231883e426d0e35ee1b12f9b69ce939f58a26bf7
[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
40 static int      hammer_unload_inode(struct hammer_inode *ip);
41 static void     hammer_free_inode(hammer_inode_t ip);
42 static void     hammer_flush_inode_core(hammer_inode_t ip,
43                                         hammer_flush_group_t flg, int flags);
44 static int      hammer_setup_child_callback(hammer_record_t rec, void *data);
45 #if 0
46 static int      hammer_syncgrp_child_callback(hammer_record_t rec, void *data);
47 #endif
48 static int      hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
49                                         hammer_flush_group_t flg);
50 static int      hammer_setup_parent_inodes_helper(hammer_record_t record,
51                                         int depth, hammer_flush_group_t flg);
52 static void     hammer_inode_wakereclaims(hammer_inode_t ip);
53 static struct hammer_inostats *hammer_inode_inostats(hammer_mount_t hmp,
54                                         pid_t pid);
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 int
82 hammer_redo_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
83 {
84         if (ip1->redo_fifo_start < ip2->redo_fifo_start)
85                 return(-1);
86         if (ip1->redo_fifo_start > ip2->redo_fifo_start)
87                 return(1);
88         return(0);
89 }
90
91 /*
92  * RB-Tree support for inode structures / special LOOKUP_INFO
93  */
94 static int
95 hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
96 {
97         if (info->obj_localization < ip->obj_localization)
98                 return(-1);
99         if (info->obj_localization > ip->obj_localization)
100                 return(1);
101         if (info->obj_id < ip->obj_id)
102                 return(-1);
103         if (info->obj_id > ip->obj_id)
104                 return(1);
105         if (info->obj_asof < ip->obj_asof)
106                 return(-1);
107         if (info->obj_asof > ip->obj_asof)
108                 return(1);
109         return(0);
110 }
111
112 /*
113  * Used by hammer_scan_inode_snapshots() to locate all of an object's
114  * snapshots.  Note that the asof field is not tested, which we can get
115  * away with because it is the lowest-priority field.
116  */
117 static int
118 hammer_inode_info_cmp_all_history(hammer_inode_t ip, void *data)
119 {
120         hammer_inode_info_t info = data;
121
122         if (ip->obj_localization > info->obj_localization)
123                 return(1);
124         if (ip->obj_localization < info->obj_localization)
125                 return(-1);
126         if (ip->obj_id > info->obj_id)
127                 return(1);
128         if (ip->obj_id < info->obj_id)
129                 return(-1);
130         return(0);
131 }
132
133 /*
134  * Used by hammer_unload_pseudofs() to locate all inodes associated with
135  * a particular PFS.
136  */
137 static int
138 hammer_inode_pfs_cmp(hammer_inode_t ip, void *data)
139 {
140         u_int32_t localization = *(u_int32_t *)data;
141         if (ip->obj_localization > localization)
142                 return(1);
143         if (ip->obj_localization < localization)
144                 return(-1);
145         return(0);
146 }
147
148 /*
149  * RB-Tree support for pseudofs structures
150  */
151 static int
152 hammer_pfs_rb_compare(hammer_pseudofs_inmem_t p1, hammer_pseudofs_inmem_t p2)
153 {
154         if (p1->localization < p2->localization)
155                 return(-1);
156         if (p1->localization > p2->localization)
157                 return(1);
158         return(0);
159 }
160
161
162 RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
163 RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
164                 hammer_inode_info_cmp, hammer_inode_info_t);
165 RB_GENERATE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
166              hammer_pfs_rb_compare, u_int32_t, localization);
167
168 /*
169  * The kernel is not actively referencing this vnode but is still holding
170  * it cached.
171  *
172  * This is called from the frontend.
173  *
174  * MPALMOSTSAFE
175  */
176 int
177 hammer_vop_inactive(struct vop_inactive_args *ap)
178 {
179         struct hammer_inode *ip = VTOI(ap->a_vp);
180         hammer_mount_t hmp;
181
182         /*
183          * Degenerate case
184          */
185         if (ip == NULL) {
186                 vrecycle(ap->a_vp);
187                 return(0);
188         }
189
190         /*
191          * If the inode no longer has visibility in the filesystem try to
192          * recycle it immediately, even if the inode is dirty.  Recycling
193          * it quickly allows the system to reclaim buffer cache and VM
194          * resources which can matter a lot in a heavily loaded system.
195          *
196          * This can deadlock in vfsync() if we aren't careful.
197          * 
198          * Do not queue the inode to the flusher if we still have visibility,
199          * otherwise namespace calls such as chmod will unnecessarily generate
200          * multiple inode updates.
201          */
202         if (ip->ino_data.nlinks == 0) {
203                 hmp = ip->hmp;
204                 lwkt_gettoken(&hmp->fs_token);
205                 hammer_inode_unloadable_check(ip, 0);
206                 if (ip->flags & HAMMER_INODE_MODMASK)
207                         hammer_flush_inode(ip, 0);
208                 lwkt_reltoken(&hmp->fs_token);
209                 vrecycle(ap->a_vp);
210         }
211         return(0);
212 }
213
214 /*
215  * Release the vnode association.  This is typically (but not always)
216  * the last reference on the inode.
217  *
218  * Once the association is lost we are on our own with regards to
219  * flushing the inode.
220  *
221  * We must interlock ip->vp so hammer_get_vnode() can avoid races.
222  */
223 int
224 hammer_vop_reclaim(struct vop_reclaim_args *ap)
225 {
226         struct hammer_inode *ip;
227         hammer_mount_t hmp;
228         struct vnode *vp;
229
230         vp = ap->a_vp;
231
232         if ((ip = vp->v_data) != NULL) {
233                 hmp = ip->hmp;
234                 lwkt_gettoken(&hmp->fs_token);
235                 hammer_lock_ex(&ip->lock);
236                 vp->v_data = NULL;
237                 ip->vp = NULL;
238
239                 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0) {
240                         ++hammer_count_reclaiming;
241                         ++hmp->inode_reclaims;
242                         ip->flags |= HAMMER_INODE_RECLAIM;
243                 }
244                 hammer_unlock(&ip->lock);
245                 hammer_rel_inode(ip, 1);
246                 lwkt_reltoken(&hmp->fs_token);
247         }
248         return(0);
249 }
250
251 /*
252  * Return a locked vnode for the specified inode.  The inode must be
253  * referenced but NOT LOCKED on entry and will remain referenced on
254  * return.
255  *
256  * Called from the frontend.
257  */
258 int
259 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
260 {
261         hammer_mount_t hmp;
262         struct vnode *vp;
263         int error = 0;
264         u_int8_t obj_type;
265
266         hmp = ip->hmp;
267
268         for (;;) {
269                 if ((vp = ip->vp) == NULL) {
270                         error = getnewvnode(VT_HAMMER, hmp->mp, vpp, 0, 0);
271                         if (error)
272                                 break;
273                         hammer_lock_ex(&ip->lock);
274                         if (ip->vp != NULL) {
275                                 hammer_unlock(&ip->lock);
276                                 vp = *vpp;
277                                 vp->v_type = VBAD;
278                                 vx_put(vp);
279                                 continue;
280                         }
281                         hammer_ref(&ip->lock);
282                         vp = *vpp;
283                         ip->vp = vp;
284
285                         obj_type = ip->ino_data.obj_type;
286                         vp->v_type = hammer_get_vnode_type(obj_type);
287
288                         hammer_inode_wakereclaims(ip);
289
290                         switch(ip->ino_data.obj_type) {
291                         case HAMMER_OBJTYPE_CDEV:
292                         case HAMMER_OBJTYPE_BDEV:
293                                 vp->v_ops = &hmp->mp->mnt_vn_spec_ops;
294                                 addaliasu(vp, ip->ino_data.rmajor,
295                                           ip->ino_data.rminor);
296                                 break;
297                         case HAMMER_OBJTYPE_FIFO:
298                                 vp->v_ops = &hmp->mp->mnt_vn_fifo_ops;
299                                 break;
300                         case HAMMER_OBJTYPE_REGFILE:
301                                 break;
302                         default:
303                                 break;
304                         }
305
306                         /*
307                          * Only mark as the root vnode if the ip is not
308                          * historical, otherwise the VFS cache will get
309                          * confused.  The other half of the special handling
310                          * is in hammer_vop_nlookupdotdot().
311                          *
312                          * Pseudo-filesystem roots can be accessed via
313                          * non-root filesystem paths and setting VROOT may
314                          * confuse the namecache.  Set VPFSROOT instead.
315                          */
316                         if (ip->obj_id == HAMMER_OBJID_ROOT &&
317                             ip->obj_asof == hmp->asof) {
318                                 if (ip->obj_localization == 0)
319                                         vsetflags(vp, VROOT);
320                                 else
321                                         vsetflags(vp, VPFSROOT);
322                         }
323
324                         vp->v_data = (void *)ip;
325                         /* vnode locked by getnewvnode() */
326                         /* make related vnode dirty if inode dirty? */
327                         hammer_unlock(&ip->lock);
328                         if (vp->v_type == VREG) {
329                                 vinitvmio(vp, ip->ino_data.size,
330                                           hammer_blocksize(ip->ino_data.size),
331                                           hammer_blockoff(ip->ino_data.size));
332                         }
333                         break;
334                 }
335
336                 /*
337                  * Interlock vnode clearing.  This does not prevent the
338                  * vnode from going into a reclaimed state but it does
339                  * prevent it from being destroyed or reused so the vget()
340                  * will properly fail.
341                  */
342                 hammer_lock_ex(&ip->lock);
343                 if ((vp = ip->vp) == NULL) {
344                         hammer_unlock(&ip->lock);
345                         continue;
346                 }
347                 vhold_interlocked(vp);
348                 hammer_unlock(&ip->lock);
349
350                 /*
351                  * loop if the vget fails (aka races), or if the vp
352                  * no longer matches ip->vp.
353                  */
354                 if (vget(vp, LK_EXCLUSIVE) == 0) {
355                         if (vp == ip->vp) {
356                                 vdrop(vp);
357                                 break;
358                         }
359                         vput(vp);
360                 }
361                 vdrop(vp);
362         }
363         *vpp = vp;
364         return(error);
365 }
366
367 /*
368  * Locate all copies of the inode for obj_id compatible with the specified
369  * asof, reference, and issue the related call-back.  This routine is used
370  * for direct-io invalidation and does not create any new inodes.
371  */
372 void
373 hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo,
374                             int (*callback)(hammer_inode_t ip, void *data),
375                             void *data)
376 {
377         hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root,
378                                    hammer_inode_info_cmp_all_history,
379                                    callback, iinfo);
380 }
381
382 /*
383  * Acquire a HAMMER inode.  The returned inode is not locked.  These functions
384  * do not attach or detach the related vnode (use hammer_get_vnode() for
385  * that).
386  *
387  * The flags argument is only applied for newly created inodes, and only
388  * certain flags are inherited.
389  *
390  * Called from the frontend.
391  */
392 struct hammer_inode *
393 hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip,
394                  int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
395                  int flags, int *errorp)
396 {
397         hammer_mount_t hmp = trans->hmp;
398         struct hammer_node_cache *cachep;
399         struct hammer_inode_info iinfo;
400         struct hammer_cursor cursor;
401         struct hammer_inode *ip;
402
403
404         /*
405          * Determine if we already have an inode cached.  If we do then
406          * we are golden.
407          *
408          * If we find an inode with no vnode we have to mark the
409          * transaction such that hammer_inode_waitreclaims() is
410          * called later on to avoid building up an infinite number
411          * of inodes.  Otherwise we can continue to * add new inodes
412          * faster then they can be disposed of, even with the tsleep
413          * delay.
414          *
415          * If we find a dummy inode we return a failure so dounlink
416          * (which does another lookup) doesn't try to mess with the
417          * link count.  hammer_vop_nresolve() uses hammer_get_dummy_inode()
418          * to ref dummy inodes.
419          */
420         iinfo.obj_id = obj_id;
421         iinfo.obj_asof = asof;
422         iinfo.obj_localization = localization;
423 loop:
424         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
425         if (ip) {
426                 if (ip->flags & HAMMER_INODE_DUMMY) {
427                         *errorp = ENOENT;
428                         return(NULL);
429                 }
430                 hammer_ref(&ip->lock);
431                 *errorp = 0;
432                 return(ip);
433         }
434
435         /*
436          * Allocate a new inode structure and deal with races later.
437          */
438         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
439         ++hammer_count_inodes;
440         ++hmp->count_inodes;
441         ip->obj_id = obj_id;
442         ip->obj_asof = iinfo.obj_asof;
443         ip->obj_localization = localization;
444         ip->hmp = hmp;
445         ip->flags = flags & HAMMER_INODE_RO;
446         ip->cache[0].ip = ip;
447         ip->cache[1].ip = ip;
448         ip->cache[2].ip = ip;
449         ip->cache[3].ip = ip;
450         if (hmp->ronly)
451                 ip->flags |= HAMMER_INODE_RO;
452         ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
453                 0x7FFFFFFFFFFFFFFFLL;
454         RB_INIT(&ip->rec_tree);
455         TAILQ_INIT(&ip->target_list);
456         hammer_ref(&ip->lock);
457
458         /*
459          * Locate the on-disk inode.  If this is a PFS root we always
460          * access the current version of the root inode and (if it is not
461          * a master) always access information under it with a snapshot
462          * TID.
463          *
464          * We cache recent inode lookups in this directory in dip->cache[2].
465          * If we can't find it we assume the inode we are looking for is
466          * close to the directory inode.
467          */
468 retry:
469         cachep = NULL;
470         if (dip) {
471                 if (dip->cache[2].node)
472                         cachep = &dip->cache[2];
473                 else
474                         cachep = &dip->cache[0];
475         }
476         hammer_init_cursor(trans, &cursor, cachep, NULL);
477         cursor.key_beg.localization = localization + HAMMER_LOCALIZE_INODE;
478         cursor.key_beg.obj_id = ip->obj_id;
479         cursor.key_beg.key = 0;
480         cursor.key_beg.create_tid = 0;
481         cursor.key_beg.delete_tid = 0;
482         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
483         cursor.key_beg.obj_type = 0;
484
485         cursor.asof = iinfo.obj_asof;
486         cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
487                        HAMMER_CURSOR_ASOF;
488
489         *errorp = hammer_btree_lookup(&cursor);
490         if (*errorp == EDEADLK) {
491                 hammer_done_cursor(&cursor);
492                 goto retry;
493         }
494
495         /*
496          * On success the B-Tree lookup will hold the appropriate
497          * buffer cache buffers and provide a pointer to the requested
498          * information.  Copy the information to the in-memory inode
499          * and cache the B-Tree node to improve future operations.
500          */
501         if (*errorp == 0) {
502                 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
503                 ip->ino_data = cursor.data->inode;
504
505                 /*
506                  * cache[0] tries to cache the location of the object inode.
507                  * The assumption is that it is near the directory inode.
508                  *
509                  * cache[1] tries to cache the location of the object data.
510                  * We might have something in the governing directory from
511                  * scan optimizations (see the strategy code in
512                  * hammer_vnops.c).
513                  *
514                  * We update dip->cache[2], if possible, with the location
515                  * of the object inode for future directory shortcuts.
516                  */
517                 hammer_cache_node(&ip->cache[0], cursor.node);
518                 if (dip) {
519                         if (dip->cache[3].node) {
520                                 hammer_cache_node(&ip->cache[1],
521                                                   dip->cache[3].node);
522                         }
523                         hammer_cache_node(&dip->cache[2], cursor.node);
524                 }
525
526                 /*
527                  * The file should not contain any data past the file size
528                  * stored in the inode.  Setting save_trunc_off to the
529                  * file size instead of max reduces B-Tree lookup overheads
530                  * on append by allowing the flusher to avoid checking for
531                  * record overwrites.
532                  */
533                 ip->save_trunc_off = ip->ino_data.size;
534
535                 /*
536                  * Locate and assign the pseudofs management structure to
537                  * the inode.
538                  */
539                 if (dip && dip->obj_localization == ip->obj_localization) {
540                         ip->pfsm = dip->pfsm;
541                         hammer_ref(&ip->pfsm->lock);
542                 } else {
543                         ip->pfsm = hammer_load_pseudofs(trans,
544                                                         ip->obj_localization,
545                                                         errorp);
546                         *errorp = 0;    /* ignore ENOENT */
547                 }
548         }
549
550         /*
551          * The inode is placed on the red-black tree and will be synced to
552          * the media when flushed or by the filesystem sync.  If this races
553          * another instantiation/lookup the insertion will fail.
554          */
555         if (*errorp == 0) {
556                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
557                         hammer_free_inode(ip);
558                         hammer_done_cursor(&cursor);
559                         goto loop;
560                 }
561                 ip->flags |= HAMMER_INODE_ONDISK;
562         } else {
563                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
564                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
565                         --hmp->rsv_inodes;
566                 }
567
568                 hammer_free_inode(ip);
569                 ip = NULL;
570         }
571         hammer_done_cursor(&cursor);
572
573         /*
574          * NEWINODE is only set if the inode becomes dirty later,
575          * setting it here just leads to unnecessary stalls.
576          *
577          * trans->flags |= HAMMER_TRANSF_NEWINODE;
578          */
579         return (ip);
580 }
581
582 /*
583  * Get a dummy inode to placemark a broken directory entry.
584  */
585 struct hammer_inode *
586 hammer_get_dummy_inode(hammer_transaction_t trans, hammer_inode_t dip,
587                  int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
588                  int flags, int *errorp)
589 {
590         hammer_mount_t hmp = trans->hmp;
591         struct hammer_inode_info iinfo;
592         struct hammer_inode *ip;
593
594         /*
595          * Determine if we already have an inode cached.  If we do then
596          * we are golden.
597          *
598          * If we find an inode with no vnode we have to mark the
599          * transaction such that hammer_inode_waitreclaims() is
600          * called later on to avoid building up an infinite number
601          * of inodes.  Otherwise we can continue to * add new inodes
602          * faster then they can be disposed of, even with the tsleep
603          * delay.
604          *
605          * If we find a non-fake inode we return an error.  Only fake
606          * inodes can be returned by this routine.
607          */
608         iinfo.obj_id = obj_id;
609         iinfo.obj_asof = asof;
610         iinfo.obj_localization = localization;
611 loop:
612         *errorp = 0;
613         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
614         if (ip) {
615                 if ((ip->flags & HAMMER_INODE_DUMMY) == 0) {
616                         *errorp = ENOENT;
617                         return(NULL);
618                 }
619                 hammer_ref(&ip->lock);
620                 return(ip);
621         }
622
623         /*
624          * Allocate a new inode structure and deal with races later.
625          */
626         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
627         ++hammer_count_inodes;
628         ++hmp->count_inodes;
629         ip->obj_id = obj_id;
630         ip->obj_asof = iinfo.obj_asof;
631         ip->obj_localization = localization;
632         ip->hmp = hmp;
633         ip->flags = flags | HAMMER_INODE_RO | HAMMER_INODE_DUMMY;
634         ip->cache[0].ip = ip;
635         ip->cache[1].ip = ip;
636         ip->cache[2].ip = ip;
637         ip->cache[3].ip = ip;
638         ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
639                 0x7FFFFFFFFFFFFFFFLL;
640         RB_INIT(&ip->rec_tree);
641         TAILQ_INIT(&ip->target_list);
642         hammer_ref(&ip->lock);
643
644         /*
645          * Populate the dummy inode.  Leave everything zero'd out.
646          *
647          * (ip->ino_leaf and ip->ino_data)
648          *
649          * Make the dummy inode a FIFO object which most copy programs
650          * will properly ignore.
651          */
652         ip->save_trunc_off = ip->ino_data.size;
653         ip->ino_data.obj_type = HAMMER_OBJTYPE_FIFO;
654
655         /*
656          * Locate and assign the pseudofs management structure to
657          * the inode.
658          */
659         if (dip && dip->obj_localization == ip->obj_localization) {
660                 ip->pfsm = dip->pfsm;
661                 hammer_ref(&ip->pfsm->lock);
662         } else {
663                 ip->pfsm = hammer_load_pseudofs(trans, ip->obj_localization,
664                                                 errorp);
665                 *errorp = 0;    /* ignore ENOENT */
666         }
667
668         /*
669          * The inode is placed on the red-black tree and will be synced to
670          * the media when flushed or by the filesystem sync.  If this races
671          * another instantiation/lookup the insertion will fail.
672          *
673          * NOTE: Do not set HAMMER_INODE_ONDISK.  The inode is a fake.
674          */
675         if (*errorp == 0) {
676                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
677                         hammer_free_inode(ip);
678                         goto loop;
679                 }
680         } else {
681                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
682                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
683                         --hmp->rsv_inodes;
684                 }
685                 hammer_free_inode(ip);
686                 ip = NULL;
687         }
688         trans->flags |= HAMMER_TRANSF_NEWINODE;
689         return (ip);
690 }
691
692 /*
693  * Return a referenced inode only if it is in our inode cache.
694  *
695  * Dummy inodes do not count.
696  */
697 struct hammer_inode *
698 hammer_find_inode(hammer_transaction_t trans, int64_t obj_id,
699                   hammer_tid_t asof, u_int32_t localization)
700 {
701         hammer_mount_t hmp = trans->hmp;
702         struct hammer_inode_info iinfo;
703         struct hammer_inode *ip;
704
705         iinfo.obj_id = obj_id;
706         iinfo.obj_asof = asof;
707         iinfo.obj_localization = localization;
708
709         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
710         if (ip) {
711                 if (ip->flags & HAMMER_INODE_DUMMY)
712                         ip = NULL;
713                 else
714                         hammer_ref(&ip->lock);
715         }
716         return(ip);
717 }
718
719 /*
720  * Create a new filesystem object, returning the inode in *ipp.  The
721  * returned inode will be referenced.  The inode is created in-memory.
722  *
723  * If pfsm is non-NULL the caller wishes to create the root inode for
724  * a master PFS.
725  */
726 int
727 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
728                     struct ucred *cred,
729                     hammer_inode_t dip, const char *name, int namelen,
730                     hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
731 {
732         hammer_mount_t hmp;
733         hammer_inode_t ip;
734         uid_t xuid;
735         int error;
736         int64_t namekey;
737         u_int32_t dummy;
738
739         hmp = trans->hmp;
740
741         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
742         ++hammer_count_inodes;
743         ++hmp->count_inodes;
744         trans->flags |= HAMMER_TRANSF_NEWINODE;
745
746         if (pfsm) {
747                 KKASSERT(pfsm->localization != 0);
748                 ip->obj_id = HAMMER_OBJID_ROOT;
749                 ip->obj_localization = pfsm->localization;
750         } else {
751                 KKASSERT(dip != NULL);
752                 namekey = hammer_directory_namekey(dip, name, namelen, &dummy);
753                 ip->obj_id = hammer_alloc_objid(hmp, dip, namekey);
754                 ip->obj_localization = dip->obj_localization;
755         }
756
757         KKASSERT(ip->obj_id != 0);
758         ip->obj_asof = hmp->asof;
759         ip->hmp = hmp;
760         ip->flush_state = HAMMER_FST_IDLE;
761         ip->flags = HAMMER_INODE_DDIRTY |
762                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
763         ip->cache[0].ip = ip;
764         ip->cache[1].ip = ip;
765         ip->cache[2].ip = ip;
766         ip->cache[3].ip = ip;
767
768         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
769         /* ip->save_trunc_off = 0; (already zero) */
770         RB_INIT(&ip->rec_tree);
771         TAILQ_INIT(&ip->target_list);
772
773         ip->ino_data.atime = trans->time;
774         ip->ino_data.mtime = trans->time;
775         ip->ino_data.size = 0;
776         ip->ino_data.nlinks = 0;
777
778         /*
779          * A nohistory designator on the parent directory is inherited by
780          * the child.  We will do this even for pseudo-fs creation... the
781          * sysad can turn it off.
782          */
783         if (dip) {
784                 ip->ino_data.uflags = dip->ino_data.uflags &
785                                       (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
786         }
787
788         ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
789         ip->ino_leaf.base.localization = ip->obj_localization +
790                                          HAMMER_LOCALIZE_INODE;
791         ip->ino_leaf.base.obj_id = ip->obj_id;
792         ip->ino_leaf.base.key = 0;
793         ip->ino_leaf.base.create_tid = 0;
794         ip->ino_leaf.base.delete_tid = 0;
795         ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
796         ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
797
798         ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
799         ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
800         ip->ino_data.mode = vap->va_mode;
801         ip->ino_data.ctime = trans->time;
802
803         /*
804          * If we are running version 2 or greater directory entries are
805          * inode-localized instead of data-localized.
806          */
807         if (trans->hmp->version >= HAMMER_VOL_VERSION_TWO) {
808                 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
809                         ip->ino_data.cap_flags |=
810                                 HAMMER_INODE_CAP_DIR_LOCAL_INO;
811                 }
812         }
813         if (trans->hmp->version >= HAMMER_VOL_VERSION_SIX) {
814                 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
815                         ip->ino_data.cap_flags |=
816                                 HAMMER_INODE_CAP_DIRHASH_ALG1;
817                 }
818         }
819
820         /*
821          * Setup the ".." pointer.  This only needs to be done for directories
822          * but we do it for all objects as a recovery aid.
823          */
824         if (dip)
825                 ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
826 #if 0
827         /*
828          * The parent_obj_localization field only applies to pseudo-fs roots.
829          * XXX this is no longer applicable, PFSs are no longer directly
830          * tied into the parent's directory structure.
831          */
832         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY &&
833             ip->obj_id == HAMMER_OBJID_ROOT) {
834                 ip->ino_data.ext.obj.parent_obj_localization = 
835                                                 dip->obj_localization;
836         }
837 #endif
838
839         switch(ip->ino_leaf.base.obj_type) {
840         case HAMMER_OBJTYPE_CDEV:
841         case HAMMER_OBJTYPE_BDEV:
842                 ip->ino_data.rmajor = vap->va_rmajor;
843                 ip->ino_data.rminor = vap->va_rminor;
844                 break;
845         default:
846                 break;
847         }
848
849         /*
850          * Calculate default uid/gid and overwrite with information from
851          * the vap.
852          */
853         if (dip) {
854                 xuid = hammer_to_unix_xid(&dip->ino_data.uid);
855                 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
856                                              xuid, cred, &vap->va_mode);
857         } else {
858                 xuid = 0;
859         }
860         ip->ino_data.mode = vap->va_mode;
861
862         if (vap->va_vaflags & VA_UID_UUID_VALID)
863                 ip->ino_data.uid = vap->va_uid_uuid;
864         else if (vap->va_uid != (uid_t)VNOVAL)
865                 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
866         else
867                 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
868
869         if (vap->va_vaflags & VA_GID_UUID_VALID)
870                 ip->ino_data.gid = vap->va_gid_uuid;
871         else if (vap->va_gid != (gid_t)VNOVAL)
872                 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
873         else if (dip)
874                 ip->ino_data.gid = dip->ino_data.gid;
875
876         hammer_ref(&ip->lock);
877
878         if (pfsm) {
879                 ip->pfsm = pfsm;
880                 hammer_ref(&pfsm->lock);
881                 error = 0;
882         } else if (dip->obj_localization == ip->obj_localization) {
883                 ip->pfsm = dip->pfsm;
884                 hammer_ref(&ip->pfsm->lock);
885                 error = 0;
886         } else {
887                 ip->pfsm = hammer_load_pseudofs(trans,
888                                                 ip->obj_localization,
889                                                 &error);
890                 error = 0;      /* ignore ENOENT */
891         }
892
893         if (error) {
894                 hammer_free_inode(ip);
895                 ip = NULL;
896         } else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
897                 panic("hammer_create_inode: duplicate obj_id %llx",
898                       (long long)ip->obj_id);
899                 /* not reached */
900                 hammer_free_inode(ip);
901         }
902         *ipp = ip;
903         return(error);
904 }
905
906 /*
907  * Final cleanup / freeing of an inode structure
908  */
909 static void
910 hammer_free_inode(hammer_inode_t ip)
911 {
912         struct hammer_mount *hmp;
913
914         hmp = ip->hmp;
915         KKASSERT(hammer_oneref(&ip->lock));
916         hammer_uncache_node(&ip->cache[0]);
917         hammer_uncache_node(&ip->cache[1]);
918         hammer_uncache_node(&ip->cache[2]);
919         hammer_uncache_node(&ip->cache[3]);
920         hammer_inode_wakereclaims(ip);
921         if (ip->objid_cache)
922                 hammer_clear_objid(ip);
923         --hammer_count_inodes;
924         --hmp->count_inodes;
925         if (ip->pfsm) {
926                 hammer_rel_pseudofs(hmp, ip->pfsm);
927                 ip->pfsm = NULL;
928         }
929         kfree(ip, hmp->m_inodes);
930         ip = NULL;
931 }
932
933 /*
934  * Retrieve pseudo-fs data.  NULL will never be returned.
935  *
936  * If an error occurs *errorp will be set and a default template is returned,
937  * otherwise *errorp is set to 0.  Typically when an error occurs it will
938  * be ENOENT.
939  */
940 hammer_pseudofs_inmem_t
941 hammer_load_pseudofs(hammer_transaction_t trans,
942                      u_int32_t localization, int *errorp)
943 {
944         hammer_mount_t hmp = trans->hmp;
945         hammer_inode_t ip;
946         hammer_pseudofs_inmem_t pfsm;
947         struct hammer_cursor cursor;
948         int bytes;
949
950 retry:
951         pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
952         if (pfsm) {
953                 hammer_ref(&pfsm->lock);
954                 *errorp = 0;
955                 return(pfsm);
956         }
957
958         /*
959          * PFS records are stored in the root inode (not the PFS root inode,
960          * but the real root).  Avoid an infinite recursion if loading
961          * the PFS for the real root.
962          */
963         if (localization) {
964                 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
965                                       HAMMER_MAX_TID,
966                                       HAMMER_DEF_LOCALIZATION, 0, errorp);
967         } else {
968                 ip = NULL;
969         }
970
971         pfsm = kmalloc(sizeof(*pfsm), hmp->m_misc, M_WAITOK | M_ZERO);
972         pfsm->localization = localization;
973         pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
974         pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
975
976         hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
977         cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION +
978                                       HAMMER_LOCALIZE_MISC;
979         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
980         cursor.key_beg.create_tid = 0;
981         cursor.key_beg.delete_tid = 0;
982         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
983         cursor.key_beg.obj_type = 0;
984         cursor.key_beg.key = localization;
985         cursor.asof = HAMMER_MAX_TID;
986         cursor.flags |= HAMMER_CURSOR_ASOF;
987
988         if (ip)
989                 *errorp = hammer_ip_lookup(&cursor);
990         else
991                 *errorp = hammer_btree_lookup(&cursor);
992         if (*errorp == 0) {
993                 *errorp = hammer_ip_resolve_data(&cursor);
994                 if (*errorp == 0) {
995                         if (cursor.data->pfsd.mirror_flags &
996                             HAMMER_PFSD_DELETED) {
997                                 *errorp = ENOENT;
998                         } else {
999                                 bytes = cursor.leaf->data_len;
1000                                 if (bytes > sizeof(pfsm->pfsd))
1001                                         bytes = sizeof(pfsm->pfsd);
1002                                 bcopy(cursor.data, &pfsm->pfsd, bytes);
1003                         }
1004                 }
1005         }
1006         hammer_done_cursor(&cursor);
1007
1008         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
1009         hammer_ref(&pfsm->lock);
1010         if (ip)
1011                 hammer_rel_inode(ip, 0);
1012         if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
1013                 kfree(pfsm, hmp->m_misc);
1014                 goto retry;
1015         }
1016         return(pfsm);
1017 }
1018
1019 /*
1020  * Store pseudo-fs data.  The backend will automatically delete any prior
1021  * on-disk pseudo-fs data but we have to delete in-memory versions.
1022  */
1023 int
1024 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
1025 {
1026         struct hammer_cursor cursor;
1027         hammer_record_t record;
1028         hammer_inode_t ip;
1029         int error;
1030
1031         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1032                               HAMMER_DEF_LOCALIZATION, 0, &error);
1033 retry:
1034         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
1035         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
1036         cursor.key_beg.localization = ip->obj_localization +
1037                                       HAMMER_LOCALIZE_MISC;
1038         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
1039         cursor.key_beg.create_tid = 0;
1040         cursor.key_beg.delete_tid = 0;
1041         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
1042         cursor.key_beg.obj_type = 0;
1043         cursor.key_beg.key = pfsm->localization;
1044         cursor.asof = HAMMER_MAX_TID;
1045         cursor.flags |= HAMMER_CURSOR_ASOF;
1046
1047         /*
1048          * Replace any in-memory version of the record.
1049          */
1050         error = hammer_ip_lookup(&cursor);
1051         if (error == 0 && hammer_cursor_inmem(&cursor)) {
1052                 record = cursor.iprec;
1053                 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
1054                         KKASSERT(cursor.deadlk_rec == NULL);
1055                         hammer_ref(&record->lock);
1056                         cursor.deadlk_rec = record;
1057                         error = EDEADLK;
1058                 } else {
1059                         record->flags |= HAMMER_RECF_DELETED_FE;
1060                         error = 0;
1061                 }
1062         }
1063
1064         /*
1065          * Allocate replacement general record.  The backend flush will
1066          * delete any on-disk version of the record.
1067          */
1068         if (error == 0 || error == ENOENT) {
1069                 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
1070                 record->type = HAMMER_MEM_RECORD_GENERAL;
1071
1072                 record->leaf.base.localization = ip->obj_localization +
1073                                                  HAMMER_LOCALIZE_MISC;
1074                 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
1075                 record->leaf.base.key = pfsm->localization;
1076                 record->leaf.data_len = sizeof(pfsm->pfsd);
1077                 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
1078                 error = hammer_ip_add_record(trans, record);
1079         }
1080         hammer_done_cursor(&cursor);
1081         if (error == EDEADLK)
1082                 goto retry;
1083         hammer_rel_inode(ip, 0);
1084         return(error);
1085 }
1086
1087 /*
1088  * Create a root directory for a PFS if one does not alredy exist.
1089  *
1090  * The PFS root stands alone so we must also bump the nlinks count
1091  * to prevent it from being destroyed on release.
1092  */
1093 int
1094 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1095                        hammer_pseudofs_inmem_t pfsm)
1096 {
1097         hammer_inode_t ip;
1098         struct vattr vap;
1099         int error;
1100
1101         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1102                               pfsm->localization, 0, &error);
1103         if (ip == NULL) {
1104                 vattr_null(&vap);
1105                 vap.va_mode = 0755;
1106                 vap.va_type = VDIR;
1107                 error = hammer_create_inode(trans, &vap, cred,
1108                                             NULL, NULL, 0,
1109                                             pfsm, &ip);
1110                 if (error == 0) {
1111                         ++ip->ino_data.nlinks;
1112                         hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
1113                 }
1114         }
1115         if (ip)
1116                 hammer_rel_inode(ip, 0);
1117         return(error);
1118 }
1119
1120 /*
1121  * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
1122  * if we are unable to disassociate all the inodes.
1123  */
1124 static
1125 int
1126 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
1127 {
1128         int res;
1129
1130         hammer_ref(&ip->lock);
1131         if (hammer_isactive(&ip->lock) == 2 && ip->vp)
1132                 vclean_unlocked(ip->vp);
1133         if (hammer_isactive(&ip->lock) == 1 && ip->vp == NULL)
1134                 res = 0;
1135         else
1136                 res = -1;       /* stop, someone is using the inode */
1137         hammer_rel_inode(ip, 0);
1138         return(res);
1139 }
1140
1141 int
1142 hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
1143 {
1144         int res;
1145         int try;
1146
1147         for (try = res = 0; try < 4; ++try) {
1148                 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
1149                                            hammer_inode_pfs_cmp,
1150                                            hammer_unload_pseudofs_callback,
1151                                            &localization);
1152                 if (res == 0 && try > 1)
1153                         break;
1154                 hammer_flusher_sync(trans->hmp);
1155         }
1156         if (res != 0)
1157                 res = ENOTEMPTY;
1158         return(res);
1159 }
1160
1161
1162 /*
1163  * Release a reference on a PFS
1164  */
1165 void
1166 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
1167 {
1168         hammer_rel(&pfsm->lock);
1169         if (hammer_norefs(&pfsm->lock)) {
1170                 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
1171                 kfree(pfsm, hmp->m_misc);
1172         }
1173 }
1174
1175 /*
1176  * Called by hammer_sync_inode().
1177  */
1178 static int
1179 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
1180 {
1181         hammer_transaction_t trans = cursor->trans;
1182         hammer_record_t record;
1183         int error;
1184         int redirty;
1185
1186 retry:
1187         error = 0;
1188
1189         /*
1190          * If the inode has a presence on-disk then locate it and mark
1191          * it deleted, setting DELONDISK.
1192          *
1193          * The record may or may not be physically deleted, depending on
1194          * the retention policy.
1195          */
1196         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
1197             HAMMER_INODE_ONDISK) {
1198                 hammer_normalize_cursor(cursor);
1199                 cursor->key_beg.localization = ip->obj_localization + 
1200                                                HAMMER_LOCALIZE_INODE;
1201                 cursor->key_beg.obj_id = ip->obj_id;
1202                 cursor->key_beg.key = 0;
1203                 cursor->key_beg.create_tid = 0;
1204                 cursor->key_beg.delete_tid = 0;
1205                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1206                 cursor->key_beg.obj_type = 0;
1207                 cursor->asof = ip->obj_asof;
1208                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1209                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
1210                 cursor->flags |= HAMMER_CURSOR_BACKEND;
1211
1212                 error = hammer_btree_lookup(cursor);
1213                 if (hammer_debug_inode)
1214                         kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
1215
1216                 if (error == 0) {
1217                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1218                         if (hammer_debug_inode)
1219                                 kprintf(" error %d\n", error);
1220                         if (error == 0) {
1221                                 ip->flags |= HAMMER_INODE_DELONDISK;
1222                         }
1223                         if (cursor->node)
1224                                 hammer_cache_node(&ip->cache[0], cursor->node);
1225                 }
1226                 if (error == EDEADLK) {
1227                         hammer_done_cursor(cursor);
1228                         error = hammer_init_cursor(trans, cursor,
1229                                                    &ip->cache[0], ip);
1230                         if (hammer_debug_inode)
1231                                 kprintf("IPDED %p %d\n", ip, error);
1232                         if (error == 0)
1233                                 goto retry;
1234                 }
1235         }
1236
1237         /*
1238          * Ok, write out the initial record or a new record (after deleting
1239          * the old one), unless the DELETED flag is set.  This routine will
1240          * clear DELONDISK if it writes out a record.
1241          *
1242          * Update our inode statistics if this is the first application of
1243          * the inode on-disk.
1244          */
1245         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
1246                 /*
1247                  * Generate a record and write it to the media.  We clean-up
1248                  * the state before releasing so we do not have to set-up
1249                  * a flush_group.
1250                  */
1251                 record = hammer_alloc_mem_record(ip, 0);
1252                 record->type = HAMMER_MEM_RECORD_INODE;
1253                 record->flush_state = HAMMER_FST_FLUSH;
1254                 record->leaf = ip->sync_ino_leaf;
1255                 record->leaf.base.create_tid = trans->tid;
1256                 record->leaf.data_len = sizeof(ip->sync_ino_data);
1257                 record->leaf.create_ts = trans->time32;
1258                 record->data = (void *)&ip->sync_ino_data;
1259                 record->flags |= HAMMER_RECF_INTERLOCK_BE;
1260
1261                 /*
1262                  * If this flag is set we cannot sync the new file size
1263                  * because we haven't finished related truncations.  The
1264                  * inode will be flushed in another flush group to finish
1265                  * the job.
1266                  */
1267                 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1268                     ip->sync_ino_data.size != ip->ino_data.size) {
1269                         redirty = 1;
1270                         ip->sync_ino_data.size = ip->ino_data.size;
1271                 } else {
1272                         redirty = 0;
1273                 }
1274
1275                 for (;;) {
1276                         error = hammer_ip_sync_record_cursor(cursor, record);
1277                         if (hammer_debug_inode)
1278                                 kprintf("GENREC %p rec %08x %d\n",      
1279                                         ip, record->flags, error);
1280                         if (error != EDEADLK)
1281                                 break;
1282                         hammer_done_cursor(cursor);
1283                         error = hammer_init_cursor(trans, cursor,
1284                                                    &ip->cache[0], ip);
1285                         if (hammer_debug_inode)
1286                                 kprintf("GENREC reinit %d\n", error);
1287                         if (error)
1288                                 break;
1289                 }
1290
1291                 /*
1292                  * Note:  The record was never on the inode's record tree
1293                  * so just wave our hands importantly and destroy it.
1294                  */
1295                 record->flags |= HAMMER_RECF_COMMITTED;
1296                 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1297                 record->flush_state = HAMMER_FST_IDLE;
1298                 ++ip->rec_generation;
1299                 hammer_rel_mem_record(record);
1300
1301                 /*
1302                  * Finish up.
1303                  */
1304                 if (error == 0) {
1305                         if (hammer_debug_inode)
1306                                 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1307                         ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1308                                             HAMMER_INODE_SDIRTY |
1309                                             HAMMER_INODE_ATIME |
1310                                             HAMMER_INODE_MTIME);
1311                         ip->flags &= ~HAMMER_INODE_DELONDISK;
1312                         if (redirty)
1313                                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1314
1315                         /*
1316                          * Root volume count of inodes
1317                          */
1318                         hammer_sync_lock_sh(trans);
1319                         if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1320                                 hammer_modify_volume_field(trans,
1321                                                            trans->rootvol,
1322                                                            vol0_stat_inodes);
1323                                 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1324                                 hammer_modify_volume_done(trans->rootvol);
1325                                 ip->flags |= HAMMER_INODE_ONDISK;
1326                                 if (hammer_debug_inode)
1327                                         kprintf("NOWONDISK %p\n", ip);
1328                         }
1329                         hammer_sync_unlock(trans);
1330                 }
1331         }
1332
1333         /*
1334          * If the inode has been destroyed, clean out any left-over flags
1335          * that may have been set by the frontend.
1336          */
1337         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) { 
1338                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1339                                     HAMMER_INODE_SDIRTY |
1340                                     HAMMER_INODE_ATIME |
1341                                     HAMMER_INODE_MTIME);
1342         }
1343         return(error);
1344 }
1345
1346 /*
1347  * Update only the itimes fields.
1348  *
1349  * ATIME can be updated without generating any UNDO.  MTIME is updated
1350  * with UNDO so it is guaranteed to be synchronized properly in case of
1351  * a crash.
1352  *
1353  * Neither field is included in the B-Tree leaf element's CRC, which is how
1354  * we can get away with updating ATIME the way we do.
1355  */
1356 static int
1357 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1358 {
1359         hammer_transaction_t trans = cursor->trans;
1360         int error;
1361
1362 retry:
1363         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1364             HAMMER_INODE_ONDISK) {
1365                 return(0);
1366         }
1367
1368         hammer_normalize_cursor(cursor);
1369         cursor->key_beg.localization = ip->obj_localization + 
1370                                        HAMMER_LOCALIZE_INODE;
1371         cursor->key_beg.obj_id = ip->obj_id;
1372         cursor->key_beg.key = 0;
1373         cursor->key_beg.create_tid = 0;
1374         cursor->key_beg.delete_tid = 0;
1375         cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1376         cursor->key_beg.obj_type = 0;
1377         cursor->asof = ip->obj_asof;
1378         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1379         cursor->flags |= HAMMER_CURSOR_ASOF;
1380         cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1381         cursor->flags |= HAMMER_CURSOR_GET_DATA;
1382         cursor->flags |= HAMMER_CURSOR_BACKEND;
1383
1384         error = hammer_btree_lookup(cursor);
1385         if (error == 0) {
1386                 hammer_cache_node(&ip->cache[0], cursor->node);
1387                 if (ip->sync_flags & HAMMER_INODE_MTIME) {
1388                         /*
1389                          * Updating MTIME requires an UNDO.  Just cover
1390                          * both atime and mtime.
1391                          */
1392                         hammer_sync_lock_sh(trans);
1393                         hammer_modify_buffer(trans, cursor->data_buffer,
1394                                      HAMMER_ITIMES_BASE(&cursor->data->inode),
1395                                      HAMMER_ITIMES_BYTES);
1396                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1397                         cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1398                         hammer_modify_buffer_done(cursor->data_buffer);
1399                         hammer_sync_unlock(trans);
1400                 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1401                         /*
1402                          * Updating atime only can be done in-place with
1403                          * no UNDO.
1404                          */
1405                         hammer_sync_lock_sh(trans);
1406                         hammer_modify_buffer(trans, cursor->data_buffer,
1407                                              NULL, 0);
1408                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1409                         hammer_modify_buffer_done(cursor->data_buffer);
1410                         hammer_sync_unlock(trans);
1411                 }
1412                 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1413         }
1414         if (error == EDEADLK) {
1415                 hammer_done_cursor(cursor);
1416                 error = hammer_init_cursor(trans, cursor,
1417                                            &ip->cache[0], ip);
1418                 if (error == 0)
1419                         goto retry;
1420         }
1421         return(error);
1422 }
1423
1424 /*
1425  * Release a reference on an inode, flush as requested.
1426  *
1427  * On the last reference we queue the inode to the flusher for its final
1428  * disposition.
1429  */
1430 void
1431 hammer_rel_inode(struct hammer_inode *ip, int flush)
1432 {
1433         /*hammer_mount_t hmp = ip->hmp;*/
1434
1435         /*
1436          * Handle disposition when dropping the last ref.
1437          */
1438         for (;;) {
1439                 if (hammer_oneref(&ip->lock)) {
1440                         /*
1441                          * Determine whether on-disk action is needed for
1442                          * the inode's final disposition.
1443                          */
1444                         KKASSERT(ip->vp == NULL);
1445                         hammer_inode_unloadable_check(ip, 0);
1446                         if (ip->flags & HAMMER_INODE_MODMASK) {
1447                                 hammer_flush_inode(ip, 0);
1448                         } else if (hammer_oneref(&ip->lock)) {
1449                                 hammer_unload_inode(ip);
1450                                 break;
1451                         }
1452                 } else {
1453                         if (flush)
1454                                 hammer_flush_inode(ip, 0);
1455
1456                         /*
1457                          * The inode still has multiple refs, try to drop
1458                          * one ref.
1459                          */
1460                         KKASSERT(hammer_isactive(&ip->lock) >= 1);
1461                         if (hammer_isactive(&ip->lock) > 1) {
1462                                 hammer_rel(&ip->lock);
1463                                 break;
1464                         }
1465                 }
1466         }
1467 }
1468
1469 /*
1470  * Unload and destroy the specified inode.  Must be called with one remaining
1471  * reference.  The reference is disposed of.
1472  *
1473  * The inode must be completely clean.
1474  */
1475 static int
1476 hammer_unload_inode(struct hammer_inode *ip)
1477 {
1478         hammer_mount_t hmp = ip->hmp;
1479
1480         KASSERT(hammer_oneref(&ip->lock),
1481                 ("hammer_unload_inode: %d refs\n", hammer_isactive(&ip->lock)));
1482         KKASSERT(ip->vp == NULL);
1483         KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1484         KKASSERT(ip->cursor_ip_refs == 0);
1485         KKASSERT(hammer_notlocked(&ip->lock));
1486         KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1487
1488         KKASSERT(RB_EMPTY(&ip->rec_tree));
1489         KKASSERT(TAILQ_EMPTY(&ip->target_list));
1490
1491         if (ip->flags & HAMMER_INODE_RDIRTY) {
1492                 RB_REMOVE(hammer_redo_rb_tree, &hmp->rb_redo_root, ip);
1493                 ip->flags &= ~HAMMER_INODE_RDIRTY;
1494         }
1495         RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1496
1497         hammer_free_inode(ip);
1498         return(0);
1499 }
1500
1501 /*
1502  * Called during unmounting if a critical error occured.  The in-memory
1503  * inode and all related structures are destroyed.
1504  *
1505  * If a critical error did not occur the unmount code calls the standard
1506  * release and asserts that the inode is gone.
1507  */
1508 int
1509 hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
1510 {
1511         hammer_record_t rec;
1512
1513         /*
1514          * Get rid of the inodes in-memory records, regardless of their
1515          * state, and clear the mod-mask.
1516          */
1517         while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1518                 TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1519                 rec->target_ip = NULL;
1520                 if (rec->flush_state == HAMMER_FST_SETUP)
1521                         rec->flush_state = HAMMER_FST_IDLE;
1522         }
1523         while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1524                 if (rec->flush_state == HAMMER_FST_FLUSH)
1525                         --rec->flush_group->refs;
1526                 else
1527                         hammer_ref(&rec->lock);
1528                 KKASSERT(hammer_oneref(&rec->lock));
1529                 rec->flush_state = HAMMER_FST_IDLE;
1530                 rec->flush_group = NULL;
1531                 rec->flags |= HAMMER_RECF_DELETED_FE; /* wave hands */
1532                 rec->flags |= HAMMER_RECF_DELETED_BE; /* wave hands */
1533                 ++ip->rec_generation;
1534                 hammer_rel_mem_record(rec);
1535         }
1536         ip->flags &= ~HAMMER_INODE_MODMASK;
1537         ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1538         KKASSERT(ip->vp == NULL);
1539
1540         /*
1541          * Remove the inode from any flush group, force it idle.  FLUSH
1542          * and SETUP states have an inode ref.
1543          */
1544         switch(ip->flush_state) {
1545         case HAMMER_FST_FLUSH:
1546                 RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
1547                 --ip->flush_group->refs;
1548                 ip->flush_group = NULL;
1549                 /* fall through */
1550         case HAMMER_FST_SETUP:
1551                 hammer_rel(&ip->lock);
1552                 ip->flush_state = HAMMER_FST_IDLE;
1553                 /* fall through */
1554         case HAMMER_FST_IDLE:
1555                 break;
1556         }
1557
1558         /*
1559          * There shouldn't be any associated vnode.  The unload needs at
1560          * least one ref, if we do have a vp steal its ip ref.
1561          */
1562         if (ip->vp) {
1563                 kprintf("hammer_destroy_inode_callback: Unexpected "
1564                         "vnode association ip %p vp %p\n", ip, ip->vp);
1565                 ip->vp->v_data = NULL;
1566                 ip->vp = NULL;
1567         } else {
1568                 hammer_ref(&ip->lock);
1569         }
1570         hammer_unload_inode(ip);
1571         return(0);
1572 }
1573
1574 /*
1575  * Called on mount -u when switching from RW to RO or vise-versa.  Adjust
1576  * the read-only flag for cached inodes.
1577  *
1578  * This routine is called from a RB_SCAN().
1579  */
1580 int
1581 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1582 {
1583         hammer_mount_t hmp = ip->hmp;
1584
1585         if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1586                 ip->flags |= HAMMER_INODE_RO;
1587         else
1588                 ip->flags &= ~HAMMER_INODE_RO;
1589         return(0);
1590 }
1591
1592 /*
1593  * A transaction has modified an inode, requiring updates as specified by
1594  * the passed flags.
1595  *
1596  * HAMMER_INODE_DDIRTY: Inode data has been updated, not incl mtime/atime,
1597  *                      and not including size changes due to write-append
1598  *                      (but other size changes are included).
1599  * HAMMER_INODE_SDIRTY: Inode data has been updated, size changes due to
1600  *                      write-append.
1601  * HAMMER_INODE_XDIRTY: Dirty in-memory records
1602  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
1603  * HAMMER_INODE_DELETED: Inode record/data must be deleted
1604  * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1605  */
1606 void
1607 hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags)
1608 {
1609         /* 
1610          * ronly of 0 or 2 does not trigger assertion.
1611          * 2 is a special error state 
1612          */
1613         KKASSERT(ip->hmp->ronly != 1 ||
1614                   (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY | 
1615                             HAMMER_INODE_SDIRTY |
1616                             HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1617                             HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1618         if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1619                 ip->flags |= HAMMER_INODE_RSV_INODES;
1620                 ++ip->hmp->rsv_inodes;
1621         }
1622
1623         /*
1624          * Set the NEWINODE flag in the transaction if the inode
1625          * transitions to a dirty state.  This is used to track
1626          * the load on the inode cache.
1627          */
1628         if (trans &&
1629             (ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1630             (flags & HAMMER_INODE_MODMASK)) {
1631                 trans->flags |= HAMMER_TRANSF_NEWINODE;
1632         }
1633
1634         ip->flags |= flags;
1635 }
1636
1637 /*
1638  * Request that an inode be flushed.  This whole mess cannot block and may
1639  * recurse (if not synchronous).  Once requested HAMMER will attempt to
1640  * actively flush the inode until the flush can be done.
1641  *
1642  * The inode may already be flushing, or may be in a setup state.  We can
1643  * place the inode in a flushing state if it is currently idle and flag it
1644  * to reflush if it is currently flushing.
1645  *
1646  * Upon return if the inode could not be flushed due to a setup
1647  * dependancy, then it will be automatically flushed when the dependancy
1648  * is satisfied.
1649  */
1650 void
1651 hammer_flush_inode(hammer_inode_t ip, int flags)
1652 {
1653         hammer_mount_t hmp;
1654         hammer_flush_group_t flg;
1655         int good;
1656
1657         /*
1658          * fill_flush_group is the first flush group we may be able to
1659          * continue filling, it may be open or closed but it will always
1660          * be past the currently flushing (running) flg.
1661          *
1662          * next_flush_group is the next open flush group.
1663          */
1664         hmp = ip->hmp;
1665         while ((flg = hmp->fill_flush_group) != NULL) {
1666                 KKASSERT(flg->running == 0);
1667                 if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit &&
1668                     flg->total_count <= hammer_autoflush) {
1669                         break;
1670                 }
1671                 hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
1672                 hammer_flusher_async(ip->hmp, flg);
1673         }
1674         if (flg == NULL) {
1675                 flg = kmalloc(sizeof(*flg), hmp->m_misc, M_WAITOK|M_ZERO);
1676                 flg->seq = hmp->flusher.next++;
1677                 if (hmp->next_flush_group == NULL)
1678                         hmp->next_flush_group = flg;
1679                 if (hmp->fill_flush_group == NULL)
1680                         hmp->fill_flush_group = flg;
1681                 RB_INIT(&flg->flush_tree);
1682                 TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1683         }
1684
1685         /*
1686          * Trivial 'nothing to flush' case.  If the inode is in a SETUP
1687          * state we have to put it back into an IDLE state so we can
1688          * drop the extra ref.
1689          *
1690          * If we have a parent dependancy we must still fall through
1691          * so we can run it.
1692          */
1693         if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1694                 if (ip->flush_state == HAMMER_FST_SETUP &&
1695                     TAILQ_EMPTY(&ip->target_list)) {
1696                         ip->flush_state = HAMMER_FST_IDLE;
1697                         hammer_rel_inode(ip, 0);
1698                 }
1699                 if (ip->flush_state == HAMMER_FST_IDLE)
1700                         return;
1701         }
1702
1703         /*
1704          * Our flush action will depend on the current state.
1705          */
1706         switch(ip->flush_state) {
1707         case HAMMER_FST_IDLE:
1708                 /*
1709                  * We have no dependancies and can flush immediately.  Some
1710                  * our children may not be flushable so we have to re-test
1711                  * with that additional knowledge.
1712                  */
1713                 hammer_flush_inode_core(ip, flg, flags);
1714                 break;
1715         case HAMMER_FST_SETUP:
1716                 /*
1717                  * Recurse upwards through dependancies via target_list
1718                  * and start their flusher actions going if possible.
1719                  *
1720                  * 'good' is our connectivity.  -1 means we have none and
1721                  * can't flush, 0 means there weren't any dependancies, and
1722                  * 1 means we have good connectivity.
1723                  */
1724                 good = hammer_setup_parent_inodes(ip, 0, flg);
1725
1726                 if (good >= 0) {
1727                         /*
1728                          * We can continue if good >= 0.  Determine how 
1729                          * many records under our inode can be flushed (and
1730                          * mark them).
1731                          */
1732                         hammer_flush_inode_core(ip, flg, flags);
1733                 } else {
1734                         /*
1735                          * Parent has no connectivity, tell it to flush
1736                          * us as soon as it does.
1737                          *
1738                          * The REFLUSH flag is also needed to trigger
1739                          * dependancy wakeups.
1740                          */
1741                         ip->flags |= HAMMER_INODE_CONN_DOWN |
1742                                      HAMMER_INODE_REFLUSH;
1743                         if (flags & HAMMER_FLUSH_SIGNAL) {
1744                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1745                                 hammer_flusher_async(ip->hmp, flg);
1746                         }
1747                 }
1748                 break;
1749         case HAMMER_FST_FLUSH:
1750                 /*
1751                  * We are already flushing, flag the inode to reflush
1752                  * if needed after it completes its current flush.
1753                  *
1754                  * The REFLUSH flag is also needed to trigger
1755                  * dependancy wakeups.
1756                  */
1757                 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1758                         ip->flags |= HAMMER_INODE_REFLUSH;
1759                 if (flags & HAMMER_FLUSH_SIGNAL) {
1760                         ip->flags |= HAMMER_INODE_RESIGNAL;
1761                         hammer_flusher_async(ip->hmp, flg);
1762                 }
1763                 break;
1764         }
1765 }
1766
1767 /*
1768  * Scan ip->target_list, which is a list of records owned by PARENTS to our
1769  * ip which reference our ip.
1770  *
1771  * XXX This is a huge mess of recursive code, but not one bit of it blocks
1772  *     so for now do not ref/deref the structures.  Note that if we use the
1773  *     ref/rel code later, the rel CAN block.
1774  */
1775 static int
1776 hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
1777                            hammer_flush_group_t flg)
1778 {
1779         hammer_record_t depend;
1780         int good;
1781         int r;
1782
1783         /*
1784          * If we hit our recursion limit and we have parent dependencies
1785          * We cannot continue.  Returning < 0 will cause us to be flagged
1786          * for reflush.  Returning -2 cuts off additional dependency checks
1787          * because they are likely to also hit the depth limit.
1788          *
1789          * We cannot return < 0 if there are no dependencies or there might
1790          * not be anything to wakeup (ip).
1791          */
1792         if (depth == 20 && TAILQ_FIRST(&ip->target_list)) {
1793                 kprintf("HAMMER Warning: depth limit reached on "
1794                         "setup recursion, inode %p %016llx\n",
1795                         ip, (long long)ip->obj_id);
1796                 return(-2);
1797         }
1798
1799         /*
1800          * Scan dependencies
1801          */
1802         good = 0;
1803         TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1804                 r = hammer_setup_parent_inodes_helper(depend, depth, flg);
1805                 KKASSERT(depend->target_ip == ip);
1806                 if (r < 0 && good == 0)
1807                         good = -1;
1808                 if (r > 0)
1809                         good = 1;
1810
1811                 /*
1812                  * If we failed due to the recursion depth limit then stop
1813                  * now.
1814                  */
1815                 if (r == -2)
1816                         break;
1817         }
1818         return(good);
1819 }
1820
1821 /*
1822  * This helper function takes a record representing the dependancy between
1823  * the parent inode and child inode.
1824  *
1825  * record->ip           = parent inode
1826  * record->target_ip    = child inode
1827  * 
1828  * We are asked to recurse upwards and convert the record from SETUP
1829  * to FLUSH if possible.
1830  *
1831  * Return 1 if the record gives us connectivity
1832  *
1833  * Return 0 if the record is not relevant 
1834  *
1835  * Return -1 if we can't resolve the dependancy and there is no connectivity.
1836  */
1837 static int
1838 hammer_setup_parent_inodes_helper(hammer_record_t record, int depth,
1839                                   hammer_flush_group_t flg)
1840 {
1841         hammer_mount_t hmp;
1842         hammer_inode_t pip;
1843         int good;
1844
1845         KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1846         pip = record->ip;
1847         hmp = pip->hmp;
1848
1849         /*
1850          * If the record is already flushing, is it in our flush group?
1851          *
1852          * If it is in our flush group but it is a general record or a 
1853          * delete-on-disk, it does not improve our connectivity (return 0),
1854          * and if the target inode is not trying to destroy itself we can't
1855          * allow the operation yet anyway (the second return -1).
1856          */
1857         if (record->flush_state == HAMMER_FST_FLUSH) {
1858                 /*
1859                  * If not in our flush group ask the parent to reflush
1860                  * us as soon as possible.
1861                  */
1862                 if (record->flush_group != flg) {
1863                         pip->flags |= HAMMER_INODE_REFLUSH;
1864                         record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1865                         return(-1);
1866                 }
1867
1868                 /*
1869                  * If in our flush group everything is already set up,
1870                  * just return whether the record will improve our
1871                  * visibility or not.
1872                  */
1873                 if (record->type == HAMMER_MEM_RECORD_ADD)
1874                         return(1);
1875                 return(0);
1876         }
1877
1878         /*
1879          * It must be a setup record.  Try to resolve the setup dependancies
1880          * by recursing upwards so we can place ip on the flush list.
1881          *
1882          * Limit ourselves to 20 levels of recursion to avoid blowing out
1883          * the kernel stack.  If we hit the recursion limit we can't flush
1884          * until the parent flushes.  The parent will flush independantly
1885          * on its own and ultimately a deep recursion will be resolved.
1886          */
1887         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1888
1889         good = hammer_setup_parent_inodes(pip, depth + 1, flg);
1890
1891         /*
1892          * If good < 0 the parent has no connectivity and we cannot safely
1893          * flush the directory entry, which also means we can't flush our
1894          * ip.  Flag us for downward recursion once the parent's
1895          * connectivity is resolved.  Flag the parent for [re]flush or it
1896          * may not check for downward recursions.
1897          */
1898         if (good < 0) {
1899                 pip->flags |= HAMMER_INODE_REFLUSH;
1900                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1901                 return(good);
1902         }
1903
1904         /*
1905          * We are go, place the parent inode in a flushing state so we can
1906          * place its record in a flushing state.  Note that the parent
1907          * may already be flushing.  The record must be in the same flush
1908          * group as the parent.
1909          */
1910         if (pip->flush_state != HAMMER_FST_FLUSH)
1911                 hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
1912         KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
1913
1914         /*
1915          * It is possible for a rename to create a loop in the recursion
1916          * and revisit a record.  This will result in the record being
1917          * placed in a flush state unexpectedly.  This check deals with
1918          * the case.
1919          */
1920         if (record->flush_state == HAMMER_FST_FLUSH) {
1921                 if (record->type == HAMMER_MEM_RECORD_ADD)
1922                         return(1);
1923                 return(0);
1924         }
1925
1926         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1927
1928 #if 0
1929         if (record->type == HAMMER_MEM_RECORD_DEL &&
1930             (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
1931                 /*
1932                  * Regardless of flushing state we cannot sync this path if the
1933                  * record represents a delete-on-disk but the target inode
1934                  * is not ready to sync its own deletion.
1935                  *
1936                  * XXX need to count effective nlinks to determine whether
1937                  * the flush is ok, otherwise removing a hardlink will
1938                  * just leave the DEL record to rot.
1939                  */
1940                 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
1941                 return(-1);
1942         } else
1943 #endif
1944         if (pip->flush_group == flg) {
1945                 /*
1946                  * Because we have not calculated nlinks yet we can just
1947                  * set records to the flush state if the parent is in
1948                  * the same flush group as we are.
1949                  */
1950                 record->flush_state = HAMMER_FST_FLUSH;
1951                 record->flush_group = flg;
1952                 ++record->flush_group->refs;
1953                 hammer_ref(&record->lock);
1954
1955                 /*
1956                  * A general directory-add contributes to our visibility.
1957                  *
1958                  * Otherwise it is probably a directory-delete or 
1959                  * delete-on-disk record and does not contribute to our
1960                  * visbility (but we can still flush it).
1961                  */
1962                 if (record->type == HAMMER_MEM_RECORD_ADD)
1963                         return(1);
1964                 return(0);
1965         } else {
1966                 /*
1967                  * If the parent is not in our flush group we cannot
1968                  * flush this record yet, there is no visibility.
1969                  * We tell the parent to reflush and mark ourselves
1970                  * so the parent knows it should flush us too.
1971                  */
1972                 pip->flags |= HAMMER_INODE_REFLUSH;
1973                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1974                 return(-1);
1975         }
1976 }
1977
1978 /*
1979  * This is the core routine placing an inode into the FST_FLUSH state.
1980  */
1981 static void
1982 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
1983 {
1984         hammer_mount_t hmp = ip->hmp;
1985         int go_count;
1986
1987         /*
1988          * Set flush state and prevent the flusher from cycling into
1989          * the next flush group.  Do not place the ip on the list yet.
1990          * Inodes not in the idle state get an extra reference.
1991          */
1992         KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
1993         if (ip->flush_state == HAMMER_FST_IDLE)
1994                 hammer_ref(&ip->lock);
1995         ip->flush_state = HAMMER_FST_FLUSH;
1996         ip->flush_group = flg;
1997         ++hmp->flusher.group_lock;
1998         ++hmp->count_iqueued;
1999         ++hammer_count_iqueued;
2000         ++flg->total_count;
2001         hammer_redo_fifo_start_flush(ip);
2002
2003 #if 0
2004         /*
2005          * We need to be able to vfsync/truncate from the backend.
2006          *
2007          * XXX Any truncation from the backend will acquire the vnode
2008          *     independently.
2009          */
2010         KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
2011         if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
2012                 ip->flags |= HAMMER_INODE_VHELD;
2013                 vref(ip->vp);
2014         }
2015 #endif
2016
2017         /*
2018          * Figure out how many in-memory records we can actually flush
2019          * (not including inode meta-data, buffers, etc).
2020          */
2021         KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
2022         if (flags & HAMMER_FLUSH_RECURSION) {
2023                 /*
2024                  * If this is a upwards recursion we do not want to
2025                  * recurse down again!
2026                  */
2027                 go_count = 1;
2028 #if 0
2029         } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2030                 /*
2031                  * No new records are added if we must complete a flush
2032                  * from a previous cycle, but we do have to move the records
2033                  * from the previous cycle to the current one.
2034                  */
2035 #if 0
2036                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2037                                    hammer_syncgrp_child_callback, NULL);
2038 #endif
2039                 go_count = 1;
2040 #endif
2041         } else {
2042                 /*
2043                  * Normal flush, scan records and bring them into the flush.
2044                  * Directory adds and deletes are usually skipped (they are
2045                  * grouped with the related inode rather then with the
2046                  * directory).
2047                  *
2048                  * go_count can be negative, which means the scan aborted
2049                  * due to the flush group being over-full and we should
2050                  * flush what we have.
2051                  */
2052                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2053                                    hammer_setup_child_callback, NULL);
2054         }
2055
2056         /*
2057          * This is a more involved test that includes go_count.  If we
2058          * can't flush, flag the inode and return.  If go_count is 0 we
2059          * were are unable to flush any records in our rec_tree and
2060          * must ignore the XDIRTY flag.
2061          */
2062         if (go_count == 0) {
2063                 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
2064                         --hmp->count_iqueued;
2065                         --hammer_count_iqueued;
2066
2067                         --flg->total_count;
2068                         ip->flush_state = HAMMER_FST_SETUP;
2069                         ip->flush_group = NULL;
2070                         if (flags & HAMMER_FLUSH_SIGNAL) {
2071                                 ip->flags |= HAMMER_INODE_REFLUSH |
2072                                              HAMMER_INODE_RESIGNAL;
2073                         } else {
2074                                 ip->flags |= HAMMER_INODE_REFLUSH;
2075                         }
2076 #if 0
2077                         if (ip->flags & HAMMER_INODE_VHELD) {
2078                                 ip->flags &= ~HAMMER_INODE_VHELD;
2079                                 vrele(ip->vp);
2080                         }
2081 #endif
2082
2083                         /*
2084                          * REFLUSH is needed to trigger dependancy wakeups
2085                          * when an inode is in SETUP.
2086                          */
2087                         ip->flags |= HAMMER_INODE_REFLUSH;
2088                         if (--hmp->flusher.group_lock == 0)
2089                                 wakeup(&hmp->flusher.group_lock);
2090                         return;
2091                 }
2092         }
2093
2094         /*
2095          * Snapshot the state of the inode for the backend flusher.
2096          *
2097          * We continue to retain save_trunc_off even when all truncations
2098          * have been resolved as an optimization to determine if we can
2099          * skip the B-Tree lookup for overwrite deletions.
2100          *
2101          * NOTE: The DELETING flag is a mod flag, but it is also sticky,
2102          * and stays in ip->flags.  Once set, it stays set until the
2103          * inode is destroyed.
2104          */
2105         if (ip->flags & HAMMER_INODE_TRUNCATED) {
2106                 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
2107                 ip->sync_trunc_off = ip->trunc_off;
2108                 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
2109                 ip->flags &= ~HAMMER_INODE_TRUNCATED;
2110                 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
2111
2112                 /*
2113                  * The save_trunc_off used to cache whether the B-Tree
2114                  * holds any records past that point is not used until
2115                  * after the truncation has succeeded, so we can safely
2116                  * set it now.
2117                  */
2118                 if (ip->save_trunc_off > ip->sync_trunc_off)
2119                         ip->save_trunc_off = ip->sync_trunc_off;
2120         }
2121         ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
2122                            ~HAMMER_INODE_TRUNCATED);
2123         ip->sync_ino_leaf = ip->ino_leaf;
2124         ip->sync_ino_data = ip->ino_data;
2125         ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
2126 #ifdef DEBUG_TRUNCATE
2127         if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
2128                 kprintf("truncateS %016llx\n", ip->sync_trunc_off);
2129 #endif
2130
2131         /*
2132          * The flusher list inherits our inode and reference.
2133          */
2134         KKASSERT(flg->running == 0);
2135         RB_INSERT(hammer_fls_rb_tree, &flg->flush_tree, ip);
2136         if (--hmp->flusher.group_lock == 0)
2137                 wakeup(&hmp->flusher.group_lock);
2138
2139         /*
2140          * Auto-flush the group if it grows too large.  Make sure the
2141          * inode reclaim wait pipeline continues to work.
2142          */
2143         if (flg->total_count >= hammer_autoflush ||
2144             flg->total_count >= hammer_limit_reclaim / 4) {
2145                 if (hmp->fill_flush_group == flg)
2146                         hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
2147                 hammer_flusher_async(hmp, flg);
2148         }
2149 }
2150
2151 /*
2152  * Callback for scan of ip->rec_tree.  Try to include each record in our
2153  * flush.  ip->flush_group has been set but the inode has not yet been
2154  * moved into a flushing state.
2155  *
2156  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
2157  * both inodes.
2158  *
2159  * We return 1 for any record placed or found in FST_FLUSH, which prevents
2160  * the caller from shortcutting the flush.
2161  */
2162 static int
2163 hammer_setup_child_callback(hammer_record_t rec, void *data)
2164 {
2165         hammer_flush_group_t flg;
2166         hammer_inode_t target_ip;
2167         hammer_inode_t ip;
2168         int r;
2169
2170         /*
2171          * Records deleted or committed by the backend are ignored.
2172          * Note that the flush detects deleted frontend records at
2173          * multiple points to deal with races.  This is just the first
2174          * line of defense.  The only time HAMMER_RECF_DELETED_FE cannot
2175          * be set is when HAMMER_RECF_INTERLOCK_BE is set, because it
2176          * messes up link-count calculations.
2177          *
2178          * NOTE: Don't get confused between record deletion and, say,
2179          * directory entry deletion.  The deletion of a directory entry
2180          * which is on-media has nothing to do with the record deletion
2181          * flags.
2182          */
2183         if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
2184                           HAMMER_RECF_COMMITTED)) {
2185                 if (rec->flush_state == HAMMER_FST_FLUSH) {
2186                         KKASSERT(rec->flush_group == rec->ip->flush_group);
2187                         r = 1;
2188                 } else {
2189                         r = 0;
2190                 }
2191                 return(r);
2192         }
2193
2194         /*
2195          * If the record is in an idle state it has no dependancies and
2196          * can be flushed.
2197          */
2198         ip = rec->ip;
2199         flg = ip->flush_group;
2200         r = 0;
2201
2202         switch(rec->flush_state) {
2203         case HAMMER_FST_IDLE:
2204                 /*
2205                  * The record has no setup dependancy, we can flush it.
2206                  */
2207                 KKASSERT(rec->target_ip == NULL);
2208                 rec->flush_state = HAMMER_FST_FLUSH;
2209                 rec->flush_group = flg;
2210                 ++flg->refs;
2211                 hammer_ref(&rec->lock);
2212                 r = 1;
2213                 break;
2214         case HAMMER_FST_SETUP:
2215                 /*
2216                  * The record has a setup dependancy.  These are typically
2217                  * directory entry adds and deletes.  Such entries will be
2218                  * flushed when their inodes are flushed so we do not
2219                  * usually have to add them to the flush here.  However,
2220                  * if the target_ip has set HAMMER_INODE_CONN_DOWN then
2221                  * it is asking us to flush this record (and it).
2222                  */
2223                 target_ip = rec->target_ip;
2224                 KKASSERT(target_ip != NULL);
2225                 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
2226
2227                 /*
2228                  * If the target IP is already flushing in our group
2229                  * we could associate the record, but target_ip has
2230                  * already synced ino_data to sync_ino_data and we
2231                  * would also have to adjust nlinks.   Plus there are
2232                  * ordering issues for adds and deletes.
2233                  *
2234                  * Reflush downward if this is an ADD, and upward if
2235                  * this is a DEL.
2236                  */
2237                 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
2238                         if (rec->flush_state == HAMMER_MEM_RECORD_ADD)
2239                                 ip->flags |= HAMMER_INODE_REFLUSH;
2240                         else
2241                                 target_ip->flags |= HAMMER_INODE_REFLUSH;
2242                         break;
2243                 } 
2244
2245                 /*
2246                  * Target IP is not yet flushing.  This can get complex
2247                  * because we have to be careful about the recursion.
2248                  *
2249                  * Directories create an issue for us in that if a flush
2250                  * of a directory is requested the expectation is to flush
2251                  * any pending directory entries, but this will cause the
2252                  * related inodes to recursively flush as well.  We can't
2253                  * really defer the operation so just get as many as we
2254                  * can and
2255                  */
2256 #if 0
2257                 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
2258                     (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
2259                         /*
2260                          * We aren't reclaiming and the target ip was not
2261                          * previously prevented from flushing due to this
2262                          * record dependancy.  Do not flush this record.
2263                          */
2264                         /*r = 0;*/
2265                 } else
2266 #endif
2267                 if (flg->total_count + flg->refs >
2268                            ip->hmp->undo_rec_limit) {
2269                         /*
2270                          * Our flush group is over-full and we risk blowing
2271                          * out the UNDO FIFO.  Stop the scan, flush what we
2272                          * have, then reflush the directory.
2273                          *
2274                          * The directory may be forced through multiple
2275                          * flush groups before it can be completely
2276                          * flushed.
2277                          */
2278                         ip->flags |= HAMMER_INODE_RESIGNAL |
2279                                      HAMMER_INODE_REFLUSH;
2280                         r = -1;
2281                 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
2282                         /*
2283                          * If the target IP is not flushing we can force
2284                          * it to flush, even if it is unable to write out
2285                          * any of its own records we have at least one in
2286                          * hand that we CAN deal with.
2287                          */
2288                         rec->flush_state = HAMMER_FST_FLUSH;
2289                         rec->flush_group = flg;
2290                         ++flg->refs;
2291                         hammer_ref(&rec->lock);
2292                         hammer_flush_inode_core(target_ip, flg,
2293                                                 HAMMER_FLUSH_RECURSION);
2294                         r = 1;
2295                 } else {
2296                         /*
2297                          * General or delete-on-disk record.
2298                          *
2299                          * XXX this needs help.  If a delete-on-disk we could
2300                          * disconnect the target.  If the target has its own
2301                          * dependancies they really need to be flushed.
2302                          *
2303                          * XXX
2304                          */
2305                         rec->flush_state = HAMMER_FST_FLUSH;
2306                         rec->flush_group = flg;
2307                         ++flg->refs;
2308                         hammer_ref(&rec->lock);
2309                         hammer_flush_inode_core(target_ip, flg,
2310                                                 HAMMER_FLUSH_RECURSION);
2311                         r = 1;
2312                 }
2313                 break;
2314         case HAMMER_FST_FLUSH:
2315                 /* 
2316                  * The record could be part of a previous flush group if the
2317                  * inode is a directory (the record being a directory entry).
2318                  * Once the flush group was closed a hammer_test_inode()
2319                  * function can cause a new flush group to be setup, placing
2320                  * the directory inode itself in a new flush group.
2321                  *
2322                  * When associated with a previous flush group we count it
2323                  * as if it were in our current flush group, since it will
2324                  * effectively be flushed by the time we flush our current
2325                  * flush group.
2326                  */
2327                 KKASSERT(
2328                     rec->ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY ||
2329                     rec->flush_group == flg);
2330                 r = 1;
2331                 break;
2332         }
2333         return(r);
2334 }
2335
2336 #if 0
2337 /*
2338  * This version just moves records already in a flush state to the new
2339  * flush group and that is it.
2340  */
2341 static int
2342 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
2343 {
2344         hammer_inode_t ip = rec->ip;
2345
2346         switch(rec->flush_state) {
2347         case HAMMER_FST_FLUSH:
2348                 KKASSERT(rec->flush_group == ip->flush_group);
2349                 break;
2350         default:
2351                 break;
2352         }
2353         return(0);
2354 }
2355 #endif
2356
2357 /*
2358  * Wait for a previously queued flush to complete.
2359  *
2360  * If a critical error occured we don't try to wait.
2361  */
2362 void
2363 hammer_wait_inode(hammer_inode_t ip)
2364 {
2365         hammer_flush_group_t flg;
2366
2367         flg = NULL;
2368         if ((ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2369                 while (ip->flush_state != HAMMER_FST_IDLE &&
2370                        (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2371                         if (ip->flush_state == HAMMER_FST_SETUP)
2372                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2373                         if (ip->flush_state != HAMMER_FST_IDLE) {
2374                                 ip->flags |= HAMMER_INODE_FLUSHW;
2375                                 tsleep(&ip->flags, 0, "hmrwin", 0);
2376                         }
2377                 }
2378         }
2379 }
2380
2381 /*
2382  * Called by the backend code when a flush has been completed.
2383  * The inode has already been removed from the flush list.
2384  *
2385  * A pipelined flush can occur, in which case we must re-enter the
2386  * inode on the list and re-copy its fields.
2387  */
2388 void
2389 hammer_flush_inode_done(hammer_inode_t ip, int error)
2390 {
2391         hammer_mount_t hmp;
2392         int dorel;
2393
2394         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2395
2396         hmp = ip->hmp;
2397
2398         /*
2399          * Auto-reflush if the backend could not completely flush
2400          * the inode.  This fixes a case where a deferred buffer flush
2401          * could cause fsync to return early.
2402          */
2403         if (ip->sync_flags & HAMMER_INODE_MODMASK)
2404                 ip->flags |= HAMMER_INODE_REFLUSH;
2405
2406         /*
2407          * Merge left-over flags back into the frontend and fix the state.
2408          * Incomplete truncations are retained by the backend.
2409          */
2410         ip->error = error;
2411         ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2412         ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2413
2414         /*
2415          * The backend may have adjusted nlinks, so if the adjusted nlinks
2416          * does not match the fronttend set the frontend's DDIRTY flag again.
2417          */
2418         if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2419                 ip->flags |= HAMMER_INODE_DDIRTY;
2420
2421         /*
2422          * Fix up the dirty buffer status.
2423          */
2424         if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2425                 ip->flags |= HAMMER_INODE_BUFS;
2426         }
2427         hammer_redo_fifo_end_flush(ip);
2428
2429         /*
2430          * Re-set the XDIRTY flag if some of the inode's in-memory records
2431          * could not be flushed.
2432          */
2433         KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2434                   (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2435                  (!RB_EMPTY(&ip->rec_tree) &&
2436                   (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2437
2438         /*
2439          * Do not lose track of inodes which no longer have vnode
2440          * assocations, otherwise they may never get flushed again.
2441          *
2442          * The reflush flag can be set superfluously, causing extra pain
2443          * for no reason.  If the inode is no longer modified it no longer
2444          * needs to be flushed.
2445          */
2446         if (ip->flags & HAMMER_INODE_MODMASK) {
2447                 if (ip->vp == NULL)
2448                         ip->flags |= HAMMER_INODE_REFLUSH;
2449         } else {
2450                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2451         }
2452
2453         /*
2454          * Adjust the flush state.
2455          */
2456         if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2457                 /*
2458                  * We were unable to flush out all our records, leave the
2459                  * inode in a flush state and in the current flush group.
2460                  * The flush group will be re-run.
2461                  *
2462                  * This occurs if the UNDO block gets too full or there is
2463                  * too much dirty meta-data and allows the flusher to
2464                  * finalize the UNDO block and then re-flush.
2465                  */
2466                 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2467                 dorel = 0;
2468         } else {
2469                 /*
2470                  * Remove from the flush_group
2471                  */
2472                 RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
2473                 ip->flush_group = NULL;
2474
2475 #if 0
2476                 /*
2477                  * Clean up the vnode ref and tracking counts.
2478                  */
2479                 if (ip->flags & HAMMER_INODE_VHELD) {
2480                         ip->flags &= ~HAMMER_INODE_VHELD;
2481                         vrele(ip->vp);
2482                 }
2483 #endif
2484                 --hmp->count_iqueued;
2485                 --hammer_count_iqueued;
2486
2487                 /*
2488                  * And adjust the state.
2489                  */
2490                 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2491                         ip->flush_state = HAMMER_FST_IDLE;
2492                         dorel = 1;
2493                 } else {
2494                         ip->flush_state = HAMMER_FST_SETUP;
2495                         dorel = 0;
2496                 }
2497
2498                 /*
2499                  * If the frontend is waiting for a flush to complete,
2500                  * wake it up.
2501                  */
2502                 if (ip->flags & HAMMER_INODE_FLUSHW) {
2503                         ip->flags &= ~HAMMER_INODE_FLUSHW;
2504                         wakeup(&ip->flags);
2505                 }
2506
2507                 /*
2508                  * If the frontend made more changes and requested another
2509                  * flush, then try to get it running.
2510                  *
2511                  * Reflushes are aborted when the inode is errored out.
2512                  */
2513                 if (ip->flags & HAMMER_INODE_REFLUSH) {
2514                         ip->flags &= ~HAMMER_INODE_REFLUSH;
2515                         if (ip->flags & HAMMER_INODE_RESIGNAL) {
2516                                 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2517                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2518                         } else {
2519                                 hammer_flush_inode(ip, 0);
2520                         }
2521                 }
2522         }
2523
2524         /*
2525          * If we have no parent dependancies we can clear CONN_DOWN
2526          */
2527         if (TAILQ_EMPTY(&ip->target_list))
2528                 ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2529
2530         /*
2531          * If the inode is now clean drop the space reservation.
2532          */
2533         if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2534             (ip->flags & HAMMER_INODE_RSV_INODES)) {
2535                 ip->flags &= ~HAMMER_INODE_RSV_INODES;
2536                 --hmp->rsv_inodes;
2537         }
2538
2539         if (dorel)
2540                 hammer_rel_inode(ip, 0);
2541 }
2542
2543 /*
2544  * Called from hammer_sync_inode() to synchronize in-memory records
2545  * to the media.
2546  */
2547 static int
2548 hammer_sync_record_callback(hammer_record_t record, void *data)
2549 {
2550         hammer_cursor_t cursor = data;
2551         hammer_transaction_t trans = cursor->trans;
2552         hammer_mount_t hmp = trans->hmp;
2553         int error;
2554
2555         /*
2556          * Skip records that do not belong to the current flush.
2557          */
2558         ++hammer_stats_record_iterations;
2559         if (record->flush_state != HAMMER_FST_FLUSH)
2560                 return(0);
2561
2562 #if 1
2563         if (record->flush_group != record->ip->flush_group) {
2564                 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2565                 if (hammer_debug_critical)
2566                         Debugger("blah2");
2567                 return(0);
2568         }
2569 #endif
2570         KKASSERT(record->flush_group == record->ip->flush_group);
2571
2572         /*
2573          * Interlock the record using the BE flag.  Once BE is set the
2574          * frontend cannot change the state of FE.
2575          *
2576          * NOTE: If FE is set prior to us setting BE we still sync the
2577          * record out, but the flush completion code converts it to 
2578          * a delete-on-disk record instead of destroying it.
2579          */
2580         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2581         record->flags |= HAMMER_RECF_INTERLOCK_BE;
2582
2583         /*
2584          * The backend has already disposed of the record.
2585          */
2586         if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
2587                 error = 0;
2588                 goto done;
2589         }
2590
2591         /*
2592          * If the whole inode is being deleting all on-disk records will
2593          * be deleted very soon, we can't sync any new records to disk
2594          * because they will be deleted in the same transaction they were
2595          * created in (delete_tid == create_tid), which will assert.
2596          *
2597          * XXX There may be a case with RECORD_ADD with DELETED_FE set
2598          * that we currently panic on.
2599          */
2600         if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2601                 switch(record->type) {
2602                 case HAMMER_MEM_RECORD_DATA:
2603                         /*
2604                          * We don't have to do anything, if the record was
2605                          * committed the space will have been accounted for
2606                          * in the blockmap.
2607                          */
2608                         /* fall through */
2609                 case HAMMER_MEM_RECORD_GENERAL:
2610                         /*
2611                          * Set deleted-by-backend flag.  Do not set the
2612                          * backend committed flag, because we are throwing
2613                          * the record away.
2614                          */
2615                         record->flags |= HAMMER_RECF_DELETED_BE;
2616                         ++record->ip->rec_generation;
2617                         error = 0;
2618                         goto done;
2619                 case HAMMER_MEM_RECORD_ADD:
2620                         panic("hammer_sync_record_callback: illegal add "
2621                               "during inode deletion record %p", record);
2622                         break; /* NOT REACHED */
2623                 case HAMMER_MEM_RECORD_INODE:
2624                         panic("hammer_sync_record_callback: attempt to "
2625                               "sync inode record %p?", record);
2626                         break; /* NOT REACHED */
2627                 case HAMMER_MEM_RECORD_DEL:
2628                         /* 
2629                          * Follow through and issue the on-disk deletion
2630                          */
2631                         break;
2632                 }
2633         }
2634
2635         /*
2636          * If DELETED_FE is set special handling is needed for directory
2637          * entries.  Dependant pieces related to the directory entry may
2638          * have already been synced to disk.  If this occurs we have to
2639          * sync the directory entry and then change the in-memory record
2640          * from an ADD to a DELETE to cover the fact that it's been
2641          * deleted by the frontend.
2642          *
2643          * A directory delete covering record (MEM_RECORD_DEL) can never
2644          * be deleted by the frontend.
2645          *
2646          * Any other record type (aka DATA) can be deleted by the frontend.
2647          * XXX At the moment the flusher must skip it because there may
2648          * be another data record in the flush group for the same block,
2649          * meaning that some frontend data changes can leak into the backend's
2650          * synchronization point.
2651          */
2652         if (record->flags & HAMMER_RECF_DELETED_FE) {
2653                 if (record->type == HAMMER_MEM_RECORD_ADD) {
2654                         /*
2655                          * Convert a front-end deleted directory-add to
2656                          * a directory-delete entry later.
2657                          */
2658                         record->flags |= HAMMER_RECF_CONVERT_DELETE;
2659                 } else {
2660                         /*
2661                          * Dispose of the record (race case).  Mark as
2662                          * deleted by backend (and not committed).
2663                          */
2664                         KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2665                         record->flags |= HAMMER_RECF_DELETED_BE;
2666                         ++record->ip->rec_generation;
2667                         error = 0;
2668                         goto done;
2669                 }
2670         }
2671
2672         /*
2673          * Assign the create_tid for new records.  Deletions already
2674          * have the record's entire key properly set up.
2675          */
2676         if (record->type != HAMMER_MEM_RECORD_DEL) {
2677                 record->leaf.base.create_tid = trans->tid;
2678                 record->leaf.create_ts = trans->time32;
2679         }
2680
2681         /*
2682          * This actually moves the record to the on-media B-Tree.  We
2683          * must also generate REDO_TERM entries in the UNDO/REDO FIFO
2684          * indicating that the related REDO_WRITE(s) have been committed.
2685          *
2686          * During recovery any REDO_TERM's within the nominal recovery span
2687          * are ignored since the related meta-data is being undone, causing
2688          * any matching REDO_WRITEs to execute.  The REDO_TERMs outside
2689          * the nominal recovery span will match against REDO_WRITEs and
2690          * prevent them from being executed (because the meta-data has
2691          * already been synchronized).
2692          */
2693         if (record->flags & HAMMER_RECF_REDO) {
2694                 KKASSERT(record->type == HAMMER_MEM_RECORD_DATA);
2695                 hammer_generate_redo(trans, record->ip,
2696                                      record->leaf.base.key -
2697                                          record->leaf.data_len,
2698                                      HAMMER_REDO_TERM_WRITE,
2699                                      NULL,
2700                                      record->leaf.data_len);
2701         }
2702         for (;;) {
2703                 error = hammer_ip_sync_record_cursor(cursor, record);
2704                 if (error != EDEADLK)
2705                         break;
2706                 hammer_done_cursor(cursor);
2707                 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2708                                            record->ip);
2709                 if (error)
2710                         break;
2711         }
2712         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2713
2714         if (error)
2715                 error = -error;
2716 done:
2717         hammer_flush_record_done(record, error);
2718
2719         /*
2720          * Do partial finalization if we have built up too many dirty
2721          * buffers.  Otherwise a buffer cache deadlock can occur when
2722          * doing things like creating tens of thousands of tiny files.
2723          *
2724          * We must release our cursor lock to avoid a 3-way deadlock
2725          * due to the exclusive sync lock the finalizer must get.
2726          *
2727          * WARNING: See warnings in hammer_unlock_cursor() function.
2728          */
2729         if (hammer_flusher_meta_limit(hmp)) {
2730                 hammer_unlock_cursor(cursor);
2731                 hammer_flusher_finalize(trans, 0);
2732                 hammer_lock_cursor(cursor);
2733         }
2734
2735         return(error);
2736 }
2737
2738 /*
2739  * Backend function called by the flusher to sync an inode to media.
2740  */
2741 int
2742 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2743 {
2744         struct hammer_cursor cursor;
2745         hammer_node_t tmp_node;
2746         hammer_record_t depend;
2747         hammer_record_t next;
2748         int error, tmp_error;
2749         u_int64_t nlinks;
2750
2751         if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2752                 return(0);
2753
2754         error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2755         if (error)
2756                 goto done;
2757
2758         /*
2759          * Any directory records referencing this inode which are not in
2760          * our current flush group must adjust our nlink count for the
2761          * purposes of synchronizating to disk.
2762          *
2763          * Records which are in our flush group can be unlinked from our
2764          * inode now, potentially allowing the inode to be physically
2765          * deleted.
2766          *
2767          * This cannot block.
2768          */
2769         nlinks = ip->ino_data.nlinks;
2770         next = TAILQ_FIRST(&ip->target_list);
2771         while ((depend = next) != NULL) {
2772                 next = TAILQ_NEXT(depend, target_entry);
2773                 if (depend->flush_state == HAMMER_FST_FLUSH &&
2774                     depend->flush_group == ip->flush_group) {
2775                         /*
2776                          * If this is an ADD that was deleted by the frontend
2777                          * the frontend nlinks count will have already been
2778                          * decremented, but the backend is going to sync its
2779                          * directory entry and must account for it.  The
2780                          * record will be converted to a delete-on-disk when
2781                          * it gets synced.
2782                          *
2783                          * If the ADD was not deleted by the frontend we
2784                          * can remove the dependancy from our target_list.
2785                          */
2786                         if (depend->flags & HAMMER_RECF_DELETED_FE) {
2787                                 ++nlinks;
2788                         } else {
2789                                 TAILQ_REMOVE(&ip->target_list, depend,
2790                                              target_entry);
2791                                 depend->target_ip = NULL;
2792                         }
2793                 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2794                         /*
2795                          * Not part of our flush group and not deleted by
2796                          * the front-end, adjust the link count synced to
2797                          * the media (undo what the frontend did when it
2798                          * queued the record).
2799                          */
2800                         KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2801                         switch(depend->type) {
2802                         case HAMMER_MEM_RECORD_ADD:
2803                                 --nlinks;
2804                                 break;
2805                         case HAMMER_MEM_RECORD_DEL:
2806                                 ++nlinks;
2807                                 break;
2808                         default:
2809                                 break;
2810                         }
2811                 }
2812         }
2813
2814         /*
2815          * Set dirty if we had to modify the link count.
2816          */
2817         if (ip->sync_ino_data.nlinks != nlinks) {
2818                 KKASSERT((int64_t)nlinks >= 0);
2819                 ip->sync_ino_data.nlinks = nlinks;
2820                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2821         }
2822
2823         /*
2824          * If there is a trunction queued destroy any data past the (aligned)
2825          * truncation point.  Userland will have dealt with the buffer
2826          * containing the truncation point for us.
2827          *
2828          * We don't flush pending frontend data buffers until after we've
2829          * dealt with the truncation.
2830          */
2831         if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2832                 /*
2833                  * Interlock trunc_off.  The VOP front-end may continue to
2834                  * make adjustments to it while we are blocked.
2835                  */
2836                 off_t trunc_off;
2837                 off_t aligned_trunc_off;
2838                 int blkmask;
2839
2840                 trunc_off = ip->sync_trunc_off;
2841                 blkmask = hammer_blocksize(trunc_off) - 1;
2842                 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
2843
2844                 /*
2845                  * Delete any whole blocks on-media.  The front-end has
2846                  * already cleaned out any partial block and made it
2847                  * pending.  The front-end may have updated trunc_off
2848                  * while we were blocked so we only use sync_trunc_off.
2849                  *
2850                  * This operation can blow out the buffer cache, EWOULDBLOCK
2851                  * means we were unable to complete the deletion.  The
2852                  * deletion will update sync_trunc_off in that case.
2853                  */
2854                 error = hammer_ip_delete_range(&cursor, ip,
2855                                                 aligned_trunc_off,
2856                                                 0x7FFFFFFFFFFFFFFFLL, 2);
2857                 if (error == EWOULDBLOCK) {
2858                         ip->flags |= HAMMER_INODE_WOULDBLOCK;
2859                         error = 0;
2860                         goto defer_buffer_flush;
2861                 }
2862
2863                 if (error)
2864                         goto done;
2865
2866                 /*
2867                  * Generate a REDO_TERM_TRUNC entry in the UNDO/REDO FIFO.
2868                  *
2869                  * XXX we do this even if we did not previously generate
2870                  * a REDO_TRUNC record.  This operation may enclosed the
2871                  * range for multiple prior truncation entries in the REDO
2872                  * log.
2873                  */
2874                 if (trans->hmp->version >= HAMMER_VOL_VERSION_FOUR &&
2875                     (ip->flags & HAMMER_INODE_RDIRTY)) {
2876                         hammer_generate_redo(trans, ip, aligned_trunc_off,
2877                                              HAMMER_REDO_TERM_TRUNC,
2878                                              NULL, 0);
2879                 }
2880
2881                 /*
2882                  * Clear the truncation flag on the backend after we have
2883                  * completed the deletions.  Backend data is now good again
2884                  * (including new records we are about to sync, below).
2885                  *
2886                  * Leave sync_trunc_off intact.  As we write additional
2887                  * records the backend will update sync_trunc_off.  This
2888                  * tells the backend whether it can skip the overwrite
2889                  * test.  This should work properly even when the backend
2890                  * writes full blocks where the truncation point straddles
2891                  * the block because the comparison is against the base
2892                  * offset of the record.
2893                  */
2894                 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2895                 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2896         } else {
2897                 error = 0;
2898         }
2899
2900         /*
2901          * Now sync related records.  These will typically be directory
2902          * entries, records tracking direct-writes, or delete-on-disk records.
2903          */
2904         if (error == 0) {
2905                 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2906                                     hammer_sync_record_callback, &cursor);
2907                 if (tmp_error < 0)
2908                         tmp_error = -error;
2909                 if (tmp_error)
2910                         error = tmp_error;
2911         }
2912         hammer_cache_node(&ip->cache[1], cursor.node);
2913
2914         /*
2915          * Re-seek for inode update, assuming our cache hasn't been ripped
2916          * out from under us.
2917          */
2918         if (error == 0) {
2919                 tmp_node = hammer_ref_node_safe(trans, &ip->cache[0], &error);
2920                 if (tmp_node) {
2921                         hammer_cursor_downgrade(&cursor);
2922                         hammer_lock_sh(&tmp_node->lock);
2923                         if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
2924                                 hammer_cursor_seek(&cursor, tmp_node, 0);
2925                         hammer_unlock(&tmp_node->lock);
2926                         hammer_rel_node(tmp_node);
2927                 }
2928                 error = 0;
2929         }
2930
2931         /*
2932          * If we are deleting the inode the frontend had better not have
2933          * any active references on elements making up the inode.
2934          *
2935          * The call to hammer_ip_delete_clean() cleans up auxillary records
2936          * but not DB or DATA records.  Those must have already been deleted
2937          * by the normal truncation mechanic.
2938          */
2939         if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
2940                 RB_EMPTY(&ip->rec_tree)  &&
2941             (ip->sync_flags & HAMMER_INODE_DELETING) &&
2942             (ip->flags & HAMMER_INODE_DELETED) == 0) {
2943                 int count1 = 0;
2944
2945                 error = hammer_ip_delete_clean(&cursor, ip, &count1);
2946                 if (error == 0) {
2947                         ip->flags |= HAMMER_INODE_DELETED;
2948                         ip->sync_flags &= ~HAMMER_INODE_DELETING;
2949                         ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2950                         KKASSERT(RB_EMPTY(&ip->rec_tree));
2951
2952                         /*
2953                          * Set delete_tid in both the frontend and backend
2954                          * copy of the inode record.  The DELETED flag handles
2955                          * this, do not set DDIRTY.
2956                          */
2957                         ip->ino_leaf.base.delete_tid = trans->tid;
2958                         ip->sync_ino_leaf.base.delete_tid = trans->tid;
2959                         ip->ino_leaf.delete_ts = trans->time32;
2960                         ip->sync_ino_leaf.delete_ts = trans->time32;
2961
2962
2963                         /*
2964                          * Adjust the inode count in the volume header
2965                          */
2966                         hammer_sync_lock_sh(trans);
2967                         if (ip->flags & HAMMER_INODE_ONDISK) {
2968                                 hammer_modify_volume_field(trans,
2969                                                            trans->rootvol,
2970                                                            vol0_stat_inodes);
2971                                 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
2972                                 hammer_modify_volume_done(trans->rootvol);
2973                         }
2974                         hammer_sync_unlock(trans);
2975                 }
2976         }
2977
2978         if (error)
2979                 goto done;
2980         ip->sync_flags &= ~HAMMER_INODE_BUFS;
2981
2982 defer_buffer_flush:
2983         /*
2984          * Now update the inode's on-disk inode-data and/or on-disk record.
2985          * DELETED and ONDISK are managed only in ip->flags.
2986          *
2987          * In the case of a defered buffer flush we still update the on-disk
2988          * inode to satisfy visibility requirements if there happen to be
2989          * directory dependancies.
2990          */
2991         switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
2992         case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
2993                 /*
2994                  * If deleted and on-disk, don't set any additional flags.
2995                  * the delete flag takes care of things.
2996                  *
2997                  * Clear flags which may have been set by the frontend.
2998                  */
2999                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
3000                                     HAMMER_INODE_SDIRTY |
3001                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
3002                                     HAMMER_INODE_DELETING);
3003                 break;
3004         case HAMMER_INODE_DELETED:
3005                 /*
3006                  * Take care of the case where a deleted inode was never
3007                  * flushed to the disk in the first place.
3008                  *
3009                  * Clear flags which may have been set by the frontend.
3010                  */
3011                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
3012                                     HAMMER_INODE_SDIRTY |
3013                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
3014                                     HAMMER_INODE_DELETING);
3015                 while (RB_ROOT(&ip->rec_tree)) {
3016                         hammer_record_t record = RB_ROOT(&ip->rec_tree);
3017                         hammer_ref(&record->lock);
3018                         KKASSERT(hammer_oneref(&record->lock));
3019                         record->flags |= HAMMER_RECF_DELETED_BE;
3020                         ++record->ip->rec_generation;
3021                         hammer_rel_mem_record(record);
3022                 }
3023                 break;
3024         case HAMMER_INODE_ONDISK:
3025                 /*
3026                  * If already on-disk, do not set any additional flags.
3027                  */
3028                 break;
3029         default:
3030                 /*
3031                  * If not on-disk and not deleted, set DDIRTY to force
3032                  * an initial record to be written.
3033                  *
3034                  * Also set the create_tid in both the frontend and backend
3035                  * copy of the inode record.
3036                  */
3037                 ip->ino_leaf.base.create_tid = trans->tid;
3038                 ip->ino_leaf.create_ts = trans->time32;
3039                 ip->sync_ino_leaf.base.create_tid = trans->tid;
3040                 ip->sync_ino_leaf.create_ts = trans->time32;
3041                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
3042                 break;
3043         }
3044
3045         /*
3046          * If DDIRTY or SDIRTY is set, write out a new record.
3047          * If the inode is already on-disk the old record is marked as
3048          * deleted.
3049          *
3050          * If DELETED is set hammer_update_inode() will delete the existing
3051          * record without writing out a new one.
3052          *
3053          * If *ONLY* the ITIMES flag is set we can update the record in-place.
3054          */
3055         if (ip->flags & HAMMER_INODE_DELETED) {
3056                 error = hammer_update_inode(&cursor, ip);
3057         } else 
3058         if (!(ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY)) &&
3059             (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
3060                 error = hammer_update_itimes(&cursor, ip);
3061         } else
3062         if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY |
3063                               HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
3064                 error = hammer_update_inode(&cursor, ip);
3065         }
3066 done:
3067         if (error) {
3068                 hammer_critical_error(ip->hmp, ip, error,
3069                                       "while syncing inode");
3070         }
3071         hammer_done_cursor(&cursor);
3072         return(error);
3073 }
3074
3075 /*
3076  * This routine is called when the OS is no longer actively referencing
3077  * the inode (but might still be keeping it cached), or when releasing
3078  * the last reference to an inode.
3079  *
3080  * At this point if the inode's nlinks count is zero we want to destroy
3081  * it, which may mean destroying it on-media too.
3082  */
3083 void
3084 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
3085 {
3086         struct vnode *vp;
3087
3088         /*
3089          * Set the DELETING flag when the link count drops to 0 and the
3090          * OS no longer has any opens on the inode.
3091          *
3092          * The backend will clear DELETING (a mod flag) and set DELETED
3093          * (a state flag) when it is actually able to perform the
3094          * operation.
3095          *
3096          * Don't reflag the deletion if the flusher is currently syncing
3097          * one that was already flagged.  A previously set DELETING flag
3098          * may bounce around flags and sync_flags until the operation is
3099          * completely done.
3100          *
3101          * Do not attempt to modify a snapshot inode (one set to read-only).
3102          */
3103         if (ip->ino_data.nlinks == 0 &&
3104             ((ip->flags | ip->sync_flags) & (HAMMER_INODE_RO|HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
3105                 ip->flags |= HAMMER_INODE_DELETING;
3106                 ip->flags |= HAMMER_INODE_TRUNCATED;
3107                 ip->trunc_off = 0;
3108                 vp = NULL;
3109                 if (getvp) {
3110                         if (hammer_get_vnode(ip, &vp) != 0)
3111                                 return;
3112                 }
3113
3114                 /*
3115                  * Final cleanup
3116                  */
3117                 if (ip->vp)
3118                         nvtruncbuf(ip->vp, 0, HAMMER_BUFSIZE, 0);
3119                 if (getvp)
3120                         vput(vp);
3121         }
3122 }
3123
3124 /*
3125  * After potentially resolving a dependancy the inode is tested
3126  * to determine whether it needs to be reflushed.
3127  */
3128 void
3129 hammer_test_inode(hammer_inode_t ip)
3130 {
3131         if (ip->flags & HAMMER_INODE_REFLUSH) {
3132                 ip->flags &= ~HAMMER_INODE_REFLUSH;
3133                 hammer_ref(&ip->lock);
3134                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
3135                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
3136                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
3137                 } else {
3138                         hammer_flush_inode(ip, 0);
3139                 }
3140                 hammer_rel_inode(ip, 0);
3141         }
3142 }
3143
3144 /*
3145  * Clear the RECLAIM flag on an inode.  This occurs when the inode is
3146  * reassociated with a vp or just before it gets freed.
3147  *
3148  * Pipeline wakeups to threads blocked due to an excessive number of
3149  * detached inodes.  This typically occurs when atime updates accumulate
3150  * while scanning a directory tree.
3151  */
3152 static void
3153 hammer_inode_wakereclaims(hammer_inode_t ip)
3154 {
3155         struct hammer_reclaim *reclaim;
3156         hammer_mount_t hmp = ip->hmp;
3157
3158         if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
3159                 return;
3160
3161         --hammer_count_reclaiming;
3162         --hmp->inode_reclaims;
3163         ip->flags &= ~HAMMER_INODE_RECLAIM;
3164
3165         if ((reclaim = TAILQ_FIRST(&hmp->reclaim_list)) != NULL) {
3166                 KKASSERT(reclaim->count > 0);
3167                 if (--reclaim->count == 0) {
3168                         TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
3169                         wakeup(reclaim);
3170                 }
3171         }
3172 }
3173
3174 /*
3175  * Setup our reclaim pipeline.  We only let so many detached (and dirty)
3176  * inodes build up before we start blocking.  This routine is called
3177  * if a new inode is created or an inode is loaded from media.
3178  *
3179  * When we block we don't care *which* inode has finished reclaiming,
3180  * as lone as one does.
3181  *
3182  * The reclaim pipeline is primary governed by the auto-flush which is
3183  * 1/4 hammer_limit_reclaim.  We don't want to block if the count is
3184  * less than 1/2 hammer_limit_reclaim.  From 1/2 to full count is
3185  * dynamically governed.
3186  */
3187 void
3188 hammer_inode_waitreclaims(hammer_transaction_t trans)
3189 {
3190         hammer_mount_t hmp = trans->hmp;
3191         struct hammer_reclaim reclaim;
3192         int lower_limit;
3193
3194         /*
3195          * Track inode load, delay if the number of reclaiming inodes is
3196          * between 2/4 and 4/4 hammer_limit_reclaim, depending.
3197          */
3198         if (curthread->td_proc) {
3199                 struct hammer_inostats *stats;
3200
3201                 stats = hammer_inode_inostats(hmp, curthread->td_proc->p_pid);
3202                 ++stats->count;
3203
3204                 if (stats->count > hammer_limit_reclaim / 2)
3205                         stats->count = hammer_limit_reclaim / 2;
3206                 lower_limit = hammer_limit_reclaim - stats->count;
3207                 if (hammer_debug_general & 0x10000) {
3208                         kprintf("pid %5d limit %d\n",
3209                                 (int)curthread->td_proc->p_pid, lower_limit);
3210                 }
3211         } else {
3212                 lower_limit = hammer_limit_reclaim * 3 / 4;
3213         }
3214         if (hmp->inode_reclaims >= lower_limit) {
3215                 reclaim.count = 1;
3216                 TAILQ_INSERT_TAIL(&hmp->reclaim_list, &reclaim, entry);
3217                 tsleep(&reclaim, 0, "hmrrcm", hz);
3218                 if (reclaim.count > 0)
3219                         TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
3220         }
3221 }
3222
3223 /*
3224  * Keep track of reclaim statistics on a per-pid basis using a loose
3225  * 4-way set associative hash table.  Collisions inherit the count of
3226  * the previous entry.
3227  *
3228  * NOTE: We want to be careful here to limit the chain size.  If the chain
3229  *       size is too large a pid will spread its stats out over too many
3230  *       entries under certain types of heavy filesystem activity and
3231  *       wind up not delaying long enough.
3232  */
3233 static
3234 struct hammer_inostats *
3235 hammer_inode_inostats(hammer_mount_t hmp, pid_t pid)
3236 {
3237         struct hammer_inostats *stats;
3238         int delta;
3239         int chain;
3240         static volatile int iterator;   /* we don't care about MP races */
3241
3242         /*
3243          * Chain up to 4 times to find our entry.
3244          */
3245         for (chain = 0; chain < 4; ++chain) {
3246                 stats = &hmp->inostats[(pid + chain) & HAMMER_INOSTATS_HMASK];
3247                 if (stats->pid == pid)
3248                         break;
3249         }
3250
3251         /*
3252          * Replace one of the four chaining entries with our new entry.
3253          */
3254         if (chain == 4) {
3255                 stats = &hmp->inostats[(pid + (iterator++ & 3)) &
3256                                        HAMMER_INOSTATS_HMASK];
3257                 stats->pid = pid;
3258         }
3259
3260         /*
3261          * Decay the entry
3262          */
3263         if (stats->count && stats->ltick != ticks) {
3264                 delta = ticks - stats->ltick;
3265                 stats->ltick = ticks;
3266                 if (delta <= 0 || delta > hz * 60)
3267                         stats->count = 0;
3268                 else
3269                         stats->count = stats->count * hz / (hz + delta);
3270         }
3271         if (hammer_debug_general & 0x10000)
3272                 kprintf("pid %5d stats %d\n", (int)pid, stats->count);
3273         return (stats);
3274 }
3275
3276 #if 0
3277
3278 /*
3279  * XXX not used, doesn't work very well due to the large batching nature
3280  * of flushes.
3281  *
3282  * A larger then normal backlog of inodes is sitting in the flusher,
3283  * enforce a general slowdown to let it catch up.  This routine is only
3284  * called on completion of a non-flusher-related transaction which
3285  * performed B-Tree node I/O.
3286  *
3287  * It is possible for the flusher to stall in a continuous load.
3288  * blogbench -i1000 -o seems to do a good job generating this sort of load.
3289  * If the flusher is unable to catch up the inode count can bloat until
3290  * we run out of kvm.
3291  *
3292  * This is a bit of a hack.
3293  */
3294 void
3295 hammer_inode_waithard(hammer_mount_t hmp)
3296 {
3297         /*
3298          * Hysteresis.
3299          */
3300         if (hmp->flags & HAMMER_MOUNT_FLUSH_RECOVERY) {
3301                 if (hmp->inode_reclaims < hammer_limit_reclaim / 2 &&
3302                     hmp->count_iqueued < hmp->count_inodes / 20) {
3303                         hmp->flags &= ~HAMMER_MOUNT_FLUSH_RECOVERY;
3304                         return;
3305                 }
3306         } else {
3307                 if (hmp->inode_reclaims < hammer_limit_reclaim ||
3308                     hmp->count_iqueued < hmp->count_inodes / 10) {
3309                         return;
3310                 }
3311                 hmp->flags |= HAMMER_MOUNT_FLUSH_RECOVERY;
3312         }
3313
3314         /*
3315          * Block for one flush cycle.
3316          */
3317         hammer_flusher_wait_next(hmp);
3318 }
3319
3320 #endif