4a4c306c95599dd5cdb8a9bb5e2752224e881637
[dragonfly.git] / sys / vfs / hammer / hammer_inode.c
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  * $DragonFly: src/sys/vfs/hammer/hammer_inode.c,v 1.114 2008/09/24 00:53:51 dillon Exp $
35  */
36
37 #include "hammer.h"
38 #include <vm/vm_extern.h>
39 #include <sys/buf.h>
40 #include <sys/buf2.h>
41
42 static int      hammer_unload_inode(struct hammer_inode *ip);
43 static void     hammer_free_inode(hammer_inode_t ip);
44 static void     hammer_flush_inode_core(hammer_inode_t ip,
45                                         hammer_flush_group_t flg, int flags);
46 static int      hammer_setup_child_callback(hammer_record_t rec, void *data);
47 #if 0
48 static int      hammer_syncgrp_child_callback(hammer_record_t rec, void *data);
49 #endif
50 static int      hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
51                                         hammer_flush_group_t flg);
52 static int      hammer_setup_parent_inodes_helper(hammer_record_t record,
53                                         int depth, hammer_flush_group_t flg);
54 static void     hammer_inode_wakereclaims(hammer_inode_t ip, int dowake);
55
56 #ifdef DEBUG_TRUNCATE
57 extern struct hammer_inode *HammerTruncIp;
58 #endif
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 /*
82  * RB-Tree support for inode structures / special LOOKUP_INFO
83  */
84 static int
85 hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
86 {
87         if (info->obj_localization < ip->obj_localization)
88                 return(-1);
89         if (info->obj_localization > ip->obj_localization)
90                 return(1);
91         if (info->obj_id < ip->obj_id)
92                 return(-1);
93         if (info->obj_id > ip->obj_id)
94                 return(1);
95         if (info->obj_asof < ip->obj_asof)
96                 return(-1);
97         if (info->obj_asof > ip->obj_asof)
98                 return(1);
99         return(0);
100 }
101
102 /*
103  * Used by hammer_scan_inode_snapshots() to locate all of an object's
104  * snapshots.  Note that the asof field is not tested, which we can get
105  * away with because it is the lowest-priority field.
106  */
107 static int
108 hammer_inode_info_cmp_all_history(hammer_inode_t ip, void *data)
109 {
110         hammer_inode_info_t info = data;
111
112         if (ip->obj_localization > info->obj_localization)
113                 return(1);
114         if (ip->obj_localization < info->obj_localization)
115                 return(-1);
116         if (ip->obj_id > info->obj_id)
117                 return(1);
118         if (ip->obj_id < info->obj_id)
119                 return(-1);
120         return(0);
121 }
122
123 /*
124  * Used by hammer_unload_pseudofs() to locate all inodes associated with
125  * a particular PFS.
126  */
127 static int
128 hammer_inode_pfs_cmp(hammer_inode_t ip, void *data)
129 {
130         u_int32_t localization = *(u_int32_t *)data;
131         if (ip->obj_localization > localization)
132                 return(1);
133         if (ip->obj_localization < localization)
134                 return(-1);
135         return(0);
136 }
137
138 /*
139  * RB-Tree support for pseudofs structures
140  */
141 static int
142 hammer_pfs_rb_compare(hammer_pseudofs_inmem_t p1, hammer_pseudofs_inmem_t p2)
143 {
144         if (p1->localization < p2->localization)
145                 return(-1);
146         if (p1->localization > p2->localization)
147                 return(1);
148         return(0);
149 }
150
151
152 RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
153 RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
154                 hammer_inode_info_cmp, hammer_inode_info_t);
155 RB_GENERATE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
156              hammer_pfs_rb_compare, u_int32_t, localization);
157
158 /*
159  * The kernel is not actively referencing this vnode but is still holding
160  * it cached.
161  *
162  * This is called from the frontend.
163  */
164 int
165 hammer_vop_inactive(struct vop_inactive_args *ap)
166 {
167         struct hammer_inode *ip = VTOI(ap->a_vp);
168
169         /*
170          * Degenerate case
171          */
172         if (ip == NULL) {
173                 vrecycle(ap->a_vp);
174                 return(0);
175         }
176
177         /*
178          * If the inode no longer has visibility in the filesystem try to
179          * recycle it immediately, even if the inode is dirty.  Recycling
180          * it quickly allows the system to reclaim buffer cache and VM
181          * resources which can matter a lot in a heavily loaded system.
182          *
183          * This can deadlock in vfsync() if we aren't careful.
184          * 
185          * Do not queue the inode to the flusher if we still have visibility,
186          * otherwise namespace calls such as chmod will unnecessarily generate
187          * multiple inode updates.
188          */
189         hammer_inode_unloadable_check(ip, 0);
190         if (ip->ino_data.nlinks == 0) {
191                 if (ip->flags & HAMMER_INODE_MODMASK)
192                         hammer_flush_inode(ip, 0);
193                 vrecycle(ap->a_vp);
194         }
195         return(0);
196 }
197
198 /*
199  * Release the vnode association.  This is typically (but not always)
200  * the last reference on the inode.
201  *
202  * Once the association is lost we are on our own with regards to
203  * flushing the inode.
204  */
205 int
206 hammer_vop_reclaim(struct vop_reclaim_args *ap)
207 {
208         struct hammer_inode *ip;
209         hammer_mount_t hmp;
210         struct vnode *vp;
211
212         vp = ap->a_vp;
213
214         if ((ip = vp->v_data) != NULL) {
215                 hmp = ip->hmp;
216                 vp->v_data = NULL;
217                 ip->vp = NULL;
218
219                 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0) {
220                         ++hammer_count_reclaiming;
221                         ++hmp->inode_reclaims;
222                         ip->flags |= HAMMER_INODE_RECLAIM;
223                 }
224                 hammer_rel_inode(ip, 1);
225         }
226         return(0);
227 }
228
229 /*
230  * Return a locked vnode for the specified inode.  The inode must be
231  * referenced but NOT LOCKED on entry and will remain referenced on
232  * return.
233  *
234  * Called from the frontend.
235  */
236 int
237 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
238 {
239         hammer_mount_t hmp;
240         struct vnode *vp;
241         int error = 0;
242         u_int8_t obj_type;
243
244         hmp = ip->hmp;
245
246         for (;;) {
247                 if ((vp = ip->vp) == NULL) {
248                         error = getnewvnode(VT_HAMMER, hmp->mp, vpp, 0, 0);
249                         if (error)
250                                 break;
251                         hammer_lock_ex(&ip->lock);
252                         if (ip->vp != NULL) {
253                                 hammer_unlock(&ip->lock);
254                                 vp = *vpp;
255                                 vp->v_type = VBAD;
256                                 vx_put(vp);
257                                 continue;
258                         }
259                         hammer_ref(&ip->lock);
260                         vp = *vpp;
261                         ip->vp = vp;
262
263                         obj_type = ip->ino_data.obj_type;
264                         vp->v_type = hammer_get_vnode_type(obj_type);
265
266                         hammer_inode_wakereclaims(ip, 0);
267
268                         switch(ip->ino_data.obj_type) {
269                         case HAMMER_OBJTYPE_CDEV:
270                         case HAMMER_OBJTYPE_BDEV:
271                                 vp->v_ops = &hmp->mp->mnt_vn_spec_ops;
272                                 addaliasu(vp, ip->ino_data.rmajor,
273                                           ip->ino_data.rminor);
274                                 break;
275                         case HAMMER_OBJTYPE_FIFO:
276                                 vp->v_ops = &hmp->mp->mnt_vn_fifo_ops;
277                                 break;
278                         default:
279                                 break;
280                         }
281
282                         /*
283                          * Only mark as the root vnode if the ip is not
284                          * historical, otherwise the VFS cache will get
285                          * confused.  The other half of the special handling
286                          * is in hammer_vop_nlookupdotdot().
287                          *
288                          * Pseudo-filesystem roots can be accessed via
289                          * non-root filesystem paths and setting VROOT may
290                          * confuse the namecache.  Set VPFSROOT instead.
291                          */
292                         if (ip->obj_id == HAMMER_OBJID_ROOT &&
293                             ip->obj_asof == hmp->asof) {
294                                 if (ip->obj_localization == 0)
295                                         vp->v_flag |= VROOT;
296                                 else
297                                         vp->v_flag |= VPFSROOT;
298                         }
299
300                         vp->v_data = (void *)ip;
301                         /* vnode locked by getnewvnode() */
302                         /* make related vnode dirty if inode dirty? */
303                         hammer_unlock(&ip->lock);
304                         if (vp->v_type == VREG)
305                                 vinitvmio(vp, ip->ino_data.size);
306                         break;
307                 }
308
309                 /*
310                  * loop if the vget fails (aka races), or if the vp
311                  * no longer matches ip->vp.
312                  */
313                 if (vget(vp, LK_EXCLUSIVE) == 0) {
314                         if (vp == ip->vp)
315                                 break;
316                         vput(vp);
317                 }
318         }
319         *vpp = vp;
320         return(error);
321 }
322
323 /*
324  * Locate all copies of the inode for obj_id compatible with the specified
325  * asof, reference, and issue the related call-back.  This routine is used
326  * for direct-io invalidation and does not create any new inodes.
327  */
328 void
329 hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo,
330                             int (*callback)(hammer_inode_t ip, void *data),
331                             void *data)
332 {
333         hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root,
334                                    hammer_inode_info_cmp_all_history,
335                                    callback, iinfo);
336 }
337
338 /*
339  * Acquire a HAMMER inode.  The returned inode is not locked.  These functions
340  * do not attach or detach the related vnode (use hammer_get_vnode() for
341  * that).
342  *
343  * The flags argument is only applied for newly created inodes, and only
344  * certain flags are inherited.
345  *
346  * Called from the frontend.
347  */
348 struct hammer_inode *
349 hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip,
350                  int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
351                  int flags, int *errorp)
352 {
353         hammer_mount_t hmp = trans->hmp;
354         struct hammer_node_cache *cachep;
355         struct hammer_inode_info iinfo;
356         struct hammer_cursor cursor;
357         struct hammer_inode *ip;
358
359
360         /*
361          * Determine if we already have an inode cached.  If we do then
362          * we are golden.
363          *
364          * If we find an inode with no vnode we have to mark the
365          * transaction such that hammer_inode_waitreclaims() is
366          * called later on to avoid building up an infinite number
367          * of inodes.  Otherwise we can continue to * add new inodes
368          * faster then they can be disposed of, even with the tsleep
369          * delay.
370          *
371          * If we find a dummy inode we return a failure so dounlink
372          * (which does another lookup) doesn't try to mess with the
373          * link count.  hammer_vop_nresolve() uses hammer_get_dummy_inode()
374          * to ref dummy inodes.
375          */
376         iinfo.obj_id = obj_id;
377         iinfo.obj_asof = asof;
378         iinfo.obj_localization = localization;
379 loop:
380         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
381         if (ip) {
382                 if (ip->flags & HAMMER_INODE_DUMMY) {
383                         *errorp = ENOENT;
384                         return(NULL);
385                 }
386                 hammer_ref(&ip->lock);
387                 *errorp = 0;
388                 return(ip);
389         }
390
391         /*
392          * Allocate a new inode structure and deal with races later.
393          */
394         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
395         ++hammer_count_inodes;
396         ++hmp->count_inodes;
397         ip->obj_id = obj_id;
398         ip->obj_asof = iinfo.obj_asof;
399         ip->obj_localization = localization;
400         ip->hmp = hmp;
401         ip->flags = flags & HAMMER_INODE_RO;
402         ip->cache[0].ip = ip;
403         ip->cache[1].ip = ip;
404         ip->cache[2].ip = ip;
405         ip->cache[3].ip = ip;
406         if (hmp->ronly)
407                 ip->flags |= HAMMER_INODE_RO;
408         ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
409                 0x7FFFFFFFFFFFFFFFLL;
410         RB_INIT(&ip->rec_tree);
411         TAILQ_INIT(&ip->target_list);
412         hammer_ref(&ip->lock);
413
414         /*
415          * Locate the on-disk inode.  If this is a PFS root we always
416          * access the current version of the root inode and (if it is not
417          * a master) always access information under it with a snapshot
418          * TID.
419          *
420          * We cache recent inode lookups in this directory in dip->cache[2].
421          * If we can't find it we assume the inode we are looking for is
422          * close to the directory inode.
423          */
424 retry:
425         cachep = NULL;
426         if (dip) {
427                 if (dip->cache[2].node)
428                         cachep = &dip->cache[2];
429                 else
430                         cachep = &dip->cache[0];
431         }
432         hammer_init_cursor(trans, &cursor, cachep, NULL);
433         cursor.key_beg.localization = localization + HAMMER_LOCALIZE_INODE;
434         cursor.key_beg.obj_id = ip->obj_id;
435         cursor.key_beg.key = 0;
436         cursor.key_beg.create_tid = 0;
437         cursor.key_beg.delete_tid = 0;
438         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
439         cursor.key_beg.obj_type = 0;
440
441         cursor.asof = iinfo.obj_asof;
442         cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
443                        HAMMER_CURSOR_ASOF;
444
445         *errorp = hammer_btree_lookup(&cursor);
446         if (*errorp == EDEADLK) {
447                 hammer_done_cursor(&cursor);
448                 goto retry;
449         }
450
451         /*
452          * On success the B-Tree lookup will hold the appropriate
453          * buffer cache buffers and provide a pointer to the requested
454          * information.  Copy the information to the in-memory inode
455          * and cache the B-Tree node to improve future operations.
456          */
457         if (*errorp == 0) {
458                 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
459                 ip->ino_data = cursor.data->inode;
460
461                 /*
462                  * cache[0] tries to cache the location of the object inode.
463                  * The assumption is that it is near the directory inode.
464                  *
465                  * cache[1] tries to cache the location of the object data.
466                  * We might have something in the governing directory from
467                  * scan optimizations (see the strategy code in
468                  * hammer_vnops.c).
469                  *
470                  * We update dip->cache[2], if possible, with the location
471                  * of the object inode for future directory shortcuts.
472                  */
473                 hammer_cache_node(&ip->cache[0], cursor.node);
474                 if (dip) {
475                         if (dip->cache[3].node) {
476                                 hammer_cache_node(&ip->cache[1],
477                                                   dip->cache[3].node);
478                         }
479                         hammer_cache_node(&dip->cache[2], cursor.node);
480                 }
481
482                 /*
483                  * The file should not contain any data past the file size
484                  * stored in the inode.  Setting save_trunc_off to the
485                  * file size instead of max reduces B-Tree lookup overheads
486                  * on append by allowing the flusher to avoid checking for
487                  * record overwrites.
488                  */
489                 ip->save_trunc_off = ip->ino_data.size;
490
491                 /*
492                  * Locate and assign the pseudofs management structure to
493                  * the inode.
494                  */
495                 if (dip && dip->obj_localization == ip->obj_localization) {
496                         ip->pfsm = dip->pfsm;
497                         hammer_ref(&ip->pfsm->lock);
498                 } else {
499                         ip->pfsm = hammer_load_pseudofs(trans,
500                                                         ip->obj_localization,
501                                                         errorp);
502                         *errorp = 0;    /* ignore ENOENT */
503                 }
504         }
505
506         /*
507          * The inode is placed on the red-black tree and will be synced to
508          * the media when flushed or by the filesystem sync.  If this races
509          * another instantiation/lookup the insertion will fail.
510          */
511         if (*errorp == 0) {
512                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
513                         hammer_free_inode(ip);
514                         hammer_done_cursor(&cursor);
515                         goto loop;
516                 }
517                 ip->flags |= HAMMER_INODE_ONDISK;
518         } else {
519                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
520                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
521                         --hmp->rsv_inodes;
522                 }
523
524                 hammer_free_inode(ip);
525                 ip = NULL;
526         }
527         hammer_done_cursor(&cursor);
528         trans->flags |= HAMMER_TRANSF_NEWINODE;
529         return (ip);
530 }
531
532 /*
533  * Get a dummy inode to placemark a broken directory entry.
534  */
535 struct hammer_inode *
536 hammer_get_dummy_inode(hammer_transaction_t trans, hammer_inode_t dip,
537                  int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
538                  int flags, int *errorp)
539 {
540         hammer_mount_t hmp = trans->hmp;
541         struct hammer_inode_info iinfo;
542         struct hammer_inode *ip;
543
544         /*
545          * Determine if we already have an inode cached.  If we do then
546          * we are golden.
547          *
548          * If we find an inode with no vnode we have to mark the
549          * transaction such that hammer_inode_waitreclaims() is
550          * called later on to avoid building up an infinite number
551          * of inodes.  Otherwise we can continue to * add new inodes
552          * faster then they can be disposed of, even with the tsleep
553          * delay.
554          *
555          * If we find a non-fake inode we return an error.  Only fake
556          * inodes can be returned by this routine.
557          */
558         iinfo.obj_id = obj_id;
559         iinfo.obj_asof = asof;
560         iinfo.obj_localization = localization;
561 loop:
562         *errorp = 0;
563         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
564         if (ip) {
565                 if ((ip->flags & HAMMER_INODE_DUMMY) == 0) {
566                         *errorp = ENOENT;
567                         return(NULL);
568                 }
569                 hammer_ref(&ip->lock);
570                 return(ip);
571         }
572
573         /*
574          * Allocate a new inode structure and deal with races later.
575          */
576         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
577         ++hammer_count_inodes;
578         ++hmp->count_inodes;
579         ip->obj_id = obj_id;
580         ip->obj_asof = iinfo.obj_asof;
581         ip->obj_localization = localization;
582         ip->hmp = hmp;
583         ip->flags = flags | HAMMER_INODE_RO | HAMMER_INODE_DUMMY;
584         ip->cache[0].ip = ip;
585         ip->cache[1].ip = ip;
586         ip->cache[2].ip = ip;
587         ip->cache[3].ip = ip;
588         ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
589                 0x7FFFFFFFFFFFFFFFLL;
590         RB_INIT(&ip->rec_tree);
591         TAILQ_INIT(&ip->target_list);
592         hammer_ref(&ip->lock);
593
594         /*
595          * Populate the dummy inode.  Leave everything zero'd out.
596          *
597          * (ip->ino_leaf and ip->ino_data)
598          *
599          * Make the dummy inode a FIFO object which most copy programs
600          * will properly ignore.
601          */
602         ip->save_trunc_off = ip->ino_data.size;
603         ip->ino_data.obj_type = HAMMER_OBJTYPE_FIFO;
604
605         /*
606          * Locate and assign the pseudofs management structure to
607          * the inode.
608          */
609         if (dip && dip->obj_localization == ip->obj_localization) {
610                 ip->pfsm = dip->pfsm;
611                 hammer_ref(&ip->pfsm->lock);
612         } else {
613                 ip->pfsm = hammer_load_pseudofs(trans, ip->obj_localization,
614                                                 errorp);
615                 *errorp = 0;    /* ignore ENOENT */
616         }
617
618         /*
619          * The inode is placed on the red-black tree and will be synced to
620          * the media when flushed or by the filesystem sync.  If this races
621          * another instantiation/lookup the insertion will fail.
622          *
623          * NOTE: Do not set HAMMER_INODE_ONDISK.  The inode is a fake.
624          */
625         if (*errorp == 0) {
626                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
627                         hammer_free_inode(ip);
628                         goto loop;
629                 }
630         } else {
631                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
632                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
633                         --hmp->rsv_inodes;
634                 }
635                 hammer_free_inode(ip);
636                 ip = NULL;
637         }
638         trans->flags |= HAMMER_TRANSF_NEWINODE;
639         return (ip);
640 }
641
642 /*
643  * Return a referenced inode only if it is in our inode cache.
644  *
645  * Dummy inodes do not count.
646  */
647 struct hammer_inode *
648 hammer_find_inode(hammer_transaction_t trans, int64_t obj_id,
649                   hammer_tid_t asof, u_int32_t localization)
650 {
651         hammer_mount_t hmp = trans->hmp;
652         struct hammer_inode_info iinfo;
653         struct hammer_inode *ip;
654
655         iinfo.obj_id = obj_id;
656         iinfo.obj_asof = asof;
657         iinfo.obj_localization = localization;
658
659         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
660         if (ip) {
661                 if (ip->flags & HAMMER_INODE_DUMMY)
662                         ip = NULL;
663                 else
664                         hammer_ref(&ip->lock);
665         }
666         return(ip);
667 }
668
669 /*
670  * Create a new filesystem object, returning the inode in *ipp.  The
671  * returned inode will be referenced.  The inode is created in-memory.
672  *
673  * If pfsm is non-NULL the caller wishes to create the root inode for
674  * a master PFS.
675  */
676 int
677 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
678                     struct ucred *cred,
679                     hammer_inode_t dip, const char *name, int namelen,
680                     hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
681 {
682         hammer_mount_t hmp;
683         hammer_inode_t ip;
684         uid_t xuid;
685         int error;
686         int64_t namekey;
687         u_int32_t dummy;
688
689         hmp = trans->hmp;
690
691         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
692         ++hammer_count_inodes;
693         ++hmp->count_inodes;
694         trans->flags |= HAMMER_TRANSF_NEWINODE;
695
696         if (pfsm) {
697                 KKASSERT(pfsm->localization != 0);
698                 ip->obj_id = HAMMER_OBJID_ROOT;
699                 ip->obj_localization = pfsm->localization;
700         } else {
701                 KKASSERT(dip != NULL);
702                 namekey = hammer_directory_namekey(dip, name, namelen, &dummy);
703                 ip->obj_id = hammer_alloc_objid(hmp, dip, namekey);
704                 ip->obj_localization = dip->obj_localization;
705         }
706
707         KKASSERT(ip->obj_id != 0);
708         ip->obj_asof = hmp->asof;
709         ip->hmp = hmp;
710         ip->flush_state = HAMMER_FST_IDLE;
711         ip->flags = HAMMER_INODE_DDIRTY |
712                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
713         ip->cache[0].ip = ip;
714         ip->cache[1].ip = ip;
715         ip->cache[2].ip = ip;
716         ip->cache[3].ip = ip;
717
718         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
719         /* ip->save_trunc_off = 0; (already zero) */
720         RB_INIT(&ip->rec_tree);
721         TAILQ_INIT(&ip->target_list);
722
723         ip->ino_data.atime = trans->time;
724         ip->ino_data.mtime = trans->time;
725         ip->ino_data.size = 0;
726         ip->ino_data.nlinks = 0;
727
728         /*
729          * A nohistory designator on the parent directory is inherited by
730          * the child.  We will do this even for pseudo-fs creation... the
731          * sysad can turn it off.
732          */
733         if (dip) {
734                 ip->ino_data.uflags = dip->ino_data.uflags &
735                                       (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
736         }
737
738         ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
739         ip->ino_leaf.base.localization = ip->obj_localization +
740                                          HAMMER_LOCALIZE_INODE;
741         ip->ino_leaf.base.obj_id = ip->obj_id;
742         ip->ino_leaf.base.key = 0;
743         ip->ino_leaf.base.create_tid = 0;
744         ip->ino_leaf.base.delete_tid = 0;
745         ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
746         ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
747
748         ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
749         ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
750         ip->ino_data.mode = vap->va_mode;
751         ip->ino_data.ctime = trans->time;
752
753         /*
754          * If we are running version 2 or greater directory entries are
755          * inode-localized instead of data-localized.
756          */
757         if (trans->hmp->version >= HAMMER_VOL_VERSION_TWO) {
758                 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
759                         ip->ino_data.cap_flags |=
760                                 HAMMER_INODE_CAP_DIR_LOCAL_INO;
761                 }
762         }
763
764         /*
765          * Setup the ".." pointer.  This only needs to be done for directories
766          * but we do it for all objects as a recovery aid.
767          */
768         if (dip)
769                 ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
770 #if 0
771         /*
772          * The parent_obj_localization field only applies to pseudo-fs roots.
773          * XXX this is no longer applicable, PFSs are no longer directly
774          * tied into the parent's directory structure.
775          */
776         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY &&
777             ip->obj_id == HAMMER_OBJID_ROOT) {
778                 ip->ino_data.ext.obj.parent_obj_localization = 
779                                                 dip->obj_localization;
780         }
781 #endif
782
783         switch(ip->ino_leaf.base.obj_type) {
784         case HAMMER_OBJTYPE_CDEV:
785         case HAMMER_OBJTYPE_BDEV:
786                 ip->ino_data.rmajor = vap->va_rmajor;
787                 ip->ino_data.rminor = vap->va_rminor;
788                 break;
789         default:
790                 break;
791         }
792
793         /*
794          * Calculate default uid/gid and overwrite with information from
795          * the vap.
796          */
797         if (dip) {
798                 xuid = hammer_to_unix_xid(&dip->ino_data.uid);
799                 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
800                                              xuid, cred, &vap->va_mode);
801         } else {
802                 xuid = 0;
803         }
804         ip->ino_data.mode = vap->va_mode;
805
806         if (vap->va_vaflags & VA_UID_UUID_VALID)
807                 ip->ino_data.uid = vap->va_uid_uuid;
808         else if (vap->va_uid != (uid_t)VNOVAL)
809                 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
810         else
811                 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
812
813         if (vap->va_vaflags & VA_GID_UUID_VALID)
814                 ip->ino_data.gid = vap->va_gid_uuid;
815         else if (vap->va_gid != (gid_t)VNOVAL)
816                 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
817         else if (dip)
818                 ip->ino_data.gid = dip->ino_data.gid;
819
820         hammer_ref(&ip->lock);
821
822         if (pfsm) {
823                 ip->pfsm = pfsm;
824                 hammer_ref(&pfsm->lock);
825                 error = 0;
826         } else if (dip->obj_localization == ip->obj_localization) {
827                 ip->pfsm = dip->pfsm;
828                 hammer_ref(&ip->pfsm->lock);
829                 error = 0;
830         } else {
831                 ip->pfsm = hammer_load_pseudofs(trans,
832                                                 ip->obj_localization,
833                                                 &error);
834                 error = 0;      /* ignore ENOENT */
835         }
836
837         if (error) {
838                 hammer_free_inode(ip);
839                 ip = NULL;
840         } else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
841                 panic("hammer_create_inode: duplicate obj_id %llx",
842                       (long long)ip->obj_id);
843                 /* not reached */
844                 hammer_free_inode(ip);
845         }
846         *ipp = ip;
847         return(error);
848 }
849
850 /*
851  * Final cleanup / freeing of an inode structure
852  */
853 static void
854 hammer_free_inode(hammer_inode_t ip)
855 {
856         struct hammer_mount *hmp;
857
858         hmp = ip->hmp;
859         KKASSERT(ip->lock.refs == 1);
860         hammer_uncache_node(&ip->cache[0]);
861         hammer_uncache_node(&ip->cache[1]);
862         hammer_uncache_node(&ip->cache[2]);
863         hammer_uncache_node(&ip->cache[3]);
864         hammer_inode_wakereclaims(ip, 1);
865         if (ip->objid_cache)
866                 hammer_clear_objid(ip);
867         --hammer_count_inodes;
868         --hmp->count_inodes;
869         if (ip->pfsm) {
870                 hammer_rel_pseudofs(hmp, ip->pfsm);
871                 ip->pfsm = NULL;
872         }
873         kfree(ip, hmp->m_inodes);
874         ip = NULL;
875 }
876
877 /*
878  * Retrieve pseudo-fs data.  NULL will never be returned.
879  *
880  * If an error occurs *errorp will be set and a default template is returned,
881  * otherwise *errorp is set to 0.  Typically when an error occurs it will
882  * be ENOENT.
883  */
884 hammer_pseudofs_inmem_t
885 hammer_load_pseudofs(hammer_transaction_t trans,
886                      u_int32_t localization, int *errorp)
887 {
888         hammer_mount_t hmp = trans->hmp;
889         hammer_inode_t ip;
890         hammer_pseudofs_inmem_t pfsm;
891         struct hammer_cursor cursor;
892         int bytes;
893
894 retry:
895         pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
896         if (pfsm) {
897                 hammer_ref(&pfsm->lock);
898                 *errorp = 0;
899                 return(pfsm);
900         }
901
902         /*
903          * PFS records are stored in the root inode (not the PFS root inode,
904          * but the real root).  Avoid an infinite recursion if loading
905          * the PFS for the real root.
906          */
907         if (localization) {
908                 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
909                                       HAMMER_MAX_TID,
910                                       HAMMER_DEF_LOCALIZATION, 0, errorp);
911         } else {
912                 ip = NULL;
913         }
914
915         pfsm = kmalloc(sizeof(*pfsm), hmp->m_misc, M_WAITOK | M_ZERO);
916         pfsm->localization = localization;
917         pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
918         pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
919
920         hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
921         cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION +
922                                       HAMMER_LOCALIZE_MISC;
923         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
924         cursor.key_beg.create_tid = 0;
925         cursor.key_beg.delete_tid = 0;
926         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
927         cursor.key_beg.obj_type = 0;
928         cursor.key_beg.key = localization;
929         cursor.asof = HAMMER_MAX_TID;
930         cursor.flags |= HAMMER_CURSOR_ASOF;
931
932         if (ip)
933                 *errorp = hammer_ip_lookup(&cursor);
934         else
935                 *errorp = hammer_btree_lookup(&cursor);
936         if (*errorp == 0) {
937                 *errorp = hammer_ip_resolve_data(&cursor);
938                 if (*errorp == 0) {
939                         if (cursor.data->pfsd.mirror_flags &
940                             HAMMER_PFSD_DELETED) {
941                                 *errorp = ENOENT;
942                         } else {
943                                 bytes = cursor.leaf->data_len;
944                                 if (bytes > sizeof(pfsm->pfsd))
945                                         bytes = sizeof(pfsm->pfsd);
946                                 bcopy(cursor.data, &pfsm->pfsd, bytes);
947                         }
948                 }
949         }
950         hammer_done_cursor(&cursor);
951
952         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
953         hammer_ref(&pfsm->lock);
954         if (ip)
955                 hammer_rel_inode(ip, 0);
956         if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
957                 kfree(pfsm, hmp->m_misc);
958                 goto retry;
959         }
960         return(pfsm);
961 }
962
963 /*
964  * Store pseudo-fs data.  The backend will automatically delete any prior
965  * on-disk pseudo-fs data but we have to delete in-memory versions.
966  */
967 int
968 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
969 {
970         struct hammer_cursor cursor;
971         hammer_record_t record;
972         hammer_inode_t ip;
973         int error;
974
975         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
976                               HAMMER_DEF_LOCALIZATION, 0, &error);
977 retry:
978         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
979         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
980         cursor.key_beg.localization = ip->obj_localization +
981                                       HAMMER_LOCALIZE_MISC;
982         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
983         cursor.key_beg.create_tid = 0;
984         cursor.key_beg.delete_tid = 0;
985         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
986         cursor.key_beg.obj_type = 0;
987         cursor.key_beg.key = pfsm->localization;
988         cursor.asof = HAMMER_MAX_TID;
989         cursor.flags |= HAMMER_CURSOR_ASOF;
990
991         /*
992          * Replace any in-memory version of the record.
993          */
994         error = hammer_ip_lookup(&cursor);
995         if (error == 0 && hammer_cursor_inmem(&cursor)) {
996                 record = cursor.iprec;
997                 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
998                         KKASSERT(cursor.deadlk_rec == NULL);
999                         hammer_ref(&record->lock);
1000                         cursor.deadlk_rec = record;
1001                         error = EDEADLK;
1002                 } else {
1003                         record->flags |= HAMMER_RECF_DELETED_FE;
1004                         error = 0;
1005                 }
1006         }
1007
1008         /*
1009          * Allocate replacement general record.  The backend flush will
1010          * delete any on-disk version of the record.
1011          */
1012         if (error == 0 || error == ENOENT) {
1013                 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
1014                 record->type = HAMMER_MEM_RECORD_GENERAL;
1015
1016                 record->leaf.base.localization = ip->obj_localization +
1017                                                  HAMMER_LOCALIZE_MISC;
1018                 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
1019                 record->leaf.base.key = pfsm->localization;
1020                 record->leaf.data_len = sizeof(pfsm->pfsd);
1021                 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
1022                 error = hammer_ip_add_record(trans, record);
1023         }
1024         hammer_done_cursor(&cursor);
1025         if (error == EDEADLK)
1026                 goto retry;
1027         hammer_rel_inode(ip, 0);
1028         return(error);
1029 }
1030
1031 /*
1032  * Create a root directory for a PFS if one does not alredy exist.
1033  *
1034  * The PFS root stands alone so we must also bump the nlinks count
1035  * to prevent it from being destroyed on release.
1036  */
1037 int
1038 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1039                        hammer_pseudofs_inmem_t pfsm)
1040 {
1041         hammer_inode_t ip;
1042         struct vattr vap;
1043         int error;
1044
1045         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1046                               pfsm->localization, 0, &error);
1047         if (ip == NULL) {
1048                 vattr_null(&vap);
1049                 vap.va_mode = 0755;
1050                 vap.va_type = VDIR;
1051                 error = hammer_create_inode(trans, &vap, cred,
1052                                             NULL, NULL, 0,
1053                                             pfsm, &ip);
1054                 if (error == 0) {
1055                         ++ip->ino_data.nlinks;
1056                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
1057                 }
1058         }
1059         if (ip)
1060                 hammer_rel_inode(ip, 0);
1061         return(error);
1062 }
1063
1064 /*
1065  * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
1066  * if we are unable to disassociate all the inodes.
1067  */
1068 static
1069 int
1070 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
1071 {
1072         int res;
1073
1074         hammer_ref(&ip->lock);
1075         if (ip->lock.refs == 2 && ip->vp)
1076                 vclean_unlocked(ip->vp);
1077         if (ip->lock.refs == 1 && ip->vp == NULL)
1078                 res = 0;
1079         else
1080                 res = -1;       /* stop, someone is using the inode */
1081         hammer_rel_inode(ip, 0);
1082         return(res);
1083 }
1084
1085 int
1086 hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
1087 {
1088         int res;
1089         int try;
1090
1091         for (try = res = 0; try < 4; ++try) {
1092                 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
1093                                            hammer_inode_pfs_cmp,
1094                                            hammer_unload_pseudofs_callback,
1095                                            &localization);
1096                 if (res == 0 && try > 1)
1097                         break;
1098                 hammer_flusher_sync(trans->hmp);
1099         }
1100         if (res != 0)
1101                 res = ENOTEMPTY;
1102         return(res);
1103 }
1104
1105
1106 /*
1107  * Release a reference on a PFS
1108  */
1109 void
1110 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
1111 {
1112         hammer_unref(&pfsm->lock);
1113         if (pfsm->lock.refs == 0) {
1114                 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
1115                 kfree(pfsm, hmp->m_misc);
1116         }
1117 }
1118
1119 /*
1120  * Called by hammer_sync_inode().
1121  */
1122 static int
1123 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
1124 {
1125         hammer_transaction_t trans = cursor->trans;
1126         hammer_record_t record;
1127         int error;
1128         int redirty;
1129
1130 retry:
1131         error = 0;
1132
1133         /*
1134          * If the inode has a presence on-disk then locate it and mark
1135          * it deleted, setting DELONDISK.
1136          *
1137          * The record may or may not be physically deleted, depending on
1138          * the retention policy.
1139          */
1140         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
1141             HAMMER_INODE_ONDISK) {
1142                 hammer_normalize_cursor(cursor);
1143                 cursor->key_beg.localization = ip->obj_localization + 
1144                                                HAMMER_LOCALIZE_INODE;
1145                 cursor->key_beg.obj_id = ip->obj_id;
1146                 cursor->key_beg.key = 0;
1147                 cursor->key_beg.create_tid = 0;
1148                 cursor->key_beg.delete_tid = 0;
1149                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1150                 cursor->key_beg.obj_type = 0;
1151                 cursor->asof = ip->obj_asof;
1152                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1153                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
1154                 cursor->flags |= HAMMER_CURSOR_BACKEND;
1155
1156                 error = hammer_btree_lookup(cursor);
1157                 if (hammer_debug_inode)
1158                         kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
1159
1160                 if (error == 0) {
1161                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1162                         if (hammer_debug_inode)
1163                                 kprintf(" error %d\n", error);
1164                         if (error == 0) {
1165                                 ip->flags |= HAMMER_INODE_DELONDISK;
1166                         }
1167                         if (cursor->node)
1168                                 hammer_cache_node(&ip->cache[0], cursor->node);
1169                 }
1170                 if (error == EDEADLK) {
1171                         hammer_done_cursor(cursor);
1172                         error = hammer_init_cursor(trans, cursor,
1173                                                    &ip->cache[0], ip);
1174                         if (hammer_debug_inode)
1175                                 kprintf("IPDED %p %d\n", ip, error);
1176                         if (error == 0)
1177                                 goto retry;
1178                 }
1179         }
1180
1181         /*
1182          * Ok, write out the initial record or a new record (after deleting
1183          * the old one), unless the DELETED flag is set.  This routine will
1184          * clear DELONDISK if it writes out a record.
1185          *
1186          * Update our inode statistics if this is the first application of
1187          * the inode on-disk.
1188          */
1189         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
1190                 /*
1191                  * Generate a record and write it to the media.  We clean-up
1192                  * the state before releasing so we do not have to set-up
1193                  * a flush_group.
1194                  */
1195                 record = hammer_alloc_mem_record(ip, 0);
1196                 record->type = HAMMER_MEM_RECORD_INODE;
1197                 record->flush_state = HAMMER_FST_FLUSH;
1198                 record->leaf = ip->sync_ino_leaf;
1199                 record->leaf.base.create_tid = trans->tid;
1200                 record->leaf.data_len = sizeof(ip->sync_ino_data);
1201                 record->leaf.create_ts = trans->time32;
1202                 record->data = (void *)&ip->sync_ino_data;
1203                 record->flags |= HAMMER_RECF_INTERLOCK_BE;
1204
1205                 /*
1206                  * If this flag is set we cannot sync the new file size
1207                  * because we haven't finished related truncations.  The
1208                  * inode will be flushed in another flush group to finish
1209                  * the job.
1210                  */
1211                 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1212                     ip->sync_ino_data.size != ip->ino_data.size) {
1213                         redirty = 1;
1214                         ip->sync_ino_data.size = ip->ino_data.size;
1215                 } else {
1216                         redirty = 0;
1217                 }
1218
1219                 for (;;) {
1220                         error = hammer_ip_sync_record_cursor(cursor, record);
1221                         if (hammer_debug_inode)
1222                                 kprintf("GENREC %p rec %08x %d\n",      
1223                                         ip, record->flags, error);
1224                         if (error != EDEADLK)
1225                                 break;
1226                         hammer_done_cursor(cursor);
1227                         error = hammer_init_cursor(trans, cursor,
1228                                                    &ip->cache[0], ip);
1229                         if (hammer_debug_inode)
1230                                 kprintf("GENREC reinit %d\n", error);
1231                         if (error)
1232                                 break;
1233                 }
1234
1235                 /*
1236                  * Note:  The record was never on the inode's record tree
1237                  * so just wave our hands importantly and destroy it.
1238                  */
1239                 record->flags |= HAMMER_RECF_COMMITTED;
1240                 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1241                 record->flush_state = HAMMER_FST_IDLE;
1242                 ++ip->rec_generation;
1243                 hammer_rel_mem_record(record);
1244
1245                 /*
1246                  * Finish up.
1247                  */
1248                 if (error == 0) {
1249                         if (hammer_debug_inode)
1250                                 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1251                         ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1252                                             HAMMER_INODE_ATIME |
1253                                             HAMMER_INODE_MTIME);
1254                         ip->flags &= ~HAMMER_INODE_DELONDISK;
1255                         if (redirty)
1256                                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1257
1258                         /*
1259                          * Root volume count of inodes
1260                          */
1261                         hammer_sync_lock_sh(trans);
1262                         if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1263                                 hammer_modify_volume_field(trans,
1264                                                            trans->rootvol,
1265                                                            vol0_stat_inodes);
1266                                 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1267                                 hammer_modify_volume_done(trans->rootvol);
1268                                 ip->flags |= HAMMER_INODE_ONDISK;
1269                                 if (hammer_debug_inode)
1270                                         kprintf("NOWONDISK %p\n", ip);
1271                         }
1272                         hammer_sync_unlock(trans);
1273                 }
1274         }
1275
1276         /*
1277          * If the inode has been destroyed, clean out any left-over flags
1278          * that may have been set by the frontend.
1279          */
1280         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) { 
1281                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1282                                     HAMMER_INODE_ATIME |
1283                                     HAMMER_INODE_MTIME);
1284         }
1285         return(error);
1286 }
1287
1288 /*
1289  * Update only the itimes fields.
1290  *
1291  * ATIME can be updated without generating any UNDO.  MTIME is updated
1292  * with UNDO so it is guaranteed to be synchronized properly in case of
1293  * a crash.
1294  *
1295  * Neither field is included in the B-Tree leaf element's CRC, which is how
1296  * we can get away with updating ATIME the way we do.
1297  */
1298 static int
1299 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1300 {
1301         hammer_transaction_t trans = cursor->trans;
1302         int error;
1303
1304 retry:
1305         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1306             HAMMER_INODE_ONDISK) {
1307                 return(0);
1308         }
1309
1310         hammer_normalize_cursor(cursor);
1311         cursor->key_beg.localization = ip->obj_localization + 
1312                                        HAMMER_LOCALIZE_INODE;
1313         cursor->key_beg.obj_id = ip->obj_id;
1314         cursor->key_beg.key = 0;
1315         cursor->key_beg.create_tid = 0;
1316         cursor->key_beg.delete_tid = 0;
1317         cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1318         cursor->key_beg.obj_type = 0;
1319         cursor->asof = ip->obj_asof;
1320         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1321         cursor->flags |= HAMMER_CURSOR_ASOF;
1322         cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1323         cursor->flags |= HAMMER_CURSOR_GET_DATA;
1324         cursor->flags |= HAMMER_CURSOR_BACKEND;
1325
1326         error = hammer_btree_lookup(cursor);
1327         if (error == 0) {
1328                 hammer_cache_node(&ip->cache[0], cursor->node);
1329                 if (ip->sync_flags & HAMMER_INODE_MTIME) {
1330                         /*
1331                          * Updating MTIME requires an UNDO.  Just cover
1332                          * both atime and mtime.
1333                          */
1334                         hammer_sync_lock_sh(trans);
1335                         hammer_modify_buffer(trans, cursor->data_buffer,
1336                                      HAMMER_ITIMES_BASE(&cursor->data->inode),
1337                                      HAMMER_ITIMES_BYTES);
1338                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1339                         cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1340                         hammer_modify_buffer_done(cursor->data_buffer);
1341                         hammer_sync_unlock(trans);
1342                 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1343                         /*
1344                          * Updating atime only can be done in-place with
1345                          * no UNDO.
1346                          */
1347                         hammer_sync_lock_sh(trans);
1348                         hammer_modify_buffer(trans, cursor->data_buffer,
1349                                              NULL, 0);
1350                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1351                         hammer_modify_buffer_done(cursor->data_buffer);
1352                         hammer_sync_unlock(trans);
1353                 }
1354                 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1355         }
1356         if (error == EDEADLK) {
1357                 hammer_done_cursor(cursor);
1358                 error = hammer_init_cursor(trans, cursor,
1359                                            &ip->cache[0], ip);
1360                 if (error == 0)
1361                         goto retry;
1362         }
1363         return(error);
1364 }
1365
1366 /*
1367  * Release a reference on an inode, flush as requested.
1368  *
1369  * On the last reference we queue the inode to the flusher for its final
1370  * disposition.
1371  */
1372 void
1373 hammer_rel_inode(struct hammer_inode *ip, int flush)
1374 {
1375         /*hammer_mount_t hmp = ip->hmp;*/
1376
1377         /*
1378          * Handle disposition when dropping the last ref.
1379          */
1380         for (;;) {
1381                 if (ip->lock.refs == 1) {
1382                         /*
1383                          * Determine whether on-disk action is needed for
1384                          * the inode's final disposition.
1385                          */
1386                         KKASSERT(ip->vp == NULL);
1387                         hammer_inode_unloadable_check(ip, 0);
1388                         if (ip->flags & HAMMER_INODE_MODMASK) {
1389                                 hammer_flush_inode(ip, 0);
1390                         } else if (ip->lock.refs == 1) {
1391                                 hammer_unload_inode(ip);
1392                                 break;
1393                         }
1394                 } else {
1395                         if (flush)
1396                                 hammer_flush_inode(ip, 0);
1397
1398                         /*
1399                          * The inode still has multiple refs, try to drop
1400                          * one ref.
1401                          */
1402                         KKASSERT(ip->lock.refs >= 1);
1403                         if (ip->lock.refs > 1) {
1404                                 hammer_unref(&ip->lock);
1405                                 break;
1406                         }
1407                 }
1408         }
1409 }
1410
1411 /*
1412  * Unload and destroy the specified inode.  Must be called with one remaining
1413  * reference.  The reference is disposed of.
1414  *
1415  * The inode must be completely clean.
1416  */
1417 static int
1418 hammer_unload_inode(struct hammer_inode *ip)
1419 {
1420         hammer_mount_t hmp = ip->hmp;
1421
1422         KASSERT(ip->lock.refs == 1,
1423                 ("hammer_unload_inode: %d refs\n", ip->lock.refs));
1424         KKASSERT(ip->vp == NULL);
1425         KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1426         KKASSERT(ip->cursor_ip_refs == 0);
1427         KKASSERT(ip->lock.lockcount == 0);
1428         KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1429
1430         KKASSERT(RB_EMPTY(&ip->rec_tree));
1431         KKASSERT(TAILQ_EMPTY(&ip->target_list));
1432
1433         RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1434
1435         hammer_free_inode(ip);
1436         return(0);
1437 }
1438
1439 /*
1440  * Called during unmounting if a critical error occured.  The in-memory
1441  * inode and all related structures are destroyed.
1442  *
1443  * If a critical error did not occur the unmount code calls the standard
1444  * release and asserts that the inode is gone.
1445  */
1446 int
1447 hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
1448 {
1449         hammer_record_t rec;
1450
1451         /*
1452          * Get rid of the inodes in-memory records, regardless of their
1453          * state, and clear the mod-mask.
1454          */
1455         while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1456                 TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1457                 rec->target_ip = NULL;
1458                 if (rec->flush_state == HAMMER_FST_SETUP)
1459                         rec->flush_state = HAMMER_FST_IDLE;
1460         }
1461         while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1462                 if (rec->flush_state == HAMMER_FST_FLUSH)
1463                         --rec->flush_group->refs;
1464                 else
1465                         hammer_ref(&rec->lock);
1466                 KKASSERT(rec->lock.refs == 1);
1467                 rec->flush_state = HAMMER_FST_IDLE;
1468                 rec->flush_group = NULL;
1469                 rec->flags |= HAMMER_RECF_DELETED_FE; /* wave hands */
1470                 rec->flags |= HAMMER_RECF_DELETED_BE; /* wave hands */
1471                 ++ip->rec_generation;
1472                 hammer_rel_mem_record(rec);
1473         }
1474         ip->flags &= ~HAMMER_INODE_MODMASK;
1475         ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1476         KKASSERT(ip->vp == NULL);
1477
1478         /*
1479          * Remove the inode from any flush group, force it idle.  FLUSH
1480          * and SETUP states have an inode ref.
1481          */
1482         switch(ip->flush_state) {
1483         case HAMMER_FST_FLUSH:
1484                 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
1485                 --ip->flush_group->refs;
1486                 ip->flush_group = NULL;
1487                 /* fall through */
1488         case HAMMER_FST_SETUP:
1489                 hammer_unref(&ip->lock);
1490                 ip->flush_state = HAMMER_FST_IDLE;
1491                 /* fall through */
1492         case HAMMER_FST_IDLE:
1493                 break;
1494         }
1495
1496         /*
1497          * There shouldn't be any associated vnode.  The unload needs at
1498          * least one ref, if we do have a vp steal its ip ref.
1499          */
1500         if (ip->vp) {
1501                 kprintf("hammer_destroy_inode_callback: Unexpected "
1502                         "vnode association ip %p vp %p\n", ip, ip->vp);
1503                 ip->vp->v_data = NULL;
1504                 ip->vp = NULL;
1505         } else {
1506                 hammer_ref(&ip->lock);
1507         }
1508         hammer_unload_inode(ip);
1509         return(0);
1510 }
1511
1512 /*
1513  * Called on mount -u when switching from RW to RO or vise-versa.  Adjust
1514  * the read-only flag for cached inodes.
1515  *
1516  * This routine is called from a RB_SCAN().
1517  */
1518 int
1519 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1520 {
1521         hammer_mount_t hmp = ip->hmp;
1522
1523         if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1524                 ip->flags |= HAMMER_INODE_RO;
1525         else
1526                 ip->flags &= ~HAMMER_INODE_RO;
1527         return(0);
1528 }
1529
1530 /*
1531  * A transaction has modified an inode, requiring updates as specified by
1532  * the passed flags.
1533  *
1534  * HAMMER_INODE_DDIRTY: Inode data has been updated
1535  * HAMMER_INODE_XDIRTY: Dirty in-memory records
1536  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
1537  * HAMMER_INODE_DELETED: Inode record/data must be deleted
1538  * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1539  */
1540 void
1541 hammer_modify_inode(hammer_inode_t ip, int flags)
1542 {
1543         /* 
1544          * ronly of 0 or 2 does not trigger assertion.
1545          * 2 is a special error state 
1546          */
1547         KKASSERT(ip->hmp->ronly != 1 ||
1548                   (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY | 
1549                             HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1550                             HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1551         if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1552                 ip->flags |= HAMMER_INODE_RSV_INODES;
1553                 ++ip->hmp->rsv_inodes;
1554         }
1555
1556         ip->flags |= flags;
1557 }
1558
1559 /*
1560  * Request that an inode be flushed.  This whole mess cannot block and may
1561  * recurse (if not synchronous).  Once requested HAMMER will attempt to
1562  * actively flush the inode until the flush can be done.
1563  *
1564  * The inode may already be flushing, or may be in a setup state.  We can
1565  * place the inode in a flushing state if it is currently idle and flag it
1566  * to reflush if it is currently flushing.
1567  *
1568  * Upon return if the inode could not be flushed due to a setup
1569  * dependancy, then it will be automatically flushed when the dependancy
1570  * is satisfied.
1571  */
1572 void
1573 hammer_flush_inode(hammer_inode_t ip, int flags)
1574 {
1575         hammer_mount_t hmp;
1576         hammer_flush_group_t flg;
1577         int good;
1578
1579         /*
1580          * next_flush_group is the first flush group we can place the inode
1581          * in.  It may be NULL.  If it becomes full we append a new flush
1582          * group and make that the next_flush_group.
1583          */
1584         hmp = ip->hmp;
1585         while ((flg = hmp->next_flush_group) != NULL) {
1586                 KKASSERT(flg->running == 0);
1587                 if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit)
1588                         break;
1589                 hmp->next_flush_group = TAILQ_NEXT(flg, flush_entry);
1590                 hammer_flusher_async(ip->hmp, flg);
1591         }
1592         if (flg == NULL) {
1593                 flg = kmalloc(sizeof(*flg), hmp->m_misc, M_WAITOK|M_ZERO);
1594                 hmp->next_flush_group = flg;
1595                 TAILQ_INIT(&flg->flush_list);
1596                 TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1597         }
1598
1599         /*
1600          * Trivial 'nothing to flush' case.  If the inode is in a SETUP
1601          * state we have to put it back into an IDLE state so we can
1602          * drop the extra ref.
1603          *
1604          * If we have a parent dependancy we must still fall through
1605          * so we can run it.
1606          */
1607         if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1608                 if (ip->flush_state == HAMMER_FST_SETUP &&
1609                     TAILQ_EMPTY(&ip->target_list)) {
1610                         ip->flush_state = HAMMER_FST_IDLE;
1611                         hammer_rel_inode(ip, 0);
1612                 }
1613                 if (ip->flush_state == HAMMER_FST_IDLE)
1614                         return;
1615         }
1616
1617         /*
1618          * Our flush action will depend on the current state.
1619          */
1620         switch(ip->flush_state) {
1621         case HAMMER_FST_IDLE:
1622                 /*
1623                  * We have no dependancies and can flush immediately.  Some
1624                  * our children may not be flushable so we have to re-test
1625                  * with that additional knowledge.
1626                  */
1627                 hammer_flush_inode_core(ip, flg, flags);
1628                 break;
1629         case HAMMER_FST_SETUP:
1630                 /*
1631                  * Recurse upwards through dependancies via target_list
1632                  * and start their flusher actions going if possible.
1633                  *
1634                  * 'good' is our connectivity.  -1 means we have none and
1635                  * can't flush, 0 means there weren't any dependancies, and
1636                  * 1 means we have good connectivity.
1637                  */
1638                 good = hammer_setup_parent_inodes(ip, 0, flg);
1639
1640                 if (good >= 0) {
1641                         /*
1642                          * We can continue if good >= 0.  Determine how 
1643                          * many records under our inode can be flushed (and
1644                          * mark them).
1645                          */
1646                         hammer_flush_inode_core(ip, flg, flags);
1647                 } else {
1648                         /*
1649                          * Parent has no connectivity, tell it to flush
1650                          * us as soon as it does.
1651                          *
1652                          * The REFLUSH flag is also needed to trigger
1653                          * dependancy wakeups.
1654                          */
1655                         ip->flags |= HAMMER_INODE_CONN_DOWN |
1656                                      HAMMER_INODE_REFLUSH;
1657                         if (flags & HAMMER_FLUSH_SIGNAL) {
1658                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1659                                 hammer_flusher_async(ip->hmp, flg);
1660                         }
1661                 }
1662                 break;
1663         case HAMMER_FST_FLUSH:
1664                 /*
1665                  * We are already flushing, flag the inode to reflush
1666                  * if needed after it completes its current flush.
1667                  *
1668                  * The REFLUSH flag is also needed to trigger
1669                  * dependancy wakeups.
1670                  */
1671                 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1672                         ip->flags |= HAMMER_INODE_REFLUSH;
1673                 if (flags & HAMMER_FLUSH_SIGNAL) {
1674                         ip->flags |= HAMMER_INODE_RESIGNAL;
1675                         hammer_flusher_async(ip->hmp, flg);
1676                 }
1677                 break;
1678         }
1679 }
1680
1681 /*
1682  * Scan ip->target_list, which is a list of records owned by PARENTS to our
1683  * ip which reference our ip.
1684  *
1685  * XXX This is a huge mess of recursive code, but not one bit of it blocks
1686  *     so for now do not ref/deref the structures.  Note that if we use the
1687  *     ref/rel code later, the rel CAN block.
1688  */
1689 static int
1690 hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
1691                            hammer_flush_group_t flg)
1692 {
1693         hammer_record_t depend;
1694         int good;
1695         int r;
1696
1697         /*
1698          * If we hit our recursion limit and we have parent dependencies
1699          * We cannot continue.  Returning < 0 will cause us to be flagged
1700          * for reflush.  Returning -2 cuts off additional dependency checks
1701          * because they are likely to also hit the depth limit.
1702          *
1703          * We cannot return < 0 if there are no dependencies or there might
1704          * not be anything to wakeup (ip).
1705          */
1706         if (depth == 20 && TAILQ_FIRST(&ip->target_list)) {
1707                 kprintf("HAMMER Warning: depth limit reached on "
1708                         "setup recursion, inode %p %016llx\n",
1709                         ip, (long long)ip->obj_id);
1710                 return(-2);
1711         }
1712
1713         /*
1714          * Scan dependencies
1715          */
1716         good = 0;
1717         TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1718                 r = hammer_setup_parent_inodes_helper(depend, depth, flg);
1719                 KKASSERT(depend->target_ip == ip);
1720                 if (r < 0 && good == 0)
1721                         good = -1;
1722                 if (r > 0)
1723                         good = 1;
1724
1725                 /*
1726                  * If we failed due to the recursion depth limit then stop
1727                  * now.
1728                  */
1729                 if (r == -2)
1730                         break;
1731         }
1732         return(good);
1733 }
1734
1735 /*
1736  * This helper function takes a record representing the dependancy between
1737  * the parent inode and child inode.
1738  *
1739  * record->ip           = parent inode
1740  * record->target_ip    = child inode
1741  * 
1742  * We are asked to recurse upwards and convert the record from SETUP
1743  * to FLUSH if possible.
1744  *
1745  * Return 1 if the record gives us connectivity
1746  *
1747  * Return 0 if the record is not relevant 
1748  *
1749  * Return -1 if we can't resolve the dependancy and there is no connectivity.
1750  */
1751 static int
1752 hammer_setup_parent_inodes_helper(hammer_record_t record, int depth,
1753                                   hammer_flush_group_t flg)
1754 {
1755         hammer_mount_t hmp;
1756         hammer_inode_t pip;
1757         int good;
1758
1759         KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1760         pip = record->ip;
1761         hmp = pip->hmp;
1762
1763         /*
1764          * If the record is already flushing, is it in our flush group?
1765          *
1766          * If it is in our flush group but it is a general record or a 
1767          * delete-on-disk, it does not improve our connectivity (return 0),
1768          * and if the target inode is not trying to destroy itself we can't
1769          * allow the operation yet anyway (the second return -1).
1770          */
1771         if (record->flush_state == HAMMER_FST_FLUSH) {
1772                 /*
1773                  * If not in our flush group ask the parent to reflush
1774                  * us as soon as possible.
1775                  */
1776                 if (record->flush_group != flg) {
1777                         pip->flags |= HAMMER_INODE_REFLUSH;
1778                         record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1779                         return(-1);
1780                 }
1781
1782                 /*
1783                  * If in our flush group everything is already set up,
1784                  * just return whether the record will improve our
1785                  * visibility or not.
1786                  */
1787                 if (record->type == HAMMER_MEM_RECORD_ADD)
1788                         return(1);
1789                 return(0);
1790         }
1791
1792         /*
1793          * It must be a setup record.  Try to resolve the setup dependancies
1794          * by recursing upwards so we can place ip on the flush list.
1795          *
1796          * Limit ourselves to 20 levels of recursion to avoid blowing out
1797          * the kernel stack.  If we hit the recursion limit we can't flush
1798          * until the parent flushes.  The parent will flush independantly
1799          * on its own and ultimately a deep recursion will be resolved.
1800          */
1801         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1802
1803         good = hammer_setup_parent_inodes(pip, depth + 1, flg);
1804
1805         /*
1806          * If good < 0 the parent has no connectivity and we cannot safely
1807          * flush the directory entry, which also means we can't flush our
1808          * ip.  Flag us for downward recursion once the parent's
1809          * connectivity is resolved.  Flag the parent for [re]flush or it
1810          * may not check for downward recursions.
1811          */
1812         if (good < 0) {
1813                 pip->flags |= HAMMER_INODE_REFLUSH;
1814                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1815                 return(good);
1816         }
1817
1818         /*
1819          * We are go, place the parent inode in a flushing state so we can
1820          * place its record in a flushing state.  Note that the parent
1821          * may already be flushing.  The record must be in the same flush
1822          * group as the parent.
1823          */
1824         if (pip->flush_state != HAMMER_FST_FLUSH)
1825                 hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
1826         KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
1827         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1828
1829 #if 0
1830         if (record->type == HAMMER_MEM_RECORD_DEL &&
1831             (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
1832                 /*
1833                  * Regardless of flushing state we cannot sync this path if the
1834                  * record represents a delete-on-disk but the target inode
1835                  * is not ready to sync its own deletion.
1836                  *
1837                  * XXX need to count effective nlinks to determine whether
1838                  * the flush is ok, otherwise removing a hardlink will
1839                  * just leave the DEL record to rot.
1840                  */
1841                 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
1842                 return(-1);
1843         } else
1844 #endif
1845         if (pip->flush_group == flg) {
1846                 /*
1847                  * Because we have not calculated nlinks yet we can just
1848                  * set records to the flush state if the parent is in
1849                  * the same flush group as we are.
1850                  */
1851                 record->flush_state = HAMMER_FST_FLUSH;
1852                 record->flush_group = flg;
1853                 ++record->flush_group->refs;
1854                 hammer_ref(&record->lock);
1855
1856                 /*
1857                  * A general directory-add contributes to our visibility.
1858                  *
1859                  * Otherwise it is probably a directory-delete or 
1860                  * delete-on-disk record and does not contribute to our
1861                  * visbility (but we can still flush it).
1862                  */
1863                 if (record->type == HAMMER_MEM_RECORD_ADD)
1864                         return(1);
1865                 return(0);
1866         } else {
1867                 /*
1868                  * If the parent is not in our flush group we cannot
1869                  * flush this record yet, there is no visibility.
1870                  * We tell the parent to reflush and mark ourselves
1871                  * so the parent knows it should flush us too.
1872                  */
1873                 pip->flags |= HAMMER_INODE_REFLUSH;
1874                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1875                 return(-1);
1876         }
1877 }
1878
1879 /*
1880  * This is the core routine placing an inode into the FST_FLUSH state.
1881  */
1882 static void
1883 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
1884 {
1885         int go_count;
1886
1887         /*
1888          * Set flush state and prevent the flusher from cycling into
1889          * the next flush group.  Do not place the ip on the list yet.
1890          * Inodes not in the idle state get an extra reference.
1891          */
1892         KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
1893         if (ip->flush_state == HAMMER_FST_IDLE)
1894                 hammer_ref(&ip->lock);
1895         ip->flush_state = HAMMER_FST_FLUSH;
1896         ip->flush_group = flg;
1897         ++ip->hmp->flusher.group_lock;
1898         ++ip->hmp->count_iqueued;
1899         ++hammer_count_iqueued;
1900         ++flg->total_count;
1901
1902         /*
1903          * If the flush group reaches the autoflush limit we want to signal
1904          * the flusher.  This is particularly important for remove()s.
1905          */
1906         if (flg->total_count == hammer_autoflush)
1907                 flags |= HAMMER_FLUSH_SIGNAL;
1908
1909         /*
1910          * We need to be able to vfsync/truncate from the backend.
1911          */
1912         KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
1913         if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
1914                 ip->flags |= HAMMER_INODE_VHELD;
1915                 vref(ip->vp);
1916         }
1917
1918         /*
1919          * Figure out how many in-memory records we can actually flush
1920          * (not including inode meta-data, buffers, etc).
1921          */
1922         KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
1923         if (flags & HAMMER_FLUSH_RECURSION) {
1924                 /*
1925                  * If this is a upwards recursion we do not want to
1926                  * recurse down again!
1927                  */
1928                 go_count = 1;
1929 #if 0
1930         } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
1931                 /*
1932                  * No new records are added if we must complete a flush
1933                  * from a previous cycle, but we do have to move the records
1934                  * from the previous cycle to the current one.
1935                  */
1936 #if 0
1937                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1938                                    hammer_syncgrp_child_callback, NULL);
1939 #endif
1940                 go_count = 1;
1941 #endif
1942         } else {
1943                 /*
1944                  * Normal flush, scan records and bring them into the flush.
1945                  * Directory adds and deletes are usually skipped (they are
1946                  * grouped with the related inode rather then with the
1947                  * directory).
1948                  *
1949                  * go_count can be negative, which means the scan aborted
1950                  * due to the flush group being over-full and we should
1951                  * flush what we have.
1952                  */
1953                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1954                                    hammer_setup_child_callback, NULL);
1955         }
1956
1957         /*
1958          * This is a more involved test that includes go_count.  If we
1959          * can't flush, flag the inode and return.  If go_count is 0 we
1960          * were are unable to flush any records in our rec_tree and
1961          * must ignore the XDIRTY flag.
1962          */
1963         if (go_count == 0) {
1964                 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
1965                         --ip->hmp->count_iqueued;
1966                         --hammer_count_iqueued;
1967
1968                         --flg->total_count;
1969                         ip->flush_state = HAMMER_FST_SETUP;
1970                         ip->flush_group = NULL;
1971                         if (ip->flags & HAMMER_INODE_VHELD) {
1972                                 ip->flags &= ~HAMMER_INODE_VHELD;
1973                                 vrele(ip->vp);
1974                         }
1975
1976                         /*
1977                          * REFLUSH is needed to trigger dependancy wakeups
1978                          * when an inode is in SETUP.
1979                          */
1980                         ip->flags |= HAMMER_INODE_REFLUSH;
1981                         if (flags & HAMMER_FLUSH_SIGNAL) {
1982                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1983                                 hammer_flusher_async(ip->hmp, flg);
1984                         }
1985                         if (--ip->hmp->flusher.group_lock == 0)
1986                                 wakeup(&ip->hmp->flusher.group_lock);
1987                         return;
1988                 }
1989         }
1990
1991         /*
1992          * Snapshot the state of the inode for the backend flusher.
1993          *
1994          * We continue to retain save_trunc_off even when all truncations
1995          * have been resolved as an optimization to determine if we can
1996          * skip the B-Tree lookup for overwrite deletions.
1997          *
1998          * NOTE: The DELETING flag is a mod flag, but it is also sticky,
1999          * and stays in ip->flags.  Once set, it stays set until the
2000          * inode is destroyed.
2001          */
2002         if (ip->flags & HAMMER_INODE_TRUNCATED) {
2003                 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
2004                 ip->sync_trunc_off = ip->trunc_off;
2005                 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
2006                 ip->flags &= ~HAMMER_INODE_TRUNCATED;
2007                 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
2008
2009                 /*
2010                  * The save_trunc_off used to cache whether the B-Tree
2011                  * holds any records past that point is not used until
2012                  * after the truncation has succeeded, so we can safely
2013                  * set it now.
2014                  */
2015                 if (ip->save_trunc_off > ip->sync_trunc_off)
2016                         ip->save_trunc_off = ip->sync_trunc_off;
2017         }
2018         ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
2019                            ~HAMMER_INODE_TRUNCATED);
2020         ip->sync_ino_leaf = ip->ino_leaf;
2021         ip->sync_ino_data = ip->ino_data;
2022         ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
2023 #ifdef DEBUG_TRUNCATE
2024         if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
2025                 kprintf("truncateS %016llx\n", ip->sync_trunc_off);
2026 #endif
2027
2028         /*
2029          * The flusher list inherits our inode and reference.
2030          */
2031         KKASSERT(flg->running == 0);
2032         TAILQ_INSERT_TAIL(&flg->flush_list, ip, flush_entry);
2033         if (--ip->hmp->flusher.group_lock == 0)
2034                 wakeup(&ip->hmp->flusher.group_lock);
2035
2036         if (flags & HAMMER_FLUSH_SIGNAL) {
2037                 hammer_flusher_async(ip->hmp, flg);
2038         }
2039 }
2040
2041 /*
2042  * Callback for scan of ip->rec_tree.  Try to include each record in our
2043  * flush.  ip->flush_group has been set but the inode has not yet been
2044  * moved into a flushing state.
2045  *
2046  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
2047  * both inodes.
2048  *
2049  * We return 1 for any record placed or found in FST_FLUSH, which prevents
2050  * the caller from shortcutting the flush.
2051  */
2052 static int
2053 hammer_setup_child_callback(hammer_record_t rec, void *data)
2054 {
2055         hammer_flush_group_t flg;
2056         hammer_inode_t target_ip;
2057         hammer_inode_t ip;
2058         int r;
2059
2060         /*
2061          * Records deleted or committed by the backend are ignored.
2062          * Note that the flush detects deleted frontend records at
2063          * multiple points to deal with races.  This is just the first
2064          * line of defense.  The only time HAMMER_RECF_DELETED_FE cannot
2065          * be set is when HAMMER_RECF_INTERLOCK_BE is set, because it
2066          * messes up link-count calculations.
2067          *
2068          * NOTE: Don't get confused between record deletion and, say,
2069          * directory entry deletion.  The deletion of a directory entry
2070          * which is on-media has nothing to do with the record deletion
2071          * flags.
2072          */
2073         if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
2074                           HAMMER_RECF_COMMITTED)) {
2075                 if (rec->flush_state == HAMMER_FST_FLUSH) {
2076                         KKASSERT(rec->flush_group == rec->ip->flush_group);
2077                         r = 1;
2078                 } else {
2079                         r = 0;
2080                 }
2081                 return(r);
2082         }
2083
2084         /*
2085          * If the record is in an idle state it has no dependancies and
2086          * can be flushed.
2087          */
2088         ip = rec->ip;
2089         flg = ip->flush_group;
2090         r = 0;
2091
2092         switch(rec->flush_state) {
2093         case HAMMER_FST_IDLE:
2094                 /*
2095                  * The record has no setup dependancy, we can flush it.
2096                  */
2097                 KKASSERT(rec->target_ip == NULL);
2098                 rec->flush_state = HAMMER_FST_FLUSH;
2099                 rec->flush_group = flg;
2100                 ++flg->refs;
2101                 hammer_ref(&rec->lock);
2102                 r = 1;
2103                 break;
2104         case HAMMER_FST_SETUP:
2105                 /*
2106                  * The record has a setup dependancy.  These are typically
2107                  * directory entry adds and deletes.  Such entries will be
2108                  * flushed when their inodes are flushed so we do not
2109                  * usually have to add them to the flush here.  However,
2110                  * if the target_ip has set HAMMER_INODE_CONN_DOWN then
2111                  * it is asking us to flush this record (and it).
2112                  */
2113                 target_ip = rec->target_ip;
2114                 KKASSERT(target_ip != NULL);
2115                 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
2116
2117                 /*
2118                  * If the target IP is already flushing in our group
2119                  * we could associate the record, but target_ip has
2120                  * already synced ino_data to sync_ino_data and we
2121                  * would also have to adjust nlinks.   Plus there are
2122                  * ordering issues for adds and deletes.
2123                  *
2124                  * Reflush downward if this is an ADD, and upward if
2125                  * this is a DEL.
2126                  */
2127                 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
2128                         if (rec->flush_state == HAMMER_MEM_RECORD_ADD)
2129                                 ip->flags |= HAMMER_INODE_REFLUSH;
2130                         else
2131                                 target_ip->flags |= HAMMER_INODE_REFLUSH;
2132                         break;
2133                 } 
2134
2135                 /*
2136                  * Target IP is not yet flushing.  This can get complex
2137                  * because we have to be careful about the recursion.
2138                  *
2139                  * Directories create an issue for us in that if a flush
2140                  * of a directory is requested the expectation is to flush
2141                  * any pending directory entries, but this will cause the
2142                  * related inodes to recursively flush as well.  We can't
2143                  * really defer the operation so just get as many as we
2144                  * can and
2145                  */
2146 #if 0
2147                 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
2148                     (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
2149                         /*
2150                          * We aren't reclaiming and the target ip was not
2151                          * previously prevented from flushing due to this
2152                          * record dependancy.  Do not flush this record.
2153                          */
2154                         /*r = 0;*/
2155                 } else
2156 #endif
2157                 if (flg->total_count + flg->refs >
2158                            ip->hmp->undo_rec_limit) {
2159                         /*
2160                          * Our flush group is over-full and we risk blowing
2161                          * out the UNDO FIFO.  Stop the scan, flush what we
2162                          * have, then reflush the directory.
2163                          *
2164                          * The directory may be forced through multiple
2165                          * flush groups before it can be completely
2166                          * flushed.
2167                          */
2168                         ip->flags |= HAMMER_INODE_RESIGNAL |
2169                                      HAMMER_INODE_REFLUSH;
2170                         r = -1;
2171                 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
2172                         /*
2173                          * If the target IP is not flushing we can force
2174                          * it to flush, even if it is unable to write out
2175                          * any of its own records we have at least one in
2176                          * hand that we CAN deal with.
2177                          */
2178                         rec->flush_state = HAMMER_FST_FLUSH;
2179                         rec->flush_group = flg;
2180                         ++flg->refs;
2181                         hammer_ref(&rec->lock);
2182                         hammer_flush_inode_core(target_ip, flg,
2183                                                 HAMMER_FLUSH_RECURSION);
2184                         r = 1;
2185                 } else {
2186                         /*
2187                          * General or delete-on-disk record.
2188                          *
2189                          * XXX this needs help.  If a delete-on-disk we could
2190                          * disconnect the target.  If the target has its own
2191                          * dependancies they really need to be flushed.
2192                          *
2193                          * XXX
2194                          */
2195                         rec->flush_state = HAMMER_FST_FLUSH;
2196                         rec->flush_group = flg;
2197                         ++flg->refs;
2198                         hammer_ref(&rec->lock);
2199                         hammer_flush_inode_core(target_ip, flg,
2200                                                 HAMMER_FLUSH_RECURSION);
2201                         r = 1;
2202                 }
2203                 break;
2204         case HAMMER_FST_FLUSH:
2205                 /* 
2206                  * The flush_group should already match.
2207                  */
2208                 KKASSERT(rec->flush_group == flg);
2209                 r = 1;
2210                 break;
2211         }
2212         return(r);
2213 }
2214
2215 #if 0
2216 /*
2217  * This version just moves records already in a flush state to the new
2218  * flush group and that is it.
2219  */
2220 static int
2221 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
2222 {
2223         hammer_inode_t ip = rec->ip;
2224
2225         switch(rec->flush_state) {
2226         case HAMMER_FST_FLUSH:
2227                 KKASSERT(rec->flush_group == ip->flush_group);
2228                 break;
2229         default:
2230                 break;
2231         }
2232         return(0);
2233 }
2234 #endif
2235
2236 /*
2237  * Wait for a previously queued flush to complete.
2238  *
2239  * If a critical error occured we don't try to wait.
2240  */
2241 void
2242 hammer_wait_inode(hammer_inode_t ip)
2243 {
2244         hammer_flush_group_t flg;
2245
2246         flg = NULL;
2247         if ((ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2248                 while (ip->flush_state != HAMMER_FST_IDLE &&
2249                        (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2250                         if (ip->flush_state == HAMMER_FST_SETUP)
2251                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2252                         if (ip->flush_state != HAMMER_FST_IDLE) {
2253                                 ip->flags |= HAMMER_INODE_FLUSHW;
2254                                 tsleep(&ip->flags, 0, "hmrwin", 0);
2255                         }
2256                 }
2257         }
2258 }
2259
2260 /*
2261  * Called by the backend code when a flush has been completed.
2262  * The inode has already been removed from the flush list.
2263  *
2264  * A pipelined flush can occur, in which case we must re-enter the
2265  * inode on the list and re-copy its fields.
2266  */
2267 void
2268 hammer_flush_inode_done(hammer_inode_t ip, int error)
2269 {
2270         hammer_mount_t hmp;
2271         int dorel;
2272
2273         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2274
2275         hmp = ip->hmp;
2276
2277         /*
2278          * Auto-reflush if the backend could not completely flush
2279          * the inode.  This fixes a case where a deferred buffer flush
2280          * could cause fsync to return early.
2281          */
2282         if (ip->sync_flags & HAMMER_INODE_MODMASK)
2283                 ip->flags |= HAMMER_INODE_REFLUSH;
2284
2285         /*
2286          * Merge left-over flags back into the frontend and fix the state.
2287          * Incomplete truncations are retained by the backend.
2288          */
2289         ip->error = error;
2290         ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2291         ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2292
2293         /*
2294          * The backend may have adjusted nlinks, so if the adjusted nlinks
2295          * does not match the fronttend set the frontend's RDIRTY flag again.
2296          */
2297         if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2298                 ip->flags |= HAMMER_INODE_DDIRTY;
2299
2300         /*
2301          * Fix up the dirty buffer status.
2302          */
2303         if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2304                 ip->flags |= HAMMER_INODE_BUFS;
2305         }
2306
2307         /*
2308          * Re-set the XDIRTY flag if some of the inode's in-memory records
2309          * could not be flushed.
2310          */
2311         KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2312                   (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2313                  (!RB_EMPTY(&ip->rec_tree) &&
2314                   (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2315
2316         /*
2317          * Do not lose track of inodes which no longer have vnode
2318          * assocations, otherwise they may never get flushed again.
2319          *
2320          * The reflush flag can be set superfluously, causing extra pain
2321          * for no reason.  If the inode is no longer modified it no longer
2322          * needs to be flushed.
2323          */
2324         if (ip->flags & HAMMER_INODE_MODMASK) {
2325                 if (ip->vp == NULL)
2326                         ip->flags |= HAMMER_INODE_REFLUSH;
2327         } else {
2328                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2329         }
2330
2331         /*
2332          * Adjust the flush state.
2333          */
2334         if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2335                 /*
2336                  * We were unable to flush out all our records, leave the
2337                  * inode in a flush state and in the current flush group.
2338                  * The flush group will be re-run.
2339                  *
2340                  * This occurs if the UNDO block gets too full or there is
2341                  * too much dirty meta-data and allows the flusher to
2342                  * finalize the UNDO block and then re-flush.
2343                  */
2344                 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2345                 dorel = 0;
2346         } else {
2347                 /*
2348                  * Remove from the flush_group
2349                  */
2350                 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
2351                 ip->flush_group = NULL;
2352
2353                 /*
2354                  * Clean up the vnode ref and tracking counts.
2355                  */
2356                 if (ip->flags & HAMMER_INODE_VHELD) {
2357                         ip->flags &= ~HAMMER_INODE_VHELD;
2358                         vrele(ip->vp);
2359                 }
2360                 --hmp->count_iqueued;
2361                 --hammer_count_iqueued;
2362
2363                 /*
2364                  * And adjust the state.
2365                  */
2366                 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2367                         ip->flush_state = HAMMER_FST_IDLE;
2368                         dorel = 1;
2369                 } else {
2370                         ip->flush_state = HAMMER_FST_SETUP;
2371                         dorel = 0;
2372                 }
2373
2374                 /*
2375                  * If the frontend is waiting for a flush to complete,
2376                  * wake it up.
2377                  */
2378                 if (ip->flags & HAMMER_INODE_FLUSHW) {
2379                         ip->flags &= ~HAMMER_INODE_FLUSHW;
2380                         wakeup(&ip->flags);
2381                 }
2382
2383                 /*
2384                  * If the frontend made more changes and requested another
2385                  * flush, then try to get it running.
2386                  *
2387                  * Reflushes are aborted when the inode is errored out.
2388                  */
2389                 if (ip->flags & HAMMER_INODE_REFLUSH) {
2390                         ip->flags &= ~HAMMER_INODE_REFLUSH;
2391                         if (ip->flags & HAMMER_INODE_RESIGNAL) {
2392                                 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2393                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2394                         } else {
2395                                 hammer_flush_inode(ip, 0);
2396                         }
2397                 }
2398         }
2399
2400         /*
2401          * If we have no parent dependancies we can clear CONN_DOWN
2402          */
2403         if (TAILQ_EMPTY(&ip->target_list))
2404                 ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2405
2406         /*
2407          * If the inode is now clean drop the space reservation.
2408          */
2409         if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2410             (ip->flags & HAMMER_INODE_RSV_INODES)) {
2411                 ip->flags &= ~HAMMER_INODE_RSV_INODES;
2412                 --hmp->rsv_inodes;
2413         }
2414
2415         if (dorel)
2416                 hammer_rel_inode(ip, 0);
2417 }
2418
2419 /*
2420  * Called from hammer_sync_inode() to synchronize in-memory records
2421  * to the media.
2422  */
2423 static int
2424 hammer_sync_record_callback(hammer_record_t record, void *data)
2425 {
2426         hammer_cursor_t cursor = data;
2427         hammer_transaction_t trans = cursor->trans;
2428         hammer_mount_t hmp = trans->hmp;
2429         int error;
2430
2431         /*
2432          * Skip records that do not belong to the current flush.
2433          */
2434         ++hammer_stats_record_iterations;
2435         if (record->flush_state != HAMMER_FST_FLUSH)
2436                 return(0);
2437
2438 #if 1
2439         if (record->flush_group != record->ip->flush_group) {
2440                 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2441                 Debugger("blah2");
2442                 return(0);
2443         }
2444 #endif
2445         KKASSERT(record->flush_group == record->ip->flush_group);
2446
2447         /*
2448          * Interlock the record using the BE flag.  Once BE is set the
2449          * frontend cannot change the state of FE.
2450          *
2451          * NOTE: If FE is set prior to us setting BE we still sync the
2452          * record out, but the flush completion code converts it to 
2453          * a delete-on-disk record instead of destroying it.
2454          */
2455         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2456         record->flags |= HAMMER_RECF_INTERLOCK_BE;
2457
2458         /*
2459          * The backend has already disposed of the record.
2460          */
2461         if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
2462                 error = 0;
2463                 goto done;
2464         }
2465
2466         /*
2467          * If the whole inode is being deleting all on-disk records will
2468          * be deleted very soon, we can't sync any new records to disk
2469          * because they will be deleted in the same transaction they were
2470          * created in (delete_tid == create_tid), which will assert.
2471          *
2472          * XXX There may be a case with RECORD_ADD with DELETED_FE set
2473          * that we currently panic on.
2474          */
2475         if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2476                 switch(record->type) {
2477                 case HAMMER_MEM_RECORD_DATA:
2478                         /*
2479                          * We don't have to do anything, if the record was
2480                          * committed the space will have been accounted for
2481                          * in the blockmap.
2482                          */
2483                         /* fall through */
2484                 case HAMMER_MEM_RECORD_GENERAL:
2485                         /*
2486                          * Set deleted-by-backend flag.  Do not set the
2487                          * backend committed flag, because we are throwing
2488                          * the record away.
2489                          */
2490                         record->flags |= HAMMER_RECF_DELETED_BE;
2491                         ++record->ip->rec_generation;
2492                         error = 0;
2493                         goto done;
2494                 case HAMMER_MEM_RECORD_ADD:
2495                         panic("hammer_sync_record_callback: illegal add "
2496                               "during inode deletion record %p", record);
2497                         break; /* NOT REACHED */
2498                 case HAMMER_MEM_RECORD_INODE:
2499                         panic("hammer_sync_record_callback: attempt to "
2500                               "sync inode record %p?", record);
2501                         break; /* NOT REACHED */
2502                 case HAMMER_MEM_RECORD_DEL:
2503                         /* 
2504                          * Follow through and issue the on-disk deletion
2505                          */
2506                         break;
2507                 }
2508         }
2509
2510         /*
2511          * If DELETED_FE is set special handling is needed for directory
2512          * entries.  Dependant pieces related to the directory entry may
2513          * have already been synced to disk.  If this occurs we have to
2514          * sync the directory entry and then change the in-memory record
2515          * from an ADD to a DELETE to cover the fact that it's been
2516          * deleted by the frontend.
2517          *
2518          * A directory delete covering record (MEM_RECORD_DEL) can never
2519          * be deleted by the frontend.
2520          *
2521          * Any other record type (aka DATA) can be deleted by the frontend.
2522          * XXX At the moment the flusher must skip it because there may
2523          * be another data record in the flush group for the same block,
2524          * meaning that some frontend data changes can leak into the backend's
2525          * synchronization point.
2526          */
2527         if (record->flags & HAMMER_RECF_DELETED_FE) {
2528                 if (record->type == HAMMER_MEM_RECORD_ADD) {
2529                         /*
2530                          * Convert a front-end deleted directory-add to
2531                          * a directory-delete entry later.
2532                          */
2533                         record->flags |= HAMMER_RECF_CONVERT_DELETE;
2534                 } else {
2535                         /*
2536                          * Dispose of the record (race case).  Mark as
2537                          * deleted by backend (and not committed).
2538                          */
2539                         KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2540                         record->flags |= HAMMER_RECF_DELETED_BE;
2541                         ++record->ip->rec_generation;
2542                         error = 0;
2543                         goto done;
2544                 }
2545         }
2546
2547         /*
2548          * Assign the create_tid for new records.  Deletions already
2549          * have the record's entire key properly set up.
2550          */
2551         if (record->type != HAMMER_MEM_RECORD_DEL) {
2552                 record->leaf.base.create_tid = trans->tid;
2553                 record->leaf.create_ts = trans->time32;
2554         }
2555         for (;;) {
2556                 error = hammer_ip_sync_record_cursor(cursor, record);
2557                 if (error != EDEADLK)
2558                         break;
2559                 hammer_done_cursor(cursor);
2560                 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2561                                            record->ip);
2562                 if (error)
2563                         break;
2564         }
2565         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2566
2567         if (error)
2568                 error = -error;
2569 done:
2570         hammer_flush_record_done(record, error);
2571
2572         /*
2573          * Do partial finalization if we have built up too many dirty
2574          * buffers.  Otherwise a buffer cache deadlock can occur when
2575          * doing things like creating tens of thousands of tiny files.
2576          *
2577          * We must release our cursor lock to avoid a 3-way deadlock
2578          * due to the exclusive sync lock the finalizer must get.
2579          */
2580         if (hammer_flusher_meta_limit(hmp)) {
2581                 hammer_unlock_cursor(cursor);
2582                 hammer_flusher_finalize(trans, 0);
2583                 hammer_lock_cursor(cursor);
2584         }
2585
2586         return(error);
2587 }
2588
2589 /*
2590  * Backend function called by the flusher to sync an inode to media.
2591  */
2592 int
2593 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2594 {
2595         struct hammer_cursor cursor;
2596         hammer_node_t tmp_node;
2597         hammer_record_t depend;
2598         hammer_record_t next;
2599         int error, tmp_error;
2600         u_int64_t nlinks;
2601
2602         if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2603                 return(0);
2604
2605         error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2606         if (error)
2607                 goto done;
2608
2609         /*
2610          * Any directory records referencing this inode which are not in
2611          * our current flush group must adjust our nlink count for the
2612          * purposes of synchronization to disk.
2613          *
2614          * Records which are in our flush group can be unlinked from our
2615          * inode now, potentially allowing the inode to be physically
2616          * deleted.
2617          *
2618          * This cannot block.
2619          */
2620         nlinks = ip->ino_data.nlinks;
2621         next = TAILQ_FIRST(&ip->target_list);
2622         while ((depend = next) != NULL) {
2623                 next = TAILQ_NEXT(depend, target_entry);
2624                 if (depend->flush_state == HAMMER_FST_FLUSH &&
2625                     depend->flush_group == ip->flush_group) {
2626                         /*
2627                          * If this is an ADD that was deleted by the frontend
2628                          * the frontend nlinks count will have already been
2629                          * decremented, but the backend is going to sync its
2630                          * directory entry and must account for it.  The
2631                          * record will be converted to a delete-on-disk when
2632                          * it gets synced.
2633                          *
2634                          * If the ADD was not deleted by the frontend we
2635                          * can remove the dependancy from our target_list.
2636                          */
2637                         if (depend->flags & HAMMER_RECF_DELETED_FE) {
2638                                 ++nlinks;
2639                         } else {
2640                                 TAILQ_REMOVE(&ip->target_list, depend,
2641                                              target_entry);
2642                                 depend->target_ip = NULL;
2643                         }
2644                 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2645                         /*
2646                          * Not part of our flush group and not deleted by
2647                          * the front-end, adjust the link count synced to
2648                          * the media (undo what the frontend did when it
2649                          * queued the record).
2650                          */
2651                         KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2652                         switch(depend->type) {
2653                         case HAMMER_MEM_RECORD_ADD:
2654                                 --nlinks;
2655                                 break;
2656                         case HAMMER_MEM_RECORD_DEL:
2657                                 ++nlinks;
2658                                 break;
2659                         default:
2660                                 break;
2661                         }
2662                 }
2663         }
2664
2665         /*
2666          * Set dirty if we had to modify the link count.
2667          */
2668         if (ip->sync_ino_data.nlinks != nlinks) {
2669                 KKASSERT((int64_t)nlinks >= 0);
2670                 ip->sync_ino_data.nlinks = nlinks;
2671                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2672         }
2673
2674         /*
2675          * If there is a trunction queued destroy any data past the (aligned)
2676          * truncation point.  Userland will have dealt with the buffer
2677          * containing the truncation point for us.
2678          *
2679          * We don't flush pending frontend data buffers until after we've
2680          * dealt with the truncation.
2681          */
2682         if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2683                 /*
2684                  * Interlock trunc_off.  The VOP front-end may continue to
2685                  * make adjustments to it while we are blocked.
2686                  */
2687                 off_t trunc_off;
2688                 off_t aligned_trunc_off;
2689                 int blkmask;
2690
2691                 trunc_off = ip->sync_trunc_off;
2692                 blkmask = hammer_blocksize(trunc_off) - 1;
2693                 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
2694
2695                 /*
2696                  * Delete any whole blocks on-media.  The front-end has
2697                  * already cleaned out any partial block and made it
2698                  * pending.  The front-end may have updated trunc_off
2699                  * while we were blocked so we only use sync_trunc_off.
2700                  *
2701                  * This operation can blow out the buffer cache, EWOULDBLOCK
2702                  * means we were unable to complete the deletion.  The
2703                  * deletion will update sync_trunc_off in that case.
2704                  */
2705                 error = hammer_ip_delete_range(&cursor, ip,
2706                                                 aligned_trunc_off,
2707                                                 0x7FFFFFFFFFFFFFFFLL, 2);
2708                 if (error == EWOULDBLOCK) {
2709                         ip->flags |= HAMMER_INODE_WOULDBLOCK;
2710                         error = 0;
2711                         goto defer_buffer_flush;
2712                 }
2713
2714                 if (error)
2715                         goto done;
2716
2717                 /*
2718                  * Clear the truncation flag on the backend after we have
2719                  * complete the deletions.  Backend data is now good again
2720                  * (including new records we are about to sync, below).
2721                  *
2722                  * Leave sync_trunc_off intact.  As we write additional
2723                  * records the backend will update sync_trunc_off.  This
2724                  * tells the backend whether it can skip the overwrite
2725                  * test.  This should work properly even when the backend
2726                  * writes full blocks where the truncation point straddles
2727                  * the block because the comparison is against the base
2728                  * offset of the record.
2729                  */
2730                 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2731                 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2732         } else {
2733                 error = 0;
2734         }
2735
2736         /*
2737          * Now sync related records.  These will typically be directory
2738          * entries, records tracking direct-writes, or delete-on-disk records.
2739          */
2740         if (error == 0) {
2741                 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2742                                     hammer_sync_record_callback, &cursor);
2743                 if (tmp_error < 0)
2744                         tmp_error = -error;
2745                 if (tmp_error)
2746                         error = tmp_error;
2747         }
2748         hammer_cache_node(&ip->cache[1], cursor.node);
2749
2750         /*
2751          * Re-seek for inode update, assuming our cache hasn't been ripped
2752          * out from under us.
2753          */
2754         if (error == 0) {
2755                 tmp_node = hammer_ref_node_safe(trans, &ip->cache[0], &error);
2756                 if (tmp_node) {
2757                         hammer_cursor_downgrade(&cursor);
2758                         hammer_lock_sh(&tmp_node->lock);
2759                         if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
2760                                 hammer_cursor_seek(&cursor, tmp_node, 0);
2761                         hammer_unlock(&tmp_node->lock);
2762                         hammer_rel_node(tmp_node);
2763                 }
2764                 error = 0;
2765         }
2766
2767         /*
2768          * If we are deleting the inode the frontend had better not have
2769          * any active references on elements making up the inode.
2770          *
2771          * The call to hammer_ip_delete_clean() cleans up auxillary records
2772          * but not DB or DATA records.  Those must have already been deleted
2773          * by the normal truncation mechanic.
2774          */
2775         if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
2776                 RB_EMPTY(&ip->rec_tree)  &&
2777             (ip->sync_flags & HAMMER_INODE_DELETING) &&
2778             (ip->flags & HAMMER_INODE_DELETED) == 0) {
2779                 int count1 = 0;
2780
2781                 error = hammer_ip_delete_clean(&cursor, ip, &count1);
2782                 if (error == 0) {
2783                         ip->flags |= HAMMER_INODE_DELETED;
2784                         ip->sync_flags &= ~HAMMER_INODE_DELETING;
2785                         ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2786                         KKASSERT(RB_EMPTY(&ip->rec_tree));
2787
2788                         /*
2789                          * Set delete_tid in both the frontend and backend
2790                          * copy of the inode record.  The DELETED flag handles
2791                          * this, do not set RDIRTY.
2792                          */
2793                         ip->ino_leaf.base.delete_tid = trans->tid;
2794                         ip->sync_ino_leaf.base.delete_tid = trans->tid;
2795                         ip->ino_leaf.delete_ts = trans->time32;
2796                         ip->sync_ino_leaf.delete_ts = trans->time32;
2797
2798
2799                         /*
2800                          * Adjust the inode count in the volume header
2801                          */
2802                         hammer_sync_lock_sh(trans);
2803                         if (ip->flags & HAMMER_INODE_ONDISK) {
2804                                 hammer_modify_volume_field(trans,
2805                                                            trans->rootvol,
2806                                                            vol0_stat_inodes);
2807                                 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
2808                                 hammer_modify_volume_done(trans->rootvol);
2809                         }
2810                         hammer_sync_unlock(trans);
2811                 }
2812         }
2813
2814         if (error)
2815                 goto done;
2816         ip->sync_flags &= ~HAMMER_INODE_BUFS;
2817
2818 defer_buffer_flush:
2819         /*
2820          * Now update the inode's on-disk inode-data and/or on-disk record.
2821          * DELETED and ONDISK are managed only in ip->flags.
2822          *
2823          * In the case of a defered buffer flush we still update the on-disk
2824          * inode to satisfy visibility requirements if there happen to be
2825          * directory dependancies.
2826          */
2827         switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
2828         case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
2829                 /*
2830                  * If deleted and on-disk, don't set any additional flags.
2831                  * the delete flag takes care of things.
2832                  *
2833                  * Clear flags which may have been set by the frontend.
2834                  */
2835                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2836                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2837                                     HAMMER_INODE_DELETING);
2838                 break;
2839         case HAMMER_INODE_DELETED:
2840                 /*
2841                  * Take care of the case where a deleted inode was never
2842                  * flushed to the disk in the first place.
2843                  *
2844                  * Clear flags which may have been set by the frontend.
2845                  */
2846                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2847                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2848                                     HAMMER_INODE_DELETING);
2849                 while (RB_ROOT(&ip->rec_tree)) {
2850                         hammer_record_t record = RB_ROOT(&ip->rec_tree);
2851                         hammer_ref(&record->lock);
2852                         KKASSERT(record->lock.refs == 1);
2853                         record->flags |= HAMMER_RECF_DELETED_BE;
2854                         ++record->ip->rec_generation;
2855                         hammer_rel_mem_record(record);
2856                 }
2857                 break;
2858         case HAMMER_INODE_ONDISK:
2859                 /*
2860                  * If already on-disk, do not set any additional flags.
2861                  */
2862                 break;
2863         default:
2864                 /*
2865                  * If not on-disk and not deleted, set DDIRTY to force
2866                  * an initial record to be written.
2867                  *
2868                  * Also set the create_tid in both the frontend and backend
2869                  * copy of the inode record.
2870                  */
2871                 ip->ino_leaf.base.create_tid = trans->tid;
2872                 ip->ino_leaf.create_ts = trans->time32;
2873                 ip->sync_ino_leaf.base.create_tid = trans->tid;
2874                 ip->sync_ino_leaf.create_ts = trans->time32;
2875                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2876                 break;
2877         }
2878
2879         /*
2880          * If RDIRTY or DDIRTY is set, write out a new record.  If the inode
2881          * is already on-disk the old record is marked as deleted.
2882          *
2883          * If DELETED is set hammer_update_inode() will delete the existing
2884          * record without writing out a new one.
2885          *
2886          * If *ONLY* the ITIMES flag is set we can update the record in-place.
2887          */
2888         if (ip->flags & HAMMER_INODE_DELETED) {
2889                 error = hammer_update_inode(&cursor, ip);
2890         } else 
2891         if ((ip->sync_flags & HAMMER_INODE_DDIRTY) == 0 &&
2892             (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
2893                 error = hammer_update_itimes(&cursor, ip);
2894         } else
2895         if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
2896                 error = hammer_update_inode(&cursor, ip);
2897         }
2898 done:
2899         if (error) {
2900                 hammer_critical_error(ip->hmp, ip, error,
2901                                       "while syncing inode");
2902         }
2903         hammer_done_cursor(&cursor);
2904         return(error);
2905 }
2906
2907 /*
2908  * This routine is called when the OS is no longer actively referencing
2909  * the inode (but might still be keeping it cached), or when releasing
2910  * the last reference to an inode.
2911  *
2912  * At this point if the inode's nlinks count is zero we want to destroy
2913  * it, which may mean destroying it on-media too.
2914  */
2915 void
2916 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
2917 {
2918         struct vnode *vp;
2919
2920         /*
2921          * Set the DELETING flag when the link count drops to 0 and the
2922          * OS no longer has any opens on the inode.
2923          *
2924          * The backend will clear DELETING (a mod flag) and set DELETED
2925          * (a state flag) when it is actually able to perform the
2926          * operation.
2927          *
2928          * Don't reflag the deletion if the flusher is currently syncing
2929          * one that was already flagged.  A previously set DELETING flag
2930          * may bounce around flags and sync_flags until the operation is
2931          * completely done.
2932          */
2933         if (ip->ino_data.nlinks == 0 &&
2934             ((ip->flags | ip->sync_flags) & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
2935                 ip->flags |= HAMMER_INODE_DELETING;
2936                 ip->flags |= HAMMER_INODE_TRUNCATED;
2937                 ip->trunc_off = 0;
2938                 vp = NULL;
2939                 if (getvp) {
2940                         if (hammer_get_vnode(ip, &vp) != 0)
2941                                 return;
2942                 }
2943
2944                 /*
2945                  * Final cleanup
2946                  */
2947                 if (ip->vp) {
2948                         vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
2949                         vnode_pager_setsize(ip->vp, 0);
2950                 }
2951                 if (getvp) {
2952                         vput(vp);
2953                 }
2954         }
2955 }
2956
2957 /*
2958  * After potentially resolving a dependancy the inode is tested
2959  * to determine whether it needs to be reflushed.
2960  */
2961 void
2962 hammer_test_inode(hammer_inode_t ip)
2963 {
2964         if (ip->flags & HAMMER_INODE_REFLUSH) {
2965                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2966                 hammer_ref(&ip->lock);
2967                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2968                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
2969                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2970                 } else {
2971                         hammer_flush_inode(ip, 0);
2972                 }
2973                 hammer_rel_inode(ip, 0);
2974         }
2975 }
2976
2977 /*
2978  * Clear the RECLAIM flag on an inode.  This occurs when the inode is
2979  * reassociated with a vp or just before it gets freed.
2980  *
2981  * Pipeline wakeups to threads blocked due to an excessive number of
2982  * detached inodes.  The reclaim count generates a bit of negative
2983  * feedback.
2984  */
2985 static void
2986 hammer_inode_wakereclaims(hammer_inode_t ip, int dowake)
2987 {
2988         struct hammer_reclaim *reclaim;
2989         hammer_mount_t hmp = ip->hmp;
2990
2991         if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
2992                 return;
2993
2994         --hammer_count_reclaiming;
2995         --hmp->inode_reclaims;
2996         ip->flags &= ~HAMMER_INODE_RECLAIM;
2997
2998         if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT || dowake) {
2999                 reclaim = TAILQ_FIRST(&hmp->reclaim_list);
3000                 if (reclaim && reclaim->count > 0 && --reclaim->count == 0) {
3001                         TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
3002                         wakeup(reclaim);
3003                 }
3004         }
3005 }
3006
3007 /*
3008  * Setup our reclaim pipeline.  We only let so many detached (and dirty)
3009  * inodes build up before we start blocking.
3010  *
3011  * When we block we don't care *which* inode has finished reclaiming,
3012  * as lone as one does.  This is somewhat heuristical... we also put a
3013  * cap on how long we are willing to wait.
3014  */
3015 void
3016 hammer_inode_waitreclaims(hammer_mount_t hmp)
3017 {
3018         struct hammer_reclaim reclaim;
3019         int delay;
3020
3021         if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT)
3022                 return;
3023         delay = (hmp->inode_reclaims - HAMMER_RECLAIM_WAIT) * hz /
3024                 (HAMMER_RECLAIM_WAIT * 3) + 1;
3025         if (delay > 0) {
3026                 reclaim.count = 2;
3027                 TAILQ_INSERT_TAIL(&hmp->reclaim_list, &reclaim, entry);
3028                 tsleep(&reclaim, 0, "hmrrcm", delay);
3029                 if (reclaim.count > 0)
3030                         TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
3031         }
3032 }
3033
3034 /*
3035  * A larger then normal backlog of inodes is sitting in the flusher,
3036  * enforce a general slowdown to let it catch up.  This routine is only
3037  * called on completion of a non-flusher-related transaction which
3038  * performed B-Tree node I/O.
3039  *
3040  * It is possible for the flusher to stall in a continuous load.
3041  * blogbench -i1000 -o seems to do a good job generating this sort of load.
3042  * If the flusher is unable to catch up the inode count can bloat until
3043  * we run out of kvm.
3044  *
3045  * This is a bit of a hack.
3046  */
3047 void
3048 hammer_inode_waithard(hammer_mount_t hmp)
3049 {
3050         /*
3051          * Hysteresis.
3052          */
3053         if (hmp->flags & HAMMER_MOUNT_FLUSH_RECOVERY) {
3054                 if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT / 2 &&
3055                     hmp->count_iqueued < hmp->count_inodes / 20) {
3056                         hmp->flags &= ~HAMMER_MOUNT_FLUSH_RECOVERY;
3057                         return;
3058                 }
3059         } else {
3060                 if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT ||
3061                     hmp->count_iqueued < hmp->count_inodes / 10) {
3062                         return;
3063                 }
3064                 hmp->flags |= HAMMER_MOUNT_FLUSH_RECOVERY;
3065         }
3066
3067         /*
3068          * Block for one flush cycle.
3069          */
3070         hammer_flusher_wait_next(hmp);
3071 }
3072