xref: /dflybsd-src/sys/vfs/hammer/hammer_inode.c (revision fc36a10bce8c5678d103e0498db849506d9dac68)
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 <vm/vm_page2.h>
36 
37 #include "hammer.h"
38 
39 static int	hammer_unload_inode(hammer_inode_t ip);
40 static void	hammer_free_inode(hammer_inode_t ip);
41 static void	hammer_flush_inode_core(hammer_inode_t ip,
42 					hammer_flush_group_t flg, int flags);
43 static int	hammer_setup_child_callback(hammer_record_t rec, void *data);
44 #if 0
45 static int	hammer_syncgrp_child_callback(hammer_record_t rec, void *data);
46 #endif
47 static int	hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
48 					hammer_flush_group_t flg);
49 static int	hammer_setup_parent_inodes_helper(hammer_record_t record,
50 					int depth, hammer_flush_group_t flg);
51 static void	hammer_inode_wakereclaims(hammer_inode_t ip);
52 static struct hammer_inostats *hammer_inode_inostats(hammer_mount_t hmp,
53 					pid_t pid);
54 static hammer_inode_t __hammer_find_inode(hammer_transaction_t trans,
55 					int64_t obj_id, hammer_tid_t asof,
56 					uint32_t localization);
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 	uint32_t localization = *(uint32_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, uint32_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 	hammer_inode_t 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 	hammer_inode_t 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(hammer_inode_t 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(hammer_inode_t ip, struct vnode **vpp)
282 {
283 	hammer_mount_t hmp;
284 	struct vnode *vp;
285 	int error = 0;
286 	uint8_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 				if (ip->obj_asof == hmp->asof) {
340 					if (ip->obj_localization ==
341 						HAMMER_DEF_LOCALIZATION)
342 						vsetflags(vp, VROOT);
343 					else
344 						vsetflags(vp, VPFSROOT);
345 				} else {
346 					vsetflags(vp, VPFSROOT);
347 				}
348 			}
349 
350 			vp->v_data = (void *)ip;
351 			/* vnode locked by getnewvnode() */
352 			/* make related vnode dirty if inode dirty? */
353 			hammer_unlock(&ip->lock);
354 			if (vp->v_type == VREG) {
355 				vinitvmio(vp, ip->ino_data.size,
356 					  hammer_blocksize(ip->ino_data.size),
357 					  hammer_blockoff(ip->ino_data.size));
358 			}
359 			vx_downgrade(vp);
360 			break;
361 		}
362 
363 		/*
364 		 * Interlock vnode clearing.  This does not prevent the
365 		 * vnode from going into a reclaimed state but it does
366 		 * prevent it from being destroyed or reused so the vget()
367 		 * will properly fail.
368 		 */
369 		hammer_lock_ex(&ip->lock);
370 		if ((vp = ip->vp) == NULL) {
371 			hammer_unlock(&ip->lock);
372 			continue;
373 		}
374 		vhold(vp);
375 		hammer_unlock(&ip->lock);
376 
377 		/*
378 		 * loop if the vget fails (aka races), or if the vp
379 		 * no longer matches ip->vp.
380 		 */
381 		if (vget(vp, LK_EXCLUSIVE) == 0) {
382 			if (vp == ip->vp) {
383 				vdrop(vp);
384 				break;
385 			}
386 			vput(vp);
387 		}
388 		vdrop(vp);
389 	}
390 	*vpp = vp;
391 	return(error);
392 }
393 
394 /*
395  * Locate all copies of the inode for obj_id compatible with the specified
396  * asof, reference, and issue the related call-back.  This routine is used
397  * for direct-io invalidation and does not create any new inodes.
398  */
399 void
400 hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo,
401 		            int (*callback)(hammer_inode_t ip, void *data),
402 			    void *data)
403 {
404 	hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root,
405 				   hammer_inode_info_cmp_all_history,
406 				   callback, iinfo);
407 }
408 
409 /*
410  * Acquire a HAMMER inode.  The returned inode is not locked.  These functions
411  * do not attach or detach the related vnode (use hammer_get_vnode() for
412  * that).
413  *
414  * The flags argument is only applied for newly created inodes, and only
415  * certain flags are inherited.
416  *
417  * Called from the frontend.
418  */
419 hammer_inode_t
420 hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip,
421 		 int64_t obj_id, hammer_tid_t asof, uint32_t localization,
422 		 int flags, int *errorp)
423 {
424 	hammer_mount_t hmp = trans->hmp;
425 	struct hammer_node_cache *cachep;
426 	struct hammer_cursor cursor;
427 	hammer_inode_t ip;
428 
429 
430 	/*
431 	 * Determine if we already have an inode cached.  If we do then
432 	 * we are golden.
433 	 *
434 	 * If we find an inode with no vnode we have to mark the
435 	 * transaction such that hammer_inode_waitreclaims() is
436 	 * called later on to avoid building up an infinite number
437 	 * of inodes.  Otherwise we can continue to * add new inodes
438 	 * faster then they can be disposed of, even with the tsleep
439 	 * delay.
440 	 *
441 	 * If we find a dummy inode we return a failure so dounlink
442 	 * (which does another lookup) doesn't try to mess with the
443 	 * link count.  hammer_vop_nresolve() uses hammer_get_dummy_inode()
444 	 * to ref dummy inodes.
445 	 */
446 loop:
447 	*errorp = 0;
448 	ip = __hammer_find_inode(trans, obj_id, asof, localization);
449 	if (ip) {
450 		if (ip->flags & HAMMER_INODE_DUMMY) {
451 			*errorp = ENOENT;
452 			return(NULL);
453 		}
454 		hammer_ref(&ip->lock);
455 		return(ip);
456 	}
457 
458 	/*
459 	 * Allocate a new inode structure and deal with races later.
460 	 */
461 	ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
462 	++hammer_count_inodes;
463 	++hmp->count_inodes;
464 	ip->obj_id = obj_id;
465 	ip->obj_asof = asof;
466 	ip->obj_localization = localization;
467 	ip->hmp = hmp;
468 	ip->flags = flags & HAMMER_INODE_RO;
469 	ip->cache[0].ip = ip;
470 	ip->cache[1].ip = ip;
471 	ip->cache[2].ip = ip;
472 	ip->cache[3].ip = ip;
473 	if (hmp->ronly)
474 		ip->flags |= HAMMER_INODE_RO;
475 	ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
476 		HAMMER_MAX_KEY;
477 	RB_INIT(&ip->rec_tree);
478 	TAILQ_INIT(&ip->target_list);
479 	hammer_ref(&ip->lock);
480 
481 	/*
482 	 * Locate the on-disk inode.  If this is a PFS root we always
483 	 * access the current version of the root inode and (if it is not
484 	 * a master) always access information under it with a snapshot
485 	 * TID.
486 	 *
487 	 * We cache recent inode lookups in this directory in dip->cache[2].
488 	 * If we can't find it we assume the inode we are looking for is
489 	 * close to the directory inode.
490 	 */
491 retry:
492 	cachep = NULL;
493 	if (dip) {
494 		if (dip->cache[2].node)
495 			cachep = &dip->cache[2];
496 		else
497 			cachep = &dip->cache[0];
498 	}
499 	hammer_init_cursor(trans, &cursor, cachep, NULL);
500 	cursor.key_beg.localization = localization | HAMMER_LOCALIZE_INODE;
501 	cursor.key_beg.obj_id = ip->obj_id;
502 	cursor.key_beg.key = 0;
503 	cursor.key_beg.create_tid = 0;
504 	cursor.key_beg.delete_tid = 0;
505 	cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
506 	cursor.key_beg.obj_type = 0;
507 
508 	cursor.asof = asof;
509 	cursor.flags = HAMMER_CURSOR_GET_DATA | 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 hammer_inode_t
608 hammer_get_dummy_inode(hammer_transaction_t trans, hammer_inode_t dip,
609 		 int64_t obj_id, hammer_tid_t asof, uint32_t localization,
610 		 int flags, int *errorp)
611 {
612 	hammer_mount_t hmp = trans->hmp;
613 	hammer_inode_t ip;
614 
615 	/*
616 	 * Determine if we already have an inode cached.  If we do then
617 	 * we are golden.
618 	 *
619 	 * If we find an inode with no vnode we have to mark the
620 	 * transaction such that hammer_inode_waitreclaims() is
621 	 * called later on to avoid building up an infinite number
622 	 * of inodes.  Otherwise we can continue to * add new inodes
623 	 * faster then they can be disposed of, even with the tsleep
624 	 * delay.
625 	 *
626 	 * If we find a non-fake inode we return an error.  Only fake
627 	 * inodes can be returned by this routine.
628 	 */
629 loop:
630 	*errorp = 0;
631 	ip = __hammer_find_inode(trans, obj_id, asof, localization);
632 	if (ip) {
633 		if ((ip->flags & HAMMER_INODE_DUMMY) == 0) {
634 			*errorp = ENOENT;
635 			return(NULL);
636 		}
637 		hammer_ref(&ip->lock);
638 		return(ip);
639 	}
640 
641 	/*
642 	 * Allocate a new inode structure and deal with races later.
643 	 */
644 	ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
645 	++hammer_count_inodes;
646 	++hmp->count_inodes;
647 	ip->obj_id = obj_id;
648 	ip->obj_asof = asof;
649 	ip->obj_localization = localization;
650 	ip->hmp = hmp;
651 	ip->flags = flags | HAMMER_INODE_RO | HAMMER_INODE_DUMMY;
652 	ip->cache[0].ip = ip;
653 	ip->cache[1].ip = ip;
654 	ip->cache[2].ip = ip;
655 	ip->cache[3].ip = ip;
656 	ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
657 		HAMMER_MAX_KEY;
658 	RB_INIT(&ip->rec_tree);
659 	TAILQ_INIT(&ip->target_list);
660 	hammer_ref(&ip->lock);
661 
662 	/*
663 	 * Populate the dummy inode.  Leave everything zero'd out.
664 	 *
665 	 * (ip->ino_leaf and ip->ino_data)
666 	 *
667 	 * Make the dummy inode a FIFO object which most copy programs
668 	 * will properly ignore.
669 	 */
670 	ip->save_trunc_off = ip->ino_data.size;
671 	ip->ino_data.obj_type = HAMMER_OBJTYPE_FIFO;
672 
673 	/*
674 	 * Locate and assign the pseudofs management structure to
675 	 * the inode.
676 	 */
677 	if (dip && dip->obj_localization == ip->obj_localization) {
678 		ip->pfsm = dip->pfsm;
679 		hammer_ref(&ip->pfsm->lock);
680 	} else {
681 		ip->pfsm = hammer_load_pseudofs(trans, ip->obj_localization,
682 						errorp);
683 		*errorp = 0;	/* ignore ENOENT */
684 	}
685 
686 	/*
687 	 * The inode is placed on the red-black tree and will be synced to
688 	 * the media when flushed or by the filesystem sync.  If this races
689 	 * another instantiation/lookup the insertion will fail.
690 	 *
691 	 * NOTE: Do not set HAMMER_INODE_ONDISK.  The inode is a fake.
692 	 */
693 	if (*errorp == 0) {
694 		if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
695 			hammer_free_inode(ip);
696 			goto loop;
697 		}
698 	} else {
699 		if (ip->flags & HAMMER_INODE_RSV_INODES) {
700 			ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
701 			--hmp->rsv_inodes;
702 		}
703 		hammer_free_inode(ip);
704 		ip = NULL;
705 	}
706 	trans->flags |= HAMMER_TRANSF_NEWINODE;
707 	return (ip);
708 }
709 
710 /*
711  * Return a referenced inode only if it is in our inode cache.
712  * Dummy inodes do not count.
713  */
714 hammer_inode_t
715 hammer_find_inode(hammer_transaction_t trans, int64_t obj_id,
716 		  hammer_tid_t asof, uint32_t localization)
717 {
718 	hammer_inode_t ip;
719 
720 	ip = __hammer_find_inode(trans, obj_id, asof, localization);
721 	if (ip) {
722 		if (ip->flags & HAMMER_INODE_DUMMY)
723 			ip = NULL;
724 		else
725 			hammer_ref(&ip->lock);
726 	}
727 	return(ip);
728 }
729 
730 /*
731  * Return a referenced inode only if it is in our inode cache.
732  * This function does not reference inode.
733  */
734 static hammer_inode_t
735 __hammer_find_inode(hammer_transaction_t trans, int64_t obj_id,
736 		  hammer_tid_t asof, uint32_t localization)
737 {
738 	hammer_mount_t hmp = trans->hmp;
739 	struct hammer_inode_info iinfo;
740 	hammer_inode_t ip;
741 
742 	iinfo.obj_id = obj_id;
743 	iinfo.obj_asof = asof;
744 	iinfo.obj_localization = localization;
745 
746 	ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
747 
748 	return(ip);
749 }
750 
751 /*
752  * Create a new filesystem object, returning the inode in *ipp.  The
753  * returned inode will be referenced.  The inode is created in-memory.
754  *
755  * If pfsm is non-NULL the caller wishes to create the root inode for
756  * a non-root PFS.
757  */
758 int
759 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
760 		    struct ucred *cred,
761 		    hammer_inode_t dip, const char *name, int namelen,
762 		    hammer_pseudofs_inmem_t pfsm, hammer_inode_t *ipp)
763 {
764 	hammer_mount_t hmp;
765 	hammer_inode_t ip;
766 	uid_t xuid;
767 	int error;
768 	int64_t namekey;
769 	uint32_t dummy;
770 
771 	hmp = trans->hmp;
772 
773 	/*
774 	 * Disallow the creation of new inodes in directories which
775 	 * have been deleted.  In HAMMER, this will cause a record
776 	 * syncing assertion later on in the flush code.
777 	 */
778 	if (dip && dip->ino_data.nlinks == 0) {
779 		*ipp = NULL;
780                 return (EINVAL);
781 	}
782 
783 	/*
784 	 * Allocate inode
785 	 */
786 	ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
787 	++hammer_count_inodes;
788 	++hmp->count_inodes;
789 	trans->flags |= HAMMER_TRANSF_NEWINODE;
790 
791 	if (pfsm) {
792 		KKASSERT(pfsm->localization != HAMMER_DEF_LOCALIZATION);
793 		ip->obj_id = HAMMER_OBJID_ROOT;
794 		ip->obj_localization = pfsm->localization;
795 	} else {
796 		KKASSERT(dip != NULL);
797 		namekey = hammer_direntry_namekey(dip, name, namelen, &dummy);
798 		ip->obj_id = hammer_alloc_objid(hmp, dip, namekey);
799 		ip->obj_localization = dip->obj_localization;
800 	}
801 
802 	KKASSERT(ip->obj_id != 0);
803 	ip->obj_asof = hmp->asof;
804 	ip->hmp = hmp;
805 	ip->flush_state = HAMMER_FST_IDLE;
806 	ip->flags = HAMMER_INODE_DDIRTY |
807 		    HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
808 	ip->cache[0].ip = ip;
809 	ip->cache[1].ip = ip;
810 	ip->cache[2].ip = ip;
811 	ip->cache[3].ip = ip;
812 
813 	ip->trunc_off = HAMMER_MAX_KEY;
814 	/* ip->save_trunc_off = 0; (already zero) */
815 	RB_INIT(&ip->rec_tree);
816 	TAILQ_INIT(&ip->target_list);
817 
818 	ip->ino_data.atime = trans->time;
819 	ip->ino_data.mtime = trans->time;
820 	ip->ino_data.size = 0;
821 	ip->ino_data.nlinks = 0;
822 
823 	/*
824 	 * A nohistory designator on the parent directory is inherited by
825 	 * the child.  We will do this even for pseudo-fs creation... the
826 	 * sysad can turn it off.
827 	 */
828 	if (dip) {
829 		ip->ino_data.uflags = dip->ino_data.uflags &
830 				      (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
831 	}
832 
833 	ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
834 	ip->ino_leaf.base.localization = ip->obj_localization |
835 					 HAMMER_LOCALIZE_INODE;
836 	ip->ino_leaf.base.obj_id = ip->obj_id;
837 	ip->ino_leaf.base.key = 0;
838 	ip->ino_leaf.base.create_tid = 0;
839 	ip->ino_leaf.base.delete_tid = 0;
840 	ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
841 	ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
842 
843 	ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
844 	ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
845 	ip->ino_data.mode = vap->va_mode;
846 	ip->ino_data.ctime = trans->time;
847 
848 	/*
849 	 * If we are running version 2 or greater directory entries are
850 	 * inode-localized instead of data-localized.
851 	 */
852 	if (trans->hmp->version >= HAMMER_VOL_VERSION_TWO) {
853 		if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
854 			ip->ino_data.cap_flags |=
855 				HAMMER_INODE_CAP_DIR_LOCAL_INO;
856 		}
857 	}
858 	if (trans->hmp->version >= HAMMER_VOL_VERSION_SIX) {
859 		if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
860 			ip->ino_data.cap_flags |=
861 				HAMMER_INODE_CAP_DIRHASH_ALG1;
862 		}
863 	}
864 
865 	/*
866 	 * Setup the ".." pointer.  This only needs to be done for directories
867 	 * but we do it for all objects as a recovery aid if dip exists.
868 	 * The inode is probably a PFS root if dip is NULL.
869 	 */
870 	if (dip)
871 		ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
872 
873 	switch(ip->ino_leaf.base.obj_type) {
874 	case HAMMER_OBJTYPE_CDEV:
875 	case HAMMER_OBJTYPE_BDEV:
876 		ip->ino_data.rmajor = vap->va_rmajor;
877 		ip->ino_data.rminor = vap->va_rminor;
878 		break;
879 	default:
880 		break;
881 	}
882 
883 	/*
884 	 * Calculate default uid/gid and overwrite with information from
885 	 * the vap.
886 	 */
887 	if (dip) {
888 		xuid = hammer_to_unix_xid(&dip->ino_data.uid);
889 		xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
890 					     xuid, cred, &vap->va_mode);
891 	} else {
892 		xuid = 0;
893 	}
894 	ip->ino_data.mode = vap->va_mode;
895 
896 	if (vap->va_vaflags & VA_UID_UUID_VALID)
897 		ip->ino_data.uid = vap->va_uid_uuid;
898 	else if (vap->va_uid != (uid_t)VNOVAL)
899 		hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
900 	else
901 		hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
902 
903 	if (vap->va_vaflags & VA_GID_UUID_VALID)
904 		ip->ino_data.gid = vap->va_gid_uuid;
905 	else if (vap->va_gid != (gid_t)VNOVAL)
906 		hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
907 	else if (dip)
908 		ip->ino_data.gid = dip->ino_data.gid;
909 
910 	hammer_ref(&ip->lock);
911 
912 	if (pfsm) {
913 		ip->pfsm = pfsm;
914 		hammer_ref(&pfsm->lock);
915 		error = 0;
916 	} else if (dip->obj_localization == ip->obj_localization) {
917 		ip->pfsm = dip->pfsm;
918 		hammer_ref(&ip->pfsm->lock);
919 		error = 0;
920 	} else {
921 		ip->pfsm = hammer_load_pseudofs(trans,
922 						ip->obj_localization,
923 						&error);
924 		error = 0;	/* ignore ENOENT */
925 	}
926 
927 	if (error) {
928 		hammer_free_inode(ip);
929 		ip = NULL;
930 	} else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
931 		hpanic("duplicate obj_id %jx", (intmax_t)ip->obj_id);
932 		/* not reached */
933 		hammer_free_inode(ip);
934 	}
935 	*ipp = ip;
936 	return(error);
937 }
938 
939 /*
940  * Final cleanup / freeing of an inode structure
941  */
942 static void
943 hammer_free_inode(hammer_inode_t ip)
944 {
945 	hammer_mount_t hmp;
946 
947 	hmp = ip->hmp;
948 	KKASSERT(hammer_oneref(&ip->lock));
949 	hammer_uncache_node(&ip->cache[0]);
950 	hammer_uncache_node(&ip->cache[1]);
951 	hammer_uncache_node(&ip->cache[2]);
952 	hammer_uncache_node(&ip->cache[3]);
953 	hammer_inode_wakereclaims(ip);
954 	if (ip->objid_cache)
955 		hammer_clear_objid(ip);
956 	--hammer_count_inodes;
957 	--hmp->count_inodes;
958 	if (ip->pfsm) {
959 		hammer_rel_pseudofs(hmp, ip->pfsm);
960 		ip->pfsm = NULL;
961 	}
962 	kfree(ip, hmp->m_inodes);
963 }
964 
965 /*
966  * Retrieve pseudo-fs data.  NULL will never be returned.
967  *
968  * If an error occurs *errorp will be set and a default template is returned,
969  * otherwise *errorp is set to 0.  Typically when an error occurs it will
970  * be ENOENT.
971  */
972 hammer_pseudofs_inmem_t
973 hammer_load_pseudofs(hammer_transaction_t trans,
974 		     uint32_t localization, int *errorp)
975 {
976 	hammer_mount_t hmp = trans->hmp;
977 	hammer_inode_t ip;
978 	hammer_pseudofs_inmem_t pfsm;
979 	struct hammer_cursor cursor;
980 	int bytes;
981 
982 retry:
983 	pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
984 	if (pfsm) {
985 		hammer_ref(&pfsm->lock);
986 		*errorp = 0;
987 		return(pfsm);
988 	}
989 
990 	/*
991 	 * PFS records are associated with the root inode (not the PFS root
992 	 * inode, but the real root).  Avoid an infinite recursion if loading
993 	 * the PFS for the real root.
994 	 */
995 	if (localization) {
996 		ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
997 				      HAMMER_MAX_TID,
998 				      HAMMER_DEF_LOCALIZATION, 0, errorp);
999 	} else {
1000 		ip = NULL;
1001 	}
1002 
1003 	pfsm = kmalloc(sizeof(*pfsm), hmp->m_misc, M_WAITOK | M_ZERO);
1004 	pfsm->localization = localization;
1005 	pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
1006 	pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
1007 
1008 	hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
1009 	cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION |
1010 				      HAMMER_LOCALIZE_MISC;
1011 	cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
1012 	cursor.key_beg.create_tid = 0;
1013 	cursor.key_beg.delete_tid = 0;
1014 	cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
1015 	cursor.key_beg.obj_type = 0;
1016 	cursor.key_beg.key = localization;
1017 	cursor.asof = HAMMER_MAX_TID;
1018 	cursor.flags |= HAMMER_CURSOR_ASOF;
1019 
1020 	if (ip)
1021 		*errorp = hammer_ip_lookup(&cursor);
1022 	else
1023 		*errorp = hammer_btree_lookup(&cursor);
1024 	if (*errorp == 0) {
1025 		*errorp = hammer_ip_resolve_data(&cursor);
1026 		if (*errorp == 0) {
1027 			if (hammer_is_pfs_deleted(&cursor.data->pfsd)) {
1028 				*errorp = ENOENT;
1029 			} else {
1030 				bytes = cursor.leaf->data_len;
1031 				if (bytes > sizeof(pfsm->pfsd))
1032 					bytes = sizeof(pfsm->pfsd);
1033 				bcopy(cursor.data, &pfsm->pfsd, bytes);
1034 			}
1035 		}
1036 	}
1037 	hammer_done_cursor(&cursor);
1038 
1039 	pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
1040 	hammer_ref(&pfsm->lock);
1041 	if (ip)
1042 		hammer_rel_inode(ip, 0);
1043 	if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
1044 		kfree(pfsm, hmp->m_misc);
1045 		goto retry;
1046 	}
1047 	return(pfsm);
1048 }
1049 
1050 /*
1051  * Store pseudo-fs data.  The backend will automatically delete any prior
1052  * on-disk pseudo-fs data but we have to delete in-memory versions.
1053  */
1054 int
1055 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
1056 {
1057 	struct hammer_cursor cursor;
1058 	hammer_record_t record;
1059 	hammer_inode_t ip;
1060 	int error;
1061 
1062 	/*
1063 	 * PFS records are associated with the root inode (not the PFS root
1064 	 * inode, but the real root).
1065 	 */
1066 	ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1067 			      HAMMER_DEF_LOCALIZATION, 0, &error);
1068 retry:
1069 	pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
1070 	hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
1071 	cursor.key_beg.localization = ip->obj_localization |
1072 				      HAMMER_LOCALIZE_MISC;
1073 	cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
1074 	cursor.key_beg.create_tid = 0;
1075 	cursor.key_beg.delete_tid = 0;
1076 	cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
1077 	cursor.key_beg.obj_type = 0;
1078 	cursor.key_beg.key = pfsm->localization;
1079 	cursor.asof = HAMMER_MAX_TID;
1080 	cursor.flags |= HAMMER_CURSOR_ASOF;
1081 
1082 	/*
1083 	 * Replace any in-memory version of the record.
1084 	 */
1085 	error = hammer_ip_lookup(&cursor);
1086 	if (error == 0 && hammer_cursor_inmem(&cursor)) {
1087 		record = cursor.iprec;
1088 		if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
1089 			KKASSERT(cursor.deadlk_rec == NULL);
1090 			hammer_ref(&record->lock);
1091 			cursor.deadlk_rec = record;
1092 			error = EDEADLK;
1093 		} else {
1094 			record->flags |= HAMMER_RECF_DELETED_FE;
1095 			error = 0;
1096 		}
1097 	}
1098 
1099 	/*
1100 	 * Allocate replacement general record.  The backend flush will
1101 	 * delete any on-disk version of the record.
1102 	 */
1103 	if (error == 0 || error == ENOENT) {
1104 		record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
1105 		record->type = HAMMER_MEM_RECORD_GENERAL;
1106 
1107 		record->leaf.base.localization = ip->obj_localization |
1108 						 HAMMER_LOCALIZE_MISC;
1109 		record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
1110 		record->leaf.base.key = pfsm->localization;
1111 		record->leaf.data_len = sizeof(pfsm->pfsd);
1112 		bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
1113 		error = hammer_ip_add_record(trans, record);
1114 	}
1115 	hammer_done_cursor(&cursor);
1116 	if (error == EDEADLK)
1117 		goto retry;
1118 	hammer_rel_inode(ip, 0);
1119 	return(error);
1120 }
1121 
1122 /*
1123  * Create a root directory for a PFS if one does not alredy exist.
1124  *
1125  * The PFS root stands alone so we must also bump the nlinks count
1126  * to prevent it from being destroyed on release.
1127  *
1128  * Make sure a caller isn't creating a PFS from non-root PFS.
1129  */
1130 int
1131 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1132 		       hammer_pseudofs_inmem_t pfsm, hammer_inode_t dip)
1133 {
1134 	hammer_inode_t ip;
1135 	struct vattr vap;
1136 	int error;
1137 
1138 	ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1139 			      pfsm->localization, 0, &error);
1140 	if (ip == NULL) {
1141 		if (lo_to_pfs(dip->obj_localization) != HAMMER_ROOT_PFSID) {
1142 			hmkprintf(trans->hmp,
1143 				"Warning: creating a PFS from non-root PFS "
1144 				"is not allowed\n");
1145 			return(EINVAL);
1146 		}
1147 		vattr_null(&vap);
1148 		vap.va_mode = 0755;
1149 		vap.va_type = VDIR;
1150 		error = hammer_create_inode(trans, &vap, cred,
1151 					    NULL, NULL, 0,
1152 					    pfsm, &ip);
1153 		if (error == 0) {
1154 			++ip->ino_data.nlinks;
1155 			hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
1156 		}
1157 	}
1158 	if (ip)
1159 		hammer_rel_inode(ip, 0);
1160 	return(error);
1161 }
1162 
1163 /*
1164  * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
1165  * if we are unable to disassociate all the inodes.
1166  */
1167 static
1168 int
1169 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
1170 {
1171 	int res;
1172 
1173 	hammer_ref(&ip->lock);
1174 	if (ip->vp && (ip->vp->v_flag & VPFSROOT)) {
1175 		/*
1176 		 * The hammer pfs-upgrade directive itself might have the
1177 		 * root of the pfs open.  Just allow it.
1178 		 */
1179 		res = 0;
1180 	} else {
1181 		/*
1182 		 * Don't allow any subdirectories or files to be open.
1183 		 */
1184 		if (hammer_isactive(&ip->lock) == 2 && ip->vp)
1185 			vclean_unlocked(ip->vp);	/* might not succeed */
1186 		if (hammer_isactive(&ip->lock) == 1 && ip->vp == NULL)
1187 			res = 0;
1188 		else
1189 			res = -1;	/* stop, someone is using the inode */
1190 	}
1191 	hammer_rel_inode(ip, 0);
1192 	return(res);
1193 }
1194 
1195 int
1196 hammer_unload_pseudofs(hammer_transaction_t trans, uint32_t localization)
1197 {
1198 	int res;
1199 	int try;
1200 
1201 	for (try = res = 0; try < 4; ++try) {
1202 		res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
1203 					   hammer_inode_pfs_cmp,
1204 					   hammer_unload_pseudofs_callback,
1205 					   &localization);
1206 		if (res == 0 && try > 1)
1207 			break;
1208 		hammer_flusher_sync(trans->hmp);
1209 	}
1210 	if (res != 0)
1211 		res = ENOTEMPTY;
1212 	return(res);
1213 }
1214 
1215 
1216 /*
1217  * Release a reference on a PFS
1218  */
1219 void
1220 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
1221 {
1222 	hammer_rel(&pfsm->lock);
1223 	if (hammer_norefs(&pfsm->lock)) {
1224 		RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
1225 		kfree(pfsm, hmp->m_misc);
1226 	}
1227 }
1228 
1229 /*
1230  * Called by hammer_sync_inode().
1231  */
1232 static int
1233 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
1234 {
1235 	hammer_transaction_t trans = cursor->trans;
1236 	hammer_record_t record;
1237 	int error;
1238 	int redirty;
1239 
1240 retry:
1241 	error = 0;
1242 
1243 	/*
1244 	 * If the inode has a presence on-disk then locate it and mark
1245 	 * it deleted, setting DELONDISK.
1246 	 *
1247 	 * The record may or may not be physically deleted, depending on
1248 	 * the retention policy.
1249 	 */
1250 	if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
1251 	    HAMMER_INODE_ONDISK) {
1252 		hammer_normalize_cursor(cursor);
1253 		cursor->key_beg.localization = ip->obj_localization |
1254 					       HAMMER_LOCALIZE_INODE;
1255 		cursor->key_beg.obj_id = ip->obj_id;
1256 		cursor->key_beg.key = 0;
1257 		cursor->key_beg.create_tid = 0;
1258 		cursor->key_beg.delete_tid = 0;
1259 		cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1260 		cursor->key_beg.obj_type = 0;
1261 		cursor->asof = ip->obj_asof;
1262 		cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1263 		cursor->flags |= HAMMER_CURSOR_ASOF;
1264 		cursor->flags |= HAMMER_CURSOR_BACKEND;
1265 
1266 		error = hammer_btree_lookup(cursor);
1267 		if (hammer_debug_inode)
1268 			hdkprintf("IPDEL %p %08x %d\n", ip, ip->flags, error);
1269 
1270 		if (error == 0) {
1271 			error = hammer_ip_delete_record(cursor, ip, trans->tid);
1272 			if (hammer_debug_inode)
1273 				hdkprintf("error %d\n", error);
1274 			if (error == 0) {
1275 				ip->flags |= HAMMER_INODE_DELONDISK;
1276 			}
1277 			if (cursor->node)
1278 				hammer_cache_node(&ip->cache[0], cursor->node);
1279 		}
1280 		if (error == EDEADLK) {
1281 			hammer_done_cursor(cursor);
1282 			error = hammer_init_cursor(trans, cursor,
1283 						   &ip->cache[0], ip);
1284 			if (hammer_debug_inode)
1285 				hdkprintf("IPDED %p %d\n", ip, error);
1286 			if (error == 0)
1287 				goto retry;
1288 		}
1289 	}
1290 
1291 	/*
1292 	 * Ok, write out the initial record or a new record (after deleting
1293 	 * the old one), unless the DELETED flag is set.  This routine will
1294 	 * clear DELONDISK if it writes out a record.
1295 	 *
1296 	 * Update our inode statistics if this is the first application of
1297 	 * the inode on-disk.
1298 	 */
1299 	if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
1300 		/*
1301 		 * Generate a record and write it to the media.  We clean-up
1302 		 * the state before releasing so we do not have to set-up
1303 		 * a flush_group.
1304 		 */
1305 		record = hammer_alloc_mem_record(ip, 0);
1306 		record->type = HAMMER_MEM_RECORD_INODE;
1307 		record->flush_state = HAMMER_FST_FLUSH;
1308 		record->leaf = ip->sync_ino_leaf;
1309 		record->leaf.base.create_tid = trans->tid;
1310 		record->leaf.data_len = sizeof(ip->sync_ino_data);
1311 		record->leaf.create_ts = trans->time32;
1312 		record->data = (void *)&ip->sync_ino_data;
1313 		record->flags |= HAMMER_RECF_INTERLOCK_BE;
1314 
1315 		/*
1316 		 * If this flag is set we cannot sync the new file size
1317 		 * because we haven't finished related truncations.  The
1318 		 * inode will be flushed in another flush group to finish
1319 		 * the job.
1320 		 */
1321 		if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1322 		    ip->sync_ino_data.size != ip->ino_data.size) {
1323 			redirty = 1;
1324 			ip->sync_ino_data.size = ip->ino_data.size;
1325 		} else {
1326 			redirty = 0;
1327 		}
1328 
1329 		for (;;) {
1330 			error = hammer_ip_sync_record_cursor(cursor, record);
1331 			if (hammer_debug_inode)
1332 				hdkprintf("GENREC %p rec %08x %d\n",
1333 					ip, record->flags, error);
1334 			if (error != EDEADLK)
1335 				break;
1336 			hammer_done_cursor(cursor);
1337 			error = hammer_init_cursor(trans, cursor,
1338 						   &ip->cache[0], ip);
1339 			if (hammer_debug_inode)
1340 				hdkprintf("GENREC reinit %d\n", error);
1341 			if (error)
1342 				break;
1343 		}
1344 
1345 		/*
1346 		 * Note:  The record was never on the inode's record tree
1347 		 * so just wave our hands importantly and destroy it.
1348 		 */
1349 		record->flags |= HAMMER_RECF_COMMITTED;
1350 		record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1351 		record->flush_state = HAMMER_FST_IDLE;
1352 		++ip->rec_generation;
1353 		hammer_rel_mem_record(record);
1354 
1355 		/*
1356 		 * Finish up.
1357 		 */
1358 		if (error == 0) {
1359 			if (hammer_debug_inode)
1360 				hdkprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1361 			ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1362 					    HAMMER_INODE_SDIRTY |
1363 					    HAMMER_INODE_ATIME |
1364 					    HAMMER_INODE_MTIME);
1365 			ip->flags &= ~HAMMER_INODE_DELONDISK;
1366 			if (redirty)
1367 				ip->sync_flags |= HAMMER_INODE_DDIRTY;
1368 
1369 			/*
1370 			 * Root volume count of inodes
1371 			 */
1372 			hammer_sync_lock_sh(trans);
1373 			if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1374 				hammer_modify_volume_field(trans,
1375 							   trans->rootvol,
1376 							   vol0_stat_inodes);
1377 				++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1378 				hammer_modify_volume_done(trans->rootvol);
1379 				ip->flags |= HAMMER_INODE_ONDISK;
1380 				if (hammer_debug_inode)
1381 					hdkprintf("NOWONDISK %p\n", ip);
1382 			}
1383 			hammer_sync_unlock(trans);
1384 		}
1385 	}
1386 
1387 	/*
1388 	 * If the inode has been destroyed, clean out any left-over flags
1389 	 * that may have been set by the frontend.
1390 	 */
1391 	if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
1392 		ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1393 				    HAMMER_INODE_SDIRTY |
1394 				    HAMMER_INODE_ATIME |
1395 				    HAMMER_INODE_MTIME);
1396 	}
1397 	return(error);
1398 }
1399 
1400 /*
1401  * Update only the itimes fields.
1402  *
1403  * ATIME can be updated without generating any UNDO.  MTIME is updated
1404  * with UNDO so it is guaranteed to be synchronized properly in case of
1405  * a crash.
1406  *
1407  * Neither field is included in the B-Tree leaf element's CRC, which is how
1408  * we can get away with updating ATIME the way we do.
1409  */
1410 static int
1411 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1412 {
1413 	hammer_transaction_t trans = cursor->trans;
1414 	int error;
1415 
1416 retry:
1417 	if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1418 	    HAMMER_INODE_ONDISK) {
1419 		return(0);
1420 	}
1421 
1422 	hammer_normalize_cursor(cursor);
1423 	cursor->key_beg.localization = ip->obj_localization |
1424 				       HAMMER_LOCALIZE_INODE;
1425 	cursor->key_beg.obj_id = ip->obj_id;
1426 	cursor->key_beg.key = 0;
1427 	cursor->key_beg.create_tid = 0;
1428 	cursor->key_beg.delete_tid = 0;
1429 	cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1430 	cursor->key_beg.obj_type = 0;
1431 	cursor->asof = ip->obj_asof;
1432 	cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1433 	cursor->flags |= HAMMER_CURSOR_ASOF;
1434 	cursor->flags |= HAMMER_CURSOR_GET_DATA;
1435 	cursor->flags |= HAMMER_CURSOR_BACKEND;
1436 
1437 	error = hammer_btree_lookup(cursor);
1438 	if (error == 0) {
1439 		hammer_cache_node(&ip->cache[0], cursor->node);
1440 		if (ip->sync_flags & HAMMER_INODE_MTIME) {
1441 			/*
1442 			 * Updating MTIME requires an UNDO.  Just cover
1443 			 * both atime and mtime.
1444 			 */
1445 			hammer_sync_lock_sh(trans);
1446 			hammer_modify_buffer(trans, cursor->data_buffer,
1447 				&cursor->data->inode.mtime,
1448 				sizeof(cursor->data->inode.atime) +
1449 				sizeof(cursor->data->inode.mtime));
1450 			cursor->data->inode.atime = ip->sync_ino_data.atime;
1451 			cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1452 			hammer_modify_buffer_done(cursor->data_buffer);
1453 			hammer_sync_unlock(trans);
1454 		} else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1455 			/*
1456 			 * Updating atime only can be done in-place with
1457 			 * no UNDO.
1458 			 */
1459 			hammer_sync_lock_sh(trans);
1460 			hammer_modify_buffer_noundo(trans, cursor->data_buffer);
1461 			cursor->data->inode.atime = ip->sync_ino_data.atime;
1462 			hammer_modify_buffer_done(cursor->data_buffer);
1463 			hammer_sync_unlock(trans);
1464 		}
1465 		ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1466 	}
1467 	if (error == EDEADLK) {
1468 		hammer_done_cursor(cursor);
1469 		error = hammer_init_cursor(trans, cursor, &ip->cache[0], ip);
1470 		if (error == 0)
1471 			goto retry;
1472 	}
1473 	return(error);
1474 }
1475 
1476 /*
1477  * Release a reference on an inode, flush as requested.
1478  *
1479  * On the last reference we queue the inode to the flusher for its final
1480  * disposition.
1481  */
1482 void
1483 hammer_rel_inode(hammer_inode_t ip, int flush)
1484 {
1485 	/*
1486 	 * Handle disposition when dropping the last ref.
1487 	 */
1488 	for (;;) {
1489 		if (hammer_oneref(&ip->lock)) {
1490 			/*
1491 			 * Determine whether on-disk action is needed for
1492 			 * the inode's final disposition.
1493 			 */
1494 			KKASSERT(ip->vp == NULL);
1495 			hammer_inode_unloadable_check(ip, 0);
1496 			if (ip->flags & HAMMER_INODE_MODMASK) {
1497 				hammer_flush_inode(ip, 0);
1498 			} else if (hammer_oneref(&ip->lock)) {
1499 				hammer_unload_inode(ip);
1500 				break;
1501 			}
1502 		} else {
1503 			if (flush)
1504 				hammer_flush_inode(ip, 0);
1505 
1506 			/*
1507 			 * The inode still has multiple refs, try to drop
1508 			 * one ref.
1509 			 */
1510 			KKASSERT(hammer_isactive(&ip->lock) >= 1);
1511 			if (hammer_isactive(&ip->lock) > 1) {
1512 				hammer_rel(&ip->lock);
1513 				break;
1514 			}
1515 		}
1516 	}
1517 }
1518 
1519 /*
1520  * Unload and destroy the specified inode.  Must be called with one remaining
1521  * reference.  The reference is disposed of.
1522  *
1523  * The inode must be completely clean.
1524  */
1525 static int
1526 hammer_unload_inode(hammer_inode_t ip)
1527 {
1528 	hammer_mount_t hmp = ip->hmp;
1529 
1530 	KASSERT(hammer_oneref(&ip->lock),
1531 		("hammer_unload_inode: %d refs", hammer_isactive(&ip->lock)));
1532 	KKASSERT(ip->vp == NULL);
1533 	KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1534 	KKASSERT(ip->cursor_ip_refs == 0);
1535 	KKASSERT(hammer_notlocked(&ip->lock));
1536 	KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1537 
1538 	KKASSERT(RB_EMPTY(&ip->rec_tree));
1539 	KKASSERT(TAILQ_EMPTY(&ip->target_list));
1540 
1541 	if (ip->flags & HAMMER_INODE_RDIRTY) {
1542 		RB_REMOVE(hammer_redo_rb_tree, &hmp->rb_redo_root, ip);
1543 		ip->flags &= ~HAMMER_INODE_RDIRTY;
1544 	}
1545 	RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1546 
1547 	hammer_free_inode(ip);
1548 	return(0);
1549 }
1550 
1551 /*
1552  * Called during unmounting if a critical error occured.  The in-memory
1553  * inode and all related structures are destroyed.
1554  *
1555  * If a critical error did not occur the unmount code calls the standard
1556  * release and asserts that the inode is gone.
1557  */
1558 int
1559 hammer_destroy_inode_callback(hammer_inode_t ip, void *data __unused)
1560 {
1561 	hammer_record_t rec;
1562 
1563 	/*
1564 	 * Get rid of the inodes in-memory records, regardless of their
1565 	 * state, and clear the mod-mask.
1566 	 */
1567 	while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1568 		TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1569 		rec->target_ip = NULL;
1570 		if (rec->flush_state == HAMMER_FST_SETUP)
1571 			rec->flush_state = HAMMER_FST_IDLE;
1572 	}
1573 	while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1574 		if (rec->flush_state == HAMMER_FST_FLUSH)
1575 			--rec->flush_group->refs;
1576 		else
1577 			hammer_ref(&rec->lock);
1578 		KKASSERT(hammer_oneref(&rec->lock));
1579 		rec->flush_state = HAMMER_FST_IDLE;
1580 		rec->flush_group = NULL;
1581 		rec->flags |= HAMMER_RECF_DELETED_FE; /* wave hands */
1582 		rec->flags |= HAMMER_RECF_DELETED_BE; /* wave hands */
1583 		++ip->rec_generation;
1584 		hammer_rel_mem_record(rec);
1585 	}
1586 	ip->flags &= ~HAMMER_INODE_MODMASK;
1587 	ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1588 	KKASSERT(ip->vp == NULL);
1589 
1590 	/*
1591 	 * Remove the inode from any flush group, force it idle.  FLUSH
1592 	 * and SETUP states have an inode ref.
1593 	 */
1594 	switch(ip->flush_state) {
1595 	case HAMMER_FST_FLUSH:
1596 		RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
1597 		--ip->flush_group->refs;
1598 		ip->flush_group = NULL;
1599 		/* fall through */
1600 	case HAMMER_FST_SETUP:
1601 		hammer_rel(&ip->lock);
1602 		ip->flush_state = HAMMER_FST_IDLE;
1603 		/* fall through */
1604 	case HAMMER_FST_IDLE:
1605 		break;
1606 	}
1607 
1608 	/*
1609 	 * There shouldn't be any associated vnode.  The unload needs at
1610 	 * least one ref, if we do have a vp steal its ip ref.
1611 	 */
1612 	if (ip->vp) {
1613 		hdkprintf("Unexpected vnode association ip %p vp %p\n",
1614 			ip, ip->vp);
1615 		ip->vp->v_data = NULL;
1616 		ip->vp = NULL;
1617 	} else {
1618 		hammer_ref(&ip->lock);
1619 	}
1620 	hammer_unload_inode(ip);
1621 	return(0);
1622 }
1623 
1624 /*
1625  * Called on mount -u when switching from RW to RO or vise-versa.  Adjust
1626  * the read-only flag for cached inodes.
1627  *
1628  * This routine is called from a RB_SCAN().
1629  */
1630 int
1631 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1632 {
1633 	hammer_mount_t hmp = ip->hmp;
1634 
1635 	if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1636 		ip->flags |= HAMMER_INODE_RO;
1637 	else
1638 		ip->flags &= ~HAMMER_INODE_RO;
1639 	return(0);
1640 }
1641 
1642 /*
1643  * A transaction has modified an inode, requiring updates as specified by
1644  * the passed flags.
1645  *
1646  * HAMMER_INODE_DDIRTY: Inode data has been updated, not incl mtime/atime,
1647  *			and not including size changes due to write-append
1648  *			(but other size changes are included).
1649  * HAMMER_INODE_SDIRTY: Inode data has been updated, size changes due to
1650  *			write-append.
1651  * HAMMER_INODE_XDIRTY: Dirty in-memory records
1652  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
1653  * HAMMER_INODE_DELETED: Inode record/data must be deleted
1654  * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1655  */
1656 void
1657 hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags)
1658 {
1659 	/*
1660 	 * ronly of 0 or 2 does not trigger assertion.
1661 	 * 2 is a special error state
1662 	 */
1663 	KKASSERT(ip->hmp->ronly != 1 ||
1664 		  (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
1665 			    HAMMER_INODE_SDIRTY |
1666 			    HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1667 			    HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1668 	if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1669 		ip->flags |= HAMMER_INODE_RSV_INODES;
1670 		++ip->hmp->rsv_inodes;
1671 	}
1672 
1673 	/*
1674 	 * Set the NEWINODE flag in the transaction if the inode
1675 	 * transitions to a dirty state.  This is used to track
1676 	 * the load on the inode cache.
1677 	 */
1678 	if (trans &&
1679 	    (ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1680 	    (flags & HAMMER_INODE_MODMASK)) {
1681 		trans->flags |= HAMMER_TRANSF_NEWINODE;
1682 	}
1683 	if (flags & HAMMER_INODE_MODMASK)
1684 		hammer_inode_dirty(ip);
1685 	ip->flags |= flags;
1686 }
1687 
1688 /*
1689  * Attempt to quickly update the atime for a hammer inode.  Return 0 on
1690  * success, -1 on failure.
1691  *
1692  * We attempt to update the atime with only the ip lock and not the
1693  * whole filesystem lock in order to improve concurrency.  We can only
1694  * do this safely if the ATIME flag is already pending on the inode.
1695  *
1696  * This function is called via a vnops path (ip pointer is stable) without
1697  * fs_token held.
1698  */
1699 int
1700 hammer_update_atime_quick(hammer_inode_t ip)
1701 {
1702 	struct timeval tv;
1703 	int res = -1;
1704 
1705 	if ((ip->flags & HAMMER_INODE_RO) ||
1706 	    (ip->hmp->mp->mnt_flag & MNT_NOATIME)) {
1707 		/*
1708 		 * Silently indicate success on read-only mount/snap
1709 		 */
1710 		res = 0;
1711 	} else if (ip->flags & HAMMER_INODE_ATIME) {
1712 		/*
1713 		 * Double check with inode lock held against backend.  This
1714 		 * is only safe if all we need to do is update
1715 		 * ino_data.atime.
1716 		 */
1717 		getmicrotime(&tv);
1718 		hammer_lock_ex(&ip->lock);
1719 		if (ip->flags & HAMMER_INODE_ATIME) {
1720 			ip->ino_data.atime =
1721 			    (unsigned long)tv.tv_sec * 1000000ULL + tv.tv_usec;
1722 			res = 0;
1723 		}
1724 		hammer_unlock(&ip->lock);
1725 	}
1726 	return res;
1727 }
1728 
1729 /*
1730  * Request that an inode be flushed.  This whole mess cannot block and may
1731  * recurse (if not synchronous).  Once requested HAMMER will attempt to
1732  * actively flush the inode until the flush can be done.
1733  *
1734  * The inode may already be flushing, or may be in a setup state.  We can
1735  * place the inode in a flushing state if it is currently idle and flag it
1736  * to reflush if it is currently flushing.
1737  *
1738  * Upon return if the inode could not be flushed due to a setup
1739  * dependancy, then it will be automatically flushed when the dependancy
1740  * is satisfied.
1741  */
1742 void
1743 hammer_flush_inode(hammer_inode_t ip, int flags)
1744 {
1745 	hammer_mount_t hmp;
1746 	hammer_flush_group_t flg;
1747 	int good;
1748 
1749 	/*
1750 	 * fill_flush_group is the first flush group we may be able to
1751 	 * continue filling, it may be open or closed but it will always
1752 	 * be past the currently flushing (running) flg.
1753 	 *
1754 	 * next_flush_group is the next open flush group.
1755 	 */
1756 	hmp = ip->hmp;
1757 	while ((flg = hmp->fill_flush_group) != NULL) {
1758 		KKASSERT(flg->running == 0);
1759 		if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit &&
1760 		    flg->total_count <= hammer_autoflush) {
1761 			break;
1762 		}
1763 		hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
1764 		hammer_flusher_async(ip->hmp, flg);
1765 	}
1766 	if (flg == NULL) {
1767 		flg = kmalloc(sizeof(*flg), hmp->m_misc, M_WAITOK|M_ZERO);
1768 		flg->seq = hmp->flusher.next++;
1769 		if (hmp->next_flush_group == NULL)
1770 			hmp->next_flush_group = flg;
1771 		if (hmp->fill_flush_group == NULL)
1772 			hmp->fill_flush_group = flg;
1773 		RB_INIT(&flg->flush_tree);
1774 		TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1775 	}
1776 
1777 	/*
1778 	 * Trivial 'nothing to flush' case.  If the inode is in a SETUP
1779 	 * state we have to put it back into an IDLE state so we can
1780 	 * drop the extra ref.
1781 	 *
1782 	 * If we have a parent dependancy we must still fall through
1783 	 * so we can run it.
1784 	 */
1785 	if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1786 		if (ip->flush_state == HAMMER_FST_SETUP &&
1787 		    TAILQ_EMPTY(&ip->target_list)) {
1788 			ip->flush_state = HAMMER_FST_IDLE;
1789 			hammer_rel_inode(ip, 0);
1790 		}
1791 		if (ip->flush_state == HAMMER_FST_IDLE)
1792 			return;
1793 	}
1794 
1795 	/*
1796 	 * Our flush action will depend on the current state.
1797 	 */
1798 	switch(ip->flush_state) {
1799 	case HAMMER_FST_IDLE:
1800 		/*
1801 		 * We have no dependancies and can flush immediately.  Some
1802 		 * our children may not be flushable so we have to re-test
1803 		 * with that additional knowledge.
1804 		 */
1805 		hammer_flush_inode_core(ip, flg, flags);
1806 		break;
1807 	case HAMMER_FST_SETUP:
1808 		/*
1809 		 * Recurse upwards through dependancies via target_list
1810 		 * and start their flusher actions going if possible.
1811 		 *
1812 		 * 'good' is our connectivity.  -1 means we have none and
1813 		 * can't flush, 0 means there weren't any dependancies, and
1814 		 * 1 means we have good connectivity.
1815 		 */
1816 		good = hammer_setup_parent_inodes(ip, 0, flg);
1817 
1818 		if (good >= 0) {
1819 			/*
1820 			 * We can continue if good >= 0.  Determine how
1821 			 * many records under our inode can be flushed (and
1822 			 * mark them).
1823 			 */
1824 			hammer_flush_inode_core(ip, flg, flags);
1825 		} else {
1826 			/*
1827 			 * Parent has no connectivity, tell it to flush
1828 			 * us as soon as it does.
1829 			 *
1830 			 * The REFLUSH flag is also needed to trigger
1831 			 * dependancy wakeups.
1832 			 */
1833 			ip->flags |= HAMMER_INODE_CONN_DOWN |
1834 				     HAMMER_INODE_REFLUSH;
1835 			if (flags & HAMMER_FLUSH_SIGNAL) {
1836 				ip->flags |= HAMMER_INODE_RESIGNAL;
1837 				hammer_flusher_async(ip->hmp, flg);
1838 			}
1839 		}
1840 		break;
1841 	case HAMMER_FST_FLUSH:
1842 		/*
1843 		 * We are already flushing, flag the inode to reflush
1844 		 * if needed after it completes its current flush.
1845 		 *
1846 		 * The REFLUSH flag is also needed to trigger
1847 		 * dependancy wakeups.
1848 		 */
1849 		if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1850 			ip->flags |= HAMMER_INODE_REFLUSH;
1851 		if (flags & HAMMER_FLUSH_SIGNAL) {
1852 			ip->flags |= HAMMER_INODE_RESIGNAL;
1853 			hammer_flusher_async(ip->hmp, flg);
1854 		}
1855 		break;
1856 	}
1857 }
1858 
1859 /*
1860  * Scan ip->target_list, which is a list of records owned by PARENTS to our
1861  * ip which reference our ip.
1862  *
1863  * XXX This is a huge mess of recursive code, but not one bit of it blocks
1864  *     so for now do not ref/deref the structures.  Note that if we use the
1865  *     ref/rel code later, the rel CAN block.
1866  */
1867 static int
1868 hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
1869 			   hammer_flush_group_t flg)
1870 {
1871 	hammer_record_t depend;
1872 	int good;
1873 	int r;
1874 
1875 	/*
1876 	 * If we hit our recursion limit and we have parent dependencies
1877 	 * We cannot continue.  Returning < 0 will cause us to be flagged
1878 	 * for reflush.  Returning -2 cuts off additional dependency checks
1879 	 * because they are likely to also hit the depth limit.
1880 	 *
1881 	 * We cannot return < 0 if there are no dependencies or there might
1882 	 * not be anything to wakeup (ip).
1883 	 */
1884 	if (depth == 20 && TAILQ_FIRST(&ip->target_list)) {
1885 		if (hammer_debug_general & 0x10000)
1886 			hkrateprintf(&hammer_gen_krate,
1887 			    "Warning: depth limit reached on "
1888 			    "setup recursion, inode %p %016jx\n",
1889 			    ip, (intmax_t)ip->obj_id);
1890 		return(-2);
1891 	}
1892 
1893 	/*
1894 	 * Scan dependencies
1895 	 */
1896 	good = 0;
1897 	TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1898 		r = hammer_setup_parent_inodes_helper(depend, depth, flg);
1899 		KKASSERT(depend->target_ip == ip);
1900 		if (r < 0 && good == 0)
1901 			good = -1;
1902 		if (r > 0)
1903 			good = 1;
1904 
1905 		/*
1906 		 * If we failed due to the recursion depth limit then stop
1907 		 * now.
1908 		 */
1909 		if (r == -2)
1910 			break;
1911 	}
1912 	return(good);
1913 }
1914 
1915 /*
1916  * This helper function takes a record representing the dependancy between
1917  * the parent inode and child inode.
1918  *
1919  * record		= record in question (*rec in below)
1920  * record->ip		= parent inode (*pip in below)
1921  * record->target_ip	= child inode (*ip in below)
1922  *
1923  * *pip--------------\
1924  *    ^               \rec_tree
1925  *     \               \
1926  *      \ip            /\\\\\ rbtree of recs from parent inode's view
1927  *       \            //\\\\\\
1928  *        \          / ........
1929  *         \        /
1930  *          \------*rec------target_ip------>*ip
1931  *               ...target_entry<----...----->target_list<---...
1932  *                                            list of recs from inode's view
1933  *
1934  * We are asked to recurse upwards and convert the record from SETUP
1935  * to FLUSH if possible.
1936  *
1937  * Return 1 if the record gives us connectivity
1938  *
1939  * Return 0 if the record is not relevant
1940  *
1941  * Return -1 if we can't resolve the dependancy and there is no connectivity.
1942  */
1943 static int
1944 hammer_setup_parent_inodes_helper(hammer_record_t record, int depth,
1945 				  hammer_flush_group_t flg)
1946 {
1947 	hammer_inode_t pip;
1948 	int good;
1949 
1950 	KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1951 	pip = record->ip;
1952 
1953 	/*
1954 	 * If the record is already flushing, is it in our flush group?
1955 	 *
1956 	 * If it is in our flush group but it is a general record or a
1957 	 * delete-on-disk, it does not improve our connectivity (return 0),
1958 	 * and if the target inode is not trying to destroy itself we can't
1959 	 * allow the operation yet anyway (the second return -1).
1960 	 */
1961 	if (record->flush_state == HAMMER_FST_FLUSH) {
1962 		/*
1963 		 * If not in our flush group ask the parent to reflush
1964 		 * us as soon as possible.
1965 		 */
1966 		if (record->flush_group != flg) {
1967 			pip->flags |= HAMMER_INODE_REFLUSH;
1968 			record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1969 			return(-1);
1970 		}
1971 
1972 		/*
1973 		 * If in our flush group everything is already set up,
1974 		 * just return whether the record will improve our
1975 		 * visibility or not.
1976 		 */
1977 		if (record->type == HAMMER_MEM_RECORD_ADD)
1978 			return(1);
1979 		return(0);
1980 	}
1981 
1982 	/*
1983 	 * It must be a setup record.  Try to resolve the setup dependancies
1984 	 * by recursing upwards so we can place ip on the flush list.
1985 	 *
1986 	 * Limit ourselves to 20 levels of recursion to avoid blowing out
1987 	 * the kernel stack.  If we hit the recursion limit we can't flush
1988 	 * until the parent flushes.  The parent will flush independantly
1989 	 * on its own and ultimately a deep recursion will be resolved.
1990 	 */
1991 	KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1992 
1993 	good = hammer_setup_parent_inodes(pip, depth + 1, flg);
1994 
1995 	/*
1996 	 * If good < 0 the parent has no connectivity and we cannot safely
1997 	 * flush the directory entry, which also means we can't flush our
1998 	 * ip.  Flag us for downward recursion once the parent's
1999 	 * connectivity is resolved.  Flag the parent for [re]flush or it
2000 	 * may not check for downward recursions.
2001 	 */
2002 	if (good < 0) {
2003 		pip->flags |= HAMMER_INODE_REFLUSH;
2004 		record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
2005 		return(good);
2006 	}
2007 
2008 	/*
2009 	 * We are go, place the parent inode in a flushing state so we can
2010 	 * place its record in a flushing state.  Note that the parent
2011 	 * may already be flushing.  The record must be in the same flush
2012 	 * group as the parent.
2013 	 */
2014 	if (pip->flush_state != HAMMER_FST_FLUSH)
2015 		hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
2016 	KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
2017 
2018 	/*
2019 	 * It is possible for a rename to create a loop in the recursion
2020 	 * and revisit a record.  This will result in the record being
2021 	 * placed in a flush state unexpectedly.  This check deals with
2022 	 * the case.
2023 	 */
2024 	if (record->flush_state == HAMMER_FST_FLUSH) {
2025 		if (record->type == HAMMER_MEM_RECORD_ADD)
2026 			return(1);
2027 		return(0);
2028 	}
2029 
2030 	KKASSERT(record->flush_state == HAMMER_FST_SETUP);
2031 
2032 #if 0
2033 	if (record->type == HAMMER_MEM_RECORD_DEL &&
2034 	    (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
2035 		/*
2036 		 * Regardless of flushing state we cannot sync this path if the
2037 		 * record represents a delete-on-disk but the target inode
2038 		 * is not ready to sync its own deletion.
2039 		 *
2040 		 * XXX need to count effective nlinks to determine whether
2041 		 * the flush is ok, otherwise removing a hardlink will
2042 		 * just leave the DEL record to rot.
2043 		 */
2044 		record->target_ip->flags |= HAMMER_INODE_REFLUSH;
2045 		return(-1);
2046 	} else
2047 #endif
2048 	if (pip->flush_group == flg) {
2049 		/*
2050 		 * Because we have not calculated nlinks yet we can just
2051 		 * set records to the flush state if the parent is in
2052 		 * the same flush group as we are.
2053 		 */
2054 		record->flush_state = HAMMER_FST_FLUSH;
2055 		record->flush_group = flg;
2056 		++record->flush_group->refs;
2057 		hammer_ref(&record->lock);
2058 
2059 		/*
2060 		 * A general directory-add contributes to our visibility.
2061 		 *
2062 		 * Otherwise it is probably a directory-delete or
2063 		 * delete-on-disk record and does not contribute to our
2064 		 * visbility (but we can still flush it).
2065 		 */
2066 		if (record->type == HAMMER_MEM_RECORD_ADD)
2067 			return(1);
2068 		return(0);
2069 	} else {
2070 		/*
2071 		 * If the parent is not in our flush group we cannot
2072 		 * flush this record yet, there is no visibility.
2073 		 * We tell the parent to reflush and mark ourselves
2074 		 * so the parent knows it should flush us too.
2075 		 */
2076 		pip->flags |= HAMMER_INODE_REFLUSH;
2077 		record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
2078 		return(-1);
2079 	}
2080 }
2081 
2082 /*
2083  * This is the core routine placing an inode into the FST_FLUSH state.
2084  */
2085 static void
2086 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
2087 {
2088 	hammer_mount_t hmp = ip->hmp;
2089 	int go_count;
2090 
2091 	/*
2092 	 * Set flush state and prevent the flusher from cycling into
2093 	 * the next flush group.  Do not place the ip on the list yet.
2094 	 * Inodes not in the idle state get an extra reference.
2095 	 */
2096 	KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
2097 	if (ip->flush_state == HAMMER_FST_IDLE)
2098 		hammer_ref(&ip->lock);
2099 	ip->flush_state = HAMMER_FST_FLUSH;
2100 	ip->flush_group = flg;
2101 	++hmp->flusher.group_lock;
2102 	++hmp->count_iqueued;
2103 	++hammer_count_iqueued;
2104 	++flg->total_count;
2105 	hammer_redo_fifo_start_flush(ip);
2106 
2107 #if 0
2108 	/*
2109 	 * We need to be able to vfsync/truncate from the backend.
2110 	 *
2111 	 * XXX Any truncation from the backend will acquire the vnode
2112 	 *     independently.
2113 	 */
2114 	KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
2115 	if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
2116 		ip->flags |= HAMMER_INODE_VHELD;
2117 		vref(ip->vp);
2118 	}
2119 #endif
2120 
2121 	/*
2122 	 * Figure out how many in-memory records we can actually flush
2123 	 * (not including inode meta-data, buffers, etc).
2124 	 */
2125 	KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
2126 	if (flags & HAMMER_FLUSH_RECURSION) {
2127 		/*
2128 		 * If this is a upwards recursion we do not want to
2129 		 * recurse down again!
2130 		 */
2131 		go_count = 1;
2132 #if 0
2133 	} else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2134 		/*
2135 		 * No new records are added if we must complete a flush
2136 		 * from a previous cycle, but we do have to move the records
2137 		 * from the previous cycle to the current one.
2138 		 */
2139 #if 0
2140 		go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2141 				   hammer_syncgrp_child_callback, NULL);
2142 #endif
2143 		go_count = 1;
2144 #endif
2145 	} else {
2146 		/*
2147 		 * Normal flush, scan records and bring them into the flush.
2148 		 * Directory adds and deletes are usually skipped (they are
2149 		 * grouped with the related inode rather then with the
2150 		 * directory).
2151 		 *
2152 		 * go_count can be negative, which means the scan aborted
2153 		 * due to the flush group being over-full and we should
2154 		 * flush what we have.
2155 		 */
2156 		go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2157 				   hammer_setup_child_callback, NULL);
2158 	}
2159 
2160 	/*
2161 	 * This is a more involved test that includes go_count.  If we
2162 	 * can't flush, flag the inode and return.  If go_count is 0 we
2163 	 * were are unable to flush any records in our rec_tree and
2164 	 * must ignore the XDIRTY flag.
2165 	 */
2166 	if (go_count == 0) {
2167 		if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
2168 			--hmp->count_iqueued;
2169 			--hammer_count_iqueued;
2170 
2171 			--flg->total_count;
2172 			ip->flush_state = HAMMER_FST_SETUP;
2173 			ip->flush_group = NULL;
2174 			if (flags & HAMMER_FLUSH_SIGNAL) {
2175 				ip->flags |= HAMMER_INODE_REFLUSH |
2176 					     HAMMER_INODE_RESIGNAL;
2177 			} else {
2178 				ip->flags |= HAMMER_INODE_REFLUSH;
2179 			}
2180 #if 0
2181 			if (ip->flags & HAMMER_INODE_VHELD) {
2182 				ip->flags &= ~HAMMER_INODE_VHELD;
2183 				vrele(ip->vp);
2184 			}
2185 #endif
2186 
2187 			/*
2188 			 * REFLUSH is needed to trigger dependancy wakeups
2189 			 * when an inode is in SETUP.
2190 			 */
2191 			ip->flags |= HAMMER_INODE_REFLUSH;
2192 			if (--hmp->flusher.group_lock == 0)
2193 				wakeup(&hmp->flusher.group_lock);
2194 			return;
2195 		}
2196 	}
2197 
2198 	/*
2199 	 * Snapshot the state of the inode for the backend flusher.
2200 	 *
2201 	 * We continue to retain save_trunc_off even when all truncations
2202 	 * have been resolved as an optimization to determine if we can
2203 	 * skip the B-Tree lookup for overwrite deletions.
2204 	 *
2205 	 * NOTE: The DELETING flag is a mod flag, but it is also sticky,
2206 	 * and stays in ip->flags.  Once set, it stays set until the
2207 	 * inode is destroyed.
2208 	 */
2209 	if (ip->flags & HAMMER_INODE_TRUNCATED) {
2210 		KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
2211 		ip->sync_trunc_off = ip->trunc_off;
2212 		ip->trunc_off = HAMMER_MAX_KEY;
2213 		ip->flags &= ~HAMMER_INODE_TRUNCATED;
2214 		ip->sync_flags |= HAMMER_INODE_TRUNCATED;
2215 
2216 		/*
2217 		 * The save_trunc_off used to cache whether the B-Tree
2218 		 * holds any records past that point is not used until
2219 		 * after the truncation has succeeded, so we can safely
2220 		 * set it now.
2221 		 */
2222 		if (ip->save_trunc_off > ip->sync_trunc_off)
2223 			ip->save_trunc_off = ip->sync_trunc_off;
2224 	}
2225 	ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
2226 			   ~HAMMER_INODE_TRUNCATED);
2227 	ip->sync_ino_leaf = ip->ino_leaf;
2228 	ip->sync_ino_data = ip->ino_data;
2229 	ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
2230 
2231 	/*
2232 	 * The flusher list inherits our inode and reference.
2233 	 */
2234 	KKASSERT(flg->running == 0);
2235 	RB_INSERT(hammer_fls_rb_tree, &flg->flush_tree, ip);
2236 	if (--hmp->flusher.group_lock == 0)
2237 		wakeup(&hmp->flusher.group_lock);
2238 
2239 	/*
2240 	 * Auto-flush the group if it grows too large.  Make sure the
2241 	 * inode reclaim wait pipeline continues to work.
2242 	 */
2243 	if (flg->total_count >= hammer_autoflush ||
2244 	    flg->total_count >= hammer_limit_reclaims / 4) {
2245 		if (hmp->fill_flush_group == flg)
2246 			hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
2247 		hammer_flusher_async(hmp, flg);
2248 	}
2249 }
2250 
2251 /*
2252  * Callback for scan of ip->rec_tree.  Try to include each record in our
2253  * flush.  ip->flush_group has been set but the inode has not yet been
2254  * moved into a flushing state.
2255  *
2256  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
2257  * both inodes.
2258  *
2259  * We return 1 for any record placed or found in FST_FLUSH, which prevents
2260  * the caller from shortcutting the flush.
2261  */
2262 static int
2263 hammer_setup_child_callback(hammer_record_t rec, void *data)
2264 {
2265 	hammer_flush_group_t flg;
2266 	hammer_inode_t target_ip;
2267 	hammer_inode_t ip;
2268 	int r;
2269 
2270 	/*
2271 	 * Records deleted or committed by the backend are ignored.
2272 	 * Note that the flush detects deleted frontend records at
2273 	 * multiple points to deal with races.  This is just the first
2274 	 * line of defense.  The only time HAMMER_RECF_DELETED_FE cannot
2275 	 * be set is when HAMMER_RECF_INTERLOCK_BE is set, because it
2276 	 * messes up link-count calculations.
2277 	 *
2278 	 * NOTE: Don't get confused between record deletion and, say,
2279 	 * directory entry deletion.  The deletion of a directory entry
2280 	 * which is on-media has nothing to do with the record deletion
2281 	 * flags.
2282 	 */
2283 	if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
2284 			  HAMMER_RECF_COMMITTED)) {
2285 		if (rec->flush_state == HAMMER_FST_FLUSH) {
2286 			KKASSERT(rec->flush_group == rec->ip->flush_group);
2287 			r = 1;
2288 		} else {
2289 			r = 0;
2290 		}
2291 		return(r);
2292 	}
2293 
2294 	/*
2295 	 * If the record is in an idle state it has no dependancies and
2296 	 * can be flushed.
2297 	 */
2298 	ip = rec->ip;
2299 	flg = ip->flush_group;
2300 	r = 0;
2301 
2302 	switch(rec->flush_state) {
2303 	case HAMMER_FST_IDLE:
2304 		/*
2305 		 * The record has no setup dependancy, we can flush it.
2306 		 */
2307 		KKASSERT(rec->target_ip == NULL);
2308 		rec->flush_state = HAMMER_FST_FLUSH;
2309 		rec->flush_group = flg;
2310 		++flg->refs;
2311 		hammer_ref(&rec->lock);
2312 		r = 1;
2313 		break;
2314 	case HAMMER_FST_SETUP:
2315 		/*
2316 		 * The record has a setup dependancy.  These are typically
2317 		 * directory entry adds and deletes.  Such entries will be
2318 		 * flushed when their inodes are flushed so we do not
2319 		 * usually have to add them to the flush here.  However,
2320 		 * if the target_ip has set HAMMER_INODE_CONN_DOWN then
2321 		 * it is asking us to flush this record (and it).
2322 		 */
2323 		target_ip = rec->target_ip;
2324 		KKASSERT(target_ip != NULL);
2325 		KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
2326 
2327 		/*
2328 		 * If the target IP is already flushing in our group
2329 		 * we could associate the record, but target_ip has
2330 		 * already synced ino_data to sync_ino_data and we
2331 		 * would also have to adjust nlinks.   Plus there are
2332 		 * ordering issues for adds and deletes.
2333 		 *
2334 		 * Reflush downward if this is an ADD, and upward if
2335 		 * this is a DEL.
2336 		 */
2337 		if (target_ip->flush_state == HAMMER_FST_FLUSH) {
2338 			if (rec->type == HAMMER_MEM_RECORD_ADD)
2339 				ip->flags |= HAMMER_INODE_REFLUSH;
2340 			else
2341 				target_ip->flags |= HAMMER_INODE_REFLUSH;
2342 			break;
2343 		}
2344 
2345 		/*
2346 		 * Target IP is not yet flushing.  This can get complex
2347 		 * because we have to be careful about the recursion.
2348 		 *
2349 		 * Directories create an issue for us in that if a flush
2350 		 * of a directory is requested the expectation is to flush
2351 		 * any pending directory entries, but this will cause the
2352 		 * related inodes to recursively flush as well.  We can't
2353 		 * really defer the operation so just get as many as we
2354 		 * can and
2355 		 */
2356 #if 0
2357 		if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
2358 		    (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
2359 			/*
2360 			 * We aren't reclaiming and the target ip was not
2361 			 * previously prevented from flushing due to this
2362 			 * record dependancy.  Do not flush this record.
2363 			 */
2364 			/*r = 0;*/
2365 		} else
2366 #endif
2367 		if (flg->total_count + flg->refs >
2368 			   ip->hmp->undo_rec_limit) {
2369 			/*
2370 			 * Our flush group is over-full and we risk blowing
2371 			 * out the UNDO FIFO.  Stop the scan, flush what we
2372 			 * have, then reflush the directory.
2373 			 *
2374 			 * The directory may be forced through multiple
2375 			 * flush groups before it can be completely
2376 			 * flushed.
2377 			 */
2378 			ip->flags |= HAMMER_INODE_RESIGNAL |
2379 				     HAMMER_INODE_REFLUSH;
2380 			r = -1;
2381 		} else if (rec->type == HAMMER_MEM_RECORD_ADD) {
2382 			/*
2383 			 * If the target IP is not flushing we can force
2384 			 * it to flush, even if it is unable to write out
2385 			 * any of its own records we have at least one in
2386 			 * hand that we CAN deal with.
2387 			 */
2388 			rec->flush_state = HAMMER_FST_FLUSH;
2389 			rec->flush_group = flg;
2390 			++flg->refs;
2391 			hammer_ref(&rec->lock);
2392 			hammer_flush_inode_core(target_ip, flg,
2393 						HAMMER_FLUSH_RECURSION);
2394 			r = 1;
2395 		} else {
2396 			/*
2397 			 * General or delete-on-disk record.
2398 			 *
2399 			 * XXX this needs help.  If a delete-on-disk we could
2400 			 * disconnect the target.  If the target has its own
2401 			 * dependancies they really need to be flushed.
2402 			 *
2403 			 * XXX
2404 			 */
2405 			rec->flush_state = HAMMER_FST_FLUSH;
2406 			rec->flush_group = flg;
2407 			++flg->refs;
2408 			hammer_ref(&rec->lock);
2409 			hammer_flush_inode_core(target_ip, flg,
2410 						HAMMER_FLUSH_RECURSION);
2411 			r = 1;
2412 		}
2413 		break;
2414 	case HAMMER_FST_FLUSH:
2415 		/*
2416 		 * The record could be part of a previous flush group if the
2417 		 * inode is a directory (the record being a directory entry).
2418 		 * Once the flush group was closed a hammer_test_inode()
2419 		 * function can cause a new flush group to be setup, placing
2420 		 * the directory inode itself in a new flush group.
2421 		 *
2422 		 * When associated with a previous flush group we count it
2423 		 * as if it were in our current flush group, since it will
2424 		 * effectively be flushed by the time we flush our current
2425 		 * flush group.
2426 		 */
2427 		KKASSERT(
2428 		    rec->ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY ||
2429 		    rec->flush_group == flg);
2430 		r = 1;
2431 		break;
2432 	}
2433 	return(r);
2434 }
2435 
2436 #if 0
2437 /*
2438  * This version just moves records already in a flush state to the new
2439  * flush group and that is it.
2440  */
2441 static int
2442 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
2443 {
2444 	hammer_inode_t ip = rec->ip;
2445 
2446 	switch(rec->flush_state) {
2447 	case HAMMER_FST_FLUSH:
2448 		KKASSERT(rec->flush_group == ip->flush_group);
2449 		break;
2450 	default:
2451 		break;
2452 	}
2453 	return(0);
2454 }
2455 #endif
2456 
2457 /*
2458  * Wait for a previously queued flush to complete.
2459  *
2460  * If a critical error occured we don't try to wait.
2461  */
2462 void
2463 hammer_wait_inode(hammer_inode_t ip)
2464 {
2465 	/*
2466 	 * The inode can be in a SETUP state in which case RESIGNAL
2467 	 * should be set.  If RESIGNAL is not set then the previous
2468 	 * flush completed and a later operation placed the inode
2469 	 * in a passive setup state again, so we're done.
2470 	 *
2471 	 * The inode can be in a FLUSH state in which case we
2472 	 * can just wait for completion.
2473 	 */
2474 	while (ip->flush_state == HAMMER_FST_FLUSH ||
2475 	    (ip->flush_state == HAMMER_FST_SETUP &&
2476 	     (ip->flags & HAMMER_INODE_RESIGNAL))) {
2477 		/*
2478 		 * Don't try to flush on a critical error
2479 		 */
2480 		if (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
2481 			break;
2482 
2483 		/*
2484 		 * If the inode was already being flushed its flg
2485 		 * may not have been queued to the backend.  We
2486 		 * have to make sure it gets queued or we can wind
2487 		 * up blocked or deadlocked (particularly if we are
2488 		 * the vnlru thread).
2489 		 */
2490 		if (ip->flush_state == HAMMER_FST_FLUSH) {
2491 			KKASSERT(ip->flush_group);
2492 			if (ip->flush_group->closed == 0) {
2493 				if (hammer_debug_inode) {
2494 					hkprintf("debug: forcing "
2495 						"async flush ip %016jx\n",
2496 						(intmax_t)ip->obj_id);
2497 				}
2498 				hammer_flusher_async(ip->hmp, ip->flush_group);
2499 				continue; /* retest */
2500 			}
2501 		}
2502 
2503 		/*
2504 		 * In a flush state with the flg queued to the backend
2505 		 * or in a setup state with RESIGNAL set, we can safely
2506 		 * wait.
2507 		 */
2508 		ip->flags |= HAMMER_INODE_FLUSHW;
2509 		tsleep(&ip->flags, 0, "hmrwin", 0);
2510 	}
2511 
2512 #if 0
2513 	/*
2514 	 * The inode may have been in a passive setup state,
2515 	 * call flush to make sure we get signaled.
2516 	 */
2517 	if (ip->flush_state == HAMMER_FST_SETUP)
2518 		hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2519 #endif
2520 
2521 }
2522 
2523 /*
2524  * Called by the backend code when a flush has been completed.
2525  * The inode has already been removed from the flush list.
2526  *
2527  * A pipelined flush can occur, in which case we must re-enter the
2528  * inode on the list and re-copy its fields.
2529  */
2530 void
2531 hammer_sync_inode_done(hammer_inode_t ip, int error)
2532 {
2533 	hammer_mount_t hmp;
2534 	int dorel;
2535 
2536 	KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2537 
2538 	hmp = ip->hmp;
2539 
2540 	/*
2541 	 * Auto-reflush if the backend could not completely flush
2542 	 * the inode.  This fixes a case where a deferred buffer flush
2543 	 * could cause fsync to return early.
2544 	 */
2545 	if (ip->sync_flags & HAMMER_INODE_MODMASK)
2546 		ip->flags |= HAMMER_INODE_REFLUSH;
2547 
2548 	/*
2549 	 * Merge left-over flags back into the frontend and fix the state.
2550 	 * Incomplete truncations are retained by the backend.
2551 	 */
2552 	ip->error = error;
2553 	ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2554 	ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2555 
2556 	/*
2557 	 * The backend may have adjusted nlinks, so if the adjusted nlinks
2558 	 * does not match the fronttend set the frontend's DDIRTY flag again.
2559 	 */
2560 	if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2561 		ip->flags |= HAMMER_INODE_DDIRTY;
2562 
2563 	/*
2564 	 * Fix up the dirty buffer status.
2565 	 */
2566 	if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2567 		ip->flags |= HAMMER_INODE_BUFS;
2568 	}
2569 	hammer_redo_fifo_end_flush(ip);
2570 
2571 	/*
2572 	 * Re-set the XDIRTY flag if some of the inode's in-memory records
2573 	 * could not be flushed.
2574 	 */
2575 	KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2576 		  (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2577 		 (!RB_EMPTY(&ip->rec_tree) &&
2578 		  (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2579 
2580 	/*
2581 	 * Do not lose track of inodes which no longer have vnode
2582 	 * assocations, otherwise they may never get flushed again.
2583 	 *
2584 	 * The reflush flag can be set superfluously, causing extra pain
2585 	 * for no reason.  If the inode is no longer modified it no longer
2586 	 * needs to be flushed.
2587 	 */
2588 	if (ip->flags & HAMMER_INODE_MODMASK) {
2589 		if (ip->vp == NULL)
2590 			ip->flags |= HAMMER_INODE_REFLUSH;
2591 	} else {
2592 		ip->flags &= ~HAMMER_INODE_REFLUSH;
2593 	}
2594 
2595 	/*
2596 	 * The fs token is held but the inode lock is not held.  Because this
2597 	 * is a backend flush it is possible that the vnode has no references
2598 	 * and cause a reclaim race inside vsetisdirty() if/when it blocks.
2599 	 *
2600 	 * Therefore, we must lock the inode around this particular dirtying
2601 	 * operation.  We don't have to around other dirtying operations
2602 	 * where the vnode is implicitly or explicitly held.
2603 	 */
2604 	if (ip->flags & HAMMER_INODE_MODMASK) {
2605 		hammer_lock_ex(&ip->lock);
2606 		hammer_inode_dirty(ip);
2607 		hammer_unlock(&ip->lock);
2608 	}
2609 
2610 	/*
2611 	 * Adjust the flush state.
2612 	 */
2613 	if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2614 		/*
2615 		 * We were unable to flush out all our records, leave the
2616 		 * inode in a flush state and in the current flush group.
2617 		 * The flush group will be re-run.
2618 		 *
2619 		 * This occurs if the UNDO block gets too full or there is
2620 		 * too much dirty meta-data and allows the flusher to
2621 		 * finalize the UNDO block and then re-flush.
2622 		 */
2623 		ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2624 		dorel = 0;
2625 	} else {
2626 		/*
2627 		 * Remove from the flush_group
2628 		 */
2629 		RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
2630 		ip->flush_group = NULL;
2631 
2632 #if 0
2633 		/*
2634 		 * Clean up the vnode ref and tracking counts.
2635 		 */
2636 		if (ip->flags & HAMMER_INODE_VHELD) {
2637 			ip->flags &= ~HAMMER_INODE_VHELD;
2638 			vrele(ip->vp);
2639 		}
2640 #endif
2641 		--hmp->count_iqueued;
2642 		--hammer_count_iqueued;
2643 
2644 		/*
2645 		 * And adjust the state.
2646 		 */
2647 		if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2648 			ip->flush_state = HAMMER_FST_IDLE;
2649 			dorel = 1;
2650 		} else {
2651 			ip->flush_state = HAMMER_FST_SETUP;
2652 			dorel = 0;
2653 		}
2654 
2655 		/*
2656 		 * If the frontend is waiting for a flush to complete,
2657 		 * wake it up.
2658 		 */
2659 		if (ip->flags & HAMMER_INODE_FLUSHW) {
2660 			ip->flags &= ~HAMMER_INODE_FLUSHW;
2661 			wakeup(&ip->flags);
2662 		}
2663 
2664 		/*
2665 		 * If the frontend made more changes and requested another
2666 		 * flush, then try to get it running.
2667 		 *
2668 		 * Reflushes are aborted when the inode is errored out.
2669 		 */
2670 		if (ip->flags & HAMMER_INODE_REFLUSH) {
2671 			ip->flags &= ~HAMMER_INODE_REFLUSH;
2672 			if (ip->flags & HAMMER_INODE_RESIGNAL) {
2673 				ip->flags &= ~HAMMER_INODE_RESIGNAL;
2674 				hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2675 			} else {
2676 				hammer_flush_inode(ip, 0);
2677 			}
2678 		}
2679 	}
2680 
2681 	/*
2682 	 * If we have no parent dependancies we can clear CONN_DOWN
2683 	 */
2684 	if (TAILQ_EMPTY(&ip->target_list))
2685 		ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2686 
2687 	/*
2688 	 * If the inode is now clean drop the space reservation.
2689 	 */
2690 	if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2691 	    (ip->flags & HAMMER_INODE_RSV_INODES)) {
2692 		ip->flags &= ~HAMMER_INODE_RSV_INODES;
2693 		--hmp->rsv_inodes;
2694 	}
2695 
2696 	ip->flags &= ~HAMMER_INODE_SLAVEFLUSH;
2697 
2698 	if (dorel)
2699 		hammer_rel_inode(ip, 0);
2700 }
2701 
2702 /*
2703  * Called from hammer_sync_inode() to synchronize in-memory records
2704  * to the media.
2705  */
2706 static int
2707 hammer_sync_record_callback(hammer_record_t record, void *data)
2708 {
2709 	hammer_cursor_t cursor = data;
2710 	hammer_transaction_t trans = cursor->trans;
2711 	hammer_mount_t hmp = trans->hmp;
2712 	int error;
2713 
2714 	/*
2715 	 * Skip records that do not belong to the current flush.
2716 	 */
2717 	++hammer_stats_record_iterations;
2718 	if (record->flush_state != HAMMER_FST_FLUSH)
2719 		return(0);
2720 
2721 	if (record->flush_group != record->ip->flush_group) {
2722 		hdkprintf("rec %p ip %p bad flush group %p %p\n",
2723 			record,
2724 			record->ip,
2725 			record->flush_group,
2726 			record->ip->flush_group);
2727 		if (hammer_debug_critical)
2728 			Debugger("blah2");
2729 		return(0);
2730 	}
2731 	KKASSERT(record->flush_group == record->ip->flush_group);
2732 
2733 	/*
2734 	 * Interlock the record using the BE flag.  Once BE is set the
2735 	 * frontend cannot change the state of FE.
2736 	 *
2737 	 * NOTE: If FE is set prior to us setting BE we still sync the
2738 	 * record out, but the flush completion code converts it to
2739 	 * a delete-on-disk record instead of destroying it.
2740 	 */
2741 	KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2742 	record->flags |= HAMMER_RECF_INTERLOCK_BE;
2743 
2744 	/*
2745 	 * The backend has already disposed of the record.
2746 	 */
2747 	if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
2748 		error = 0;
2749 		goto done;
2750 	}
2751 
2752 	/*
2753 	 * If the whole inode is being deleted and all on-disk records will
2754 	 * be deleted very soon, we can't sync any new records to disk
2755 	 * because they will be deleted in the same transaction they were
2756 	 * created in (delete_tid == create_tid), which will assert.
2757 	 *
2758 	 * XXX There may be a case with RECORD_ADD with DELETED_FE set
2759 	 * that we currently panic on.
2760 	 */
2761 	if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2762 		switch(record->type) {
2763 		case HAMMER_MEM_RECORD_DATA:
2764 			/*
2765 			 * We don't have to do anything, if the record was
2766 			 * committed the space will have been accounted for
2767 			 * in the blockmap.
2768 			 */
2769 			/* fall through */
2770 		case HAMMER_MEM_RECORD_GENERAL:
2771 			/*
2772 			 * Set deleted-by-backend flag.  Do not set the
2773 			 * backend committed flag, because we are throwing
2774 			 * the record away.
2775 			 */
2776 			record->flags |= HAMMER_RECF_DELETED_BE;
2777 			++record->ip->rec_generation;
2778 			error = 0;
2779 			goto done;
2780 		case HAMMER_MEM_RECORD_ADD:
2781 			hpanic("illegal add during inode deletion record %p",
2782 				record);
2783 			break; /* NOT REACHED */
2784 		case HAMMER_MEM_RECORD_INODE:
2785 			hpanic("attempt to sync inode record %p?", record);
2786 			break; /* NOT REACHED */
2787 		case HAMMER_MEM_RECORD_DEL:
2788 			/*
2789 			 * Follow through and issue the on-disk deletion
2790 			 */
2791 			break;
2792 		}
2793 	}
2794 
2795 	/*
2796 	 * If DELETED_FE is set special handling is needed for directory
2797 	 * entries.  Dependant pieces related to the directory entry may
2798 	 * have already been synced to disk.  If this occurs we have to
2799 	 * sync the directory entry and then change the in-memory record
2800 	 * from an ADD to a DELETE to cover the fact that it's been
2801 	 * deleted by the frontend.
2802 	 *
2803 	 * A directory delete covering record (MEM_RECORD_DEL) can never
2804 	 * be deleted by the frontend.
2805 	 *
2806 	 * Any other record type (aka DATA) can be deleted by the frontend.
2807 	 * XXX At the moment the flusher must skip it because there may
2808 	 * be another data record in the flush group for the same block,
2809 	 * meaning that some frontend data changes can leak into the backend's
2810 	 * synchronization point.
2811 	 */
2812 	if (record->flags & HAMMER_RECF_DELETED_FE) {
2813 		if (record->type == HAMMER_MEM_RECORD_ADD) {
2814 			/*
2815 			 * Convert a front-end deleted directory-add to
2816 			 * a directory-delete entry later.
2817 			 */
2818 			record->flags |= HAMMER_RECF_CONVERT_DELETE;
2819 		} else {
2820 			/*
2821 			 * Dispose of the record (race case).  Mark as
2822 			 * deleted by backend (and not committed).
2823 			 */
2824 			KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2825 			record->flags |= HAMMER_RECF_DELETED_BE;
2826 			++record->ip->rec_generation;
2827 			error = 0;
2828 			goto done;
2829 		}
2830 	}
2831 
2832 	/*
2833 	 * Assign the create_tid for new records.  Deletions already
2834 	 * have the record's entire key properly set up.
2835 	 */
2836 	if (record->type != HAMMER_MEM_RECORD_DEL) {
2837 		record->leaf.base.create_tid = trans->tid;
2838 		record->leaf.create_ts = trans->time32;
2839 	}
2840 
2841 	/*
2842 	 * This actually moves the record to the on-media B-Tree.  We
2843 	 * must also generate REDO_TERM entries in the UNDO/REDO FIFO
2844 	 * indicating that the related REDO_WRITE(s) have been committed.
2845 	 *
2846 	 * During recovery any REDO_TERM's within the nominal recovery span
2847 	 * are ignored since the related meta-data is being undone, causing
2848 	 * any matching REDO_WRITEs to execute.  The REDO_TERMs outside
2849 	 * the nominal recovery span will match against REDO_WRITEs and
2850 	 * prevent them from being executed (because the meta-data has
2851 	 * already been synchronized).
2852 	 */
2853 	if (record->flags & HAMMER_RECF_REDO) {
2854 		KKASSERT(record->type == HAMMER_MEM_RECORD_DATA);
2855 		hammer_generate_redo(trans, record->ip,
2856 				     record->leaf.base.key -
2857 					 record->leaf.data_len,
2858 				     HAMMER_REDO_TERM_WRITE,
2859 				     NULL,
2860 				     record->leaf.data_len);
2861 	}
2862 
2863 	for (;;) {
2864 		error = hammer_ip_sync_record_cursor(cursor, record);
2865 		if (error != EDEADLK)
2866 			break;
2867 		hammer_done_cursor(cursor);
2868 		error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2869 					   record->ip);
2870 		if (error)
2871 			break;
2872 	}
2873 	record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2874 
2875 	if (error)
2876 		error = -error;
2877 done:
2878 	hammer_flush_record_done(record, error);
2879 
2880 	/*
2881 	 * Do partial finalization if we have built up too many dirty
2882 	 * buffers.  Otherwise a buffer cache deadlock can occur when
2883 	 * doing things like creating tens of thousands of tiny files.
2884 	 *
2885 	 * We must release our cursor lock to avoid a 3-way deadlock
2886 	 * due to the exclusive sync lock the finalizer must get.
2887 	 *
2888 	 * WARNING: See warnings in hammer_unlock_cursor() function.
2889 	 */
2890         if (hammer_flusher_meta_limit(hmp) ||
2891 	    vm_page_count_severe()) {
2892 		hammer_unlock_cursor(cursor);
2893                 hammer_flusher_finalize(trans, 0);
2894 		hammer_lock_cursor(cursor);
2895 	}
2896 	return(error);
2897 }
2898 
2899 /*
2900  * Backend function called by the flusher to sync an inode to media.
2901  */
2902 int
2903 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2904 {
2905 	struct hammer_cursor cursor;
2906 	hammer_node_t tmp_node;
2907 	hammer_record_t depend;
2908 	hammer_record_t next;
2909 	int error, tmp_error;
2910 	uint64_t nlinks;
2911 
2912 	if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2913 		return(0);
2914 
2915 	error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2916 	if (error)
2917 		goto done;
2918 
2919 	/*
2920 	 * Any directory records referencing this inode which are not in
2921 	 * our current flush group must adjust our nlink count for the
2922 	 * purposes of synchronizating to disk.
2923 	 *
2924 	 * Records which are in our flush group can be unlinked from our
2925 	 * inode now, potentially allowing the inode to be physically
2926 	 * deleted.
2927 	 *
2928 	 * This cannot block.
2929 	 */
2930 	nlinks = ip->ino_data.nlinks;
2931 	next = TAILQ_FIRST(&ip->target_list);
2932 	while ((depend = next) != NULL) {
2933 		next = TAILQ_NEXT(depend, target_entry);
2934 		if (depend->flush_state == HAMMER_FST_FLUSH &&
2935 		    depend->flush_group == ip->flush_group) {
2936 			/*
2937 			 * If this is an ADD that was deleted by the frontend
2938 			 * the frontend nlinks count will have already been
2939 			 * decremented, but the backend is going to sync its
2940 			 * directory entry and must account for it.  The
2941 			 * record will be converted to a delete-on-disk when
2942 			 * it gets synced.
2943 			 *
2944 			 * If the ADD was not deleted by the frontend we
2945 			 * can remove the dependancy from our target_list.
2946 			 */
2947 			if (depend->flags & HAMMER_RECF_DELETED_FE) {
2948 				++nlinks;
2949 			} else {
2950 				TAILQ_REMOVE(&ip->target_list, depend,
2951 					     target_entry);
2952 				depend->target_ip = NULL;
2953 			}
2954 		} else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2955 			/*
2956 			 * Not part of our flush group and not deleted by
2957 			 * the front-end, adjust the link count synced to
2958 			 * the media (undo what the frontend did when it
2959 			 * queued the record).
2960 			 */
2961 			KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2962 			switch(depend->type) {
2963 			case HAMMER_MEM_RECORD_ADD:
2964 				--nlinks;
2965 				break;
2966 			case HAMMER_MEM_RECORD_DEL:
2967 				++nlinks;
2968 				break;
2969 			default:
2970 				break;
2971 			}
2972 		}
2973 	}
2974 
2975 	/*
2976 	 * Set dirty if we had to modify the link count.
2977 	 */
2978 	if (ip->sync_ino_data.nlinks != nlinks) {
2979 		KKASSERT((int64_t)nlinks >= 0);
2980 		ip->sync_ino_data.nlinks = nlinks;
2981 		ip->sync_flags |= HAMMER_INODE_DDIRTY;
2982 	}
2983 
2984 	/*
2985 	 * If there is a trunction queued destroy any data past the (aligned)
2986 	 * truncation point.  Userland will have dealt with the buffer
2987 	 * containing the truncation point for us.
2988 	 *
2989 	 * We don't flush pending frontend data buffers until after we've
2990 	 * dealt with the truncation.
2991 	 */
2992 	if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2993 		/*
2994 		 * Interlock trunc_off.  The VOP front-end may continue to
2995 		 * make adjustments to it while we are blocked.
2996 		 */
2997 		off_t trunc_off;
2998 		off_t aligned_trunc_off;
2999 		int blkmask;
3000 
3001 		trunc_off = ip->sync_trunc_off;
3002 		blkmask = hammer_blocksize(trunc_off) - 1;
3003 		aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
3004 
3005 		/*
3006 		 * Delete any whole blocks on-media.  The front-end has
3007 		 * already cleaned out any partial block and made it
3008 		 * pending.  The front-end may have updated trunc_off
3009 		 * while we were blocked so we only use sync_trunc_off.
3010 		 *
3011 		 * This operation can blow out the buffer cache, EWOULDBLOCK
3012 		 * means we were unable to complete the deletion.  The
3013 		 * deletion will update sync_trunc_off in that case.
3014 		 */
3015 		error = hammer_ip_delete_range(&cursor, ip,
3016 						aligned_trunc_off,
3017 						HAMMER_MAX_KEY, 2);
3018 		if (error == EWOULDBLOCK) {
3019 			ip->flags |= HAMMER_INODE_WOULDBLOCK;
3020 			error = 0;
3021 			goto defer_buffer_flush;
3022 		}
3023 
3024 		if (error)
3025 			goto done;
3026 
3027 		/*
3028 		 * Generate a REDO_TERM_TRUNC entry in the UNDO/REDO FIFO.
3029 		 *
3030 		 * XXX we do this even if we did not previously generate
3031 		 * a REDO_TRUNC record.  This operation may enclosed the
3032 		 * range for multiple prior truncation entries in the REDO
3033 		 * log.
3034 		 */
3035 		if (trans->hmp->version >= HAMMER_VOL_VERSION_FOUR &&
3036 		    (ip->flags & HAMMER_INODE_RDIRTY)) {
3037 			hammer_generate_redo(trans, ip, aligned_trunc_off,
3038 					     HAMMER_REDO_TERM_TRUNC,
3039 					     NULL, 0);
3040 		}
3041 
3042 		/*
3043 		 * Clear the truncation flag on the backend after we have
3044 		 * completed the deletions.  Backend data is now good again
3045 		 * (including new records we are about to sync, below).
3046 		 *
3047 		 * Leave sync_trunc_off intact.  As we write additional
3048 		 * records the backend will update sync_trunc_off.  This
3049 		 * tells the backend whether it can skip the overwrite
3050 		 * test.  This should work properly even when the backend
3051 		 * writes full blocks where the truncation point straddles
3052 		 * the block because the comparison is against the base
3053 		 * offset of the record.
3054 		 */
3055 		ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
3056 		/* ip->sync_trunc_off = HAMMER_MAX_KEY; */
3057 	} else {
3058 		error = 0;
3059 	}
3060 
3061 	/*
3062 	 * Now sync related records.  These will typically be directory
3063 	 * entries, records tracking direct-writes, or delete-on-disk records.
3064 	 */
3065 	if (error == 0) {
3066 		tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
3067 				    hammer_sync_record_callback, &cursor);
3068 		if (tmp_error < 0)
3069 			tmp_error = -error;
3070 		if (tmp_error)
3071 			error = tmp_error;
3072 	}
3073 	hammer_cache_node(&ip->cache[1], cursor.node);
3074 
3075 	/*
3076 	 * Re-seek for inode update, assuming our cache hasn't been ripped
3077 	 * out from under us.
3078 	 */
3079 	if (error == 0) {
3080 		tmp_node = hammer_ref_node_safe(trans, &ip->cache[0], &error);
3081 		if (tmp_node) {
3082 			hammer_cursor_downgrade(&cursor);
3083 			hammer_lock_sh(&tmp_node->lock);
3084 			if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
3085 				hammer_cursor_seek(&cursor, tmp_node, 0);
3086 			hammer_unlock(&tmp_node->lock);
3087 			hammer_rel_node(tmp_node);
3088 		}
3089 		error = 0;
3090 	}
3091 
3092 	/*
3093 	 * If we are deleting the inode the frontend had better not have
3094 	 * any active references on elements making up the inode.
3095 	 *
3096 	 * The call to hammer_ip_delete_clean() cleans up auxillary records
3097 	 * but not DB or DATA records.  Those must have already been deleted
3098 	 * by the normal truncation mechanic.
3099 	 */
3100 	if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
3101 		RB_EMPTY(&ip->rec_tree)  &&
3102 	    (ip->sync_flags & HAMMER_INODE_DELETING) &&
3103 	    (ip->flags & HAMMER_INODE_DELETED) == 0) {
3104 		int count1 = 0;
3105 
3106 		error = hammer_ip_delete_clean(&cursor, ip, &count1);
3107 		if (error == 0) {
3108 			ip->flags |= HAMMER_INODE_DELETED;
3109 			ip->sync_flags &= ~HAMMER_INODE_DELETING;
3110 			ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
3111 			KKASSERT(RB_EMPTY(&ip->rec_tree));
3112 
3113 			/*
3114 			 * Set delete_tid in both the frontend and backend
3115 			 * copy of the inode record.  The DELETED flag handles
3116 			 * this, do not set DDIRTY.
3117 			 */
3118 			ip->ino_leaf.base.delete_tid = trans->tid;
3119 			ip->sync_ino_leaf.base.delete_tid = trans->tid;
3120 			ip->ino_leaf.delete_ts = trans->time32;
3121 			ip->sync_ino_leaf.delete_ts = trans->time32;
3122 
3123 
3124 			/*
3125 			 * Adjust the inode count in the volume header
3126 			 */
3127 			hammer_sync_lock_sh(trans);
3128 			if (ip->flags & HAMMER_INODE_ONDISK) {
3129 				hammer_modify_volume_field(trans,
3130 							   trans->rootvol,
3131 							   vol0_stat_inodes);
3132 				--ip->hmp->rootvol->ondisk->vol0_stat_inodes;
3133 				hammer_modify_volume_done(trans->rootvol);
3134 			}
3135 			hammer_sync_unlock(trans);
3136 		}
3137 	}
3138 
3139 	if (error)
3140 		goto done;
3141 	ip->sync_flags &= ~HAMMER_INODE_BUFS;
3142 
3143 defer_buffer_flush:
3144 	/*
3145 	 * Now update the inode's on-disk inode-data and/or on-disk record.
3146 	 * DELETED and ONDISK are managed only in ip->flags.
3147 	 *
3148 	 * In the case of a defered buffer flush we still update the on-disk
3149 	 * inode to satisfy visibility requirements if there happen to be
3150 	 * directory dependancies.
3151 	 */
3152 	switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
3153 	case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
3154 		/*
3155 		 * If deleted and on-disk, don't set any additional flags.
3156 		 * the delete flag takes care of things.
3157 		 *
3158 		 * Clear flags which may have been set by the frontend.
3159 		 */
3160 		ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
3161 				    HAMMER_INODE_SDIRTY |
3162 				    HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
3163 				    HAMMER_INODE_DELETING);
3164 		break;
3165 	case HAMMER_INODE_DELETED:
3166 		/*
3167 		 * Take care of the case where a deleted inode was never
3168 		 * flushed to the disk in the first place.
3169 		 *
3170 		 * Clear flags which may have been set by the frontend.
3171 		 */
3172 		ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
3173 				    HAMMER_INODE_SDIRTY |
3174 				    HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
3175 				    HAMMER_INODE_DELETING);
3176 		while (RB_ROOT(&ip->rec_tree)) {
3177 			hammer_record_t record = RB_ROOT(&ip->rec_tree);
3178 			hammer_ref(&record->lock);
3179 			KKASSERT(hammer_oneref(&record->lock));
3180 			record->flags |= HAMMER_RECF_DELETED_BE;
3181 			++record->ip->rec_generation;
3182 			hammer_rel_mem_record(record);
3183 		}
3184 		break;
3185 	case HAMMER_INODE_ONDISK:
3186 		/*
3187 		 * If already on-disk, do not set any additional flags.
3188 		 */
3189 		break;
3190 	default:
3191 		/*
3192 		 * If not on-disk and not deleted, set DDIRTY to force
3193 		 * an initial record to be written.
3194 		 *
3195 		 * Also set the create_tid in both the frontend and backend
3196 		 * copy of the inode record.
3197 		 */
3198 		ip->ino_leaf.base.create_tid = trans->tid;
3199 		ip->ino_leaf.create_ts = trans->time32;
3200 		ip->sync_ino_leaf.base.create_tid = trans->tid;
3201 		ip->sync_ino_leaf.create_ts = trans->time32;
3202 		ip->sync_flags |= HAMMER_INODE_DDIRTY;
3203 		break;
3204 	}
3205 
3206 	/*
3207 	 * If DDIRTY or SDIRTY is set, write out a new record.
3208 	 * If the inode is already on-disk the old record is marked as
3209 	 * deleted.
3210 	 *
3211 	 * If DELETED is set hammer_update_inode() will delete the existing
3212 	 * record without writing out a new one.
3213 	 */
3214 	if (ip->flags & HAMMER_INODE_DELETED) {
3215 		error = hammer_update_inode(&cursor, ip);
3216 	} else
3217 	if (!(ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY)) &&
3218 	    (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
3219 		error = hammer_update_itimes(&cursor, ip);
3220 	} else
3221 	if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY |
3222 			      HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
3223 		error = hammer_update_inode(&cursor, ip);
3224 	}
3225 done:
3226 	if (ip->flags & HAMMER_INODE_MODMASK)
3227 		hammer_inode_dirty(ip);
3228 	if (error) {
3229 		hammer_critical_error(ip->hmp, ip, error,
3230 				      "while syncing inode");
3231 	}
3232 	hammer_done_cursor(&cursor);
3233 	return(error);
3234 }
3235 
3236 /*
3237  * This routine is called when the OS is no longer actively referencing
3238  * the inode (but might still be keeping it cached), or when releasing
3239  * the last reference to an inode.
3240  *
3241  * At this point if the inode's nlinks count is zero we want to destroy
3242  * it, which may mean destroying it on-media too.
3243  */
3244 void
3245 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
3246 {
3247 	struct vnode *vp;
3248 
3249 	/*
3250 	 * Set the DELETING flag when the link count drops to 0 and the
3251 	 * OS no longer has any opens on the inode.
3252 	 *
3253 	 * The backend will clear DELETING (a mod flag) and set DELETED
3254 	 * (a state flag) when it is actually able to perform the
3255 	 * operation.
3256 	 *
3257 	 * Don't reflag the deletion if the flusher is currently syncing
3258 	 * one that was already flagged.  A previously set DELETING flag
3259 	 * may bounce around flags and sync_flags until the operation is
3260 	 * completely done.
3261 	 *
3262 	 * Do not attempt to modify a snapshot inode (one set to read-only).
3263 	 */
3264 	if (ip->ino_data.nlinks == 0 &&
3265 	    ((ip->flags | ip->sync_flags) & (HAMMER_INODE_RO|HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
3266 		ip->flags |= HAMMER_INODE_DELETING;
3267 		ip->flags |= HAMMER_INODE_TRUNCATED;
3268 		ip->trunc_off = 0;
3269 		vp = NULL;
3270 		if (getvp) {
3271 			if (hammer_get_vnode(ip, &vp) != 0)
3272 				return;
3273 		}
3274 
3275 		/*
3276 		 * Final cleanup
3277 		 */
3278 		if (ip->vp)
3279 			nvtruncbuf(ip->vp, 0, HAMMER_BUFSIZE, 0, 0);
3280 		if (ip->flags & HAMMER_INODE_MODMASK)
3281 			hammer_inode_dirty(ip);
3282 		if (getvp)
3283 			vput(vp);
3284 	}
3285 }
3286 
3287 /*
3288  * After potentially resolving a dependancy the inode is tested
3289  * to determine whether it needs to be reflushed.
3290  */
3291 void
3292 hammer_test_inode(hammer_inode_t ip)
3293 {
3294 	if (ip->flags & HAMMER_INODE_REFLUSH) {
3295 		ip->flags &= ~HAMMER_INODE_REFLUSH;
3296 		hammer_ref(&ip->lock);
3297 		if (ip->flags & HAMMER_INODE_RESIGNAL) {
3298 			ip->flags &= ~HAMMER_INODE_RESIGNAL;
3299 			hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
3300 		} else {
3301 			hammer_flush_inode(ip, 0);
3302 		}
3303 		hammer_rel_inode(ip, 0);
3304 	}
3305 }
3306 
3307 /*
3308  * Clear the RECLAIM flag on an inode.  This occurs when the inode is
3309  * reassociated with a vp or just before it gets freed.
3310  *
3311  * Pipeline wakeups to threads blocked due to an excessive number of
3312  * detached inodes.  This typically occurs when atime updates accumulate
3313  * while scanning a directory tree.
3314  */
3315 static void
3316 hammer_inode_wakereclaims(hammer_inode_t ip)
3317 {
3318 	struct hammer_reclaim *reclaim;
3319 	hammer_mount_t hmp = ip->hmp;
3320 
3321 	if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
3322 		return;
3323 
3324 	--hammer_count_reclaims;
3325 	--hmp->count_reclaims;
3326 	ip->flags &= ~HAMMER_INODE_RECLAIM;
3327 
3328 	if ((reclaim = TAILQ_FIRST(&hmp->reclaim_list)) != NULL) {
3329 		KKASSERT(reclaim->count > 0);
3330 		if (--reclaim->count == 0) {
3331 			TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
3332 			wakeup(reclaim);
3333 		}
3334 	}
3335 }
3336 
3337 /*
3338  * Setup our reclaim pipeline.  We only let so many detached (and dirty)
3339  * inodes build up before we start blocking.  This routine is called
3340  * if a new inode is created or an inode is loaded from media.
3341  *
3342  * When we block we don't care *which* inode has finished reclaiming,
3343  * as long as one does.
3344  *
3345  * The reclaim pipeline is primarily governed by the auto-flush which is
3346  * 1/4 hammer_limit_reclaims.  We don't want to block if the count is
3347  * less than 1/2 hammer_limit_reclaims.  From 1/2 to full count is
3348  * dynamically governed.
3349  */
3350 void
3351 hammer_inode_waitreclaims(hammer_transaction_t trans)
3352 {
3353 	hammer_mount_t hmp = trans->hmp;
3354 	struct hammer_reclaim reclaim;
3355 	int lower_limit;
3356 
3357 	/*
3358 	 * Track inode load, delay if the number of reclaiming inodes is
3359 	 * between 2/4 and 4/4 hammer_limit_reclaims, depending.
3360 	 */
3361 	if (curthread->td_proc) {
3362 		struct hammer_inostats *stats;
3363 
3364 		stats = hammer_inode_inostats(hmp, curthread->td_proc->p_pid);
3365 		++stats->count;
3366 
3367 		if (stats->count > hammer_limit_reclaims / 2)
3368 			stats->count = hammer_limit_reclaims / 2;
3369 		lower_limit = hammer_limit_reclaims - stats->count;
3370 		if (hammer_debug_general & 0x10000) {
3371 			hdkprintf("pid %5d limit %d\n",
3372 				(int)curthread->td_proc->p_pid, lower_limit);
3373 		}
3374 	} else {
3375 		lower_limit = hammer_limit_reclaims * 3 / 4;
3376 	}
3377 	if (hmp->count_reclaims >= lower_limit) {
3378 		reclaim.count = 1;
3379 		TAILQ_INSERT_TAIL(&hmp->reclaim_list, &reclaim, entry);
3380 		tsleep(&reclaim, 0, "hmrrcm", hz);
3381 		if (reclaim.count > 0)
3382 			TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
3383 	}
3384 }
3385 
3386 /*
3387  * Keep track of reclaim statistics on a per-pid basis using a loose
3388  * 4-way set associative hash table.  Collisions inherit the count of
3389  * the previous entry.
3390  *
3391  * NOTE: We want to be careful here to limit the chain size.  If the chain
3392  *	 size is too large a pid will spread its stats out over too many
3393  *	 entries under certain types of heavy filesystem activity and
3394  *	 wind up not delaying long enough.
3395  */
3396 static
3397 struct hammer_inostats *
3398 hammer_inode_inostats(hammer_mount_t hmp, pid_t pid)
3399 {
3400 	struct hammer_inostats *stats;
3401 	int delta;
3402 	int chain;
3403 	static volatile int iterator;	/* we don't care about MP races */
3404 
3405 	/*
3406 	 * Chain up to 4 times to find our entry.
3407 	 */
3408 	for (chain = 0; chain < 4; ++chain) {
3409 		stats = &hmp->inostats[(pid + chain) & HAMMER_INOSTATS_HMASK];
3410 		if (stats->pid == pid)
3411 			break;
3412 	}
3413 
3414 	/*
3415 	 * Replace one of the four chaining entries with our new entry.
3416 	 */
3417 	if (chain == 4) {
3418 		stats = &hmp->inostats[(pid + (iterator++ & 3)) &
3419 				       HAMMER_INOSTATS_HMASK];
3420 		stats->pid = pid;
3421 	}
3422 
3423 	/*
3424 	 * Decay the entry
3425 	 */
3426 	if (stats->count && stats->ltick != ticks) {
3427 		delta = ticks - stats->ltick;
3428 		stats->ltick = ticks;
3429 		if (delta <= 0 || delta > hz * 60)
3430 			stats->count = 0;
3431 		else
3432 			stats->count = stats->count * hz / (hz + delta);
3433 	}
3434 	if (hammer_debug_general & 0x10000)
3435 		hdkprintf("pid %5d stats %d\n", (int)pid, stats->count);
3436 	return (stats);
3437 }
3438 
3439 #if 0
3440 
3441 /*
3442  * XXX not used, doesn't work very well due to the large batching nature
3443  * of flushes.
3444  *
3445  * A larger then normal backlog of inodes is sitting in the flusher,
3446  * enforce a general slowdown to let it catch up.  This routine is only
3447  * called on completion of a non-flusher-related transaction which
3448  * performed B-Tree node I/O.
3449  *
3450  * It is possible for the flusher to stall in a continuous load.
3451  * blogbench -i1000 -o seems to do a good job generating this sort of load.
3452  * If the flusher is unable to catch up the inode count can bloat until
3453  * we run out of kvm.
3454  *
3455  * This is a bit of a hack.
3456  */
3457 void
3458 hammer_inode_waithard(hammer_mount_t hmp)
3459 {
3460 	/*
3461 	 * Hysteresis.
3462 	 */
3463 	if (hmp->flags & HAMMER_MOUNT_FLUSH_RECOVERY) {
3464 		if (hmp->count_reclaims < hammer_limit_reclaims / 2 &&
3465 		    hmp->count_iqueued < hmp->count_inodes / 20) {
3466 			hmp->flags &= ~HAMMER_MOUNT_FLUSH_RECOVERY;
3467 			return;
3468 		}
3469 	} else {
3470 		if (hmp->count_reclaims < hammer_limit_reclaims ||
3471 		    hmp->count_iqueued < hmp->count_inodes / 10) {
3472 			return;
3473 		}
3474 		hmp->flags |= HAMMER_MOUNT_FLUSH_RECOVERY;
3475 	}
3476 
3477 	/*
3478 	 * Block for one flush cycle.
3479 	 */
3480 	hammer_flusher_wait_next(hmp);
3481 }
3482 
3483 #endif
3484