Fix a simple code bug when dealing with a raced vnode in hammer_get_vnode().
[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_inode_info iinfo;
355         struct hammer_cursor cursor;
356         struct hammer_inode *ip;
357
358
359         /*
360          * Determine if we already have an inode cached.  If we do then
361          * we are golden.
362          *
363          * If we find an inode with no vnode we have to mark the
364          * transaction such that hammer_inode_waitreclaims() is
365          * called later on to avoid building up an infinite number
366          * of inodes.  Otherwise we can continue to * add new inodes
367          * faster then they can be disposed of, even with the tsleep
368          * delay.
369          *
370          * If we find a dummy inode we return a failure so dounlink
371          * (which does another lookup) doesn't try to mess with the
372          * link count.  hammer_vop_nresolve() uses hammer_get_dummy_inode()
373          * to ref dummy inodes.
374          */
375         iinfo.obj_id = obj_id;
376         iinfo.obj_asof = asof;
377         iinfo.obj_localization = localization;
378 loop:
379         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
380         if (ip) {
381                 if (ip->flags & HAMMER_INODE_DUMMY) {
382                         *errorp = ENOENT;
383                         return(NULL);
384                 }
385                 hammer_ref(&ip->lock);
386                 *errorp = 0;
387                 return(ip);
388         }
389
390         /*
391          * Allocate a new inode structure and deal with races later.
392          */
393         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
394         ++hammer_count_inodes;
395         ++hmp->count_inodes;
396         ip->obj_id = obj_id;
397         ip->obj_asof = iinfo.obj_asof;
398         ip->obj_localization = localization;
399         ip->hmp = hmp;
400         ip->flags = flags & HAMMER_INODE_RO;
401         ip->cache[0].ip = ip;
402         ip->cache[1].ip = ip;
403         if (hmp->ronly)
404                 ip->flags |= HAMMER_INODE_RO;
405         ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
406                 0x7FFFFFFFFFFFFFFFLL;
407         RB_INIT(&ip->rec_tree);
408         TAILQ_INIT(&ip->target_list);
409         hammer_ref(&ip->lock);
410
411         /*
412          * Locate the on-disk inode.  If this is a PFS root we always
413          * access the current version of the root inode and (if it is not
414          * a master) always access information under it with a snapshot
415          * TID.
416          */
417 retry:
418         hammer_init_cursor(trans, &cursor, (dip ? &dip->cache[0] : NULL), NULL);
419         cursor.key_beg.localization = localization + HAMMER_LOCALIZE_INODE;
420         cursor.key_beg.obj_id = ip->obj_id;
421         cursor.key_beg.key = 0;
422         cursor.key_beg.create_tid = 0;
423         cursor.key_beg.delete_tid = 0;
424         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
425         cursor.key_beg.obj_type = 0;
426
427         cursor.asof = iinfo.obj_asof;
428         cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
429                        HAMMER_CURSOR_ASOF;
430
431         *errorp = hammer_btree_lookup(&cursor);
432         if (*errorp == EDEADLK) {
433                 hammer_done_cursor(&cursor);
434                 goto retry;
435         }
436
437         /*
438          * On success the B-Tree lookup will hold the appropriate
439          * buffer cache buffers and provide a pointer to the requested
440          * information.  Copy the information to the in-memory inode
441          * and cache the B-Tree node to improve future operations.
442          */
443         if (*errorp == 0) {
444                 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
445                 ip->ino_data = cursor.data->inode;
446
447                 /*
448                  * cache[0] tries to cache the location of the object inode.
449                  * The assumption is that it is near the directory inode.
450                  *
451                  * cache[1] tries to cache the location of the object data.
452                  * The assumption is that it is near the directory data.
453                  */
454                 hammer_cache_node(&ip->cache[0], cursor.node);
455                 if (dip && dip->cache[1].node)
456                         hammer_cache_node(&ip->cache[1], dip->cache[1].node);
457
458                 /*
459                  * The file should not contain any data past the file size
460                  * stored in the inode.  Setting save_trunc_off to the
461                  * file size instead of max reduces B-Tree lookup overheads
462                  * on append by allowing the flusher to avoid checking for
463                  * record overwrites.
464                  */
465                 ip->save_trunc_off = ip->ino_data.size;
466
467                 /*
468                  * Locate and assign the pseudofs management structure to
469                  * the inode.
470                  */
471                 if (dip && dip->obj_localization == ip->obj_localization) {
472                         ip->pfsm = dip->pfsm;
473                         hammer_ref(&ip->pfsm->lock);
474                 } else {
475                         ip->pfsm = hammer_load_pseudofs(trans,
476                                                         ip->obj_localization,
477                                                         errorp);
478                         *errorp = 0;    /* ignore ENOENT */
479                 }
480         }
481
482         /*
483          * The inode is placed on the red-black tree and will be synced to
484          * the media when flushed or by the filesystem sync.  If this races
485          * another instantiation/lookup the insertion will fail.
486          */
487         if (*errorp == 0) {
488                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
489                         hammer_free_inode(ip);
490                         hammer_done_cursor(&cursor);
491                         goto loop;
492                 }
493                 ip->flags |= HAMMER_INODE_ONDISK;
494         } else {
495                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
496                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
497                         --hmp->rsv_inodes;
498                 }
499
500                 hammer_free_inode(ip);
501                 ip = NULL;
502         }
503         hammer_done_cursor(&cursor);
504         trans->flags |= HAMMER_TRANSF_NEWINODE;
505         return (ip);
506 }
507
508 /*
509  * Get a dummy inode to placemark a broken directory entry.
510  */
511 struct hammer_inode *
512 hammer_get_dummy_inode(hammer_transaction_t trans, hammer_inode_t dip,
513                  int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
514                  int flags, int *errorp)
515 {
516         hammer_mount_t hmp = trans->hmp;
517         struct hammer_inode_info iinfo;
518         struct hammer_inode *ip;
519
520         /*
521          * Determine if we already have an inode cached.  If we do then
522          * we are golden.
523          *
524          * If we find an inode with no vnode we have to mark the
525          * transaction such that hammer_inode_waitreclaims() is
526          * called later on to avoid building up an infinite number
527          * of inodes.  Otherwise we can continue to * add new inodes
528          * faster then they can be disposed of, even with the tsleep
529          * delay.
530          *
531          * If we find a non-fake inode we return an error.  Only fake
532          * inodes can be returned by this routine.
533          */
534         iinfo.obj_id = obj_id;
535         iinfo.obj_asof = asof;
536         iinfo.obj_localization = localization;
537 loop:
538         *errorp = 0;
539         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
540         if (ip) {
541                 if ((ip->flags & HAMMER_INODE_DUMMY) == 0) {
542                         *errorp = ENOENT;
543                         return(NULL);
544                 }
545                 hammer_ref(&ip->lock);
546                 return(ip);
547         }
548
549         /*
550          * Allocate a new inode structure and deal with races later.
551          */
552         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
553         ++hammer_count_inodes;
554         ++hmp->count_inodes;
555         ip->obj_id = obj_id;
556         ip->obj_asof = iinfo.obj_asof;
557         ip->obj_localization = localization;
558         ip->hmp = hmp;
559         ip->flags = flags | HAMMER_INODE_RO | HAMMER_INODE_DUMMY;
560         ip->cache[0].ip = ip;
561         ip->cache[1].ip = ip;
562         ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
563                 0x7FFFFFFFFFFFFFFFLL;
564         RB_INIT(&ip->rec_tree);
565         TAILQ_INIT(&ip->target_list);
566         hammer_ref(&ip->lock);
567
568         /*
569          * Populate the dummy inode.  Leave everything zero'd out.
570          *
571          * (ip->ino_leaf and ip->ino_data)
572          *
573          * Make the dummy inode a FIFO object which most copy programs
574          * will properly ignore.
575          */
576         ip->save_trunc_off = ip->ino_data.size;
577         ip->ino_data.obj_type = HAMMER_OBJTYPE_FIFO;
578
579         /*
580          * Locate and assign the pseudofs management structure to
581          * the inode.
582          */
583         if (dip && dip->obj_localization == ip->obj_localization) {
584                 ip->pfsm = dip->pfsm;
585                 hammer_ref(&ip->pfsm->lock);
586         } else {
587                 ip->pfsm = hammer_load_pseudofs(trans, ip->obj_localization,
588                                                 errorp);
589                 *errorp = 0;    /* ignore ENOENT */
590         }
591
592         /*
593          * The inode is placed on the red-black tree and will be synced to
594          * the media when flushed or by the filesystem sync.  If this races
595          * another instantiation/lookup the insertion will fail.
596          *
597          * NOTE: Do not set HAMMER_INODE_ONDISK.  The inode is a fake.
598          */
599         if (*errorp == 0) {
600                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
601                         hammer_free_inode(ip);
602                         goto loop;
603                 }
604         } else {
605                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
606                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
607                         --hmp->rsv_inodes;
608                 }
609                 hammer_free_inode(ip);
610                 ip = NULL;
611         }
612         trans->flags |= HAMMER_TRANSF_NEWINODE;
613         return (ip);
614 }
615
616 /*
617  * Create a new filesystem object, returning the inode in *ipp.  The
618  * returned inode will be referenced.  The inode is created in-memory.
619  *
620  * If pfsm is non-NULL the caller wishes to create the root inode for
621  * a master PFS.
622  */
623 int
624 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
625                     struct ucred *cred, hammer_inode_t dip,
626                     hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
627 {
628         hammer_mount_t hmp;
629         hammer_inode_t ip;
630         uid_t xuid;
631         int error;
632
633         hmp = trans->hmp;
634
635         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
636         ++hammer_count_inodes;
637         ++hmp->count_inodes;
638         trans->flags |= HAMMER_TRANSF_NEWINODE;
639
640         if (pfsm) {
641                 KKASSERT(pfsm->localization != 0);
642                 ip->obj_id = HAMMER_OBJID_ROOT;
643                 ip->obj_localization = pfsm->localization;
644         } else {
645                 KKASSERT(dip != NULL);
646                 ip->obj_id = hammer_alloc_objid(hmp, dip);
647                 ip->obj_localization = dip->obj_localization;
648         }
649
650         KKASSERT(ip->obj_id != 0);
651         ip->obj_asof = hmp->asof;
652         ip->hmp = hmp;
653         ip->flush_state = HAMMER_FST_IDLE;
654         ip->flags = HAMMER_INODE_DDIRTY |
655                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
656         ip->cache[0].ip = ip;
657         ip->cache[1].ip = ip;
658
659         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
660         /* ip->save_trunc_off = 0; (already zero) */
661         RB_INIT(&ip->rec_tree);
662         TAILQ_INIT(&ip->target_list);
663
664         ip->ino_data.atime = trans->time;
665         ip->ino_data.mtime = trans->time;
666         ip->ino_data.size = 0;
667         ip->ino_data.nlinks = 0;
668
669         /*
670          * A nohistory designator on the parent directory is inherited by
671          * the child.  We will do this even for pseudo-fs creation... the
672          * sysad can turn it off.
673          */
674         if (dip) {
675                 ip->ino_data.uflags = dip->ino_data.uflags &
676                                       (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
677         }
678
679         ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
680         ip->ino_leaf.base.localization = ip->obj_localization +
681                                          HAMMER_LOCALIZE_INODE;
682         ip->ino_leaf.base.obj_id = ip->obj_id;
683         ip->ino_leaf.base.key = 0;
684         ip->ino_leaf.base.create_tid = 0;
685         ip->ino_leaf.base.delete_tid = 0;
686         ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
687         ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
688
689         ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
690         ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
691         ip->ino_data.mode = vap->va_mode;
692         ip->ino_data.ctime = trans->time;
693
694         /*
695          * If we are running version 2 or greater we use dirhash algorithm #1
696          * which is semi-sorted.  Algorithm #0 was just a pure crc.
697          */
698         if (trans->hmp->version >= HAMMER_VOL_VERSION_TWO) {
699                 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
700                         ip->ino_data.cap_flags |= HAMMER_INODE_CAP_DIRHASH_ALG1;
701                 }
702         }
703
704         /*
705          * Setup the ".." pointer.  This only needs to be done for directories
706          * but we do it for all objects as a recovery aid.
707          */
708         if (dip)
709                 ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
710 #if 0
711         /*
712          * The parent_obj_localization field only applies to pseudo-fs roots.
713          * XXX this is no longer applicable, PFSs are no longer directly
714          * tied into the parent's directory structure.
715          */
716         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY &&
717             ip->obj_id == HAMMER_OBJID_ROOT) {
718                 ip->ino_data.ext.obj.parent_obj_localization = 
719                                                 dip->obj_localization;
720         }
721 #endif
722
723         switch(ip->ino_leaf.base.obj_type) {
724         case HAMMER_OBJTYPE_CDEV:
725         case HAMMER_OBJTYPE_BDEV:
726                 ip->ino_data.rmajor = vap->va_rmajor;
727                 ip->ino_data.rminor = vap->va_rminor;
728                 break;
729         default:
730                 break;
731         }
732
733         /*
734          * Calculate default uid/gid and overwrite with information from
735          * the vap.
736          */
737         if (dip) {
738                 xuid = hammer_to_unix_xid(&dip->ino_data.uid);
739                 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
740                                              xuid, cred, &vap->va_mode);
741         } else {
742                 xuid = 0;
743         }
744         ip->ino_data.mode = vap->va_mode;
745
746         if (vap->va_vaflags & VA_UID_UUID_VALID)
747                 ip->ino_data.uid = vap->va_uid_uuid;
748         else if (vap->va_uid != (uid_t)VNOVAL)
749                 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
750         else
751                 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
752
753         if (vap->va_vaflags & VA_GID_UUID_VALID)
754                 ip->ino_data.gid = vap->va_gid_uuid;
755         else if (vap->va_gid != (gid_t)VNOVAL)
756                 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
757         else if (dip)
758                 ip->ino_data.gid = dip->ino_data.gid;
759
760         hammer_ref(&ip->lock);
761
762         if (pfsm) {
763                 ip->pfsm = pfsm;
764                 hammer_ref(&pfsm->lock);
765                 error = 0;
766         } else if (dip->obj_localization == ip->obj_localization) {
767                 ip->pfsm = dip->pfsm;
768                 hammer_ref(&ip->pfsm->lock);
769                 error = 0;
770         } else {
771                 ip->pfsm = hammer_load_pseudofs(trans,
772                                                 ip->obj_localization,
773                                                 &error);
774                 error = 0;      /* ignore ENOENT */
775         }
776
777         if (error) {
778                 hammer_free_inode(ip);
779                 ip = NULL;
780         } else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
781                 panic("hammer_create_inode: duplicate obj_id %llx", ip->obj_id);
782                 /* not reached */
783                 hammer_free_inode(ip);
784         }
785         *ipp = ip;
786         return(error);
787 }
788
789 /*
790  * Final cleanup / freeing of an inode structure
791  */
792 static void
793 hammer_free_inode(hammer_inode_t ip)
794 {
795         struct hammer_mount *hmp;
796
797         hmp = ip->hmp;
798         KKASSERT(ip->lock.refs == 1);
799         hammer_uncache_node(&ip->cache[0]);
800         hammer_uncache_node(&ip->cache[1]);
801         hammer_inode_wakereclaims(ip, 1);
802         if (ip->objid_cache)
803                 hammer_clear_objid(ip);
804         --hammer_count_inodes;
805         --hmp->count_inodes;
806         if (ip->pfsm) {
807                 hammer_rel_pseudofs(hmp, ip->pfsm);
808                 ip->pfsm = NULL;
809         }
810         kfree(ip, hmp->m_inodes);
811         ip = NULL;
812 }
813
814 /*
815  * Retrieve pseudo-fs data.  NULL will never be returned.
816  *
817  * If an error occurs *errorp will be set and a default template is returned,
818  * otherwise *errorp is set to 0.  Typically when an error occurs it will
819  * be ENOENT.
820  */
821 hammer_pseudofs_inmem_t
822 hammer_load_pseudofs(hammer_transaction_t trans,
823                      u_int32_t localization, int *errorp)
824 {
825         hammer_mount_t hmp = trans->hmp;
826         hammer_inode_t ip;
827         hammer_pseudofs_inmem_t pfsm;
828         struct hammer_cursor cursor;
829         int bytes;
830
831 retry:
832         pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
833         if (pfsm) {
834                 hammer_ref(&pfsm->lock);
835                 *errorp = 0;
836                 return(pfsm);
837         }
838
839         /*
840          * PFS records are stored in the root inode (not the PFS root inode,
841          * but the real root).  Avoid an infinite recursion if loading
842          * the PFS for the real root.
843          */
844         if (localization) {
845                 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
846                                       HAMMER_MAX_TID,
847                                       HAMMER_DEF_LOCALIZATION, 0, errorp);
848         } else {
849                 ip = NULL;
850         }
851
852         pfsm = kmalloc(sizeof(*pfsm), hmp->m_misc, M_WAITOK | M_ZERO);
853         pfsm->localization = localization;
854         pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
855         pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
856
857         hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
858         cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION +
859                                       HAMMER_LOCALIZE_MISC;
860         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
861         cursor.key_beg.create_tid = 0;
862         cursor.key_beg.delete_tid = 0;
863         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
864         cursor.key_beg.obj_type = 0;
865         cursor.key_beg.key = localization;
866         cursor.asof = HAMMER_MAX_TID;
867         cursor.flags |= HAMMER_CURSOR_ASOF;
868
869         if (ip)
870                 *errorp = hammer_ip_lookup(&cursor);
871         else
872                 *errorp = hammer_btree_lookup(&cursor);
873         if (*errorp == 0) {
874                 *errorp = hammer_ip_resolve_data(&cursor);
875                 if (*errorp == 0) {
876                         if (cursor.data->pfsd.mirror_flags &
877                             HAMMER_PFSD_DELETED) {
878                                 *errorp = ENOENT;
879                         } else {
880                                 bytes = cursor.leaf->data_len;
881                                 if (bytes > sizeof(pfsm->pfsd))
882                                         bytes = sizeof(pfsm->pfsd);
883                                 bcopy(cursor.data, &pfsm->pfsd, bytes);
884                         }
885                 }
886         }
887         hammer_done_cursor(&cursor);
888
889         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
890         hammer_ref(&pfsm->lock);
891         if (ip)
892                 hammer_rel_inode(ip, 0);
893         if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
894                 kfree(pfsm, hmp->m_misc);
895                 goto retry;
896         }
897         return(pfsm);
898 }
899
900 /*
901  * Store pseudo-fs data.  The backend will automatically delete any prior
902  * on-disk pseudo-fs data but we have to delete in-memory versions.
903  */
904 int
905 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
906 {
907         struct hammer_cursor cursor;
908         hammer_record_t record;
909         hammer_inode_t ip;
910         int error;
911
912         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
913                               HAMMER_DEF_LOCALIZATION, 0, &error);
914 retry:
915         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
916         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
917         cursor.key_beg.localization = ip->obj_localization +
918                                       HAMMER_LOCALIZE_MISC;
919         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
920         cursor.key_beg.create_tid = 0;
921         cursor.key_beg.delete_tid = 0;
922         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
923         cursor.key_beg.obj_type = 0;
924         cursor.key_beg.key = pfsm->localization;
925         cursor.asof = HAMMER_MAX_TID;
926         cursor.flags |= HAMMER_CURSOR_ASOF;
927
928         /*
929          * Replace any in-memory version of the record.
930          */
931         error = hammer_ip_lookup(&cursor);
932         if (error == 0 && hammer_cursor_inmem(&cursor)) {
933                 record = cursor.iprec;
934                 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
935                         KKASSERT(cursor.deadlk_rec == NULL);
936                         hammer_ref(&record->lock);
937                         cursor.deadlk_rec = record;
938                         error = EDEADLK;
939                 } else {
940                         record->flags |= HAMMER_RECF_DELETED_FE;
941                         error = 0;
942                 }
943         }
944
945         /*
946          * Allocate replacement general record.  The backend flush will
947          * delete any on-disk version of the record.
948          */
949         if (error == 0 || error == ENOENT) {
950                 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
951                 record->type = HAMMER_MEM_RECORD_GENERAL;
952
953                 record->leaf.base.localization = ip->obj_localization +
954                                                  HAMMER_LOCALIZE_MISC;
955                 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
956                 record->leaf.base.key = pfsm->localization;
957                 record->leaf.data_len = sizeof(pfsm->pfsd);
958                 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
959                 error = hammer_ip_add_record(trans, record);
960         }
961         hammer_done_cursor(&cursor);
962         if (error == EDEADLK)
963                 goto retry;
964         hammer_rel_inode(ip, 0);
965         return(error);
966 }
967
968 /*
969  * Create a root directory for a PFS if one does not alredy exist.
970  *
971  * The PFS root stands alone so we must also bump the nlinks count
972  * to prevent it from being destroyed on release.
973  */
974 int
975 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
976                        hammer_pseudofs_inmem_t pfsm)
977 {
978         hammer_inode_t ip;
979         struct vattr vap;
980         int error;
981
982         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
983                               pfsm->localization, 0, &error);
984         if (ip == NULL) {
985                 vattr_null(&vap);
986                 vap.va_mode = 0755;
987                 vap.va_type = VDIR;
988                 error = hammer_create_inode(trans, &vap, cred, NULL, pfsm, &ip);
989                 if (error == 0) {
990                         ++ip->ino_data.nlinks;
991                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
992                 }
993         }
994         if (ip)
995                 hammer_rel_inode(ip, 0);
996         return(error);
997 }
998
999 /*
1000  * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
1001  * if we are unable to disassociate all the inodes.
1002  */
1003 static
1004 int
1005 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
1006 {
1007         int res;
1008
1009         hammer_ref(&ip->lock);
1010         if (ip->lock.refs == 2 && ip->vp)
1011                 vclean_unlocked(ip->vp);
1012         if (ip->lock.refs == 1 && ip->vp == NULL)
1013                 res = 0;
1014         else
1015                 res = -1;       /* stop, someone is using the inode */
1016         hammer_rel_inode(ip, 0);
1017         return(res);
1018 }
1019
1020 int
1021 hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
1022 {
1023         int res;
1024         int try;
1025
1026         for (try = res = 0; try < 4; ++try) {
1027                 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
1028                                            hammer_inode_pfs_cmp,
1029                                            hammer_unload_pseudofs_callback,
1030                                            &localization);
1031                 if (res == 0 && try > 1)
1032                         break;
1033                 hammer_flusher_sync(trans->hmp);
1034         }
1035         if (res != 0)
1036                 res = ENOTEMPTY;
1037         return(res);
1038 }
1039
1040
1041 /*
1042  * Release a reference on a PFS
1043  */
1044 void
1045 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
1046 {
1047         hammer_unref(&pfsm->lock);
1048         if (pfsm->lock.refs == 0) {
1049                 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
1050                 kfree(pfsm, hmp->m_misc);
1051         }
1052 }
1053
1054 /*
1055  * Called by hammer_sync_inode().
1056  */
1057 static int
1058 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
1059 {
1060         hammer_transaction_t trans = cursor->trans;
1061         hammer_record_t record;
1062         int error;
1063         int redirty;
1064
1065 retry:
1066         error = 0;
1067
1068         /*
1069          * If the inode has a presence on-disk then locate it and mark
1070          * it deleted, setting DELONDISK.
1071          *
1072          * The record may or may not be physically deleted, depending on
1073          * the retention policy.
1074          */
1075         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
1076             HAMMER_INODE_ONDISK) {
1077                 hammer_normalize_cursor(cursor);
1078                 cursor->key_beg.localization = ip->obj_localization + 
1079                                                HAMMER_LOCALIZE_INODE;
1080                 cursor->key_beg.obj_id = ip->obj_id;
1081                 cursor->key_beg.key = 0;
1082                 cursor->key_beg.create_tid = 0;
1083                 cursor->key_beg.delete_tid = 0;
1084                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1085                 cursor->key_beg.obj_type = 0;
1086                 cursor->asof = ip->obj_asof;
1087                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1088                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
1089                 cursor->flags |= HAMMER_CURSOR_BACKEND;
1090
1091                 error = hammer_btree_lookup(cursor);
1092                 if (hammer_debug_inode)
1093                         kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
1094
1095                 if (error == 0) {
1096                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1097                         if (hammer_debug_inode)
1098                                 kprintf(" error %d\n", error);
1099                         if (error == 0) {
1100                                 ip->flags |= HAMMER_INODE_DELONDISK;
1101                         }
1102                         if (cursor->node)
1103                                 hammer_cache_node(&ip->cache[0], cursor->node);
1104                 }
1105                 if (error == EDEADLK) {
1106                         hammer_done_cursor(cursor);
1107                         error = hammer_init_cursor(trans, cursor,
1108                                                    &ip->cache[0], ip);
1109                         if (hammer_debug_inode)
1110                                 kprintf("IPDED %p %d\n", ip, error);
1111                         if (error == 0)
1112                                 goto retry;
1113                 }
1114         }
1115
1116         /*
1117          * Ok, write out the initial record or a new record (after deleting
1118          * the old one), unless the DELETED flag is set.  This routine will
1119          * clear DELONDISK if it writes out a record.
1120          *
1121          * Update our inode statistics if this is the first application of
1122          * the inode on-disk.
1123          */
1124         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
1125                 /*
1126                  * Generate a record and write it to the media.  We clean-up
1127                  * the state before releasing so we do not have to set-up
1128                  * a flush_group.
1129                  */
1130                 record = hammer_alloc_mem_record(ip, 0);
1131                 record->type = HAMMER_MEM_RECORD_INODE;
1132                 record->flush_state = HAMMER_FST_FLUSH;
1133                 record->leaf = ip->sync_ino_leaf;
1134                 record->leaf.base.create_tid = trans->tid;
1135                 record->leaf.data_len = sizeof(ip->sync_ino_data);
1136                 record->leaf.create_ts = trans->time32;
1137                 record->data = (void *)&ip->sync_ino_data;
1138                 record->flags |= HAMMER_RECF_INTERLOCK_BE;
1139
1140                 /*
1141                  * If this flag is set we cannot sync the new file size
1142                  * because we haven't finished related truncations.  The
1143                  * inode will be flushed in another flush group to finish
1144                  * the job.
1145                  */
1146                 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1147                     ip->sync_ino_data.size != ip->ino_data.size) {
1148                         redirty = 1;
1149                         ip->sync_ino_data.size = ip->ino_data.size;
1150                 } else {
1151                         redirty = 0;
1152                 }
1153
1154                 for (;;) {
1155                         error = hammer_ip_sync_record_cursor(cursor, record);
1156                         if (hammer_debug_inode)
1157                                 kprintf("GENREC %p rec %08x %d\n",      
1158                                         ip, record->flags, error);
1159                         if (error != EDEADLK)
1160                                 break;
1161                         hammer_done_cursor(cursor);
1162                         error = hammer_init_cursor(trans, cursor,
1163                                                    &ip->cache[0], ip);
1164                         if (hammer_debug_inode)
1165                                 kprintf("GENREC reinit %d\n", error);
1166                         if (error)
1167                                 break;
1168                 }
1169
1170                 /*
1171                  * Note:  The record was never on the inode's record tree
1172                  * so just wave our hands importantly and destroy it.
1173                  */
1174                 record->flags |= HAMMER_RECF_COMMITTED;
1175                 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1176                 record->flush_state = HAMMER_FST_IDLE;
1177                 ++ip->rec_generation;
1178                 hammer_rel_mem_record(record);
1179
1180                 /*
1181                  * Finish up.
1182                  */
1183                 if (error == 0) {
1184                         if (hammer_debug_inode)
1185                                 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1186                         ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1187                                             HAMMER_INODE_ATIME |
1188                                             HAMMER_INODE_MTIME);
1189                         ip->flags &= ~HAMMER_INODE_DELONDISK;
1190                         if (redirty)
1191                                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1192
1193                         /*
1194                          * Root volume count of inodes
1195                          */
1196                         hammer_sync_lock_sh(trans);
1197                         if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1198                                 hammer_modify_volume_field(trans,
1199                                                            trans->rootvol,
1200                                                            vol0_stat_inodes);
1201                                 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1202                                 hammer_modify_volume_done(trans->rootvol);
1203                                 ip->flags |= HAMMER_INODE_ONDISK;
1204                                 if (hammer_debug_inode)
1205                                         kprintf("NOWONDISK %p\n", ip);
1206                         }
1207                         hammer_sync_unlock(trans);
1208                 }
1209         }
1210
1211         /*
1212          * If the inode has been destroyed, clean out any left-over flags
1213          * that may have been set by the frontend.
1214          */
1215         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) { 
1216                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1217                                     HAMMER_INODE_ATIME |
1218                                     HAMMER_INODE_MTIME);
1219         }
1220         return(error);
1221 }
1222
1223 /*
1224  * Update only the itimes fields.
1225  *
1226  * ATIME can be updated without generating any UNDO.  MTIME is updated
1227  * with UNDO so it is guaranteed to be synchronized properly in case of
1228  * a crash.
1229  *
1230  * Neither field is included in the B-Tree leaf element's CRC, which is how
1231  * we can get away with updating ATIME the way we do.
1232  */
1233 static int
1234 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1235 {
1236         hammer_transaction_t trans = cursor->trans;
1237         int error;
1238
1239 retry:
1240         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1241             HAMMER_INODE_ONDISK) {
1242                 return(0);
1243         }
1244
1245         hammer_normalize_cursor(cursor);
1246         cursor->key_beg.localization = ip->obj_localization + 
1247                                        HAMMER_LOCALIZE_INODE;
1248         cursor->key_beg.obj_id = ip->obj_id;
1249         cursor->key_beg.key = 0;
1250         cursor->key_beg.create_tid = 0;
1251         cursor->key_beg.delete_tid = 0;
1252         cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1253         cursor->key_beg.obj_type = 0;
1254         cursor->asof = ip->obj_asof;
1255         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1256         cursor->flags |= HAMMER_CURSOR_ASOF;
1257         cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1258         cursor->flags |= HAMMER_CURSOR_GET_DATA;
1259         cursor->flags |= HAMMER_CURSOR_BACKEND;
1260
1261         error = hammer_btree_lookup(cursor);
1262         if (error == 0) {
1263                 hammer_cache_node(&ip->cache[0], cursor->node);
1264                 if (ip->sync_flags & HAMMER_INODE_MTIME) {
1265                         /*
1266                          * Updating MTIME requires an UNDO.  Just cover
1267                          * both atime and mtime.
1268                          */
1269                         hammer_sync_lock_sh(trans);
1270                         hammer_modify_buffer(trans, cursor->data_buffer,
1271                                      HAMMER_ITIMES_BASE(&cursor->data->inode),
1272                                      HAMMER_ITIMES_BYTES);
1273                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1274                         cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1275                         hammer_modify_buffer_done(cursor->data_buffer);
1276                         hammer_sync_unlock(trans);
1277                 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1278                         /*
1279                          * Updating atime only can be done in-place with
1280                          * no UNDO.
1281                          */
1282                         hammer_sync_lock_sh(trans);
1283                         hammer_modify_buffer(trans, cursor->data_buffer,
1284                                              NULL, 0);
1285                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1286                         hammer_modify_buffer_done(cursor->data_buffer);
1287                         hammer_sync_unlock(trans);
1288                 }
1289                 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1290         }
1291         if (error == EDEADLK) {
1292                 hammer_done_cursor(cursor);
1293                 error = hammer_init_cursor(trans, cursor,
1294                                            &ip->cache[0], ip);
1295                 if (error == 0)
1296                         goto retry;
1297         }
1298         return(error);
1299 }
1300
1301 /*
1302  * Release a reference on an inode, flush as requested.
1303  *
1304  * On the last reference we queue the inode to the flusher for its final
1305  * disposition.
1306  */
1307 void
1308 hammer_rel_inode(struct hammer_inode *ip, int flush)
1309 {
1310         /*hammer_mount_t hmp = ip->hmp;*/
1311
1312         /*
1313          * Handle disposition when dropping the last ref.
1314          */
1315         for (;;) {
1316                 if (ip->lock.refs == 1) {
1317                         /*
1318                          * Determine whether on-disk action is needed for
1319                          * the inode's final disposition.
1320                          */
1321                         KKASSERT(ip->vp == NULL);
1322                         hammer_inode_unloadable_check(ip, 0);
1323                         if (ip->flags & HAMMER_INODE_MODMASK) {
1324                                 hammer_flush_inode(ip, 0);
1325                         } else if (ip->lock.refs == 1) {
1326                                 hammer_unload_inode(ip);
1327                                 break;
1328                         }
1329                 } else {
1330                         if (flush)
1331                                 hammer_flush_inode(ip, 0);
1332
1333                         /*
1334                          * The inode still has multiple refs, try to drop
1335                          * one ref.
1336                          */
1337                         KKASSERT(ip->lock.refs >= 1);
1338                         if (ip->lock.refs > 1) {
1339                                 hammer_unref(&ip->lock);
1340                                 break;
1341                         }
1342                 }
1343         }
1344 }
1345
1346 /*
1347  * Unload and destroy the specified inode.  Must be called with one remaining
1348  * reference.  The reference is disposed of.
1349  *
1350  * The inode must be completely clean.
1351  */
1352 static int
1353 hammer_unload_inode(struct hammer_inode *ip)
1354 {
1355         hammer_mount_t hmp = ip->hmp;
1356
1357         KASSERT(ip->lock.refs == 1,
1358                 ("hammer_unload_inode: %d refs\n", ip->lock.refs));
1359         KKASSERT(ip->vp == NULL);
1360         KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1361         KKASSERT(ip->cursor_ip_refs == 0);
1362         KKASSERT(ip->lock.lockcount == 0);
1363         KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1364
1365         KKASSERT(RB_EMPTY(&ip->rec_tree));
1366         KKASSERT(TAILQ_EMPTY(&ip->target_list));
1367
1368         RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1369
1370         hammer_free_inode(ip);
1371         return(0);
1372 }
1373
1374 /*
1375  * Called during unmounting if a critical error occured.  The in-memory
1376  * inode and all related structures are destroyed.
1377  *
1378  * If a critical error did not occur the unmount code calls the standard
1379  * release and asserts that the inode is gone.
1380  */
1381 int
1382 hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
1383 {
1384         hammer_record_t rec;
1385
1386         /*
1387          * Get rid of the inodes in-memory records, regardless of their
1388          * state, and clear the mod-mask.
1389          */
1390         while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1391                 TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1392                 rec->target_ip = NULL;
1393                 if (rec->flush_state == HAMMER_FST_SETUP)
1394                         rec->flush_state = HAMMER_FST_IDLE;
1395         }
1396         while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1397                 if (rec->flush_state == HAMMER_FST_FLUSH)
1398                         --rec->flush_group->refs;
1399                 else
1400                         hammer_ref(&rec->lock);
1401                 KKASSERT(rec->lock.refs == 1);
1402                 rec->flush_state = HAMMER_FST_IDLE;
1403                 rec->flush_group = NULL;
1404                 rec->flags |= HAMMER_RECF_DELETED_FE; /* wave hands */
1405                 rec->flags |= HAMMER_RECF_DELETED_BE; /* wave hands */
1406                 ++ip->rec_generation;
1407                 hammer_rel_mem_record(rec);
1408         }
1409         ip->flags &= ~HAMMER_INODE_MODMASK;
1410         ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1411         KKASSERT(ip->vp == NULL);
1412
1413         /*
1414          * Remove the inode from any flush group, force it idle.  FLUSH
1415          * and SETUP states have an inode ref.
1416          */
1417         switch(ip->flush_state) {
1418         case HAMMER_FST_FLUSH:
1419                 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
1420                 --ip->flush_group->refs;
1421                 ip->flush_group = NULL;
1422                 /* fall through */
1423         case HAMMER_FST_SETUP:
1424                 hammer_unref(&ip->lock);
1425                 ip->flush_state = HAMMER_FST_IDLE;
1426                 /* fall through */
1427         case HAMMER_FST_IDLE:
1428                 break;
1429         }
1430
1431         /*
1432          * There shouldn't be any associated vnode.  The unload needs at
1433          * least one ref, if we do have a vp steal its ip ref.
1434          */
1435         if (ip->vp) {
1436                 kprintf("hammer_destroy_inode_callback: Unexpected "
1437                         "vnode association ip %p vp %p\n", ip, ip->vp);
1438                 ip->vp->v_data = NULL;
1439                 ip->vp = NULL;
1440         } else {
1441                 hammer_ref(&ip->lock);
1442         }
1443         hammer_unload_inode(ip);
1444         return(0);
1445 }
1446
1447 /*
1448  * Called on mount -u when switching from RW to RO or vise-versa.  Adjust
1449  * the read-only flag for cached inodes.
1450  *
1451  * This routine is called from a RB_SCAN().
1452  */
1453 int
1454 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1455 {
1456         hammer_mount_t hmp = ip->hmp;
1457
1458         if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1459                 ip->flags |= HAMMER_INODE_RO;
1460         else
1461                 ip->flags &= ~HAMMER_INODE_RO;
1462         return(0);
1463 }
1464
1465 /*
1466  * A transaction has modified an inode, requiring updates as specified by
1467  * the passed flags.
1468  *
1469  * HAMMER_INODE_DDIRTY: Inode data has been updated
1470  * HAMMER_INODE_XDIRTY: Dirty in-memory records
1471  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
1472  * HAMMER_INODE_DELETED: Inode record/data must be deleted
1473  * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1474  */
1475 void
1476 hammer_modify_inode(hammer_inode_t ip, int flags)
1477 {
1478         /* 
1479          * ronly of 0 or 2 does not trigger assertion.
1480          * 2 is a special error state 
1481          */
1482         KKASSERT(ip->hmp->ronly != 1 ||
1483                   (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY | 
1484                             HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1485                             HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1486         if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1487                 ip->flags |= HAMMER_INODE_RSV_INODES;
1488                 ++ip->hmp->rsv_inodes;
1489         }
1490
1491         ip->flags |= flags;
1492 }
1493
1494 /*
1495  * Request that an inode be flushed.  This whole mess cannot block and may
1496  * recurse (if not synchronous).  Once requested HAMMER will attempt to
1497  * actively flush the inode until the flush can be done.
1498  *
1499  * The inode may already be flushing, or may be in a setup state.  We can
1500  * place the inode in a flushing state if it is currently idle and flag it
1501  * to reflush if it is currently flushing.
1502  *
1503  * Upon return if the inode could not be flushed due to a setup
1504  * dependancy, then it will be automatically flushed when the dependancy
1505  * is satisfied.
1506  */
1507 void
1508 hammer_flush_inode(hammer_inode_t ip, int flags)
1509 {
1510         hammer_mount_t hmp;
1511         hammer_flush_group_t flg;
1512         int good;
1513
1514         /*
1515          * next_flush_group is the first flush group we can place the inode
1516          * in.  It may be NULL.  If it becomes full we append a new flush
1517          * group and make that the next_flush_group.
1518          */
1519         hmp = ip->hmp;
1520         while ((flg = hmp->next_flush_group) != NULL) {
1521                 KKASSERT(flg->running == 0);
1522                 if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit)
1523                         break;
1524                 hmp->next_flush_group = TAILQ_NEXT(flg, flush_entry);
1525                 hammer_flusher_async(ip->hmp, flg);
1526         }
1527         if (flg == NULL) {
1528                 flg = kmalloc(sizeof(*flg), hmp->m_misc, M_WAITOK|M_ZERO);
1529                 hmp->next_flush_group = flg;
1530                 TAILQ_INIT(&flg->flush_list);
1531                 TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1532         }
1533
1534         /*
1535          * Trivial 'nothing to flush' case.  If the inode is in a SETUP
1536          * state we have to put it back into an IDLE state so we can
1537          * drop the extra ref.
1538          *
1539          * If we have a parent dependancy we must still fall through
1540          * so we can run it.
1541          */
1542         if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1543                 if (ip->flush_state == HAMMER_FST_SETUP &&
1544                     TAILQ_EMPTY(&ip->target_list)) {
1545                         ip->flush_state = HAMMER_FST_IDLE;
1546                         hammer_rel_inode(ip, 0);
1547                 }
1548                 if (ip->flush_state == HAMMER_FST_IDLE)
1549                         return;
1550         }
1551
1552         /*
1553          * Our flush action will depend on the current state.
1554          */
1555         switch(ip->flush_state) {
1556         case HAMMER_FST_IDLE:
1557                 /*
1558                  * We have no dependancies and can flush immediately.  Some
1559                  * our children may not be flushable so we have to re-test
1560                  * with that additional knowledge.
1561                  */
1562                 hammer_flush_inode_core(ip, flg, flags);
1563                 break;
1564         case HAMMER_FST_SETUP:
1565                 /*
1566                  * Recurse upwards through dependancies via target_list
1567                  * and start their flusher actions going if possible.
1568                  *
1569                  * 'good' is our connectivity.  -1 means we have none and
1570                  * can't flush, 0 means there weren't any dependancies, and
1571                  * 1 means we have good connectivity.
1572                  */
1573                 good = hammer_setup_parent_inodes(ip, 0, flg);
1574
1575                 if (good >= 0) {
1576                         /*
1577                          * We can continue if good >= 0.  Determine how 
1578                          * many records under our inode can be flushed (and
1579                          * mark them).
1580                          */
1581                         hammer_flush_inode_core(ip, flg, flags);
1582                 } else {
1583                         /*
1584                          * Parent has no connectivity, tell it to flush
1585                          * us as soon as it does.
1586                          *
1587                          * The REFLUSH flag is also needed to trigger
1588                          * dependancy wakeups.
1589                          */
1590                         ip->flags |= HAMMER_INODE_CONN_DOWN |
1591                                      HAMMER_INODE_REFLUSH;
1592                         if (flags & HAMMER_FLUSH_SIGNAL) {
1593                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1594                                 hammer_flusher_async(ip->hmp, flg);
1595                         }
1596                 }
1597                 break;
1598         case HAMMER_FST_FLUSH:
1599                 /*
1600                  * We are already flushing, flag the inode to reflush
1601                  * if needed after it completes its current flush.
1602                  *
1603                  * The REFLUSH flag is also needed to trigger
1604                  * dependancy wakeups.
1605                  */
1606                 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1607                         ip->flags |= HAMMER_INODE_REFLUSH;
1608                 if (flags & HAMMER_FLUSH_SIGNAL) {
1609                         ip->flags |= HAMMER_INODE_RESIGNAL;
1610                         hammer_flusher_async(ip->hmp, flg);
1611                 }
1612                 break;
1613         }
1614 }
1615
1616 /*
1617  * Scan ip->target_list, which is a list of records owned by PARENTS to our
1618  * ip which reference our ip.
1619  *
1620  * XXX This is a huge mess of recursive code, but not one bit of it blocks
1621  *     so for now do not ref/deref the structures.  Note that if we use the
1622  *     ref/rel code later, the rel CAN block.
1623  */
1624 static int
1625 hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
1626                            hammer_flush_group_t flg)
1627 {
1628         hammer_record_t depend;
1629         int good;
1630         int r;
1631
1632         /*
1633          * If we hit our recursion limit and we have parent dependencies
1634          * We cannot continue.  Returning < 0 will cause us to be flagged
1635          * for reflush.  Returning -2 cuts off additional dependency checks
1636          * because they are likely to also hit the depth limit.
1637          *
1638          * We cannot return < 0 if there are no dependencies or there might
1639          * not be anything to wakeup (ip).
1640          */
1641         if (depth == 20 && TAILQ_FIRST(&ip->target_list)) {
1642                 kprintf("HAMMER Warning: depth limit reached on "
1643                         "setup recursion, inode %p %016llx\n",
1644                         ip, (long long)ip->obj_id);
1645                 return(-2);
1646         }
1647
1648         /*
1649          * Scan dependencies
1650          */
1651         good = 0;
1652         TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1653                 r = hammer_setup_parent_inodes_helper(depend, depth, flg);
1654                 KKASSERT(depend->target_ip == ip);
1655                 if (r < 0 && good == 0)
1656                         good = -1;
1657                 if (r > 0)
1658                         good = 1;
1659
1660                 /*
1661                  * If we failed due to the recursion depth limit then stop
1662                  * now.
1663                  */
1664                 if (r == -2)
1665                         break;
1666         }
1667         return(good);
1668 }
1669
1670 /*
1671  * This helper function takes a record representing the dependancy between
1672  * the parent inode and child inode.
1673  *
1674  * record->ip           = parent inode
1675  * record->target_ip    = child inode
1676  * 
1677  * We are asked to recurse upwards and convert the record from SETUP
1678  * to FLUSH if possible.
1679  *
1680  * Return 1 if the record gives us connectivity
1681  *
1682  * Return 0 if the record is not relevant 
1683  *
1684  * Return -1 if we can't resolve the dependancy and there is no connectivity.
1685  */
1686 static int
1687 hammer_setup_parent_inodes_helper(hammer_record_t record, int depth,
1688                                   hammer_flush_group_t flg)
1689 {
1690         hammer_mount_t hmp;
1691         hammer_inode_t pip;
1692         int good;
1693
1694         KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1695         pip = record->ip;
1696         hmp = pip->hmp;
1697
1698         /*
1699          * If the record is already flushing, is it in our flush group?
1700          *
1701          * If it is in our flush group but it is a general record or a 
1702          * delete-on-disk, it does not improve our connectivity (return 0),
1703          * and if the target inode is not trying to destroy itself we can't
1704          * allow the operation yet anyway (the second return -1).
1705          */
1706         if (record->flush_state == HAMMER_FST_FLUSH) {
1707                 /*
1708                  * If not in our flush group ask the parent to reflush
1709                  * us as soon as possible.
1710                  */
1711                 if (record->flush_group != flg) {
1712                         pip->flags |= HAMMER_INODE_REFLUSH;
1713                         record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1714                         return(-1);
1715                 }
1716
1717                 /*
1718                  * If in our flush group everything is already set up,
1719                  * just return whether the record will improve our
1720                  * visibility or not.
1721                  */
1722                 if (record->type == HAMMER_MEM_RECORD_ADD)
1723                         return(1);
1724                 return(0);
1725         }
1726
1727         /*
1728          * It must be a setup record.  Try to resolve the setup dependancies
1729          * by recursing upwards so we can place ip on the flush list.
1730          *
1731          * Limit ourselves to 20 levels of recursion to avoid blowing out
1732          * the kernel stack.  If we hit the recursion limit we can't flush
1733          * until the parent flushes.  The parent will flush independantly
1734          * on its own and ultimately a deep recursion will be resolved.
1735          */
1736         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1737
1738         good = hammer_setup_parent_inodes(pip, depth + 1, flg);
1739
1740         /*
1741          * If good < 0 the parent has no connectivity and we cannot safely
1742          * flush the directory entry, which also means we can't flush our
1743          * ip.  Flag us for downward recursion once the parent's
1744          * connectivity is resolved.  Flag the parent for [re]flush or it
1745          * may not check for downward recursions.
1746          */
1747         if (good < 0) {
1748                 pip->flags |= HAMMER_INODE_REFLUSH;
1749                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1750                 return(good);
1751         }
1752
1753         /*
1754          * We are go, place the parent inode in a flushing state so we can
1755          * place its record in a flushing state.  Note that the parent
1756          * may already be flushing.  The record must be in the same flush
1757          * group as the parent.
1758          */
1759         if (pip->flush_state != HAMMER_FST_FLUSH)
1760                 hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
1761         KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
1762         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1763
1764 #if 0
1765         if (record->type == HAMMER_MEM_RECORD_DEL &&
1766             (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
1767                 /*
1768                  * Regardless of flushing state we cannot sync this path if the
1769                  * record represents a delete-on-disk but the target inode
1770                  * is not ready to sync its own deletion.
1771                  *
1772                  * XXX need to count effective nlinks to determine whether
1773                  * the flush is ok, otherwise removing a hardlink will
1774                  * just leave the DEL record to rot.
1775                  */
1776                 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
1777                 return(-1);
1778         } else
1779 #endif
1780         if (pip->flush_group == flg) {
1781                 /*
1782                  * Because we have not calculated nlinks yet we can just
1783                  * set records to the flush state if the parent is in
1784                  * the same flush group as we are.
1785                  */
1786                 record->flush_state = HAMMER_FST_FLUSH;
1787                 record->flush_group = flg;
1788                 ++record->flush_group->refs;
1789                 hammer_ref(&record->lock);
1790
1791                 /*
1792                  * A general directory-add contributes to our visibility.
1793                  *
1794                  * Otherwise it is probably a directory-delete or 
1795                  * delete-on-disk record and does not contribute to our
1796                  * visbility (but we can still flush it).
1797                  */
1798                 if (record->type == HAMMER_MEM_RECORD_ADD)
1799                         return(1);
1800                 return(0);
1801         } else {
1802                 /*
1803                  * If the parent is not in our flush group we cannot
1804                  * flush this record yet, there is no visibility.
1805                  * We tell the parent to reflush and mark ourselves
1806                  * so the parent knows it should flush us too.
1807                  */
1808                 pip->flags |= HAMMER_INODE_REFLUSH;
1809                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1810                 return(-1);
1811         }
1812 }
1813
1814 /*
1815  * This is the core routine placing an inode into the FST_FLUSH state.
1816  */
1817 static void
1818 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
1819 {
1820         int go_count;
1821
1822         /*
1823          * Set flush state and prevent the flusher from cycling into
1824          * the next flush group.  Do not place the ip on the list yet.
1825          * Inodes not in the idle state get an extra reference.
1826          */
1827         KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
1828         if (ip->flush_state == HAMMER_FST_IDLE)
1829                 hammer_ref(&ip->lock);
1830         ip->flush_state = HAMMER_FST_FLUSH;
1831         ip->flush_group = flg;
1832         ++ip->hmp->flusher.group_lock;
1833         ++ip->hmp->count_iqueued;
1834         ++hammer_count_iqueued;
1835         ++flg->total_count;
1836
1837         /*
1838          * If the flush group reaches the autoflush limit we want to signal
1839          * the flusher.  This is particularly important for remove()s.
1840          */
1841         if (flg->total_count == hammer_autoflush)
1842                 flags |= HAMMER_FLUSH_SIGNAL;
1843
1844         /*
1845          * We need to be able to vfsync/truncate from the backend.
1846          */
1847         KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
1848         if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
1849                 ip->flags |= HAMMER_INODE_VHELD;
1850                 vref(ip->vp);
1851         }
1852
1853         /*
1854          * Figure out how many in-memory records we can actually flush
1855          * (not including inode meta-data, buffers, etc).
1856          */
1857         KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
1858         if (flags & HAMMER_FLUSH_RECURSION) {
1859                 /*
1860                  * If this is a upwards recursion we do not want to
1861                  * recurse down again!
1862                  */
1863                 go_count = 1;
1864 #if 0
1865         } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
1866                 /*
1867                  * No new records are added if we must complete a flush
1868                  * from a previous cycle, but we do have to move the records
1869                  * from the previous cycle to the current one.
1870                  */
1871 #if 0
1872                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1873                                    hammer_syncgrp_child_callback, NULL);
1874 #endif
1875                 go_count = 1;
1876 #endif
1877         } else {
1878                 /*
1879                  * Normal flush, scan records and bring them into the flush.
1880                  * Directory adds and deletes are usually skipped (they are
1881                  * grouped with the related inode rather then with the
1882                  * directory).
1883                  *
1884                  * go_count can be negative, which means the scan aborted
1885                  * due to the flush group being over-full and we should
1886                  * flush what we have.
1887                  */
1888                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1889                                    hammer_setup_child_callback, NULL);
1890         }
1891
1892         /*
1893          * This is a more involved test that includes go_count.  If we
1894          * can't flush, flag the inode and return.  If go_count is 0 we
1895          * were are unable to flush any records in our rec_tree and
1896          * must ignore the XDIRTY flag.
1897          */
1898         if (go_count == 0) {
1899                 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
1900                         --ip->hmp->count_iqueued;
1901                         --hammer_count_iqueued;
1902
1903                         --flg->total_count;
1904                         ip->flush_state = HAMMER_FST_SETUP;
1905                         ip->flush_group = NULL;
1906                         if (ip->flags & HAMMER_INODE_VHELD) {
1907                                 ip->flags &= ~HAMMER_INODE_VHELD;
1908                                 vrele(ip->vp);
1909                         }
1910
1911                         /*
1912                          * REFLUSH is needed to trigger dependancy wakeups
1913                          * when an inode is in SETUP.
1914                          */
1915                         ip->flags |= HAMMER_INODE_REFLUSH;
1916                         if (flags & HAMMER_FLUSH_SIGNAL) {
1917                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1918                                 hammer_flusher_async(ip->hmp, flg);
1919                         }
1920                         if (--ip->hmp->flusher.group_lock == 0)
1921                                 wakeup(&ip->hmp->flusher.group_lock);
1922                         return;
1923                 }
1924         }
1925
1926         /*
1927          * Snapshot the state of the inode for the backend flusher.
1928          *
1929          * We continue to retain save_trunc_off even when all truncations
1930          * have been resolved as an optimization to determine if we can
1931          * skip the B-Tree lookup for overwrite deletions.
1932          *
1933          * NOTE: The DELETING flag is a mod flag, but it is also sticky,
1934          * and stays in ip->flags.  Once set, it stays set until the
1935          * inode is destroyed.
1936          */
1937         if (ip->flags & HAMMER_INODE_TRUNCATED) {
1938                 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
1939                 ip->sync_trunc_off = ip->trunc_off;
1940                 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
1941                 ip->flags &= ~HAMMER_INODE_TRUNCATED;
1942                 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
1943
1944                 /*
1945                  * The save_trunc_off used to cache whether the B-Tree
1946                  * holds any records past that point is not used until
1947                  * after the truncation has succeeded, so we can safely
1948                  * set it now.
1949                  */
1950                 if (ip->save_trunc_off > ip->sync_trunc_off)
1951                         ip->save_trunc_off = ip->sync_trunc_off;
1952         }
1953         ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
1954                            ~HAMMER_INODE_TRUNCATED);
1955         ip->sync_ino_leaf = ip->ino_leaf;
1956         ip->sync_ino_data = ip->ino_data;
1957         ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
1958 #ifdef DEBUG_TRUNCATE
1959         if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
1960                 kprintf("truncateS %016llx\n", ip->sync_trunc_off);
1961 #endif
1962
1963         /*
1964          * The flusher list inherits our inode and reference.
1965          */
1966         KKASSERT(flg->running == 0);
1967         TAILQ_INSERT_TAIL(&flg->flush_list, ip, flush_entry);
1968         if (--ip->hmp->flusher.group_lock == 0)
1969                 wakeup(&ip->hmp->flusher.group_lock);
1970
1971         if (flags & HAMMER_FLUSH_SIGNAL) {
1972                 hammer_flusher_async(ip->hmp, flg);
1973         }
1974 }
1975
1976 /*
1977  * Callback for scan of ip->rec_tree.  Try to include each record in our
1978  * flush.  ip->flush_group has been set but the inode has not yet been
1979  * moved into a flushing state.
1980  *
1981  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1982  * both inodes.
1983  *
1984  * We return 1 for any record placed or found in FST_FLUSH, which prevents
1985  * the caller from shortcutting the flush.
1986  */
1987 static int
1988 hammer_setup_child_callback(hammer_record_t rec, void *data)
1989 {
1990         hammer_flush_group_t flg;
1991         hammer_inode_t target_ip;
1992         hammer_inode_t ip;
1993         int r;
1994
1995         /*
1996          * Records deleted or committed by the backend are ignored.
1997          * Note that the flush detects deleted frontend records at
1998          * multiple points to deal with races.  This is just the first
1999          * line of defense.  The only time HAMMER_RECF_DELETED_FE cannot
2000          * be set is when HAMMER_RECF_INTERLOCK_BE is set, because it
2001          * messes up link-count calculations.
2002          *
2003          * NOTE: Don't get confused between record deletion and, say,
2004          * directory entry deletion.  The deletion of a directory entry
2005          * which is on-media has nothing to do with the record deletion
2006          * flags.
2007          */
2008         if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
2009                           HAMMER_RECF_COMMITTED)) {
2010                 if (rec->flush_state == HAMMER_FST_FLUSH) {
2011                         KKASSERT(rec->flush_group == rec->ip->flush_group);
2012                         r = 1;
2013                 } else {
2014                         r = 0;
2015                 }
2016                 return(r);
2017         }
2018
2019         /*
2020          * If the record is in an idle state it has no dependancies and
2021          * can be flushed.
2022          */
2023         ip = rec->ip;
2024         flg = ip->flush_group;
2025         r = 0;
2026
2027         switch(rec->flush_state) {
2028         case HAMMER_FST_IDLE:
2029                 /*
2030                  * The record has no setup dependancy, we can flush it.
2031                  */
2032                 KKASSERT(rec->target_ip == NULL);
2033                 rec->flush_state = HAMMER_FST_FLUSH;
2034                 rec->flush_group = flg;
2035                 ++flg->refs;
2036                 hammer_ref(&rec->lock);
2037                 r = 1;
2038                 break;
2039         case HAMMER_FST_SETUP:
2040                 /*
2041                  * The record has a setup dependancy.  These are typically
2042                  * directory entry adds and deletes.  Such entries will be
2043                  * flushed when their inodes are flushed so we do not
2044                  * usually have to add them to the flush here.  However,
2045                  * if the target_ip has set HAMMER_INODE_CONN_DOWN then
2046                  * it is asking us to flush this record (and it).
2047                  */
2048                 target_ip = rec->target_ip;
2049                 KKASSERT(target_ip != NULL);
2050                 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
2051
2052                 /*
2053                  * If the target IP is already flushing in our group
2054                  * we could associate the record, but target_ip has
2055                  * already synced ino_data to sync_ino_data and we
2056                  * would also have to adjust nlinks.   Plus there are
2057                  * ordering issues for adds and deletes.
2058                  *
2059                  * Reflush downward if this is an ADD, and upward if
2060                  * this is a DEL.
2061                  */
2062                 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
2063                         if (rec->flush_state == HAMMER_MEM_RECORD_ADD)
2064                                 ip->flags |= HAMMER_INODE_REFLUSH;
2065                         else
2066                                 target_ip->flags |= HAMMER_INODE_REFLUSH;
2067                         break;
2068                 } 
2069
2070                 /*
2071                  * Target IP is not yet flushing.  This can get complex
2072                  * because we have to be careful about the recursion.
2073                  *
2074                  * Directories create an issue for us in that if a flush
2075                  * of a directory is requested the expectation is to flush
2076                  * any pending directory entries, but this will cause the
2077                  * related inodes to recursively flush as well.  We can't
2078                  * really defer the operation so just get as many as we
2079                  * can and
2080                  */
2081 #if 0
2082                 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
2083                     (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
2084                         /*
2085                          * We aren't reclaiming and the target ip was not
2086                          * previously prevented from flushing due to this
2087                          * record dependancy.  Do not flush this record.
2088                          */
2089                         /*r = 0;*/
2090                 } else
2091 #endif
2092                 if (flg->total_count + flg->refs >
2093                            ip->hmp->undo_rec_limit) {
2094                         /*
2095                          * Our flush group is over-full and we risk blowing
2096                          * out the UNDO FIFO.  Stop the scan, flush what we
2097                          * have, then reflush the directory.
2098                          *
2099                          * The directory may be forced through multiple
2100                          * flush groups before it can be completely
2101                          * flushed.
2102                          */
2103                         ip->flags |= HAMMER_INODE_RESIGNAL |
2104                                      HAMMER_INODE_REFLUSH;
2105                         r = -1;
2106                 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
2107                         /*
2108                          * If the target IP is not flushing we can force
2109                          * it to flush, even if it is unable to write out
2110                          * any of its own records we have at least one in
2111                          * hand that we CAN deal with.
2112                          */
2113                         rec->flush_state = HAMMER_FST_FLUSH;
2114                         rec->flush_group = flg;
2115                         ++flg->refs;
2116                         hammer_ref(&rec->lock);
2117                         hammer_flush_inode_core(target_ip, flg,
2118                                                 HAMMER_FLUSH_RECURSION);
2119                         r = 1;
2120                 } else {
2121                         /*
2122                          * General or delete-on-disk record.
2123                          *
2124                          * XXX this needs help.  If a delete-on-disk we could
2125                          * disconnect the target.  If the target has its own
2126                          * dependancies they really need to be flushed.
2127                          *
2128                          * XXX
2129                          */
2130                         rec->flush_state = HAMMER_FST_FLUSH;
2131                         rec->flush_group = flg;
2132                         ++flg->refs;
2133                         hammer_ref(&rec->lock);
2134                         hammer_flush_inode_core(target_ip, flg,
2135                                                 HAMMER_FLUSH_RECURSION);
2136                         r = 1;
2137                 }
2138                 break;
2139         case HAMMER_FST_FLUSH:
2140                 /* 
2141                  * The flush_group should already match.
2142                  */
2143                 KKASSERT(rec->flush_group == flg);
2144                 r = 1;
2145                 break;
2146         }
2147         return(r);
2148 }
2149
2150 #if 0
2151 /*
2152  * This version just moves records already in a flush state to the new
2153  * flush group and that is it.
2154  */
2155 static int
2156 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
2157 {
2158         hammer_inode_t ip = rec->ip;
2159
2160         switch(rec->flush_state) {
2161         case HAMMER_FST_FLUSH:
2162                 KKASSERT(rec->flush_group == ip->flush_group);
2163                 break;
2164         default:
2165                 break;
2166         }
2167         return(0);
2168 }
2169 #endif
2170
2171 /*
2172  * Wait for a previously queued flush to complete.
2173  *
2174  * If a critical error occured we don't try to wait.
2175  */
2176 void
2177 hammer_wait_inode(hammer_inode_t ip)
2178 {
2179         hammer_flush_group_t flg;
2180
2181         flg = NULL;
2182         if ((ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2183                 while (ip->flush_state != HAMMER_FST_IDLE &&
2184                        (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2185                         if (ip->flush_state == HAMMER_FST_SETUP)
2186                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2187                         if (ip->flush_state != HAMMER_FST_IDLE) {
2188                                 ip->flags |= HAMMER_INODE_FLUSHW;
2189                                 tsleep(&ip->flags, 0, "hmrwin", 0);
2190                         }
2191                 }
2192         }
2193 }
2194
2195 /*
2196  * Called by the backend code when a flush has been completed.
2197  * The inode has already been removed from the flush list.
2198  *
2199  * A pipelined flush can occur, in which case we must re-enter the
2200  * inode on the list and re-copy its fields.
2201  */
2202 void
2203 hammer_flush_inode_done(hammer_inode_t ip, int error)
2204 {
2205         hammer_mount_t hmp;
2206         int dorel;
2207
2208         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2209
2210         hmp = ip->hmp;
2211
2212         /*
2213          * Auto-reflush if the backend could not completely flush
2214          * the inode.  This fixes a case where a deferred buffer flush
2215          * could cause fsync to return early.
2216          */
2217         if (ip->sync_flags & HAMMER_INODE_MODMASK)
2218                 ip->flags |= HAMMER_INODE_REFLUSH;
2219
2220         /*
2221          * Merge left-over flags back into the frontend and fix the state.
2222          * Incomplete truncations are retained by the backend.
2223          */
2224         ip->error = error;
2225         ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2226         ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2227
2228         /*
2229          * The backend may have adjusted nlinks, so if the adjusted nlinks
2230          * does not match the fronttend set the frontend's RDIRTY flag again.
2231          */
2232         if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2233                 ip->flags |= HAMMER_INODE_DDIRTY;
2234
2235         /*
2236          * Fix up the dirty buffer status.
2237          */
2238         if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2239                 ip->flags |= HAMMER_INODE_BUFS;
2240         }
2241
2242         /*
2243          * Re-set the XDIRTY flag if some of the inode's in-memory records
2244          * could not be flushed.
2245          */
2246         KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2247                   (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2248                  (!RB_EMPTY(&ip->rec_tree) &&
2249                   (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2250
2251         /*
2252          * Do not lose track of inodes which no longer have vnode
2253          * assocations, otherwise they may never get flushed again.
2254          *
2255          * The reflush flag can be set superfluously, causing extra pain
2256          * for no reason.  If the inode is no longer modified it no longer
2257          * needs to be flushed.
2258          */
2259         if (ip->flags & HAMMER_INODE_MODMASK) {
2260                 if (ip->vp == NULL)
2261                         ip->flags |= HAMMER_INODE_REFLUSH;
2262         } else {
2263                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2264         }
2265
2266         /*
2267          * Adjust the flush state.
2268          */
2269         if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2270                 /*
2271                  * We were unable to flush out all our records, leave the
2272                  * inode in a flush state and in the current flush group.
2273                  * The flush group will be re-run.
2274                  *
2275                  * This occurs if the UNDO block gets too full or there is
2276                  * too much dirty meta-data and allows the flusher to
2277                  * finalize the UNDO block and then re-flush.
2278                  */
2279                 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2280                 dorel = 0;
2281         } else {
2282                 /*
2283                  * Remove from the flush_group
2284                  */
2285                 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
2286                 ip->flush_group = NULL;
2287
2288                 /*
2289                  * Clean up the vnode ref and tracking counts.
2290                  */
2291                 if (ip->flags & HAMMER_INODE_VHELD) {
2292                         ip->flags &= ~HAMMER_INODE_VHELD;
2293                         vrele(ip->vp);
2294                 }
2295                 --hmp->count_iqueued;
2296                 --hammer_count_iqueued;
2297
2298                 /*
2299                  * And adjust the state.
2300                  */
2301                 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2302                         ip->flush_state = HAMMER_FST_IDLE;
2303                         dorel = 1;
2304                 } else {
2305                         ip->flush_state = HAMMER_FST_SETUP;
2306                         dorel = 0;
2307                 }
2308
2309                 /*
2310                  * If the frontend is waiting for a flush to complete,
2311                  * wake it up.
2312                  */
2313                 if (ip->flags & HAMMER_INODE_FLUSHW) {
2314                         ip->flags &= ~HAMMER_INODE_FLUSHW;
2315                         wakeup(&ip->flags);
2316                 }
2317
2318                 /*
2319                  * If the frontend made more changes and requested another
2320                  * flush, then try to get it running.
2321                  *
2322                  * Reflushes are aborted when the inode is errored out.
2323                  */
2324                 if (ip->flags & HAMMER_INODE_REFLUSH) {
2325                         ip->flags &= ~HAMMER_INODE_REFLUSH;
2326                         if (ip->flags & HAMMER_INODE_RESIGNAL) {
2327                                 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2328                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2329                         } else {
2330                                 hammer_flush_inode(ip, 0);
2331                         }
2332                 }
2333         }
2334
2335         /*
2336          * If we have no parent dependancies we can clear CONN_DOWN
2337          */
2338         if (TAILQ_EMPTY(&ip->target_list))
2339                 ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2340
2341         /*
2342          * If the inode is now clean drop the space reservation.
2343          */
2344         if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2345             (ip->flags & HAMMER_INODE_RSV_INODES)) {
2346                 ip->flags &= ~HAMMER_INODE_RSV_INODES;
2347                 --hmp->rsv_inodes;
2348         }
2349
2350         if (dorel)
2351                 hammer_rel_inode(ip, 0);
2352 }
2353
2354 /*
2355  * Called from hammer_sync_inode() to synchronize in-memory records
2356  * to the media.
2357  */
2358 static int
2359 hammer_sync_record_callback(hammer_record_t record, void *data)
2360 {
2361         hammer_cursor_t cursor = data;
2362         hammer_transaction_t trans = cursor->trans;
2363         hammer_mount_t hmp = trans->hmp;
2364         int error;
2365
2366         /*
2367          * Skip records that do not belong to the current flush.
2368          */
2369         ++hammer_stats_record_iterations;
2370         if (record->flush_state != HAMMER_FST_FLUSH)
2371                 return(0);
2372
2373 #if 1
2374         if (record->flush_group != record->ip->flush_group) {
2375                 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2376                 Debugger("blah2");
2377                 return(0);
2378         }
2379 #endif
2380         KKASSERT(record->flush_group == record->ip->flush_group);
2381
2382         /*
2383          * Interlock the record using the BE flag.  Once BE is set the
2384          * frontend cannot change the state of FE.
2385          *
2386          * NOTE: If FE is set prior to us setting BE we still sync the
2387          * record out, but the flush completion code converts it to 
2388          * a delete-on-disk record instead of destroying it.
2389          */
2390         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2391         record->flags |= HAMMER_RECF_INTERLOCK_BE;
2392
2393         /*
2394          * The backend has already disposed of the record.
2395          */
2396         if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
2397                 error = 0;
2398                 goto done;
2399         }
2400
2401         /*
2402          * If the whole inode is being deleting all on-disk records will
2403          * be deleted very soon, we can't sync any new records to disk
2404          * because they will be deleted in the same transaction they were
2405          * created in (delete_tid == create_tid), which will assert.
2406          *
2407          * XXX There may be a case with RECORD_ADD with DELETED_FE set
2408          * that we currently panic on.
2409          */
2410         if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2411                 switch(record->type) {
2412                 case HAMMER_MEM_RECORD_DATA:
2413                         /*
2414                          * We don't have to do anything, if the record was
2415                          * committed the space will have been accounted for
2416                          * in the blockmap.
2417                          */
2418                         /* fall through */
2419                 case HAMMER_MEM_RECORD_GENERAL:
2420                         /*
2421                          * Set deleted-by-backend flag.  Do not set the
2422                          * backend committed flag, because we are throwing
2423                          * the record away.
2424                          */
2425                         record->flags |= HAMMER_RECF_DELETED_BE;
2426                         ++record->ip->rec_generation;
2427                         error = 0;
2428                         goto done;
2429                 case HAMMER_MEM_RECORD_ADD:
2430                         panic("hammer_sync_record_callback: illegal add "
2431                               "during inode deletion record %p", record);
2432                         break; /* NOT REACHED */
2433                 case HAMMER_MEM_RECORD_INODE:
2434                         panic("hammer_sync_record_callback: attempt to "
2435                               "sync inode record %p?", record);
2436                         break; /* NOT REACHED */
2437                 case HAMMER_MEM_RECORD_DEL:
2438                         /* 
2439                          * Follow through and issue the on-disk deletion
2440                          */
2441                         break;
2442                 }
2443         }
2444
2445         /*
2446          * If DELETED_FE is set special handling is needed for directory
2447          * entries.  Dependant pieces related to the directory entry may
2448          * have already been synced to disk.  If this occurs we have to
2449          * sync the directory entry and then change the in-memory record
2450          * from an ADD to a DELETE to cover the fact that it's been
2451          * deleted by the frontend.
2452          *
2453          * A directory delete covering record (MEM_RECORD_DEL) can never
2454          * be deleted by the frontend.
2455          *
2456          * Any other record type (aka DATA) can be deleted by the frontend.
2457          * XXX At the moment the flusher must skip it because there may
2458          * be another data record in the flush group for the same block,
2459          * meaning that some frontend data changes can leak into the backend's
2460          * synchronization point.
2461          */
2462         if (record->flags & HAMMER_RECF_DELETED_FE) {
2463                 if (record->type == HAMMER_MEM_RECORD_ADD) {
2464                         /*
2465                          * Convert a front-end deleted directory-add to
2466                          * a directory-delete entry later.
2467                          */
2468                         record->flags |= HAMMER_RECF_CONVERT_DELETE;
2469                 } else {
2470                         /*
2471                          * Dispose of the record (race case).  Mark as
2472                          * deleted by backend (and not committed).
2473                          */
2474                         KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2475                         record->flags |= HAMMER_RECF_DELETED_BE;
2476                         ++record->ip->rec_generation;
2477                         error = 0;
2478                         goto done;
2479                 }
2480         }
2481
2482         /*
2483          * Assign the create_tid for new records.  Deletions already
2484          * have the record's entire key properly set up.
2485          */
2486         if (record->type != HAMMER_MEM_RECORD_DEL) {
2487                 record->leaf.base.create_tid = trans->tid;
2488                 record->leaf.create_ts = trans->time32;
2489         }
2490         for (;;) {
2491                 error = hammer_ip_sync_record_cursor(cursor, record);
2492                 if (error != EDEADLK)
2493                         break;
2494                 hammer_done_cursor(cursor);
2495                 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2496                                            record->ip);
2497                 if (error)
2498                         break;
2499         }
2500         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2501
2502         if (error)
2503                 error = -error;
2504 done:
2505         hammer_flush_record_done(record, error);
2506
2507         /*
2508          * Do partial finalization if we have built up too many dirty
2509          * buffers.  Otherwise a buffer cache deadlock can occur when
2510          * doing things like creating tens of thousands of tiny files.
2511          *
2512          * We must release our cursor lock to avoid a 3-way deadlock
2513          * due to the exclusive sync lock the finalizer must get.
2514          */
2515         if (hammer_flusher_meta_limit(hmp)) {
2516                 hammer_unlock_cursor(cursor);
2517                 hammer_flusher_finalize(trans, 0);
2518                 hammer_lock_cursor(cursor);
2519         }
2520
2521         return(error);
2522 }
2523
2524 /*
2525  * Backend function called by the flusher to sync an inode to media.
2526  */
2527 int
2528 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2529 {
2530         struct hammer_cursor cursor;
2531         hammer_node_t tmp_node;
2532         hammer_record_t depend;
2533         hammer_record_t next;
2534         int error, tmp_error;
2535         u_int64_t nlinks;
2536
2537         if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2538                 return(0);
2539
2540         error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2541         if (error)
2542                 goto done;
2543
2544         /*
2545          * Any directory records referencing this inode which are not in
2546          * our current flush group must adjust our nlink count for the
2547          * purposes of synchronization to disk.
2548          *
2549          * Records which are in our flush group can be unlinked from our
2550          * inode now, potentially allowing the inode to be physically
2551          * deleted.
2552          *
2553          * This cannot block.
2554          */
2555         nlinks = ip->ino_data.nlinks;
2556         next = TAILQ_FIRST(&ip->target_list);
2557         while ((depend = next) != NULL) {
2558                 next = TAILQ_NEXT(depend, target_entry);
2559                 if (depend->flush_state == HAMMER_FST_FLUSH &&
2560                     depend->flush_group == ip->flush_group) {
2561                         /*
2562                          * If this is an ADD that was deleted by the frontend
2563                          * the frontend nlinks count will have already been
2564                          * decremented, but the backend is going to sync its
2565                          * directory entry and must account for it.  The
2566                          * record will be converted to a delete-on-disk when
2567                          * it gets synced.
2568                          *
2569                          * If the ADD was not deleted by the frontend we
2570                          * can remove the dependancy from our target_list.
2571                          */
2572                         if (depend->flags & HAMMER_RECF_DELETED_FE) {
2573                                 ++nlinks;
2574                         } else {
2575                                 TAILQ_REMOVE(&ip->target_list, depend,
2576                                              target_entry);
2577                                 depend->target_ip = NULL;
2578                         }
2579                 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2580                         /*
2581                          * Not part of our flush group and not deleted by
2582                          * the front-end, adjust the link count synced to
2583                          * the media (undo what the frontend did when it
2584                          * queued the record).
2585                          */
2586                         KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2587                         switch(depend->type) {
2588                         case HAMMER_MEM_RECORD_ADD:
2589                                 --nlinks;
2590                                 break;
2591                         case HAMMER_MEM_RECORD_DEL:
2592                                 ++nlinks;
2593                                 break;
2594                         default:
2595                                 break;
2596                         }
2597                 }
2598         }
2599
2600         /*
2601          * Set dirty if we had to modify the link count.
2602          */
2603         if (ip->sync_ino_data.nlinks != nlinks) {
2604                 KKASSERT((int64_t)nlinks >= 0);
2605                 ip->sync_ino_data.nlinks = nlinks;
2606                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2607         }
2608
2609         /*
2610          * If there is a trunction queued destroy any data past the (aligned)
2611          * truncation point.  Userland will have dealt with the buffer
2612          * containing the truncation point for us.
2613          *
2614          * We don't flush pending frontend data buffers until after we've
2615          * dealt with the truncation.
2616          */
2617         if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2618                 /*
2619                  * Interlock trunc_off.  The VOP front-end may continue to
2620                  * make adjustments to it while we are blocked.
2621                  */
2622                 off_t trunc_off;
2623                 off_t aligned_trunc_off;
2624                 int blkmask;
2625
2626                 trunc_off = ip->sync_trunc_off;
2627                 blkmask = hammer_blocksize(trunc_off) - 1;
2628                 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
2629
2630                 /*
2631                  * Delete any whole blocks on-media.  The front-end has
2632                  * already cleaned out any partial block and made it
2633                  * pending.  The front-end may have updated trunc_off
2634                  * while we were blocked so we only use sync_trunc_off.
2635                  *
2636                  * This operation can blow out the buffer cache, EWOULDBLOCK
2637                  * means we were unable to complete the deletion.  The
2638                  * deletion will update sync_trunc_off in that case.
2639                  */
2640                 error = hammer_ip_delete_range(&cursor, ip,
2641                                                 aligned_trunc_off,
2642                                                 0x7FFFFFFFFFFFFFFFLL, 2);
2643                 if (error == EWOULDBLOCK) {
2644                         ip->flags |= HAMMER_INODE_WOULDBLOCK;
2645                         error = 0;
2646                         goto defer_buffer_flush;
2647                 }
2648
2649                 if (error)
2650                         goto done;
2651
2652                 /*
2653                  * Clear the truncation flag on the backend after we have
2654                  * complete the deletions.  Backend data is now good again
2655                  * (including new records we are about to sync, below).
2656                  *
2657                  * Leave sync_trunc_off intact.  As we write additional
2658                  * records the backend will update sync_trunc_off.  This
2659                  * tells the backend whether it can skip the overwrite
2660                  * test.  This should work properly even when the backend
2661                  * writes full blocks where the truncation point straddles
2662                  * the block because the comparison is against the base
2663                  * offset of the record.
2664                  */
2665                 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2666                 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2667         } else {
2668                 error = 0;
2669         }
2670
2671         /*
2672          * Now sync related records.  These will typically be directory
2673          * entries, records tracking direct-writes, or delete-on-disk records.
2674          */
2675         if (error == 0) {
2676                 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2677                                     hammer_sync_record_callback, &cursor);
2678                 if (tmp_error < 0)
2679                         tmp_error = -error;
2680                 if (tmp_error)
2681                         error = tmp_error;
2682         }
2683         hammer_cache_node(&ip->cache[1], cursor.node);
2684
2685         /*
2686          * Re-seek for inode update, assuming our cache hasn't been ripped
2687          * out from under us.
2688          */
2689         if (error == 0) {
2690                 tmp_node = hammer_ref_node_safe(trans, &ip->cache[0], &error);
2691                 if (tmp_node) {
2692                         hammer_cursor_downgrade(&cursor);
2693                         hammer_lock_sh(&tmp_node->lock);
2694                         if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
2695                                 hammer_cursor_seek(&cursor, tmp_node, 0);
2696                         hammer_unlock(&tmp_node->lock);
2697                         hammer_rel_node(tmp_node);
2698                 }
2699                 error = 0;
2700         }
2701
2702         /*
2703          * If we are deleting the inode the frontend had better not have
2704          * any active references on elements making up the inode.
2705          *
2706          * The call to hammer_ip_delete_clean() cleans up auxillary records
2707          * but not DB or DATA records.  Those must have already been deleted
2708          * by the normal truncation mechanic.
2709          */
2710         if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
2711                 RB_EMPTY(&ip->rec_tree)  &&
2712             (ip->sync_flags & HAMMER_INODE_DELETING) &&
2713             (ip->flags & HAMMER_INODE_DELETED) == 0) {
2714                 int count1 = 0;
2715
2716                 error = hammer_ip_delete_clean(&cursor, ip, &count1);
2717                 if (error == 0) {
2718                         ip->flags |= HAMMER_INODE_DELETED;
2719                         ip->sync_flags &= ~HAMMER_INODE_DELETING;
2720                         ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2721                         KKASSERT(RB_EMPTY(&ip->rec_tree));
2722
2723                         /*
2724                          * Set delete_tid in both the frontend and backend
2725                          * copy of the inode record.  The DELETED flag handles
2726                          * this, do not set RDIRTY.
2727                          */
2728                         ip->ino_leaf.base.delete_tid = trans->tid;
2729                         ip->sync_ino_leaf.base.delete_tid = trans->tid;
2730                         ip->ino_leaf.delete_ts = trans->time32;
2731                         ip->sync_ino_leaf.delete_ts = trans->time32;
2732
2733
2734                         /*
2735                          * Adjust the inode count in the volume header
2736                          */
2737                         hammer_sync_lock_sh(trans);
2738                         if (ip->flags & HAMMER_INODE_ONDISK) {
2739                                 hammer_modify_volume_field(trans,
2740                                                            trans->rootvol,
2741                                                            vol0_stat_inodes);
2742                                 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
2743                                 hammer_modify_volume_done(trans->rootvol);
2744                         }
2745                         hammer_sync_unlock(trans);
2746                 }
2747         }
2748
2749         if (error)
2750                 goto done;
2751         ip->sync_flags &= ~HAMMER_INODE_BUFS;
2752
2753 defer_buffer_flush:
2754         /*
2755          * Now update the inode's on-disk inode-data and/or on-disk record.
2756          * DELETED and ONDISK are managed only in ip->flags.
2757          *
2758          * In the case of a defered buffer flush we still update the on-disk
2759          * inode to satisfy visibility requirements if there happen to be
2760          * directory dependancies.
2761          */
2762         switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
2763         case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
2764                 /*
2765                  * If deleted and on-disk, don't set any additional flags.
2766                  * the delete flag takes care of things.
2767                  *
2768                  * Clear flags which may have been set by the frontend.
2769                  */
2770                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2771                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2772                                     HAMMER_INODE_DELETING);
2773                 break;
2774         case HAMMER_INODE_DELETED:
2775                 /*
2776                  * Take care of the case where a deleted inode was never
2777                  * flushed to the disk in the first place.
2778                  *
2779                  * Clear flags which may have been set by the frontend.
2780                  */
2781                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2782                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2783                                     HAMMER_INODE_DELETING);
2784                 while (RB_ROOT(&ip->rec_tree)) {
2785                         hammer_record_t record = RB_ROOT(&ip->rec_tree);
2786                         hammer_ref(&record->lock);
2787                         KKASSERT(record->lock.refs == 1);
2788                         record->flags |= HAMMER_RECF_DELETED_BE;
2789                         ++record->ip->rec_generation;
2790                         hammer_rel_mem_record(record);
2791                 }
2792                 break;
2793         case HAMMER_INODE_ONDISK:
2794                 /*
2795                  * If already on-disk, do not set any additional flags.
2796                  */
2797                 break;
2798         default:
2799                 /*
2800                  * If not on-disk and not deleted, set DDIRTY to force
2801                  * an initial record to be written.
2802                  *
2803                  * Also set the create_tid in both the frontend and backend
2804                  * copy of the inode record.
2805                  */
2806                 ip->ino_leaf.base.create_tid = trans->tid;
2807                 ip->ino_leaf.create_ts = trans->time32;
2808                 ip->sync_ino_leaf.base.create_tid = trans->tid;
2809                 ip->sync_ino_leaf.create_ts = trans->time32;
2810                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2811                 break;
2812         }
2813
2814         /*
2815          * If RDIRTY or DDIRTY is set, write out a new record.  If the inode
2816          * is already on-disk the old record is marked as deleted.
2817          *
2818          * If DELETED is set hammer_update_inode() will delete the existing
2819          * record without writing out a new one.
2820          *
2821          * If *ONLY* the ITIMES flag is set we can update the record in-place.
2822          */
2823         if (ip->flags & HAMMER_INODE_DELETED) {
2824                 error = hammer_update_inode(&cursor, ip);
2825         } else 
2826         if ((ip->sync_flags & HAMMER_INODE_DDIRTY) == 0 &&
2827             (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
2828                 error = hammer_update_itimes(&cursor, ip);
2829         } else
2830         if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
2831                 error = hammer_update_inode(&cursor, ip);
2832         }
2833 done:
2834         if (error) {
2835                 hammer_critical_error(ip->hmp, ip, error,
2836                                       "while syncing inode");
2837         }
2838         hammer_done_cursor(&cursor);
2839         return(error);
2840 }
2841
2842 /*
2843  * This routine is called when the OS is no longer actively referencing
2844  * the inode (but might still be keeping it cached), or when releasing
2845  * the last reference to an inode.
2846  *
2847  * At this point if the inode's nlinks count is zero we want to destroy
2848  * it, which may mean destroying it on-media too.
2849  */
2850 void
2851 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
2852 {
2853         struct vnode *vp;
2854
2855         /*
2856          * Set the DELETING flag when the link count drops to 0 and the
2857          * OS no longer has any opens on the inode.
2858          *
2859          * The backend will clear DELETING (a mod flag) and set DELETED
2860          * (a state flag) when it is actually able to perform the
2861          * operation.
2862          *
2863          * Don't reflag the deletion if the flusher is currently syncing
2864          * one that was already flagged.  A previously set DELETING flag
2865          * may bounce around flags and sync_flags until the operation is
2866          * completely done.
2867          */
2868         if (ip->ino_data.nlinks == 0 &&
2869             ((ip->flags | ip->sync_flags) & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
2870                 ip->flags |= HAMMER_INODE_DELETING;
2871                 ip->flags |= HAMMER_INODE_TRUNCATED;
2872                 ip->trunc_off = 0;
2873                 vp = NULL;
2874                 if (getvp) {
2875                         if (hammer_get_vnode(ip, &vp) != 0)
2876                                 return;
2877                 }
2878
2879                 /*
2880                  * Final cleanup
2881                  */
2882                 if (ip->vp) {
2883                         vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
2884                         vnode_pager_setsize(ip->vp, 0);
2885                 }
2886                 if (getvp) {
2887                         vput(vp);
2888                 }
2889         }
2890 }
2891
2892 /*
2893  * After potentially resolving a dependancy the inode is tested
2894  * to determine whether it needs to be reflushed.
2895  */
2896 void
2897 hammer_test_inode(hammer_inode_t ip)
2898 {
2899         if (ip->flags & HAMMER_INODE_REFLUSH) {
2900                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2901                 hammer_ref(&ip->lock);
2902                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2903                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
2904                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2905                 } else {
2906                         hammer_flush_inode(ip, 0);
2907                 }
2908                 hammer_rel_inode(ip, 0);
2909         }
2910 }
2911
2912 /*
2913  * Clear the RECLAIM flag on an inode.  This occurs when the inode is
2914  * reassociated with a vp or just before it gets freed.
2915  *
2916  * Pipeline wakeups to threads blocked due to an excessive number of
2917  * detached inodes.  The reclaim count generates a bit of negative
2918  * feedback.
2919  */
2920 static void
2921 hammer_inode_wakereclaims(hammer_inode_t ip, int dowake)
2922 {
2923         struct hammer_reclaim *reclaim;
2924         hammer_mount_t hmp = ip->hmp;
2925
2926         if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
2927                 return;
2928
2929         --hammer_count_reclaiming;
2930         --hmp->inode_reclaims;
2931         ip->flags &= ~HAMMER_INODE_RECLAIM;
2932
2933         if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT || dowake) {
2934                 reclaim = TAILQ_FIRST(&hmp->reclaim_list);
2935                 if (reclaim && reclaim->count > 0 && --reclaim->count == 0) {
2936                         TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
2937                         wakeup(reclaim);
2938                 }
2939         }
2940 }
2941
2942 /*
2943  * Setup our reclaim pipeline.  We only let so many detached (and dirty)
2944  * inodes build up before we start blocking.
2945  *
2946  * When we block we don't care *which* inode has finished reclaiming,
2947  * as lone as one does.  This is somewhat heuristical... we also put a
2948  * cap on how long we are willing to wait.
2949  */
2950 void
2951 hammer_inode_waitreclaims(hammer_mount_t hmp)
2952 {
2953         struct hammer_reclaim reclaim;
2954         int delay;
2955
2956         if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT)
2957                 return;
2958         delay = (hmp->inode_reclaims - HAMMER_RECLAIM_WAIT) * hz /
2959                 (HAMMER_RECLAIM_WAIT * 3) + 1;
2960         if (delay > 0) {
2961                 reclaim.count = 2;
2962                 TAILQ_INSERT_TAIL(&hmp->reclaim_list, &reclaim, entry);
2963                 tsleep(&reclaim, 0, "hmrrcm", delay);
2964                 if (reclaim.count > 0)
2965                         TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
2966         }
2967 }
2968
2969 /*
2970  * A larger then normal backlog of inodes is sitting in the flusher,
2971  * enforce a general slowdown to let it catch up.  This routine is only
2972  * called on completion of a non-flusher-related transaction which
2973  * performed B-Tree node I/O.
2974  *
2975  * It is possible for the flusher to stall in a continuous load.
2976  * blogbench -i1000 -o seems to do a good job generating this sort of load.
2977  * If the flusher is unable to catch up the inode count can bloat until
2978  * we run out of kvm.
2979  *
2980  * This is a bit of a hack.
2981  */
2982 void
2983 hammer_inode_waithard(hammer_mount_t hmp)
2984 {
2985         /*
2986          * Hysteresis.
2987          */
2988         if (hmp->flags & HAMMER_MOUNT_FLUSH_RECOVERY) {
2989                 if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT / 2 &&
2990                     hmp->count_iqueued < hmp->count_inodes / 20) {
2991                         hmp->flags &= ~HAMMER_MOUNT_FLUSH_RECOVERY;
2992                         return;
2993                 }
2994         } else {
2995                 if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT ||
2996                     hmp->count_iqueued < hmp->count_inodes / 10) {
2997                         return;
2998                 }
2999                 hmp->flags |= HAMMER_MOUNT_FLUSH_RECOVERY;
3000         }
3001
3002         /*
3003          * Block for one flush cycle.
3004          */
3005         hammer_flusher_wait_next(hmp);
3006 }
3007