->ip->flush_group) {
2705 		kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2706 		if (hammer_debug_critical)
2707 			Debugger("blah2");
2708 		return(0);
2709 	}
2710 #endif
2711 	KKASSERT(record->flush_group == record->ip->flush_group);
2712 
2713 	/*
2714 	 * Interlock the record using the BE flag.  Once BE is set the
2715 	 * frontend cannot change the state of FE.
2716 	 *
2717 	 * NOTE: If FE is set prior to us setting BE we still sync the
2718 	 * record out, but the flush completion code converts it to
2719 	 * a delete-on-disk record instead of destroying it.
2720 	 */
2721 	KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2722 	record->flags |= HAMMER_RECF_INTERLOCK_BE;
2723 
2724 	/*
2725 	 * The backend has already disposed of the record.
2726 	 */
2727 	if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
2728 		error = 0;
2729 		goto done;
2730 	}
2731 
2732 	/*
2733 	 * If the whole inode is being deleted and all on-disk records will
2734 	 * be deleted very soon, we can't sync any new records to disk
2735 	 * because they will be deleted in the same transaction they were
2736 	 * created in (delete_tid == create_tid), which will assert.
2737 	 *
2738 	 * XXX There may be a case with RECORD_ADD with DELETED_FE set
2739 	 * that we currently panic on.
2740 	 */
2741 	if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2742 		switch(record->type) {
2743 		case HAMMER_MEM_RECORD_DATA:
2744 			/*
2745 			 * We don't have to do anything, if the record was
2746 			 * committed the space will have been accounted for
2747 			 * in the blockmap.
2748 			 */
2749 			/* fall through */
2750 		case HAMMER_MEM_RECORD_GENERAL:
2751 			/*
2752 			 * Set deleted-by-backend flag.  Do not set the
2753 			 * backend committed flag, because we are throwing
2754 			 * the record away.
2755 			 */
2756 			record->flags |= HAMMER_RECF_DELETED_BE;
2757 			++record->ip->rec_generation;
2758 			error = 0;
2759 			goto done;
2760 		case HAMMER_MEM_RECORD_ADD:
2761 			panic("hammer_sync_record_callback: illegal add "
2762 			      "during inode deletion record %p", record);
2763 			break; /* NOT REACHED */
2764 		case HAMMER_MEM_RECORD_INODE:
2765 			panic("hammer_sync_record_callback: attempt to "
2766 			      "sync inode record %p?", record);
2767 			break; /* NOT REACHED */
2768 		case HAMMER_MEM_RECORD_DEL:
2769 			/*
2770 			 * Follow through and issue the on-disk deletion
2771 			 */
2772 			break;
2773 		}
2774 	}
2775 
2776 	/*
2777 	 * If DELETED_FE is set special handling is needed for directory
2778 	 * entries.  Dependant pieces related to the directory entry may
2779 	 * have already been synced to disk.  If this occurs we have to
2780 	 * sync the directory entry and then change the in-memory record
2781 	 * from an ADD to a DELETE to cover the fact that it's been
2782 	 * deleted by the frontend.
2783 	 *
2784 	 * A directory delete covering record (MEM_RECORD_DEL) can never
2785 	 * be deleted by the frontend.
2786 	 *
2787 	 * Any other record type (aka DATA) can be deleted by the frontend.
2788 	 * XXX At the moment the flusher must skip it because there may
2789 	 * be another data record in the flush group for the same block,
2790 	 * meaning that some frontend data changes can leak into the backend's
2791 	 * synchronization point.
2792 	 */
2793 	if (record->flags & HAMMER_RECF_DELETED_FE) {
2794 		if (record->type == HAMMER_MEM_RECORD_ADD) {
2795 			/*
2796 			 * Convert a front-end deleted directory-add to
2797 			 * a directory-delete entry later.
2798 			 */
2799 			record->flags |= HAMMER_RECF_CONVERT_DELETE;
2800 		} else {
2801 			/*
2802 			 * Dispose of the record (race case).  Mark as
2803 			 * deleted by backend (and not committed).
2804 			 */
2805 			KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2806 			record->flags |= HAMMER_RECF_DELETED_BE;
2807 			++record->ip->rec_generation;
2808 			error = 0;
2809 			goto done;
2810 		}
2811 	}
2812 
2813 	/*
2814 	 * Assign the create_tid for new records.  Deletions already
2815 	 * have the record's entire key properly set up.
2816 	 */
2817 	if (record->type != HAMMER_MEM_RECORD_DEL) {
2818 		record->leaf.base.create_tid = trans->tid;
2819 		record->leaf.create_ts = trans->time32;
2820 	}
2821 
2822 	/*
2823 	 * This actually moves the record to the on-media B-Tree.  We
2824 	 * must also generate REDO_TERM entries in the UNDO/REDO FIFO
2825 	 * indicating that the related REDO_WRITE(s) have been committed.
2826 	 *
2827 	 * During recovery any REDO_TERM's within the nominal recovery span
2828 	 * are ignored since the related meta-data is being undone, causing
2829 	 * any matching REDO_WRITEs to execute.  The REDO_TERMs outside
2830 	 * the nominal recovery span will match against REDO_WRITEs and
2831 	 * prevent them from being executed (because the meta-data has
2832 	 * already been synchronized).
2833 	 */
2834 	if (record->flags & HAMMER_RECF_REDO) {
2835 		KKASSERT(record->type == HAMMER_MEM_RECORD_DATA);
2836 		hammer_generate_redo(trans, record->ip,
2837 				     record->leaf.base.key -
2838 					 record->leaf.data_len,
2839 				     HAMMER_REDO_TERM_WRITE,
2840 				     NULL,
2841 				     record->leaf.data_len);
2842 	}
2843 
2844 	for (;;) {
2845 		error = hammer_ip_sync_record_cursor(cursor, record);
2846 		if (error != EDEADLK)
2847 			break;
2848 		hammer_done_cursor(cursor);
2849 		error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2850 					   record->ip);
2851 		if (error)
2852 			break;
2853 	}
2854 	record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2855 
2856 	if (error)
2857 		error = -error;
2858 done:
2859 	hammer_flush_record_done(record, error);
2860 
2861 	/*
2862 	 * Do partial finalization if we have built up too many dirty
2863 	 * buffers.  Otherwise a buffer cache deadlock can occur when
2864 	 * doing things like creating tens of thousands of tiny files.
2865 	 *
2866 	 * We must release our cursor lock to avoid a 3-way deadlock
2867 	 * due to the exclusive sync lock the finalizer must get.
2868 	 *
2869 	 * WARNING: See warnings in hammer_unlock_cursor() function.
2870 	 */
2871         if (hammer_flusher_meta_limit(hmp) ||
2872 	    vm_page_count_severe()) {
2873 		hammer_unlock_cursor(cursor);
2874                 hammer_flusher_finalize(trans, 0);
2875 		hammer_lock_cursor(cursor);
2876 	}
2877 	return(error);
2878 }
2879 
2880 /*
2881  * Backend function called by the flusher to sync an inode to media.
2882  */
2883 int
2884 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2885 {
2886 	struct hammer_cursor cursor;
2887 	hammer_node_t tmp_node;
2888 	hammer_record_t depend;
2889 	hammer_record_t next;
2890 	int error, tmp_error;
2891 	u_int64_t nlinks;
2892 
2893 	if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2894 		return(0);
2895 
2896 	error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2897 	if (error)
2898 		goto done;
2899 
2900 	/*
2901 	 * Any directory records referencing this inode which are not in
2902 	 * our current flush group must adjust our nlink count for the
2903 	 * purposes of synchronizating to disk.
2904 	 *
2905 	 * Records which are in our flush group can be unlinked from our
2906 	 * inode now, potentially allowing the inode to be physically
2907 	 * deleted.
2908 	 *
2909 	 * This cannot block.
2910 	 */
2911 	nlinks = ip->ino_data.nlinks;
2912 	next = TAILQ_FIRST(&ip->target_list);
2913 	while ((depend = next) != NULL) {
2914 		next = TAILQ_NEXT(depend, target_entry);
2915 		if (depend->flush_state == HAMMER_FST_FLUSH &&
2916 		    depend->flush_group == ip->flush_group) {
2917 			/*
2918 			 * If this is an ADD that was deleted by the frontend
2919 			 * the frontend nlinks count will have already been
2920 			 * decremented, but the backend is going to sync its
2921 			 * directory entry and must account for it.  The
2922 			 * record will be converted to a delete-on-disk when
2923 			 * it gets synced.
2924 			 *
2925 			 * If the ADD was not deleted by the frontend we
2926 			 * can remove the dependancy from our target_list.
2927 			 */
2928 			if (depend->flags & HAMMER_RECF_DELETED_FE) {
2929 				++nlinks;
2930 			} else {
2931 				TAILQ_REMOVE(&ip->target_list, depend,
2932 					     target_entry);
2933 				depend->target_ip = NULL;
2934 			}
2935 		} else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2936 			/*
2937 			 * Not part of our flush group and not deleted by
2938 			 * the front-end, adjust the link count synced to
2939 			 * the media (undo what the frontend did when it
2940 			 * queued the record).
2941 			 */
2942 			KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2943 			switch(depend->type) {
2944 			case HAMMER_MEM_RECORD_ADD:
2945 				--nlinks;
2946 				break;
2947 			case HAMMER_MEM_RECORD_DEL:
2948 				++nlinks;
2949 				break;
2950 			default:
2951 				break;
2952 			}
2953 		}
2954 	}
2955 
2956 	/*
2957 	 * Set dirty if we had to modify the link count.
2958 	 */
2959 	if (ip->sync_ino_data.nlinks != nlinks) {
2960 		KKASSERT((int64_t)nlinks >= 0);
2961 		ip->sync_ino_data.nlinks = nlinks;
2962 		ip->sync_flags |= HAMMER_INODE_DDIRTY;
2963 	}
2964 
2965 	/*
2966 	 * If there is a trunction queued destroy any data past the (aligned)
2967 	 * truncation point.  Userland will have dealt with the buffer
2968 	 * containing the truncation point for us.
2969 	 *
2970 	 * We don't flush pending frontend data buffers until after we've
2971 	 * dealt with the truncation.
2972 	 */
2973 	if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2974 		/*
2975 		 * Interlock trunc_off.  The VOP front-end may continue to
2976 		 * make adjustments to it while we are blocked.
2977 		 */
2978 		off_t trunc_off;
2979 		off_t aligned_trunc_off;
2980 		int blkmask;
2981 
2982 		trunc_off = ip->sync_trunc_off;
2983 		blkmask = hammer_blocksize(trunc_off) - 1;
2984 		aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
2985 
2986 		/*
2987 		 * Delete any whole blocks on-media.  The front-end has
2988 		 * already cleaned out any partial block and made it
2989 		 * pending.  The front-end may have updated trunc_off
2990 		 * while we were blocked so we only use sync_trunc_off.
2991 		 *
2992 		 * This operation can blow out the buffer cache, EWOULDBLOCK
2993 		 * means we were unable to complete the deletion.  The
2994 		 * deletion will update sync_trunc_off in that case.
2995 		 */
2996 		error = hammer_ip_delete_range(&cursor, ip,
2997 						aligned_trunc_off,
2998 						0x7FFFFFFFFFFFFFFFLL, 2);
2999 		if (error == EWOULDBLOCK) {
3000 			ip->flags |= HAMMER_INODE_WOULDBLOCK;
3001 			error = 0;
3002 			goto defer_buffer_flush;
3003 		}
3004 
3005 		if (error)
3006 			goto done;
3007 
3008 		/*
3009 		 * Generate a REDO_TERM_TRUNC entry in the UNDO/REDO FIFO.
3010 		 *
3011 		 * XXX we do this even if we did not previously generate
3012 		 * a REDO_TRUNC record.  This operation may enclosed the
3013 		 * range for multiple prior truncation entries in the REDO
3014 		 * log.
3015 		 */
3016 		if (trans->hmp->version >= HAMMER_VOL_VERSION_FOUR &&
3017 		    (ip->flags & HAMMER_INODE_RDIRTY)) {
3018 			hammer_generate_redo(trans, ip, aligned_trunc_off,
3019 					     HAMMER_REDO_TERM_TRUNC,
3020 					     NULL, 0);
3021 		}
3022 
3023 		/*
3024 		 * Clear the truncation flag on the backend after we have
3025 		 * completed the deletions.  Backend data is now good again
3026 		 * (including new records we are about to sync, below).
3027 		 *
3028 		 * Leave sync_trunc_off intact.  As we write additional
3029 		 * records the backend will update sync_trunc_off.  This
3030 		 * tells the backend whether it can skip the overwrite
3031 		 * test.  This should work properly even when the backend
3032 		 * writes full blocks where the truncation point straddles
3033 		 * the block because the comparison is against the base
3034 		 * offset of the record.
3035 		 */
3036 		ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
3037 		/* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
3038 	} else {
3039 		error = 0;
3040 	}
3041 
3042 	/*
3043 	 * Now sync related records.  These will typically be directory
3044 	 * entries, records tracking direct-writes, or delete-on-disk records.
3045 	 */
3046 	if (error == 0) {
3047 		tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
3048 				    hammer_sync_record_callback, &cursor);
3049 		if (tmp_error < 0)
3050 			tmp_error = -error;
3051 		if (tmp_error)
3052 			error = tmp_error;
3053 	}
3054 	hammer_cache_node(&ip->cache[1], cursor.node);
3055 
3056 	/*
3057 	 * Re-seek for inode update, assuming our cache hasn't been ripped
3058 	 * out from under us.
3059 	 */
3060 	if (error == 0) {
3061 		tmp_node = hammer_ref_node_safe(trans, &ip->cache[0], &error);
3062 		if (tmp_node) {
3063 			hammer_cursor_downgrade(&cursor);
3064 			hammer_lock_sh(&tmp_node->lock);
3065 			if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
3066 				hammer_cursor_seek(&cursor, tmp_node, 0);
3067 			hammer_unlock(&tmp_node->lock);
3068 			hammer_rel_node(tmp_node);
3069 		}
3070 		error = 0;
3071 	}
3072 
3073 	/*
3074 	 * If we are deleting the inode the frontend had better not have
3075 	 * any active references on elements making up the inode.
3076 	 *
3077 	 * The call to hammer_ip_delete_clean() cleans up auxillary records
3078 	 * but not DB or DATA records.  Those must have already been deleted
3079 	 * by the normal truncation mechanic.
3080 	 */
3081 	if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
3082 		RB_EMPTY(&ip->rec_tree)  &&
3083 	    (ip->sync_flags & HAMMER_INODE_DELETING) &&
3084 	    (ip->flags & HAMMER_INODE_DELETED) == 0) {
3085 		int count1 = 0;
3086 
3087 		error = hammer_ip_delete_clean(&cursor, ip, &count1);
3088 		if (error == 0) {
3089 			ip->flags |= HAMMER_INODE_DELETED;
3090 			ip->sync_flags &= ~HAMMER_INODE_DELETING;
3091 			ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
3092 			KKASSERT(RB_EMPTY(&ip->rec_tree));
3093 
3094 			/*
3095 			 * Set delete_tid in both the frontend and backend
3096 			 * copy of the inode record.  The DELETED flag handles
3097 			 * this, do not set DDIRTY.
3098 			 */
3099 			ip->ino_leaf.base.delete_tid = trans->tid;
3100 			ip->sync_ino_leaf.base.delete_tid = trans->tid;
3101 			ip->ino_leaf.delete_ts = trans->time32;
3102 			ip->sync_ino_leaf.delete_ts = trans->time32;
3103 
3104 
3105 			/*
3106 			 * Adjust the inode count in the volume header
3107 			 */
3108 			hammer_sync_lock_sh(trans);
3109 			if (ip->flags & HAMMER_INODE_ONDISK) {
3110 				hammer_modify_volume_field(trans,
3111 							   trans->rootvol,
3112 							   vol0_stat_inodes);
3113 				--ip->hmp->rootvol->ondisk->vol0_stat_inodes;
3114 				hammer_modify_volume_done(trans->rootvol);
3115 			}
3116 			hammer_sync_unlock(trans);
3117 		}
3118 	}
3119 
3120 	if (error)
3121 		goto done;
3122 	ip->sync_flags &= ~HAMMER_INODE_BUFS;
3123 
3124 defer_buffer_flush:
3125 	/*
3126 	 * Now update the inode's on-disk inode-data and/or on-disk record.
3127 	 * DELETED and ONDISK are managed only in ip->flags.
3128 	 *
3129 	 * In the case of a defered buffer flush we still update the on-disk
3130 	 * inode to satisfy visibility requirements if there happen to be
3131 	 * directory dependancies.
3132 	 */
3133 	switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
3134 	case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
3135 		/*
3136 		 * If deleted and on-disk, don't set any additional flags.
3137 		 * the delete flag takes care of things.
3138 		 *
3139 		 * Clear flags which may have been set by the frontend.
3140 		 */
3141 		ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
3142 				    HAMMER_INODE_SDIRTY |
3143 				    HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
3144 				    HAMMER_INODE_DELETING);
3145 		break;
3146 	case HAMMER_INODE_DELETED:
3147 		/*
3148 		 * Take care of the case where a deleted inode was never
3149 		 * flushed to the disk in the first place.
3150 		 *
3151 		 * Clear flags which may have been set by the frontend.
3152 		 */
3153 		ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
3154 				    HAMMER_INODE_SDIRTY |
3155 				    HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
3156 				    HAMMER_INODE_DELETING);
3157 		while (RB_ROOT(&ip->rec_tree)) {
3158 			hammer_record_t record = RB_ROOT(&ip->rec_tree);
3159 			hammer_ref(&record->lock);
3160 			KKASSERT(hammer_oneref(&record->lock));
3161 			record->flags |= HAMMER_RECF_DELETED_BE;
3162 			++record->ip->rec_generation;
3163 			hammer_rel_mem_record(record);
3164 		}
3165 		break;
3166 	case HAMMER_INODE_ONDISK:
3167 		/*
3168 		 * If already on-disk, do not set any additional flags.
3169 		 */
3170 		break;
3171 	default:
3172 		/*
3173 		 * If not on-disk and not deleted, set DDIRTY to force
3174 		 * an initial record to be written.
3175 		 *
3176 		 * Also set the create_tid in both the frontend and backend
3177 		 * copy of the inode record.
3178 		 */
3179 		ip->ino_leaf.base.create_tid = trans->tid;
3180 		ip->ino_leaf.create_ts = trans->time32;
3181 		ip->sync_ino_leaf.base.create_tid = trans->tid;
3182 		ip->sync_ino_leaf.create_ts = trans->time32;
3183 		ip->sync_flags |= HAMMER_INODE_DDIRTY;
3184 		break;
3185 	}
3186 
3187 	/*
3188 	 * If DDIRTY or SDIRTY is set, write out a new record.
3189 	 * If the inode is already on-disk the old record is marked as
3190 	 * deleted.
3191 	 *
3192 	 * If DELETED is set hammer_update_inode() will delete the existing
3193 	 * record without writing out a new one.
3194 	 *
3195 	 * If *ONLY* the ITIMES flag is set we can update the record in-place.
3196 	 */
3197 	if (ip->flags & HAMMER_INODE_DELETED) {
3198 		error = hammer_update_inode(&cursor, ip);
3199 	} else
3200 	if (!(ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY)) &&
3201 	    (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
3202 		error = hammer_update_itimes(&cursor, ip);
3203 	} else
3204 	if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY |
3205 			      HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
3206 		error = hammer_update_inode(&cursor, ip);
3207 	}
3208 done:
3209 	if (ip->flags & HAMMER_INODE_MODMASK)
3210 		hammer_inode_dirty(ip);
3211 	if (error) {
3212 		hammer_critical_error(ip->hmp, ip, error,
3213 				      "while syncing inode");
3214 	}
3215 	hammer_done_cursor(&cursor);
3216 	return(error);
3217 }
3218 
3219 /*
3220  * This routine is called when the OS is no longer actively referencing
3221  * the inode (but might still be keeping it cached), or when releasing
3222  * the last reference to an inode.
3223  *
3224  * At this point if the inode's nlinks count is zero we want to destroy
3225  * it, which may mean destroying it on-media too.
3226  */
3227 void
3228 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
3229 {
3230 	struct vnode *vp;
3231 
3232 	/*
3233 	 * Set the DELETING flag when the link count drops to 0 and the
3234 	 * OS no longer has any opens on the inode.
3235 	 *
3236 	 * The backend will clear DELETING (a mod flag) and set DELETED
3237 	 * (a state flag) when it is actually able to perform the
3238 	 * operation.
3239 	 *
3240 	 * Don't reflag the deletion if the flusher is currently syncing
3241 	 * one that was already flagged.  A previously set DELETING flag
3242 	 * may bounce around flags and sync_flags until the operation is
3243 	 * completely done.
3244 	 *
3245 	 * Do not attempt to modify a snapshot inode (one set to read-only).
3246 	 */
3247 	if (ip->ino_data.nlinks == 0 &&
3248 	    ((ip->flags | ip->sync_flags) & (HAMMER_INODE_RO|HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
3249 		ip->flags |= HAMMER_INODE_DELETING;
3250 		ip->flags |= HAMMER_INODE_TRUNCATED;
3251 		ip->trunc_off = 0;
3252 		vp = NULL;
3253 		if (getvp) {
3254 			if (hammer_get_vnode(ip, &vp) != 0)
3255 				return;
3256 		}
3257 
3258 		/*
3259 		 * Final cleanup
3260 		 */
3261 		if (ip->vp)
3262 			nvtruncbuf(ip->vp, 0, HAMMER_BUFSIZE, 0, 0);
3263 		if (ip->flags & HAMMER_INODE_MODMASK)
3264 			hammer_inode_dirty(ip);
3265 		if (getvp)
3266 			vput(vp);
3267 	}
3268 }
3269 
3270 /*
3271  * After potentially resolving a dependancy the inode is tested
3272  * to determine whether it needs to be reflushed.
3273  */
3274 void
3275 hammer_test_inode(hammer_inode_t ip)
3276 {
3277 	if (ip->flags & HAMMER_INODE_REFLUSH) {
3278 		ip->flags &= ~HAMMER_INODE_REFLUSH;
3279 		hammer_ref(&ip->lock);
3280 		if (ip->flags & HAMMER_INODE_RESIGNAL) {
3281 			ip->flags &= ~HAMMER_INODE_RESIGNAL;
3282 			hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
3283 		} else {
3284 			hammer_flush_inode(ip, 0);
3285 		}
3286 		hammer_rel_inode(ip, 0);
3287 	}
3288 }
3289 
3290 /*
3291  * Clear the RECLAIM flag on an inode.  This occurs when the inode is
3292  * reassociated with a vp or just before it gets freed.
3293  *
3294  * Pipeline wakeups to threads blocked due to an excessive number of
3295  * detached inodes.  This typically occurs when atime updates accumulate
3296  * while scanning a directory tree.
3297  */
3298 static void
3299 hammer_inode_wakereclaims(hammer_inode_t ip)
3300 {
3301 	struct hammer_reclaim *reclaim;
3302 	hammer_mount_t hmp = ip->hmp;
3303 
3304 	if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
3305 		return;
3306 
3307 	--hammer_count_reclaims;
3308 	--hmp->count_reclaims;
3309 	ip->flags &= ~HAMMER_INODE_RECLAIM;
3310 
3311 	if ((reclaim = TAILQ_FIRST(&hmp->reclaim_list)) != NULL) {
3312 		KKASSERT(reclaim->count > 0);
3313 		if (--reclaim->count == 0) {
3314 			TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
3315 			wakeup(reclaim);
3316 		}
3317 	}
3318 }
3319 
3320 /*
3321  * Setup our reclaim pipeline.  We only let so many detached (and dirty)
3322  * inodes build up before we start blocking.  This routine is called
3323  * if a new inode is created or an inode is loaded from media.
3324  *
3325  * When we block we don't care *which* inode has finished reclaiming,
3326  * as long as one does.
3327  *
3328  * The reclaim pipeline is primarily governed by the auto-flush which is
3329  * 1/4 hammer_limit_reclaims.  We don't want to block if the count is
3330  * less than 1/2 hammer_limit_reclaims.  From 1/2 to full count is
3331  * dynamically governed.
3332  */
3333 void
3334 hammer_inode_waitreclaims(hammer_transaction_t trans)
3335 {
3336 	hammer_mount_t hmp = trans->hmp;
3337 	struct hammer_reclaim reclaim;
3338 	int lower_limit;
3339 
3340 	/*
3341 	 * Track inode load, delay if the number of reclaiming inodes is
3342 	 * between 2/4 and 4/4 hammer_limit_reclaims, depending.
3343 	 */
3344 	if (curthread->td_proc) {
3345 		struct hammer_inostats *stats;
3346 
3347 		stats = hammer_inode_inostats(hmp, curthread->td_proc->p_pid);
3348 		++stats->count;
3349 
3350 		if (stats->count > hammer_limit_reclaims / 2)
3351 			stats->count = hammer_limit_reclaims / 2;
3352 		lower_limit = hammer_limit_reclaims - stats->count;
3353 		if (hammer_debug_general & 0x10000) {
3354 			kprintf("pid %5d limit %d\n",
3355 				(int)curthread->td_proc->p_pid, lower_limit);
3356 		}
3357 	} else {
3358 		lower_limit = hammer_limit_reclaims * 3 / 4;
3359 	}
3360 	if (hmp->count_reclaims >= lower_limit) {
3361 		reclaim.count = 1;
3362 		TAILQ_INSERT_TAIL(&hmp->reclaim_list, &reclaim, entry);
3363 		tsleep(&reclaim, 0, "hmrrcm", hz);
3364 		if (reclaim.count > 0)
3365 			TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
3366 	}
3367 }
3368 
3369 /*
3370  * Keep track of reclaim statistics on a per-pid basis using a loose
3371  * 4-way set associative hash table.  Collisions inherit the count of
3372  * the previous entry.
3373  *
3374  * NOTE: We want to be careful here to limit the chain size.  If the chain
3375  *	 size is too large a pid will spread its stats out over too many
3376  *	 entries under certain types of heavy filesystem activity and
3377  *	 wind up not delaying long enough.
3378  */
3379 static
3380 struct hammer_inostats *
3381 hammer_inode_inostats(hammer_mount_t hmp, pid_t pid)
3382 {
3383 	struct hammer_inostats *stats;
3384 	int delta;
3385 	int chain;
3386 	static volatile int iterator;	/* we don't care about MP races */
3387 
3388 	/*
3389 	 * Chain up to 4 times to find our entry.
3390 	 */
3391 	for (chain = 0; chain < 4; ++chain) {
3392 		stats = &hmp->inostats[(pid + chain) & HAMMER_INOSTATS_HMASK];
3393 		if (stats->pid == pid)
3394 			break;
3395 	}
3396 
3397 	/*
3398 	 * Replace one of the four chaining entries with our new entry.
3399 	 */
3400 	if (chain == 4) {
3401 		stats = &hmp->inostats[(pid + (iterator++ & 3)) &
3402 				       HAMMER_INOSTATS_HMASK];
3403 		stats->pid = pid;
3404 	}
3405 
3406 	/*
3407 	 * Decay the entry
3408 	 */
3409 	if (stats->count && stats->ltick != ticks) {
3410 		delta = ticks - stats->ltick;
3411 		stats->ltick = ticks;
3412 		if (delta <= 0 || delta > hz * 60)
3413 			stats->count = 0;
3414 		else
3415 			stats->count = stats->count * hz / (hz + delta);
3416 	}
3417 	if (hammer_debug_general & 0x10000)
3418 		kprintf("pid %5d stats %d\n", (int)pid, stats->count);
3419 	return (stats);
3420 }
3421 
3422 #if 0
3423 
3424 /*
3425  * XXX not used, doesn't work very well due to the large batching nature
3426  * of flushes.
3427  *
3428  * A larger then normal backlog of inodes is sitting in the flusher,
3429  * enforce a general slowdown to let it catch up.  This routine is only
3430  * called on completion of a non-flusher-related transaction which
3431  * performed B-Tree node I/O.
3432  *
3433  * It is possible for the flusher to stall in a continuous load.
3434  * blogbench -i1000 -o seems to do a good job generating this sort of load.
3435  * If the flusher is unable to catch up the inode count can bloat until
3436  * we run out of kvm.
3437  *
3438  * This is a bit of a hack.
3439  */
3440 void
3441 hammer_inode_waithard(hammer_mount_t hmp)
3442 {
3443 	/*
3444 	 * Hysteresis.
3445 	 */
3446 	if (hmp->flags & HAMMER_MOUNT_FLUSH_RECOVERY) {
3447 		if (hmp->count_reclaims < hammer_limit_reclaims / 2 &&
3448 		    hmp->count_iqueued < hmp->count_inodes / 20) {
3449 			hmp->flags &= ~HAMMER_MOUNT_FLUSH_RECOVERY;
3450 			return;
3451 		}
3452 	} else {
3453 		if (hmp->count_reclaims < hammer_limit_reclaims ||
3454 		    hmp->count_iqueued < hmp->count_inodes / 10) {
3455 			return;
3456 		}
3457 		hmp->flags |= HAMMER_MOUNT_FLUSH_RECOVERY;
3458 	}
3459 
3460 	/*
3461 	 * Block for one flush cycle.
3462 	 */
3463 	hammer_flusher_wait_next(hmp);
3464 }
3465 
3466 #endif
3467