27e6bc792cd42f320cdac25895461e8f6de4eca2
[dragonfly.git] / sys / vfs / hammer / hammer_inode.c
1 /*
2  * Copyright (c) 2007-2008 The DragonFly Project.  All rights reserved.
3  * 
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  * 
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  * 
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  * 
34  * $DragonFly: src/sys/vfs/hammer/hammer_inode.c,v 1.114 2008/09/24 00:53:51 dillon Exp $
35  */
36
37 #include "hammer.h"
38 #include <vm/vm_extern.h>
39 #include <sys/buf.h>
40 #include <sys/buf2.h>
41
42 static int      hammer_unload_inode(struct hammer_inode *ip);
43 static void     hammer_free_inode(hammer_inode_t ip);
44 static void     hammer_flush_inode_core(hammer_inode_t ip,
45                                         hammer_flush_group_t flg, int flags);
46 static int      hammer_setup_child_callback(hammer_record_t rec, void *data);
47 #if 0
48 static int      hammer_syncgrp_child_callback(hammer_record_t rec, void *data);
49 #endif
50 static int      hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
51                                         hammer_flush_group_t flg);
52 static int      hammer_setup_parent_inodes_helper(hammer_record_t record,
53                                         int depth, hammer_flush_group_t flg);
54 static void     hammer_inode_wakereclaims(hammer_inode_t ip, int dowake);
55
56 #ifdef DEBUG_TRUNCATE
57 extern struct hammer_inode *HammerTruncIp;
58 #endif
59
60 /*
61  * RB-Tree support for inode structures
62  */
63 int
64 hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
65 {
66         if (ip1->obj_localization < ip2->obj_localization)
67                 return(-1);
68         if (ip1->obj_localization > ip2->obj_localization)
69                 return(1);
70         if (ip1->obj_id < ip2->obj_id)
71                 return(-1);
72         if (ip1->obj_id > ip2->obj_id)
73                 return(1);
74         if (ip1->obj_asof < ip2->obj_asof)
75                 return(-1);
76         if (ip1->obj_asof > ip2->obj_asof)
77                 return(1);
78         return(0);
79 }
80
81 /*
82  * RB-Tree support for inode structures / special LOOKUP_INFO
83  */
84 static int
85 hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
86 {
87         if (info->obj_localization < ip->obj_localization)
88                 return(-1);
89         if (info->obj_localization > ip->obj_localization)
90                 return(1);
91         if (info->obj_id < ip->obj_id)
92                 return(-1);
93         if (info->obj_id > ip->obj_id)
94                 return(1);
95         if (info->obj_asof < ip->obj_asof)
96                 return(-1);
97         if (info->obj_asof > ip->obj_asof)
98                 return(1);
99         return(0);
100 }
101
102 /*
103  * Used by hammer_scan_inode_snapshots() to locate all of an object's
104  * snapshots.  Note that the asof field is not tested, which we can get
105  * away with because it is the lowest-priority field.
106  */
107 static int
108 hammer_inode_info_cmp_all_history(hammer_inode_t ip, void *data)
109 {
110         hammer_inode_info_t info = data;
111
112         if (ip->obj_localization > info->obj_localization)
113                 return(1);
114         if (ip->obj_localization < info->obj_localization)
115                 return(-1);
116         if (ip->obj_id > info->obj_id)
117                 return(1);
118         if (ip->obj_id < info->obj_id)
119                 return(-1);
120         return(0);
121 }
122
123 /*
124  * Used by hammer_unload_pseudofs() to locate all inodes associated with
125  * a particular PFS.
126  */
127 static int
128 hammer_inode_pfs_cmp(hammer_inode_t ip, void *data)
129 {
130         u_int32_t localization = *(u_int32_t *)data;
131         if (ip->obj_localization > localization)
132                 return(1);
133         if (ip->obj_localization < localization)
134                 return(-1);
135         return(0);
136 }
137
138 /*
139  * RB-Tree support for pseudofs structures
140  */
141 static int
142 hammer_pfs_rb_compare(hammer_pseudofs_inmem_t p1, hammer_pseudofs_inmem_t p2)
143 {
144         if (p1->localization < p2->localization)
145                 return(-1);
146         if (p1->localization > p2->localization)
147                 return(1);
148         return(0);
149 }
150
151
152 RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
153 RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
154                 hammer_inode_info_cmp, hammer_inode_info_t);
155 RB_GENERATE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
156              hammer_pfs_rb_compare, u_int32_t, localization);
157
158 /*
159  * The kernel is not actively referencing this vnode but is still holding
160  * it cached.
161  *
162  * This is called from the frontend.
163  */
164 int
165 hammer_vop_inactive(struct vop_inactive_args *ap)
166 {
167         struct hammer_inode *ip = VTOI(ap->a_vp);
168
169         /*
170          * Degenerate case
171          */
172         if (ip == NULL) {
173                 vrecycle(ap->a_vp);
174                 return(0);
175         }
176
177         /*
178          * If the inode no longer has visibility in the filesystem try to
179          * recycle it immediately, even if the inode is dirty.  Recycling
180          * it quickly allows the system to reclaim buffer cache and VM
181          * resources which can matter a lot in a heavily loaded system.
182          *
183          * This can deadlock in vfsync() if we aren't careful.
184          * 
185          * Do not queue the inode to the flusher if we still have visibility,
186          * otherwise namespace calls such as chmod will unnecessarily generate
187          * multiple inode updates.
188          */
189         hammer_inode_unloadable_check(ip, 0);
190         if (ip->ino_data.nlinks == 0) {
191                 if (ip->flags & HAMMER_INODE_MODMASK)
192                         hammer_flush_inode(ip, 0);
193                 vrecycle(ap->a_vp);
194         }
195         return(0);
196 }
197
198 /*
199  * Release the vnode association.  This is typically (but not always)
200  * the last reference on the inode.
201  *
202  * Once the association is lost we are on our own with regards to
203  * flushing the inode.
204  */
205 int
206 hammer_vop_reclaim(struct vop_reclaim_args *ap)
207 {
208         struct hammer_inode *ip;
209         hammer_mount_t hmp;
210         struct vnode *vp;
211
212         vp = ap->a_vp;
213
214         if ((ip = vp->v_data) != NULL) {
215                 hmp = ip->hmp;
216                 vp->v_data = NULL;
217                 ip->vp = NULL;
218
219                 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0) {
220                         ++hammer_count_reclaiming;
221                         ++hmp->inode_reclaims;
222                         ip->flags |= HAMMER_INODE_RECLAIM;
223                 }
224                 hammer_rel_inode(ip, 1);
225         }
226         return(0);
227 }
228
229 /*
230  * Return a locked vnode for the specified inode.  The inode must be
231  * referenced but NOT LOCKED on entry and will remain referenced on
232  * return.
233  *
234  * Called from the frontend.
235  */
236 int
237 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
238 {
239         hammer_mount_t hmp;
240         struct vnode *vp;
241         int error = 0;
242         u_int8_t obj_type;
243
244         hmp = ip->hmp;
245
246         for (;;) {
247                 if ((vp = ip->vp) == NULL) {
248                         error = getnewvnode(VT_HAMMER, hmp->mp, vpp, 0, 0);
249                         if (error)
250                                 break;
251                         hammer_lock_ex(&ip->lock);
252                         if (ip->vp != NULL) {
253                                 hammer_unlock(&ip->lock);
254                                 vp = *vpp;
255                                 vp->v_type = VBAD;
256                                 vx_put(vp);
257                                 continue;
258                         }
259                         hammer_ref(&ip->lock);
260                         vp = *vpp;
261                         ip->vp = vp;
262
263                         obj_type = ip->ino_data.obj_type;
264                         vp->v_type = hammer_get_vnode_type(obj_type);
265
266                         hammer_inode_wakereclaims(ip, 0);
267
268                         switch(ip->ino_data.obj_type) {
269                         case HAMMER_OBJTYPE_CDEV:
270                         case HAMMER_OBJTYPE_BDEV:
271                                 vp->v_ops = &hmp->mp->mnt_vn_spec_ops;
272                                 addaliasu(vp, ip->ino_data.rmajor,
273                                           ip->ino_data.rminor);
274                                 break;
275                         case HAMMER_OBJTYPE_FIFO:
276                                 vp->v_ops = &hmp->mp->mnt_vn_fifo_ops;
277                                 break;
278                         default:
279                                 break;
280                         }
281
282                         /*
283                          * Only mark as the root vnode if the ip is not
284                          * historical, otherwise the VFS cache will get
285                          * confused.  The other half of the special handling
286                          * is in hammer_vop_nlookupdotdot().
287                          *
288                          * Pseudo-filesystem roots can be accessed via
289                          * non-root filesystem paths and setting VROOT may
290                          * confuse the namecache.  Set VPFSROOT instead.
291                          */
292                         if (ip->obj_id == HAMMER_OBJID_ROOT &&
293                             ip->obj_asof == hmp->asof) {
294                                 if (ip->obj_localization == 0)
295                                         vp->v_flag |= VROOT;
296                                 else
297                                         vp->v_flag |= VPFSROOT;
298                         }
299
300                         vp->v_data = (void *)ip;
301                         /* vnode locked by getnewvnode() */
302                         /* make related vnode dirty if inode dirty? */
303                         hammer_unlock(&ip->lock);
304                         if (vp->v_type == VREG)
305                                 vinitvmio(vp, ip->ino_data.size);
306                         break;
307                 }
308
309                 /*
310                  * loop if the vget fails (aka races), or if the vp
311                  * no longer matches ip->vp.
312                  */
313                 if (vget(vp, LK_EXCLUSIVE) == 0) {
314                         if (vp == ip->vp)
315                                 break;
316                         vput(vp);
317                 }
318         }
319         *vpp = vp;
320         return(error);
321 }
322
323 /*
324  * Locate all copies of the inode for obj_id compatible with the specified
325  * asof, reference, and issue the related call-back.  This routine is used
326  * for direct-io invalidation and does not create any new inodes.
327  */
328 void
329 hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo,
330                             int (*callback)(hammer_inode_t ip, void *data),
331                             void *data)
332 {
333         hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root,
334                                    hammer_inode_info_cmp_all_history,
335                                    callback, iinfo);
336 }
337
338 /*
339  * Acquire a HAMMER inode.  The returned inode is not locked.  These functions
340  * do not attach or detach the related vnode (use hammer_get_vnode() for
341  * that).
342  *
343  * The flags argument is only applied for newly created inodes, and only
344  * certain flags are inherited.
345  *
346  * Called from the frontend.
347  */
348 struct hammer_inode *
349 hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip,
350                  int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
351                  int flags, int *errorp)
352 {
353         hammer_mount_t hmp = trans->hmp;
354         struct hammer_node_cache *cachep;
355         struct hammer_inode_info iinfo;
356         struct hammer_cursor cursor;
357         struct hammer_inode *ip;
358
359
360         /*
361          * Determine if we already have an inode cached.  If we do then
362          * we are golden.
363          *
364          * If we find an inode with no vnode we have to mark the
365          * transaction such that hammer_inode_waitreclaims() is
366          * called later on to avoid building up an infinite number
367          * of inodes.  Otherwise we can continue to * add new inodes
368          * faster then they can be disposed of, even with the tsleep
369          * delay.
370          *
371          * If we find a dummy inode we return a failure so dounlink
372          * (which does another lookup) doesn't try to mess with the
373          * link count.  hammer_vop_nresolve() uses hammer_get_dummy_inode()
374          * to ref dummy inodes.
375          */
376         iinfo.obj_id = obj_id;
377         iinfo.obj_asof = asof;
378         iinfo.obj_localization = localization;
379 loop:
380         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
381         if (ip) {
382                 if (ip->flags & HAMMER_INODE_DUMMY) {
383                         *errorp = ENOENT;
384                         return(NULL);
385                 }
386                 hammer_ref(&ip->lock);
387                 *errorp = 0;
388                 return(ip);
389         }
390
391         /*
392          * Allocate a new inode structure and deal with races later.
393          */
394         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
395         ++hammer_count_inodes;
396         ++hmp->count_inodes;
397         ip->obj_id = obj_id;
398         ip->obj_asof = iinfo.obj_asof;
399         ip->obj_localization = localization;
400         ip->hmp = hmp;
401         ip->flags = flags & HAMMER_INODE_RO;
402         ip->cache[0].ip = ip;
403         ip->cache[1].ip = ip;
404         ip->cache[2].ip = ip;
405         ip->cache[3].ip = ip;
406         if (hmp->ronly)
407                 ip->flags |= HAMMER_INODE_RO;
408         ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
409                 0x7FFFFFFFFFFFFFFFLL;
410         RB_INIT(&ip->rec_tree);
411         TAILQ_INIT(&ip->target_list);
412         hammer_ref(&ip->lock);
413
414         /*
415          * Locate the on-disk inode.  If this is a PFS root we always
416          * access the current version of the root inode and (if it is not
417          * a master) always access information under it with a snapshot
418          * TID.
419          *
420          * We cache recent inode lookups in this directory in dip->cache[2].
421          * If we can't find it we assume the inode we are looking for is
422          * close to the directory inode.
423          */
424 retry:
425         cachep = NULL;
426         if (dip) {
427                 if (dip->cache[2].node)
428                         cachep = &dip->cache[2];
429                 else
430                         cachep = &dip->cache[0];
431         }
432         hammer_init_cursor(trans, &cursor, cachep, NULL);
433         cursor.key_beg.localization = localization + HAMMER_LOCALIZE_INODE;
434         cursor.key_beg.obj_id = ip->obj_id;
435         cursor.key_beg.key = 0;
436         cursor.key_beg.create_tid = 0;
437         cursor.key_beg.delete_tid = 0;
438         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
439         cursor.key_beg.obj_type = 0;
440
441         cursor.asof = iinfo.obj_asof;
442         cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
443                        HAMMER_CURSOR_ASOF;
444
445         *errorp = hammer_btree_lookup(&cursor);
446         if (*errorp == EDEADLK) {
447                 hammer_done_cursor(&cursor);
448                 goto retry;
449         }
450
451         /*
452          * On success the B-Tree lookup will hold the appropriate
453          * buffer cache buffers and provide a pointer to the requested
454          * information.  Copy the information to the in-memory inode
455          * and cache the B-Tree node to improve future operations.
456          */
457         if (*errorp == 0) {
458                 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
459                 ip->ino_data = cursor.data->inode;
460
461                 /*
462                  * cache[0] tries to cache the location of the object inode.
463                  * The assumption is that it is near the directory inode.
464                  *
465                  * cache[1] tries to cache the location of the object data.
466                  * We might have something in the governing directory from
467                  * scan optimizations (see the strategy code in
468                  * hammer_vnops.c).
469                  *
470                  * We update dip->cache[2], if possible, with the location
471                  * of the object inode for future directory shortcuts.
472                  */
473                 hammer_cache_node(&ip->cache[0], cursor.node);
474                 if (dip) {
475                         if (dip->cache[3].node) {
476                                 hammer_cache_node(&ip->cache[1],
477                                                   dip->cache[3].node);
478                         }
479                         hammer_cache_node(&dip->cache[2], cursor.node);
480                 }
481
482                 /*
483                  * The file should not contain any data past the file size
484                  * stored in the inode.  Setting save_trunc_off to the
485                  * file size instead of max reduces B-Tree lookup overheads
486                  * on append by allowing the flusher to avoid checking for
487                  * record overwrites.
488                  */
489                 ip->save_trunc_off = ip->ino_data.size;
490
491                 /*
492                  * Locate and assign the pseudofs management structure to
493                  * the inode.
494                  */
495                 if (dip && dip->obj_localization == ip->obj_localization) {
496                         ip->pfsm = dip->pfsm;
497                         hammer_ref(&ip->pfsm->lock);
498                 } else {
499                         ip->pfsm = hammer_load_pseudofs(trans,
500                                                         ip->obj_localization,
501                                                         errorp);
502                         *errorp = 0;    /* ignore ENOENT */
503                 }
504         }
505
506         /*
507          * The inode is placed on the red-black tree and will be synced to
508          * the media when flushed or by the filesystem sync.  If this races
509          * another instantiation/lookup the insertion will fail.
510          */
511         if (*errorp == 0) {
512                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
513                         hammer_free_inode(ip);
514                         hammer_done_cursor(&cursor);
515                         goto loop;
516                 }
517                 ip->flags |= HAMMER_INODE_ONDISK;
518         } else {
519                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
520                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
521                         --hmp->rsv_inodes;
522                 }
523
524                 hammer_free_inode(ip);
525                 ip = NULL;
526         }
527         hammer_done_cursor(&cursor);
528         trans->flags |= HAMMER_TRANSF_NEWINODE;
529         return (ip);
530 }
531
532 /*
533  * Get a dummy inode to placemark a broken directory entry.
534  */
535 struct hammer_inode *
536 hammer_get_dummy_inode(hammer_transaction_t trans, hammer_inode_t dip,
537                  int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
538                  int flags, int *errorp)
539 {
540         hammer_mount_t hmp = trans->hmp;
541         struct hammer_inode_info iinfo;
542         struct hammer_inode *ip;
543
544         /*
545          * Determine if we already have an inode cached.  If we do then
546          * we are golden.
547          *
548          * If we find an inode with no vnode we have to mark the
549          * transaction such that hammer_inode_waitreclaims() is
550          * called later on to avoid building up an infinite number
551          * of inodes.  Otherwise we can continue to * add new inodes
552          * faster then they can be disposed of, even with the tsleep
553          * delay.
554          *
555          * If we find a non-fake inode we return an error.  Only fake
556          * inodes can be returned by this routine.
557          */
558         iinfo.obj_id = obj_id;
559         iinfo.obj_asof = asof;
560         iinfo.obj_localization = localization;
561 loop:
562         *errorp = 0;
563         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
564         if (ip) {
565                 if ((ip->flags & HAMMER_INODE_DUMMY) == 0) {
566                         *errorp = ENOENT;
567                         return(NULL);
568                 }
569                 hammer_ref(&ip->lock);
570                 return(ip);
571         }
572
573         /*
574          * Allocate a new inode structure and deal with races later.
575          */
576         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
577         ++hammer_count_inodes;
578         ++hmp->count_inodes;
579         ip->obj_id = obj_id;
580         ip->obj_asof = iinfo.obj_asof;
581         ip->obj_localization = localization;
582         ip->hmp = hmp;
583         ip->flags = flags | HAMMER_INODE_RO | HAMMER_INODE_DUMMY;
584         ip->cache[0].ip = ip;
585         ip->cache[1].ip = ip;
586         ip->cache[2].ip = ip;
587         ip->cache[3].ip = ip;
588         ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
589                 0x7FFFFFFFFFFFFFFFLL;
590         RB_INIT(&ip->rec_tree);
591         TAILQ_INIT(&ip->target_list);
592         hammer_ref(&ip->lock);
593
594         /*
595          * Populate the dummy inode.  Leave everything zero'd out.
596          *
597          * (ip->ino_leaf and ip->ino_data)
598          *
599          * Make the dummy inode a FIFO object which most copy programs
600          * will properly ignore.
601          */
602         ip->save_trunc_off = ip->ino_data.size;
603         ip->ino_data.obj_type = HAMMER_OBJTYPE_FIFO;
604
605         /*
606          * Locate and assign the pseudofs management structure to
607          * the inode.
608          */
609         if (dip && dip->obj_localization == ip->obj_localization) {
610                 ip->pfsm = dip->pfsm;
611                 hammer_ref(&ip->pfsm->lock);
612         } else {
613                 ip->pfsm = hammer_load_pseudofs(trans, ip->obj_localization,
614                                                 errorp);
615                 *errorp = 0;    /* ignore ENOENT */
616         }
617
618         /*
619          * The inode is placed on the red-black tree and will be synced to
620          * the media when flushed or by the filesystem sync.  If this races
621          * another instantiation/lookup the insertion will fail.
622          *
623          * NOTE: Do not set HAMMER_INODE_ONDISK.  The inode is a fake.
624          */
625         if (*errorp == 0) {
626                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
627                         hammer_free_inode(ip);
628                         goto loop;
629                 }
630         } else {
631                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
632                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
633                         --hmp->rsv_inodes;
634                 }
635                 hammer_free_inode(ip);
636                 ip = NULL;
637         }
638         trans->flags |= HAMMER_TRANSF_NEWINODE;
639         return (ip);
640 }
641
642 /*
643  * Return a referenced inode only if it is in our inode cache.
644  *
645  * Dummy inodes do not count.
646  */
647 struct hammer_inode *
648 hammer_find_inode(hammer_transaction_t trans, int64_t obj_id,
649                   hammer_tid_t asof, u_int32_t localization)
650 {
651         hammer_mount_t hmp = trans->hmp;
652         struct hammer_inode_info iinfo;
653         struct hammer_inode *ip;
654
655         iinfo.obj_id = obj_id;
656         iinfo.obj_asof = asof;
657         iinfo.obj_localization = localization;
658
659         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
660         if (ip) {
661                 if (ip->flags & HAMMER_INODE_DUMMY)
662                         ip = NULL;
663                 else
664                         hammer_ref(&ip->lock);
665         }
666         return(ip);
667 }
668
669 /*
670  * Create a new filesystem object, returning the inode in *ipp.  The
671  * returned inode will be referenced.  The inode is created in-memory.
672  *
673  * If pfsm is non-NULL the caller wishes to create the root inode for
674  * a master PFS.
675  */
676 int
677 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
678                     struct ucred *cred,
679                     hammer_inode_t dip, const char *name, int namelen,
680                     hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
681 {
682         hammer_mount_t hmp;
683         hammer_inode_t ip;
684         uid_t xuid;
685         int error;
686         int64_t namekey;
687         u_int32_t dummy;
688
689         hmp = trans->hmp;
690
691         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
692         ++hammer_count_inodes;
693         ++hmp->count_inodes;
694         trans->flags |= HAMMER_TRANSF_NEWINODE;
695
696         if (pfsm) {
697                 KKASSERT(pfsm->localization != 0);
698                 ip->obj_id = HAMMER_OBJID_ROOT;
699                 ip->obj_localization = pfsm->localization;
700         } else {
701                 KKASSERT(dip != NULL);
702                 namekey = hammer_directory_namekey(dip, name, namelen, &dummy);
703                 ip->obj_id = hammer_alloc_objid(hmp, dip, namekey);
704                 ip->obj_localization = dip->obj_localization;
705         }
706
707         KKASSERT(ip->obj_id != 0);
708         ip->obj_asof = hmp->asof;
709         ip->hmp = hmp;
710         ip->flush_state = HAMMER_FST_IDLE;
711         ip->flags = HAMMER_INODE_DDIRTY |
712                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
713         ip->cache[0].ip = ip;
714         ip->cache[1].ip = ip;
715         ip->cache[2].ip = ip;
716         ip->cache[3].ip = ip;
717
718         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
719         /* ip->save_trunc_off = 0; (already zero) */
720         RB_INIT(&ip->rec_tree);
721         TAILQ_INIT(&ip->target_list);
722
723         ip->ino_data.atime = trans->time;
724         ip->ino_data.mtime = trans->time;
725         ip->ino_data.size = 0;
726         ip->ino_data.nlinks = 0;
727
728         /*
729          * A nohistory designator on the parent directory is inherited by
730          * the child.  We will do this even for pseudo-fs creation... the
731          * sysad can turn it off.
732          */
733         if (dip) {
734                 ip->ino_data.uflags = dip->ino_data.uflags &
735                                       (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
736         }
737
738         ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
739         ip->ino_leaf.base.localization = ip->obj_localization +
740                                          HAMMER_LOCALIZE_INODE;
741         ip->ino_leaf.base.obj_id = ip->obj_id;
742         ip->ino_leaf.base.key = 0;
743         ip->ino_leaf.base.create_tid = 0;
744         ip->ino_leaf.base.delete_tid = 0;
745         ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
746         ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
747
748         ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
749         ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
750         ip->ino_data.mode = vap->va_mode;
751         ip->ino_data.ctime = trans->time;
752
753         /*
754          * If we are running version 2 or greater we use dirhash algorithm #1
755          * which is semi-sorted.  Algorithm #0 was just a pure crc.
756          */
757         if (trans->hmp->version >= HAMMER_VOL_VERSION_TWO) {
758                 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
759                         ip->ino_data.cap_flags |= HAMMER_INODE_CAP_DIRHASH_ALG1;
760                 }
761         }
762
763         /*
764          * Setup the ".." pointer.  This only needs to be done for directories
765          * but we do it for all objects as a recovery aid.
766          */
767         if (dip)
768                 ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
769 #if 0
770         /*
771          * The parent_obj_localization field only applies to pseudo-fs roots.
772          * XXX this is no longer applicable, PFSs are no longer directly
773          * tied into the parent's directory structure.
774          */
775         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY &&
776             ip->obj_id == HAMMER_OBJID_ROOT) {
777                 ip->ino_data.ext.obj.parent_obj_localization = 
778                                                 dip->obj_localization;
779         }
780 #endif
781
782         switch(ip->ino_leaf.base.obj_type) {
783         case HAMMER_OBJTYPE_CDEV:
784         case HAMMER_OBJTYPE_BDEV:
785                 ip->ino_data.rmajor = vap->va_rmajor;
786                 ip->ino_data.rminor = vap->va_rminor;
787                 break;
788         default:
789                 break;
790         }
791
792         /*
793          * Calculate default uid/gid and overwrite with information from
794          * the vap.
795          */
796         if (dip) {
797                 xuid = hammer_to_unix_xid(&dip->ino_data.uid);
798                 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
799                                              xuid, cred, &vap->va_mode);
800         } else {
801                 xuid = 0;
802         }
803         ip->ino_data.mode = vap->va_mode;
804
805         if (vap->va_vaflags & VA_UID_UUID_VALID)
806                 ip->ino_data.uid = vap->va_uid_uuid;
807         else if (vap->va_uid != (uid_t)VNOVAL)
808                 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
809         else
810                 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
811
812         if (vap->va_vaflags & VA_GID_UUID_VALID)
813                 ip->ino_data.gid = vap->va_gid_uuid;
814         else if (vap->va_gid != (gid_t)VNOVAL)
815                 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
816         else if (dip)
817                 ip->ino_data.gid = dip->ino_data.gid;
818
819         hammer_ref(&ip->lock);
820
821         if (pfsm) {
822                 ip->pfsm = pfsm;
823                 hammer_ref(&pfsm->lock);
824                 error = 0;
825         } else if (dip->obj_localization == ip->obj_localization) {
826                 ip->pfsm = dip->pfsm;
827                 hammer_ref(&ip->pfsm->lock);
828                 error = 0;
829         } else {
830                 ip->pfsm = hammer_load_pseudofs(trans,
831                                                 ip->obj_localization,
832                                                 &error);
833                 error = 0;      /* ignore ENOENT */
834         }
835
836         if (error) {
837                 hammer_free_inode(ip);
838                 ip = NULL;
839         } else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
840                 panic("hammer_create_inode: duplicate obj_id %llx", ip->obj_id);
841                 /* not reached */
842                 hammer_free_inode(ip);
843         }
844         *ipp = ip;
845         return(error);
846 }
847
848 /*
849  * Final cleanup / freeing of an inode structure
850  */
851 static void
852 hammer_free_inode(hammer_inode_t ip)
853 {
854         struct hammer_mount *hmp;
855
856         hmp = ip->hmp;
857         KKASSERT(ip->lock.refs == 1);
858         hammer_uncache_node(&ip->cache[0]);
859         hammer_uncache_node(&ip->cache[1]);
860         hammer_uncache_node(&ip->cache[2]);
861         hammer_uncache_node(&ip->cache[3]);
862         hammer_inode_wakereclaims(ip, 1);
863         if (ip->objid_cache)
864                 hammer_clear_objid(ip);
865         --hammer_count_inodes;
866         --hmp->count_inodes;
867         if (ip->pfsm) {
868                 hammer_rel_pseudofs(hmp, ip->pfsm);
869                 ip->pfsm = NULL;
870         }
871         kfree(ip, hmp->m_inodes);
872         ip = NULL;
873 }
874
875 /*
876  * Retrieve pseudo-fs data.  NULL will never be returned.
877  *
878  * If an error occurs *errorp will be set and a default template is returned,
879  * otherwise *errorp is set to 0.  Typically when an error occurs it will
880  * be ENOENT.
881  */
882 hammer_pseudofs_inmem_t
883 hammer_load_pseudofs(hammer_transaction_t trans,
884                      u_int32_t localization, int *errorp)
885 {
886         hammer_mount_t hmp = trans->hmp;
887         hammer_inode_t ip;
888         hammer_pseudofs_inmem_t pfsm;
889         struct hammer_cursor cursor;
890         int bytes;
891
892 retry:
893         pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
894         if (pfsm) {
895                 hammer_ref(&pfsm->lock);
896                 *errorp = 0;
897                 return(pfsm);
898         }
899
900         /*
901          * PFS records are stored in the root inode (not the PFS root inode,
902          * but the real root).  Avoid an infinite recursion if loading
903          * the PFS for the real root.
904          */
905         if (localization) {
906                 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
907                                       HAMMER_MAX_TID,
908                                       HAMMER_DEF_LOCALIZATION, 0, errorp);
909         } else {
910                 ip = NULL;
911         }
912
913         pfsm = kmalloc(sizeof(*pfsm), hmp->m_misc, M_WAITOK | M_ZERO);
914         pfsm->localization = localization;
915         pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
916         pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
917
918         hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
919         cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION +
920                                       HAMMER_LOCALIZE_MISC;
921         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
922         cursor.key_beg.create_tid = 0;
923         cursor.key_beg.delete_tid = 0;
924         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
925         cursor.key_beg.obj_type = 0;
926         cursor.key_beg.key = localization;
927         cursor.asof = HAMMER_MAX_TID;
928         cursor.flags |= HAMMER_CURSOR_ASOF;
929
930         if (ip)
931                 *errorp = hammer_ip_lookup(&cursor);
932         else
933                 *errorp = hammer_btree_lookup(&cursor);
934         if (*errorp == 0) {
935                 *errorp = hammer_ip_resolve_data(&cursor);
936                 if (*errorp == 0) {
937                         if (cursor.data->pfsd.mirror_flags &
938                             HAMMER_PFSD_DELETED) {
939                                 *errorp = ENOENT;
940                         } else {
941                                 bytes = cursor.leaf->data_len;
942                                 if (bytes > sizeof(pfsm->pfsd))
943                                         bytes = sizeof(pfsm->pfsd);
944                                 bcopy(cursor.data, &pfsm->pfsd, bytes);
945                         }
946                 }
947         }
948         hammer_done_cursor(&cursor);
949
950         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
951         hammer_ref(&pfsm->lock);
952         if (ip)
953                 hammer_rel_inode(ip, 0);
954         if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
955                 kfree(pfsm, hmp->m_misc);
956                 goto retry;
957         }
958         return(pfsm);
959 }
960
961 /*
962  * Store pseudo-fs data.  The backend will automatically delete any prior
963  * on-disk pseudo-fs data but we have to delete in-memory versions.
964  */
965 int
966 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
967 {
968         struct hammer_cursor cursor;
969         hammer_record_t record;
970         hammer_inode_t ip;
971         int error;
972
973         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
974                               HAMMER_DEF_LOCALIZATION, 0, &error);
975 retry:
976         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
977         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
978         cursor.key_beg.localization = ip->obj_localization +
979                                       HAMMER_LOCALIZE_MISC;
980         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
981         cursor.key_beg.create_tid = 0;
982         cursor.key_beg.delete_tid = 0;
983         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
984         cursor.key_beg.obj_type = 0;
985         cursor.key_beg.key = pfsm->localization;
986         cursor.asof = HAMMER_MAX_TID;
987         cursor.flags |= HAMMER_CURSOR_ASOF;
988
989         /*
990          * Replace any in-memory version of the record.
991          */
992         error = hammer_ip_lookup(&cursor);
993         if (error == 0 && hammer_cursor_inmem(&cursor)) {
994                 record = cursor.iprec;
995                 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
996                         KKASSERT(cursor.deadlk_rec == NULL);
997                         hammer_ref(&record->lock);
998                         cursor.deadlk_rec = record;
999                         error = EDEADLK;
1000                 } else {
1001                         record->flags |= HAMMER_RECF_DELETED_FE;
1002                         error = 0;
1003                 }
1004         }
1005
1006         /*
1007          * Allocate replacement general record.  The backend flush will
1008          * delete any on-disk version of the record.
1009          */
1010         if (error == 0 || error == ENOENT) {
1011                 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
1012                 record->type = HAMMER_MEM_RECORD_GENERAL;
1013
1014                 record->leaf.base.localization = ip->obj_localization +
1015                                                  HAMMER_LOCALIZE_MISC;
1016                 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
1017                 record->leaf.base.key = pfsm->localization;
1018                 record->leaf.data_len = sizeof(pfsm->pfsd);
1019                 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
1020                 error = hammer_ip_add_record(trans, record);
1021         }
1022         hammer_done_cursor(&cursor);
1023         if (error == EDEADLK)
1024                 goto retry;
1025         hammer_rel_inode(ip, 0);
1026         return(error);
1027 }
1028
1029 /*
1030  * Create a root directory for a PFS if one does not alredy exist.
1031  *
1032  * The PFS root stands alone so we must also bump the nlinks count
1033  * to prevent it from being destroyed on release.
1034  */
1035 int
1036 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1037                        hammer_pseudofs_inmem_t pfsm)
1038 {
1039         hammer_inode_t ip;
1040         struct vattr vap;
1041         int error;
1042
1043         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1044                               pfsm->localization, 0, &error);
1045         if (ip == NULL) {
1046                 vattr_null(&vap);
1047                 vap.va_mode = 0755;
1048                 vap.va_type = VDIR;
1049                 error = hammer_create_inode(trans, &vap, cred,
1050                                             NULL, NULL, 0,
1051                                             pfsm, &ip);
1052                 if (error == 0) {
1053                         ++ip->ino_data.nlinks;
1054                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
1055                 }
1056         }
1057         if (ip)
1058                 hammer_rel_inode(ip, 0);
1059         return(error);
1060 }
1061
1062 /*
1063  * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
1064  * if we are unable to disassociate all the inodes.
1065  */
1066 static
1067 int
1068 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
1069 {
1070         int res;
1071
1072         hammer_ref(&ip->lock);
1073         if (ip->lock.refs == 2 && ip->vp)
1074                 vclean_unlocked(ip->vp);
1075         if (ip->lock.refs == 1 && ip->vp == NULL)
1076                 res = 0;
1077         else
1078                 res = -1;       /* stop, someone is using the inode */
1079         hammer_rel_inode(ip, 0);
1080         return(res);
1081 }
1082
1083 int
1084 hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
1085 {
1086         int res;
1087         int try;
1088
1089         for (try = res = 0; try < 4; ++try) {
1090                 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
1091                                            hammer_inode_pfs_cmp,
1092                                            hammer_unload_pseudofs_callback,
1093                                            &localization);
1094                 if (res == 0 && try > 1)
1095                         break;
1096                 hammer_flusher_sync(trans->hmp);
1097         }
1098         if (res != 0)
1099                 res = ENOTEMPTY;
1100         return(res);
1101 }
1102
1103
1104 /*
1105  * Release a reference on a PFS
1106  */
1107 void
1108 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
1109 {
1110         hammer_unref(&pfsm->lock);
1111         if (pfsm->lock.refs == 0) {
1112                 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
1113                 kfree(pfsm, hmp->m_misc);
1114         }
1115 }
1116
1117 /*
1118  * Called by hammer_sync_inode().
1119  */
1120 static int
1121 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
1122 {
1123         hammer_transaction_t trans = cursor->trans;
1124         hammer_record_t record;
1125         int error;
1126         int redirty;
1127
1128 retry:
1129         error = 0;
1130
1131         /*
1132          * If the inode has a presence on-disk then locate it and mark
1133          * it deleted, setting DELONDISK.
1134          *
1135          * The record may or may not be physically deleted, depending on
1136          * the retention policy.
1137          */
1138         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
1139             HAMMER_INODE_ONDISK) {
1140                 hammer_normalize_cursor(cursor);
1141                 cursor->key_beg.localization = ip->obj_localization + 
1142                                                HAMMER_LOCALIZE_INODE;
1143                 cursor->key_beg.obj_id = ip->obj_id;
1144                 cursor->key_beg.key = 0;
1145                 cursor->key_beg.create_tid = 0;
1146                 cursor->key_beg.delete_tid = 0;
1147                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1148                 cursor->key_beg.obj_type = 0;
1149                 cursor->asof = ip->obj_asof;
1150                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1151                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
1152                 cursor->flags |= HAMMER_CURSOR_BACKEND;
1153
1154                 error = hammer_btree_lookup(cursor);
1155                 if (hammer_debug_inode)
1156                         kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
1157
1158                 if (error == 0) {
1159                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1160                         if (hammer_debug_inode)
1161                                 kprintf(" error %d\n", error);
1162                         if (error == 0) {
1163                                 ip->flags |= HAMMER_INODE_DELONDISK;
1164                         }
1165                         if (cursor->node)
1166                                 hammer_cache_node(&ip->cache[0], cursor->node);
1167                 }
1168                 if (error == EDEADLK) {
1169                         hammer_done_cursor(cursor);
1170                         error = hammer_init_cursor(trans, cursor,
1171                                                    &ip->cache[0], ip);
1172                         if (hammer_debug_inode)
1173                                 kprintf("IPDED %p %d\n", ip, error);
1174                         if (error == 0)
1175                                 goto retry;
1176                 }
1177         }
1178
1179         /*
1180          * Ok, write out the initial record or a new record (after deleting
1181          * the old one), unless the DELETED flag is set.  This routine will
1182          * clear DELONDISK if it writes out a record.
1183          *
1184          * Update our inode statistics if this is the first application of
1185          * the inode on-disk.
1186          */
1187         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
1188                 /*
1189                  * Generate a record and write it to the media.  We clean-up
1190                  * the state before releasing so we do not have to set-up
1191                  * a flush_group.
1192                  */
1193                 record = hammer_alloc_mem_record(ip, 0);
1194                 record->type = HAMMER_MEM_RECORD_INODE;
1195                 record->flush_state = HAMMER_FST_FLUSH;
1196                 record->leaf = ip->sync_ino_leaf;
1197                 record->leaf.base.create_tid = trans->tid;
1198                 record->leaf.data_len = sizeof(ip->sync_ino_data);
1199                 record->leaf.create_ts = trans->time32;
1200                 record->data = (void *)&ip->sync_ino_data;
1201                 record->flags |= HAMMER_RECF_INTERLOCK_BE;
1202
1203                 /*
1204                  * If this flag is set we cannot sync the new file size
1205                  * because we haven't finished related truncations.  The
1206                  * inode will be flushed in another flush group to finish
1207                  * the job.
1208                  */
1209                 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1210                     ip->sync_ino_data.size != ip->ino_data.size) {
1211                         redirty = 1;
1212                         ip->sync_ino_data.size = ip->ino_data.size;
1213                 } else {
1214                         redirty = 0;
1215                 }
1216
1217                 for (;;) {
1218                         error = hammer_ip_sync_record_cursor(cursor, record);
1219                         if (hammer_debug_inode)
1220                                 kprintf("GENREC %p rec %08x %d\n",      
1221                                         ip, record->flags, error);
1222                         if (error != EDEADLK)
1223                                 break;
1224                         hammer_done_cursor(cursor);
1225                         error = hammer_init_cursor(trans, cursor,
1226                                                    &ip->cache[0], ip);
1227                         if (hammer_debug_inode)
1228                                 kprintf("GENREC reinit %d\n", error);
1229                         if (error)
1230                                 break;
1231                 }
1232
1233                 /*
1234                  * Note:  The record was never on the inode's record tree
1235                  * so just wave our hands importantly and destroy it.
1236                  */
1237                 record->flags |= HAMMER_RECF_COMMITTED;
1238                 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1239                 record->flush_state = HAMMER_FST_IDLE;
1240                 ++ip->rec_generation;
1241                 hammer_rel_mem_record(record);
1242
1243                 /*
1244                  * Finish up.
1245                  */
1246                 if (error == 0) {
1247                         if (hammer_debug_inode)
1248                                 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1249                         ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1250                                             HAMMER_INODE_ATIME |
1251                                             HAMMER_INODE_MTIME);
1252                         ip->flags &= ~HAMMER_INODE_DELONDISK;
1253                         if (redirty)
1254                                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1255
1256                         /*
1257                          * Root volume count of inodes
1258                          */
1259                         hammer_sync_lock_sh(trans);
1260                         if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1261                                 hammer_modify_volume_field(trans,
1262                                                            trans->rootvol,
1263                                                            vol0_stat_inodes);
1264                                 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1265                                 hammer_modify_volume_done(trans->rootvol);
1266                                 ip->flags |= HAMMER_INODE_ONDISK;
1267                                 if (hammer_debug_inode)
1268                                         kprintf("NOWONDISK %p\n", ip);
1269                         }
1270                         hammer_sync_unlock(trans);
1271                 }
1272         }
1273
1274         /*
1275          * If the inode has been destroyed, clean out any left-over flags
1276          * that may have been set by the frontend.
1277          */
1278         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) { 
1279                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1280                                     HAMMER_INODE_ATIME |
1281                                     HAMMER_INODE_MTIME);
1282         }
1283         return(error);
1284 }
1285
1286 /*
1287  * Update only the itimes fields.
1288  *
1289  * ATIME can be updated without generating any UNDO.  MTIME is updated
1290  * with UNDO so it is guaranteed to be synchronized properly in case of
1291  * a crash.
1292  *
1293  * Neither field is included in the B-Tree leaf element's CRC, which is how
1294  * we can get away with updating ATIME the way we do.
1295  */
1296 static int
1297 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1298 {
1299         hammer_transaction_t trans = cursor->trans;
1300         int error;
1301
1302 retry:
1303         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1304             HAMMER_INODE_ONDISK) {
1305                 return(0);
1306         }
1307
1308         hammer_normalize_cursor(cursor);
1309         cursor->key_beg.localization = ip->obj_localization + 
1310                                        HAMMER_LOCALIZE_INODE;
1311         cursor->key_beg.obj_id = ip->obj_id;
1312         cursor->key_beg.key = 0;
1313         cursor->key_beg.create_tid = 0;
1314         cursor->key_beg.delete_tid = 0;
1315         cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1316         cursor->key_beg.obj_type = 0;
1317         cursor->asof = ip->obj_asof;
1318         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1319         cursor->flags |= HAMMER_CURSOR_ASOF;
1320         cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1321         cursor->flags |= HAMMER_CURSOR_GET_DATA;
1322         cursor->flags |= HAMMER_CURSOR_BACKEND;
1323
1324         error = hammer_btree_lookup(cursor);
1325         if (error == 0) {
1326                 hammer_cache_node(&ip->cache[0], cursor->node);
1327                 if (ip->sync_flags & HAMMER_INODE_MTIME) {
1328                         /*
1329                          * Updating MTIME requires an UNDO.  Just cover
1330                          * both atime and mtime.
1331                          */
1332                         hammer_sync_lock_sh(trans);
1333                         hammer_modify_buffer(trans, cursor->data_buffer,
1334                                      HAMMER_ITIMES_BASE(&cursor->data->inode),
1335                                      HAMMER_ITIMES_BYTES);
1336                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1337                         cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1338                         hammer_modify_buffer_done(cursor->data_buffer);
1339                         hammer_sync_unlock(trans);
1340                 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1341                         /*
1342                          * Updating atime only can be done in-place with
1343                          * no UNDO.
1344                          */
1345                         hammer_sync_lock_sh(trans);
1346                         hammer_modify_buffer(trans, cursor->data_buffer,
1347                                              NULL, 0);
1348                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1349                         hammer_modify_buffer_done(cursor->data_buffer);
1350                         hammer_sync_unlock(trans);
1351                 }
1352                 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1353         }
1354         if (error == EDEADLK) {
1355                 hammer_done_cursor(cursor);
1356                 error = hammer_init_cursor(trans, cursor,
1357                                            &ip->cache[0], ip);
1358                 if (error == 0)
1359                         goto retry;
1360         }
1361         return(error);
1362 }
1363
1364 /*
1365  * Release a reference on an inode, flush as requested.
1366  *
1367  * On the last reference we queue the inode to the flusher for its final
1368  * disposition.
1369  */
1370 void
1371 hammer_rel_inode(struct hammer_inode *ip, int flush)
1372 {
1373         /*hammer_mount_t hmp = ip->hmp;*/
1374
1375         /*
1376          * Handle disposition when dropping the last ref.
1377          */
1378         for (;;) {
1379                 if (ip->lock.refs == 1) {
1380                         /*
1381                          * Determine whether on-disk action is needed for
1382                          * the inode's final disposition.
1383                          */
1384                         KKASSERT(ip->vp == NULL);
1385                         hammer_inode_unloadable_check(ip, 0);
1386                         if (ip->flags & HAMMER_INODE_MODMASK) {
1387                                 hammer_flush_inode(ip, 0);
1388                         } else if (ip->lock.refs == 1) {
1389                                 hammer_unload_inode(ip);
1390                                 break;
1391                         }
1392                 } else {
1393                         if (flush)
1394                                 hammer_flush_inode(ip, 0);
1395
1396                         /*
1397                          * The inode still has multiple refs, try to drop
1398                          * one ref.
1399                          */
1400                         KKASSERT(ip->lock.refs >= 1);
1401                         if (ip->lock.refs > 1) {
1402                                 hammer_unref(&ip->lock);
1403                                 break;
1404                         }
1405                 }
1406         }
1407 }
1408
1409 /*
1410  * Unload and destroy the specified inode.  Must be called with one remaining
1411  * reference.  The reference is disposed of.
1412  *
1413  * The inode must be completely clean.
1414  */
1415 static int
1416 hammer_unload_inode(struct hammer_inode *ip)
1417 {
1418         hammer_mount_t hmp = ip->hmp;
1419
1420         KASSERT(ip->lock.refs == 1,
1421                 ("hammer_unload_inode: %d refs\n", ip->lock.refs));
1422         KKASSERT(ip->vp == NULL);
1423         KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1424         KKASSERT(ip->cursor_ip_refs == 0);
1425         KKASSERT(ip->lock.lockcount == 0);
1426         KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1427
1428         KKASSERT(RB_EMPTY(&ip->rec_tree));
1429         KKASSERT(TAILQ_EMPTY(&ip->target_list));
1430
1431         RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1432
1433         hammer_free_inode(ip);
1434         return(0);
1435 }
1436
1437 /*
1438  * Called during unmounting if a critical error occured.  The in-memory
1439  * inode and all related structures are destroyed.
1440  *
1441  * If a critical error did not occur the unmount code calls the standard
1442  * release and asserts that the inode is gone.
1443  */
1444 int
1445 hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
1446 {
1447         hammer_record_t rec;
1448
1449         /*
1450          * Get rid of the inodes in-memory records, regardless of their
1451          * state, and clear the mod-mask.
1452          */
1453         while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1454                 TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1455                 rec->target_ip = NULL;
1456                 if (rec->flush_state == HAMMER_FST_SETUP)
1457                         rec->flush_state = HAMMER_FST_IDLE;
1458         }
1459         while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1460                 if (rec->flush_state == HAMMER_FST_FLUSH)
1461                         --rec->flush_group->refs;
1462                 else
1463                         hammer_ref(&rec->lock);
1464                 KKASSERT(rec->lock.refs == 1);
1465                 rec->flush_state = HAMMER_FST_IDLE;
1466                 rec->flush_group = NULL;
1467                 rec->flags |= HAMMER_RECF_DELETED_FE; /* wave hands */
1468                 rec->flags |= HAMMER_RECF_DELETED_BE; /* wave hands */
1469                 ++ip->rec_generation;
1470                 hammer_rel_mem_record(rec);
1471         }
1472         ip->flags &= ~HAMMER_INODE_MODMASK;
1473         ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1474         KKASSERT(ip->vp == NULL);
1475
1476         /*
1477          * Remove the inode from any flush group, force it idle.  FLUSH
1478          * and SETUP states have an inode ref.
1479          */
1480         switch(ip->flush_state) {
1481         case HAMMER_FST_FLUSH:
1482                 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
1483                 --ip->flush_group->refs;
1484                 ip->flush_group = NULL;
1485                 /* fall through */
1486         case HAMMER_FST_SETUP:
1487                 hammer_unref(&ip->lock);
1488                 ip->flush_state = HAMMER_FST_IDLE;
1489                 /* fall through */
1490         case HAMMER_FST_IDLE:
1491                 break;
1492         }
1493
1494         /*
1495          * There shouldn't be any associated vnode.  The unload needs at
1496          * least one ref, if we do have a vp steal its ip ref.
1497          */
1498         if (ip->vp) {
1499                 kprintf("hammer_destroy_inode_callback: Unexpected "
1500                         "vnode association ip %p vp %p\n", ip, ip->vp);
1501                 ip->vp->v_data = NULL;
1502                 ip->vp = NULL;
1503         } else {
1504                 hammer_ref(&ip->lock);
1505         }
1506         hammer_unload_inode(ip);
1507         return(0);
1508 }
1509
1510 /*
1511  * Called on mount -u when switching from RW to RO or vise-versa.  Adjust
1512  * the read-only flag for cached inodes.
1513  *
1514  * This routine is called from a RB_SCAN().
1515  */
1516 int
1517 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1518 {
1519         hammer_mount_t hmp = ip->hmp;
1520
1521         if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1522                 ip->flags |= HAMMER_INODE_RO;
1523         else
1524                 ip->flags &= ~HAMMER_INODE_RO;
1525         return(0);
1526 }
1527
1528 /*
1529  * A transaction has modified an inode, requiring updates as specified by
1530  * the passed flags.
1531  *
1532  * HAMMER_INODE_DDIRTY: Inode data has been updated
1533  * HAMMER_INODE_XDIRTY: Dirty in-memory records
1534  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
1535  * HAMMER_INODE_DELETED: Inode record/data must be deleted
1536  * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1537  */
1538 void
1539 hammer_modify_inode(hammer_inode_t ip, int flags)
1540 {
1541         /* 
1542          * ronly of 0 or 2 does not trigger assertion.
1543          * 2 is a special error state 
1544          */
1545         KKASSERT(ip->hmp->ronly != 1 ||
1546                   (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY | 
1547                             HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1548                             HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1549         if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1550                 ip->flags |= HAMMER_INODE_RSV_INODES;
1551                 ++ip->hmp->rsv_inodes;
1552         }
1553
1554         ip->flags |= flags;
1555 }
1556
1557 /*
1558  * Request that an inode be flushed.  This whole mess cannot block and may
1559  * recurse (if not synchronous).  Once requested HAMMER will attempt to
1560  * actively flush the inode until the flush can be done.
1561  *
1562  * The inode may already be flushing, or may be in a setup state.  We can
1563  * place the inode in a flushing state if it is currently idle and flag it
1564  * to reflush if it is currently flushing.
1565  *
1566  * Upon return if the inode could not be flushed due to a setup
1567  * dependancy, then it will be automatically flushed when the dependancy
1568  * is satisfied.
1569  */
1570 void
1571 hammer_flush_inode(hammer_inode_t ip, int flags)
1572 {
1573         hammer_mount_t hmp;
1574         hammer_flush_group_t flg;
1575         int good;
1576
1577         /*
1578          * next_flush_group is the first flush group we can place the inode
1579          * in.  It may be NULL.  If it becomes full we append a new flush
1580          * group and make that the next_flush_group.
1581          */
1582         hmp = ip->hmp;
1583         while ((flg = hmp->next_flush_group) != NULL) {
1584                 KKASSERT(flg->running == 0);
1585                 if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit)
1586                         break;
1587                 hmp->next_flush_group = TAILQ_NEXT(flg, flush_entry);
1588                 hammer_flusher_async(ip->hmp, flg);
1589         }
1590         if (flg == NULL) {
1591                 flg = kmalloc(sizeof(*flg), hmp->m_misc, M_WAITOK|M_ZERO);
1592                 hmp->next_flush_group = flg;
1593                 TAILQ_INIT(&flg->flush_list);
1594                 TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1595         }
1596
1597         /*
1598          * Trivial 'nothing to flush' case.  If the inode is in a SETUP
1599          * state we have to put it back into an IDLE state so we can
1600          * drop the extra ref.
1601          *
1602          * If we have a parent dependancy we must still fall through
1603          * so we can run it.
1604          */
1605         if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1606                 if (ip->flush_state == HAMMER_FST_SETUP &&
1607                     TAILQ_EMPTY(&ip->target_list)) {
1608                         ip->flush_state = HAMMER_FST_IDLE;
1609                         hammer_rel_inode(ip, 0);
1610                 }
1611                 if (ip->flush_state == HAMMER_FST_IDLE)
1612                         return;
1613         }
1614
1615         /*
1616          * Our flush action will depend on the current state.
1617          */
1618         switch(ip->flush_state) {
1619         case HAMMER_FST_IDLE:
1620                 /*
1621                  * We have no dependancies and can flush immediately.  Some
1622                  * our children may not be flushable so we have to re-test
1623                  * with that additional knowledge.
1624                  */
1625                 hammer_flush_inode_core(ip, flg, flags);
1626                 break;
1627         case HAMMER_FST_SETUP:
1628                 /*
1629                  * Recurse upwards through dependancies via target_list
1630                  * and start their flusher actions going if possible.
1631                  *
1632                  * 'good' is our connectivity.  -1 means we have none and
1633                  * can't flush, 0 means there weren't any dependancies, and
1634                  * 1 means we have good connectivity.
1635                  */
1636                 good = hammer_setup_parent_inodes(ip, 0, flg);
1637
1638                 if (good >= 0) {
1639                         /*
1640                          * We can continue if good >= 0.  Determine how 
1641                          * many records under our inode can be flushed (and
1642                          * mark them).
1643                          */
1644                         hammer_flush_inode_core(ip, flg, flags);
1645                 } else {
1646                         /*
1647                          * Parent has no connectivity, tell it to flush
1648                          * us as soon as it does.
1649                          *
1650                          * The REFLUSH flag is also needed to trigger
1651                          * dependancy wakeups.
1652                          */
1653                         ip->flags |= HAMMER_INODE_CONN_DOWN |
1654                                      HAMMER_INODE_REFLUSH;
1655                         if (flags & HAMMER_FLUSH_SIGNAL) {
1656                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1657                                 hammer_flusher_async(ip->hmp, flg);
1658                         }
1659                 }
1660                 break;
1661         case HAMMER_FST_FLUSH:
1662                 /*
1663                  * We are already flushing, flag the inode to reflush
1664                  * if needed after it completes its current flush.
1665                  *
1666                  * The REFLUSH flag is also needed to trigger
1667                  * dependancy wakeups.
1668                  */
1669                 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1670                         ip->flags |= HAMMER_INODE_REFLUSH;
1671                 if (flags & HAMMER_FLUSH_SIGNAL) {
1672                         ip->flags |= HAMMER_INODE_RESIGNAL;
1673                         hammer_flusher_async(ip->hmp, flg);
1674                 }
1675                 break;
1676         }
1677 }
1678
1679 /*
1680  * Scan ip->target_list, which is a list of records owned by PARENTS to our
1681  * ip which reference our ip.
1682  *
1683  * XXX This is a huge mess of recursive code, but not one bit of it blocks
1684  *     so for now do not ref/deref the structures.  Note that if we use the
1685  *     ref/rel code later, the rel CAN block.
1686  */
1687 static int
1688 hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
1689                            hammer_flush_group_t flg)
1690 {
1691         hammer_record_t depend;
1692         int good;
1693         int r;
1694
1695         /*
1696          * If we hit our recursion limit and we have parent dependencies
1697          * We cannot continue.  Returning < 0 will cause us to be flagged
1698          * for reflush.  Returning -2 cuts off additional dependency checks
1699          * because they are likely to also hit the depth limit.
1700          *
1701          * We cannot return < 0 if there are no dependencies or there might
1702          * not be anything to wakeup (ip).
1703          */
1704         if (depth == 20 && TAILQ_FIRST(&ip->target_list)) {
1705                 kprintf("HAMMER Warning: depth limit reached on "
1706                         "setup recursion, inode %p %016llx\n",
1707                         ip, (long long)ip->obj_id);
1708                 return(-2);
1709         }
1710
1711         /*
1712          * Scan dependencies
1713          */
1714         good = 0;
1715         TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1716                 r = hammer_setup_parent_inodes_helper(depend, depth, flg);
1717                 KKASSERT(depend->target_ip == ip);
1718                 if (r < 0 && good == 0)
1719                         good = -1;
1720                 if (r > 0)
1721                         good = 1;
1722
1723                 /*
1724                  * If we failed due to the recursion depth limit then stop
1725                  * now.
1726                  */
1727                 if (r == -2)
1728                         break;
1729         }
1730         return(good);
1731 }
1732
1733 /*
1734  * This helper function takes a record representing the dependancy between
1735  * the parent inode and child inode.
1736  *
1737  * record->ip           = parent inode
1738  * record->target_ip    = child inode
1739  * 
1740  * We are asked to recurse upwards and convert the record from SETUP
1741  * to FLUSH if possible.
1742  *
1743  * Return 1 if the record gives us connectivity
1744  *
1745  * Return 0 if the record is not relevant 
1746  *
1747  * Return -1 if we can't resolve the dependancy and there is no connectivity.
1748  */
1749 static int
1750 hammer_setup_parent_inodes_helper(hammer_record_t record, int depth,
1751                                   hammer_flush_group_t flg)
1752 {
1753         hammer_mount_t hmp;
1754         hammer_inode_t pip;
1755         int good;
1756
1757         KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1758         pip = record->ip;
1759         hmp = pip->hmp;
1760
1761         /*
1762          * If the record is already flushing, is it in our flush group?
1763          *
1764          * If it is in our flush group but it is a general record or a 
1765          * delete-on-disk, it does not improve our connectivity (return 0),
1766          * and if the target inode is not trying to destroy itself we can't
1767          * allow the operation yet anyway (the second return -1).
1768          */
1769         if (record->flush_state == HAMMER_FST_FLUSH) {
1770                 /*
1771                  * If not in our flush group ask the parent to reflush
1772                  * us as soon as possible.
1773                  */
1774                 if (record->flush_group != flg) {
1775                         pip->flags |= HAMMER_INODE_REFLUSH;
1776                         record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1777                         return(-1);
1778                 }
1779
1780                 /*
1781                  * If in our flush group everything is already set up,
1782                  * just return whether the record will improve our
1783                  * visibility or not.
1784                  */
1785                 if (record->type == HAMMER_MEM_RECORD_ADD)
1786                         return(1);
1787                 return(0);
1788         }
1789
1790         /*
1791          * It must be a setup record.  Try to resolve the setup dependancies
1792          * by recursing upwards so we can place ip on the flush list.
1793          *
1794          * Limit ourselves to 20 levels of recursion to avoid blowing out
1795          * the kernel stack.  If we hit the recursion limit we can't flush
1796          * until the parent flushes.  The parent will flush independantly
1797          * on its own and ultimately a deep recursion will be resolved.
1798          */
1799         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1800
1801         good = hammer_setup_parent_inodes(pip, depth + 1, flg);
1802
1803         /*
1804          * If good < 0 the parent has no connectivity and we cannot safely
1805          * flush the directory entry, which also means we can't flush our
1806          * ip.  Flag us for downward recursion once the parent's
1807          * connectivity is resolved.  Flag the parent for [re]flush or it
1808          * may not check for downward recursions.
1809          */
1810         if (good < 0) {
1811                 pip->flags |= HAMMER_INODE_REFLUSH;
1812                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1813                 return(good);
1814         }
1815
1816         /*
1817          * We are go, place the parent inode in a flushing state so we can
1818          * place its record in a flushing state.  Note that the parent
1819          * may already be flushing.  The record must be in the same flush
1820          * group as the parent.
1821          */
1822         if (pip->flush_state != HAMMER_FST_FLUSH)
1823                 hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
1824         KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
1825         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1826
1827 #if 0
1828         if (record->type == HAMMER_MEM_RECORD_DEL &&
1829             (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
1830                 /*
1831                  * Regardless of flushing state we cannot sync this path if the
1832                  * record represents a delete-on-disk but the target inode
1833                  * is not ready to sync its own deletion.
1834                  *
1835                  * XXX need to count effective nlinks to determine whether
1836                  * the flush is ok, otherwise removing a hardlink will
1837                  * just leave the DEL record to rot.
1838                  */
1839                 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
1840                 return(-1);
1841         } else
1842 #endif
1843         if (pip->flush_group == flg) {
1844                 /*
1845                  * Because we have not calculated nlinks yet we can just
1846                  * set records to the flush state if the parent is in
1847                  * the same flush group as we are.
1848                  */
1849                 record->flush_state = HAMMER_FST_FLUSH;
1850                 record->flush_group = flg;
1851                 ++record->flush_group->refs;
1852                 hammer_ref(&record->lock);
1853
1854                 /*
1855                  * A general directory-add contributes to our visibility.
1856                  *
1857                  * Otherwise it is probably a directory-delete or 
1858                  * delete-on-disk record and does not contribute to our
1859                  * visbility (but we can still flush it).
1860                  */
1861                 if (record->type == HAMMER_MEM_RECORD_ADD)
1862                         return(1);
1863                 return(0);
1864         } else {
1865                 /*
1866                  * If the parent is not in our flush group we cannot
1867                  * flush this record yet, there is no visibility.
1868                  * We tell the parent to reflush and mark ourselves
1869                  * so the parent knows it should flush us too.
1870                  */
1871                 pip->flags |= HAMMER_INODE_REFLUSH;
1872                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1873                 return(-1);
1874         }
1875 }
1876
1877 /*
1878  * This is the core routine placing an inode into the FST_FLUSH state.
1879  */
1880 static void
1881 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
1882 {
1883         int go_count;
1884
1885         /*
1886          * Set flush state and prevent the flusher from cycling into
1887          * the next flush group.  Do not place the ip on the list yet.
1888          * Inodes not in the idle state get an extra reference.
1889          */
1890         KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
1891         if (ip->flush_state == HAMMER_FST_IDLE)
1892                 hammer_ref(&ip->lock);
1893         ip->flush_state = HAMMER_FST_FLUSH;
1894         ip->flush_group = flg;
1895         ++ip->hmp->flusher.group_lock;
1896         ++ip->hmp->count_iqueued;
1897         ++hammer_count_iqueued;
1898         ++flg->total_count;
1899
1900         /*
1901          * If the flush group reaches the autoflush limit we want to signal
1902          * the flusher.  This is particularly important for remove()s.
1903          */
1904         if (flg->total_count == hammer_autoflush)
1905                 flags |= HAMMER_FLUSH_SIGNAL;
1906
1907         /*
1908          * We need to be able to vfsync/truncate from the backend.
1909          */
1910         KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
1911         if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
1912                 ip->flags |= HAMMER_INODE_VHELD;
1913                 vref(ip->vp);
1914         }
1915
1916         /*
1917          * Figure out how many in-memory records we can actually flush
1918          * (not including inode meta-data, buffers, etc).
1919          */
1920         KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
1921         if (flags & HAMMER_FLUSH_RECURSION) {
1922                 /*
1923                  * If this is a upwards recursion we do not want to
1924                  * recurse down again!
1925                  */
1926                 go_count = 1;
1927 #if 0
1928         } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
1929                 /*
1930                  * No new records are added if we must complete a flush
1931                  * from a previous cycle, but we do have to move the records
1932                  * from the previous cycle to the current one.
1933                  */
1934 #if 0
1935                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1936                                    hammer_syncgrp_child_callback, NULL);
1937 #endif
1938                 go_count = 1;
1939 #endif
1940         } else {
1941                 /*
1942                  * Normal flush, scan records and bring them into the flush.
1943                  * Directory adds and deletes are usually skipped (they are
1944                  * grouped with the related inode rather then with the
1945                  * directory).
1946                  *
1947                  * go_count can be negative, which means the scan aborted
1948                  * due to the flush group being over-full and we should
1949                  * flush what we have.
1950                  */
1951                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1952                                    hammer_setup_child_callback, NULL);
1953         }
1954
1955         /*
1956          * This is a more involved test that includes go_count.  If we
1957          * can't flush, flag the inode and return.  If go_count is 0 we
1958          * were are unable to flush any records in our rec_tree and
1959          * must ignore the XDIRTY flag.
1960          */
1961         if (go_count == 0) {
1962                 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
1963                         --ip->hmp->count_iqueued;
1964                         --hammer_count_iqueued;
1965
1966                         --flg->total_count;
1967                         ip->flush_state = HAMMER_FST_SETUP;
1968                         ip->flush_group = NULL;
1969                         if (ip->flags & HAMMER_INODE_VHELD) {
1970                                 ip->flags &= ~HAMMER_INODE_VHELD;
1971                                 vrele(ip->vp);
1972                         }
1973
1974                         /*
1975                          * REFLUSH is needed to trigger dependancy wakeups
1976                          * when an inode is in SETUP.
1977                          */
1978                         ip->flags |= HAMMER_INODE_REFLUSH;
1979                         if (flags & HAMMER_FLUSH_SIGNAL) {
1980                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1981                                 hammer_flusher_async(ip->hmp, flg);
1982                         }
1983                         if (--ip->hmp->flusher.group_lock == 0)
1984                                 wakeup(&ip->hmp->flusher.group_lock);
1985                         return;
1986                 }
1987         }
1988
1989         /*
1990          * Snapshot the state of the inode for the backend flusher.
1991          *
1992          * We continue to retain save_trunc_off even when all truncations
1993          * have been resolved as an optimization to determine if we can
1994          * skip the B-Tree lookup for overwrite deletions.
1995          *
1996          * NOTE: The DELETING flag is a mod flag, but it is also sticky,
1997          * and stays in ip->flags.  Once set, it stays set until the
1998          * inode is destroyed.
1999          */
2000         if (ip->flags & HAMMER_INODE_TRUNCATED) {
2001                 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
2002                 ip->sync_trunc_off = ip->trunc_off;
2003                 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
2004                 ip->flags &= ~HAMMER_INODE_TRUNCATED;
2005                 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
2006
2007                 /*
2008                  * The save_trunc_off used to cache whether the B-Tree
2009                  * holds any records past that point is not used until
2010                  * after the truncation has succeeded, so we can safely
2011                  * set it now.
2012                  */
2013                 if (ip->save_trunc_off > ip->sync_trunc_off)
2014                         ip->save_trunc_off = ip->sync_trunc_off;
2015         }
2016         ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
2017                            ~HAMMER_INODE_TRUNCATED);
2018         ip->sync_ino_leaf = ip->ino_leaf;
2019         ip->sync_ino_data = ip->ino_data;
2020         ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
2021 #ifdef DEBUG_TRUNCATE
2022         if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
2023                 kprintf("truncateS %016llx\n", ip->sync_trunc_off);
2024 #endif
2025
2026         /*
2027          * The flusher list inherits our inode and reference.
2028          */
2029         KKASSERT(flg->running == 0);
2030         TAILQ_INSERT_TAIL(&flg->flush_list, ip, flush_entry);
2031         if (--ip->hmp->flusher.group_lock == 0)
2032                 wakeup(&ip->hmp->flusher.group_lock);
2033
2034         if (flags & HAMMER_FLUSH_SIGNAL) {
2035                 hammer_flusher_async(ip->hmp, flg);
2036         }
2037 }
2038
2039 /*
2040  * Callback for scan of ip->rec_tree.  Try to include each record in our
2041  * flush.  ip->flush_group has been set but the inode has not yet been
2042  * moved into a flushing state.
2043  *
2044  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
2045  * both inodes.
2046  *
2047  * We return 1 for any record placed or found in FST_FLUSH, which prevents
2048  * the caller from shortcutting the flush.
2049  */
2050 static int
2051 hammer_setup_child_callback(hammer_record_t rec, void *data)
2052 {
2053         hammer_flush_group_t flg;
2054         hammer_inode_t target_ip;
2055         hammer_inode_t ip;
2056         int r;
2057
2058         /*
2059          * Records deleted or committed by the backend are ignored.
2060          * Note that the flush detects deleted frontend records at
2061          * multiple points to deal with races.  This is just the first
2062          * line of defense.  The only time HAMMER_RECF_DELETED_FE cannot
2063          * be set is when HAMMER_RECF_INTERLOCK_BE is set, because it
2064          * messes up link-count calculations.
2065          *
2066          * NOTE: Don't get confused between record deletion and, say,
2067          * directory entry deletion.  The deletion of a directory entry
2068          * which is on-media has nothing to do with the record deletion
2069          * flags.
2070          */
2071         if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
2072                           HAMMER_RECF_COMMITTED)) {
2073                 if (rec->flush_state == HAMMER_FST_FLUSH) {
2074                         KKASSERT(rec->flush_group == rec->ip->flush_group);
2075                         r = 1;
2076                 } else {
2077                         r = 0;
2078                 }
2079                 return(r);
2080         }
2081
2082         /*
2083          * If the record is in an idle state it has no dependancies and
2084          * can be flushed.
2085          */
2086         ip = rec->ip;
2087         flg = ip->flush_group;
2088         r = 0;
2089
2090         switch(rec->flush_state) {
2091         case HAMMER_FST_IDLE:
2092                 /*
2093                  * The record has no setup dependancy, we can flush it.
2094                  */
2095                 KKASSERT(rec->target_ip == NULL);
2096                 rec->flush_state = HAMMER_FST_FLUSH;
2097                 rec->flush_group = flg;
2098                 ++flg->refs;
2099                 hammer_ref(&rec->lock);
2100                 r = 1;
2101                 break;
2102         case HAMMER_FST_SETUP:
2103                 /*
2104                  * The record has a setup dependancy.  These are typically
2105                  * directory entry adds and deletes.  Such entries will be
2106                  * flushed when their inodes are flushed so we do not
2107                  * usually have to add them to the flush here.  However,
2108                  * if the target_ip has set HAMMER_INODE_CONN_DOWN then
2109                  * it is asking us to flush this record (and it).
2110                  */
2111                 target_ip = rec->target_ip;
2112                 KKASSERT(target_ip != NULL);
2113                 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
2114
2115                 /*
2116                  * If the target IP is already flushing in our group
2117                  * we could associate the record, but target_ip has
2118                  * already synced ino_data to sync_ino_data and we
2119                  * would also have to adjust nlinks.   Plus there are
2120                  * ordering issues for adds and deletes.
2121                  *
2122                  * Reflush downward if this is an ADD, and upward if
2123                  * this is a DEL.
2124                  */
2125                 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
2126                         if (rec->flush_state == HAMMER_MEM_RECORD_ADD)
2127                                 ip->flags |= HAMMER_INODE_REFLUSH;
2128                         else
2129                                 target_ip->flags |= HAMMER_INODE_REFLUSH;
2130                         break;
2131                 } 
2132
2133                 /*
2134                  * Target IP is not yet flushing.  This can get complex
2135                  * because we have to be careful about the recursion.
2136                  *
2137                  * Directories create an issue for us in that if a flush
2138                  * of a directory is requested the expectation is to flush
2139                  * any pending directory entries, but this will cause the
2140                  * related inodes to recursively flush as well.  We can't
2141                  * really defer the operation so just get as many as we
2142                  * can and
2143                  */
2144 #if 0
2145                 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
2146                     (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
2147                         /*
2148                          * We aren't reclaiming and the target ip was not
2149                          * previously prevented from flushing due to this
2150                          * record dependancy.  Do not flush this record.
2151                          */
2152                         /*r = 0;*/
2153                 } else
2154 #endif
2155                 if (flg->total_count + flg->refs >
2156                            ip->hmp->undo_rec_limit) {
2157                         /*
2158                          * Our flush group is over-full and we risk blowing
2159                          * out the UNDO FIFO.  Stop the scan, flush what we
2160                          * have, then reflush the directory.
2161                          *
2162                          * The directory may be forced through multiple
2163                          * flush groups before it can be completely
2164                          * flushed.
2165                          */
2166                         ip->flags |= HAMMER_INODE_RESIGNAL |
2167                                      HAMMER_INODE_REFLUSH;
2168                         r = -1;
2169                 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
2170                         /*
2171                          * If the target IP is not flushing we can force
2172                          * it to flush, even if it is unable to write out
2173                          * any of its own records we have at least one in
2174                          * hand that we CAN deal with.
2175                          */
2176                         rec->flush_state = HAMMER_FST_FLUSH;
2177                         rec->flush_group = flg;
2178                         ++flg->refs;
2179                         hammer_ref(&rec->lock);
2180                         hammer_flush_inode_core(target_ip, flg,
2181                                                 HAMMER_FLUSH_RECURSION);
2182                         r = 1;
2183                 } else {
2184                         /*
2185                          * General or delete-on-disk record.
2186                          *
2187                          * XXX this needs help.  If a delete-on-disk we could
2188                          * disconnect the target.  If the target has its own
2189                          * dependancies they really need to be flushed.
2190                          *
2191                          * XXX
2192                          */
2193                         rec->flush_state = HAMMER_FST_FLUSH;
2194                         rec->flush_group = flg;
2195                         ++flg->refs;
2196                         hammer_ref(&rec->lock);
2197                         hammer_flush_inode_core(target_ip, flg,
2198                                                 HAMMER_FLUSH_RECURSION);
2199                         r = 1;
2200                 }
2201                 break;
2202         case HAMMER_FST_FLUSH:
2203                 /* 
2204                  * The flush_group should already match.
2205                  */
2206                 KKASSERT(rec->flush_group == flg);
2207                 r = 1;
2208                 break;
2209         }
2210         return(r);
2211 }
2212
2213 #if 0
2214 /*
2215  * This version just moves records already in a flush state to the new
2216  * flush group and that is it.
2217  */
2218 static int
2219 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
2220 {
2221         hammer_inode_t ip = rec->ip;
2222
2223         switch(rec->flush_state) {
2224         case HAMMER_FST_FLUSH:
2225                 KKASSERT(rec->flush_group == ip->flush_group);
2226                 break;
2227         default:
2228                 break;
2229         }
2230         return(0);
2231 }
2232 #endif
2233
2234 /*
2235  * Wait for a previously queued flush to complete.
2236  *
2237  * If a critical error occured we don't try to wait.
2238  */
2239 void
2240 hammer_wait_inode(hammer_inode_t ip)
2241 {
2242         hammer_flush_group_t flg;
2243
2244         flg = NULL;
2245         if ((ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2246                 while (ip->flush_state != HAMMER_FST_IDLE &&
2247                        (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2248                         if (ip->flush_state == HAMMER_FST_SETUP)
2249                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2250                         if (ip->flush_state != HAMMER_FST_IDLE) {
2251                                 ip->flags |= HAMMER_INODE_FLUSHW;
2252                                 tsleep(&ip->flags, 0, "hmrwin", 0);
2253                         }
2254                 }
2255         }
2256 }
2257
2258 /*
2259  * Called by the backend code when a flush has been completed.
2260  * The inode has already been removed from the flush list.
2261  *
2262  * A pipelined flush can occur, in which case we must re-enter the
2263  * inode on the list and re-copy its fields.
2264  */
2265 void
2266 hammer_flush_inode_done(hammer_inode_t ip, int error)
2267 {
2268         hammer_mount_t hmp;
2269         int dorel;
2270
2271         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2272
2273         hmp = ip->hmp;
2274
2275         /*
2276          * Auto-reflush if the backend could not completely flush
2277          * the inode.  This fixes a case where a deferred buffer flush
2278          * could cause fsync to return early.
2279          */
2280         if (ip->sync_flags & HAMMER_INODE_MODMASK)
2281                 ip->flags |= HAMMER_INODE_REFLUSH;
2282
2283         /*
2284          * Merge left-over flags back into the frontend and fix the state.
2285          * Incomplete truncations are retained by the backend.
2286          */
2287         ip->error = error;
2288         ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2289         ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2290
2291         /*
2292          * The backend may have adjusted nlinks, so if the adjusted nlinks
2293          * does not match the fronttend set the frontend's RDIRTY flag again.
2294          */
2295         if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2296                 ip->flags |= HAMMER_INODE_DDIRTY;
2297
2298         /*
2299          * Fix up the dirty buffer status.
2300          */
2301         if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2302                 ip->flags |= HAMMER_INODE_BUFS;
2303         }
2304
2305         /*
2306          * Re-set the XDIRTY flag if some of the inode's in-memory records
2307          * could not be flushed.
2308          */
2309         KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2310                   (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2311                  (!RB_EMPTY(&ip->rec_tree) &&
2312                   (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2313
2314         /*
2315          * Do not lose track of inodes which no longer have vnode
2316          * assocations, otherwise they may never get flushed again.
2317          *
2318          * The reflush flag can be set superfluously, causing extra pain
2319          * for no reason.  If the inode is no longer modified it no longer
2320          * needs to be flushed.
2321          */
2322         if (ip->flags & HAMMER_INODE_MODMASK) {
2323                 if (ip->vp == NULL)
2324                         ip->flags |= HAMMER_INODE_REFLUSH;
2325         } else {
2326                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2327         }
2328
2329         /*
2330          * Adjust the flush state.
2331          */
2332         if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2333                 /*
2334                  * We were unable to flush out all our records, leave the
2335                  * inode in a flush state and in the current flush group.
2336                  * The flush group will be re-run.
2337                  *
2338                  * This occurs if the UNDO block gets too full or there is
2339                  * too much dirty meta-data and allows the flusher to
2340                  * finalize the UNDO block and then re-flush.
2341                  */
2342                 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2343                 dorel = 0;
2344         } else {
2345                 /*
2346                  * Remove from the flush_group
2347                  */
2348                 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
2349                 ip->flush_group = NULL;
2350
2351                 /*
2352                  * Clean up the vnode ref and tracking counts.
2353                  */
2354                 if (ip->flags & HAMMER_INODE_VHELD) {
2355                         ip->flags &= ~HAMMER_INODE_VHELD;
2356                         vrele(ip->vp);
2357                 }
2358                 --hmp->count_iqueued;
2359                 --hammer_count_iqueued;
2360
2361                 /*
2362                  * And adjust the state.
2363                  */
2364                 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2365                         ip->flush_state = HAMMER_FST_IDLE;
2366                         dorel = 1;
2367                 } else {
2368                         ip->flush_state = HAMMER_FST_SETUP;
2369                         dorel = 0;
2370                 }
2371
2372                 /*
2373                  * If the frontend is waiting for a flush to complete,
2374                  * wake it up.
2375                  */
2376                 if (ip->flags & HAMMER_INODE_FLUSHW) {
2377                         ip->flags &= ~HAMMER_INODE_FLUSHW;
2378                         wakeup(&ip->flags);
2379                 }
2380
2381                 /*
2382                  * If the frontend made more changes and requested another
2383                  * flush, then try to get it running.
2384                  *
2385                  * Reflushes are aborted when the inode is errored out.
2386                  */
2387                 if (ip->flags & HAMMER_INODE_REFLUSH) {
2388                         ip->flags &= ~HAMMER_INODE_REFLUSH;
2389                         if (ip->flags & HAMMER_INODE_RESIGNAL) {
2390                                 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2391                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2392                         } else {
2393                                 hammer_flush_inode(ip, 0);
2394                         }
2395                 }
2396         }
2397
2398         /*
2399          * If we have no parent dependancies we can clear CONN_DOWN
2400          */
2401         if (TAILQ_EMPTY(&ip->target_list))
2402                 ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2403
2404         /*
2405          * If the inode is now clean drop the space reservation.
2406          */
2407         if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2408             (ip->flags & HAMMER_INODE_RSV_INODES)) {
2409                 ip->flags &= ~HAMMER_INODE_RSV_INODES;
2410                 --hmp->rsv_inodes;
2411         }
2412
2413         if (dorel)
2414                 hammer_rel_inode(ip, 0);
2415 }
2416
2417 /*
2418  * Called from hammer_sync_inode() to synchronize in-memory records
2419  * to the media.
2420  */
2421 static int
2422 hammer_sync_record_callback(hammer_record_t record, void *data)
2423 {
2424         hammer_cursor_t cursor = data;
2425         hammer_transaction_t trans = cursor->trans;
2426         hammer_mount_t hmp = trans->hmp;
2427         int error;
2428
2429         /*
2430          * Skip records that do not belong to the current flush.
2431          */
2432         ++hammer_stats_record_iterations;
2433         if (record->flush_state != HAMMER_FST_FLUSH)
2434                 return(0);
2435
2436 #if 1
2437         if (record->flush_group != record->ip->flush_group) {
2438                 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2439                 Debugger("blah2");
2440                 return(0);
2441         }
2442 #endif
2443         KKASSERT(record->flush_group == record->ip->flush_group);
2444
2445         /*
2446          * Interlock the record using the BE flag.  Once BE is set the
2447          * frontend cannot change the state of FE.
2448          *
2449          * NOTE: If FE is set prior to us setting BE we still sync the
2450          * record out, but the flush completion code converts it to 
2451          * a delete-on-disk record instead of destroying it.
2452          */
2453         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2454         record->flags |= HAMMER_RECF_INTERLOCK_BE;
2455
2456         /*
2457          * The backend has already disposed of the record.
2458          */
2459         if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
2460                 error = 0;
2461                 goto done;
2462         }
2463
2464         /*
2465          * If the whole inode is being deleting all on-disk records will
2466          * be deleted very soon, we can't sync any new records to disk
2467          * because they will be deleted in the same transaction they were
2468          * created in (delete_tid == create_tid), which will assert.
2469          *
2470          * XXX There may be a case with RECORD_ADD with DELETED_FE set
2471          * that we currently panic on.
2472          */
2473         if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2474                 switch(record->type) {
2475                 case HAMMER_MEM_RECORD_DATA:
2476                         /*
2477                          * We don't have to do anything, if the record was
2478                          * committed the space will have been accounted for
2479                          * in the blockmap.
2480                          */
2481                         /* fall through */
2482                 case HAMMER_MEM_RECORD_GENERAL:
2483                         /*
2484                          * Set deleted-by-backend flag.  Do not set the
2485                          * backend committed flag, because we are throwing
2486                          * the record away.
2487                          */
2488                         record->flags |= HAMMER_RECF_DELETED_BE;
2489                         ++record->ip->rec_generation;
2490                         error = 0;
2491                         goto done;
2492                 case HAMMER_MEM_RECORD_ADD:
2493                         panic("hammer_sync_record_callback: illegal add "
2494                               "during inode deletion record %p", record);
2495                         break; /* NOT REACHED */
2496                 case HAMMER_MEM_RECORD_INODE:
2497                         panic("hammer_sync_record_callback: attempt to "
2498                               "sync inode record %p?", record);
2499                         break; /* NOT REACHED */
2500                 case HAMMER_MEM_RECORD_DEL:
2501                         /* 
2502                          * Follow through and issue the on-disk deletion
2503                          */
2504                         break;
2505                 }
2506         }
2507
2508         /*
2509          * If DELETED_FE is set special handling is needed for directory
2510          * entries.  Dependant pieces related to the directory entry may
2511          * have already been synced to disk.  If this occurs we have to
2512          * sync the directory entry and then change the in-memory record
2513          * from an ADD to a DELETE to cover the fact that it's been
2514          * deleted by the frontend.
2515          *
2516          * A directory delete covering record (MEM_RECORD_DEL) can never
2517          * be deleted by the frontend.
2518          *
2519          * Any other record type (aka DATA) can be deleted by the frontend.
2520          * XXX At the moment the flusher must skip it because there may
2521          * be another data record in the flush group for the same block,
2522          * meaning that some frontend data changes can leak into the backend's
2523          * synchronization point.
2524          */
2525         if (record->flags & HAMMER_RECF_DELETED_FE) {
2526                 if (record->type == HAMMER_MEM_RECORD_ADD) {
2527                         /*
2528                          * Convert a front-end deleted directory-add to
2529                          * a directory-delete entry later.
2530                          */
2531                         record->flags |= HAMMER_RECF_CONVERT_DELETE;
2532                 } else {
2533                         /*
2534                          * Dispose of the record (race case).  Mark as
2535                          * deleted by backend (and not committed).
2536                          */
2537                         KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2538                         record->flags |= HAMMER_RECF_DELETED_BE;
2539                         ++record->ip->rec_generation;
2540                         error = 0;
2541                         goto done;
2542                 }
2543         }
2544
2545         /*
2546          * Assign the create_tid for new records.  Deletions already
2547          * have the record's entire key properly set up.
2548          */
2549         if (record->type != HAMMER_MEM_RECORD_DEL) {
2550                 record->leaf.base.create_tid = trans->tid;
2551                 record->leaf.create_ts = trans->time32;
2552         }
2553         for (;;) {
2554                 error = hammer_ip_sync_record_cursor(cursor, record);
2555                 if (error != EDEADLK)
2556                         break;
2557                 hammer_done_cursor(cursor);
2558                 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2559                                            record->ip);
2560                 if (error)
2561                         break;
2562         }
2563         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2564
2565         if (error)
2566                 error = -error;
2567 done:
2568         hammer_flush_record_done(record, error);
2569
2570         /*
2571          * Do partial finalization if we have built up too many dirty
2572          * buffers.  Otherwise a buffer cache deadlock can occur when
2573          * doing things like creating tens of thousands of tiny files.
2574          *
2575          * We must release our cursor lock to avoid a 3-way deadlock
2576          * due to the exclusive sync lock the finalizer must get.
2577          */
2578         if (hammer_flusher_meta_limit(hmp)) {
2579                 hammer_unlock_cursor(cursor);
2580                 hammer_flusher_finalize(trans, 0);
2581                 hammer_lock_cursor(cursor);
2582         }
2583
2584         return(error);
2585 }
2586
2587 /*
2588  * Backend function called by the flusher to sync an inode to media.
2589  */
2590 int
2591 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2592 {
2593         struct hammer_cursor cursor;
2594         hammer_node_t tmp_node;
2595         hammer_record_t depend;
2596         hammer_record_t next;
2597         int error, tmp_error;
2598         u_int64_t nlinks;
2599
2600         if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2601                 return(0);
2602
2603         error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2604         if (error)
2605                 goto done;
2606
2607         /*
2608          * Any directory records referencing this inode which are not in
2609          * our current flush group must adjust our nlink count for the
2610          * purposes of synchronization to disk.
2611          *
2612          * Records which are in our flush group can be unlinked from our
2613          * inode now, potentially allowing the inode to be physically
2614          * deleted.
2615          *
2616          * This cannot block.
2617          */
2618         nlinks = ip->ino_data.nlinks;
2619         next = TAILQ_FIRST(&ip->target_list);
2620         while ((depend = next) != NULL) {
2621                 next = TAILQ_NEXT(depend, target_entry);
2622                 if (depend->flush_state == HAMMER_FST_FLUSH &&
2623                     depend->flush_group == ip->flush_group) {
2624                         /*
2625                          * If this is an ADD that was deleted by the frontend
2626                          * the frontend nlinks count will have already been
2627                          * decremented, but the backend is going to sync its
2628                          * directory entry and must account for it.  The
2629                          * record will be converted to a delete-on-disk when
2630                          * it gets synced.
2631                          *
2632                          * If the ADD was not deleted by the frontend we
2633                          * can remove the dependancy from our target_list.
2634                          */
2635                         if (depend->flags & HAMMER_RECF_DELETED_FE) {
2636                                 ++nlinks;
2637                         } else {
2638                                 TAILQ_REMOVE(&ip->target_list, depend,
2639                                              target_entry);
2640                                 depend->target_ip = NULL;
2641                         }
2642                 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2643                         /*
2644                          * Not part of our flush group and not deleted by
2645                          * the front-end, adjust the link count synced to
2646                          * the media (undo what the frontend did when it
2647                          * queued the record).
2648                          */
2649                         KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2650                         switch(depend->type) {
2651                         case HAMMER_MEM_RECORD_ADD:
2652                                 --nlinks;
2653                                 break;
2654                         case HAMMER_MEM_RECORD_DEL:
2655                                 ++nlinks;
2656                                 break;
2657                         default:
2658                                 break;
2659                         }
2660                 }
2661         }
2662
2663         /*
2664          * Set dirty if we had to modify the link count.
2665          */
2666         if (ip->sync_ino_data.nlinks != nlinks) {
2667                 KKASSERT((int64_t)nlinks >= 0);
2668                 ip->sync_ino_data.nlinks = nlinks;
2669                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2670         }
2671
2672         /*
2673          * If there is a trunction queued destroy any data past the (aligned)
2674          * truncation point.  Userland will have dealt with the buffer
2675          * containing the truncation point for us.
2676          *
2677          * We don't flush pending frontend data buffers until after we've
2678          * dealt with the truncation.
2679          */
2680         if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2681                 /*
2682                  * Interlock trunc_off.  The VOP front-end may continue to
2683                  * make adjustments to it while we are blocked.
2684                  */
2685                 off_t trunc_off;
2686                 off_t aligned_trunc_off;
2687                 int blkmask;
2688
2689                 trunc_off = ip->sync_trunc_off;
2690                 blkmask = hammer_blocksize(trunc_off) - 1;
2691                 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
2692
2693                 /*
2694                  * Delete any whole blocks on-media.  The front-end has
2695                  * already cleaned out any partial block and made it
2696                  * pending.  The front-end may have updated trunc_off
2697                  * while we were blocked so we only use sync_trunc_off.
2698                  *
2699                  * This operation can blow out the buffer cache, EWOULDBLOCK
2700                  * means we were unable to complete the deletion.  The
2701                  * deletion will update sync_trunc_off in that case.
2702                  */
2703                 error = hammer_ip_delete_range(&cursor, ip,
2704                                                 aligned_trunc_off,
2705                                                 0x7FFFFFFFFFFFFFFFLL, 2);
2706                 if (error == EWOULDBLOCK) {
2707                         ip->flags |= HAMMER_INODE_WOULDBLOCK;
2708                         error = 0;
2709                         goto defer_buffer_flush;
2710                 }
2711
2712                 if (error)
2713                         goto done;
2714
2715                 /*
2716                  * Clear the truncation flag on the backend after we have
2717                  * complete the deletions.  Backend data is now good again
2718                  * (including new records we are about to sync, below).
2719                  *
2720                  * Leave sync_trunc_off intact.  As we write additional
2721                  * records the backend will update sync_trunc_off.  This
2722                  * tells the backend whether it can skip the overwrite
2723                  * test.  This should work properly even when the backend
2724                  * writes full blocks where the truncation point straddles
2725                  * the block because the comparison is against the base
2726                  * offset of the record.
2727                  */
2728                 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2729                 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2730         } else {
2731                 error = 0;
2732         }
2733
2734         /*
2735          * Now sync related records.  These will typically be directory
2736          * entries, records tracking direct-writes, or delete-on-disk records.
2737          */
2738         if (error == 0) {
2739                 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2740                                     hammer_sync_record_callback, &cursor);
2741                 if (tmp_error < 0)
2742                         tmp_error = -error;
2743                 if (tmp_error)
2744                         error = tmp_error;
2745         }
2746         hammer_cache_node(&ip->cache[1], cursor.node);
2747
2748         /*
2749          * Re-seek for inode update, assuming our cache hasn't been ripped
2750          * out from under us.
2751          */
2752         if (error == 0) {
2753                 tmp_node = hammer_ref_node_safe(trans, &ip->cache[0], &error);
2754                 if (tmp_node) {
2755                         hammer_cursor_downgrade(&cursor);
2756                         hammer_lock_sh(&tmp_node->lock);
2757                         if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
2758                                 hammer_cursor_seek(&cursor, tmp_node, 0);
2759                         hammer_unlock(&tmp_node->lock);
2760                         hammer_rel_node(tmp_node);
2761                 }
2762                 error = 0;
2763         }
2764
2765         /*
2766          * If we are deleting the inode the frontend had better not have
2767          * any active references on elements making up the inode.
2768          *
2769          * The call to hammer_ip_delete_clean() cleans up auxillary records
2770          * but not DB or DATA records.  Those must have already been deleted
2771          * by the normal truncation mechanic.
2772          */
2773         if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
2774                 RB_EMPTY(&ip->rec_tree)  &&
2775             (ip->sync_flags & HAMMER_INODE_DELETING) &&
2776             (ip->flags & HAMMER_INODE_DELETED) == 0) {
2777                 int count1 = 0;
2778
2779                 error = hammer_ip_delete_clean(&cursor, ip, &count1);
2780                 if (error == 0) {
2781                         ip->flags |= HAMMER_INODE_DELETED;
2782                         ip->sync_flags &= ~HAMMER_INODE_DELETING;
2783                         ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2784                         KKASSERT(RB_EMPTY(&ip->rec_tree));
2785
2786                         /*
2787                          * Set delete_tid in both the frontend and backend
2788                          * copy of the inode record.  The DELETED flag handles
2789                          * this, do not set RDIRTY.
2790                          */
2791                         ip->ino_leaf.base.delete_tid = trans->tid;
2792                         ip->sync_ino_leaf.base.delete_tid = trans->tid;
2793                         ip->ino_leaf.delete_ts = trans->time32;
2794                         ip->sync_ino_leaf.delete_ts = trans->time32;
2795
2796
2797                         /*
2798                          * Adjust the inode count in the volume header
2799                          */
2800                         hammer_sync_lock_sh(trans);
2801                         if (ip->flags & HAMMER_INODE_ONDISK) {
2802                                 hammer_modify_volume_field(trans,
2803                                                            trans->rootvol,
2804                                                            vol0_stat_inodes);
2805                                 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
2806                                 hammer_modify_volume_done(trans->rootvol);
2807                         }
2808                         hammer_sync_unlock(trans);
2809                 }
2810         }
2811
2812         if (error)
2813                 goto done;
2814         ip->sync_flags &= ~HAMMER_INODE_BUFS;
2815
2816 defer_buffer_flush:
2817         /*
2818          * Now update the inode's on-disk inode-data and/or on-disk record.
2819          * DELETED and ONDISK are managed only in ip->flags.
2820          *
2821          * In the case of a defered buffer flush we still update the on-disk
2822          * inode to satisfy visibility requirements if there happen to be
2823          * directory dependancies.
2824          */
2825         switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
2826         case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
2827                 /*
2828                  * If deleted and on-disk, don't set any additional flags.
2829                  * the delete flag takes care of things.
2830                  *
2831                  * Clear flags which may have been set by the frontend.
2832                  */
2833                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2834                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2835                                     HAMMER_INODE_DELETING);
2836                 break;
2837         case HAMMER_INODE_DELETED:
2838                 /*
2839                  * Take care of the case where a deleted inode was never
2840                  * flushed to the disk in the first place.
2841                  *
2842                  * Clear flags which may have been set by the frontend.
2843                  */
2844                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2845                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2846                                     HAMMER_INODE_DELETING);
2847                 while (RB_ROOT(&ip->rec_tree)) {
2848                         hammer_record_t record = RB_ROOT(&ip->rec_tree);
2849                         hammer_ref(&record->lock);
2850                         KKASSERT(record->lock.refs == 1);
2851                         record->flags |= HAMMER_RECF_DELETED_BE;
2852                         ++record->ip->rec_generation;
2853                         hammer_rel_mem_record(record);
2854                 }
2855                 break;
2856         case HAMMER_INODE_ONDISK:
2857                 /*
2858                  * If already on-disk, do not set any additional flags.
2859                  */
2860                 break;
2861         default:
2862                 /*
2863                  * If not on-disk and not deleted, set DDIRTY to force
2864                  * an initial record to be written.
2865                  *
2866                  * Also set the create_tid in both the frontend and backend
2867                  * copy of the inode record.
2868                  */
2869                 ip->ino_leaf.base.create_tid = trans->tid;
2870                 ip->ino_leaf.create_ts = trans->time32;
2871                 ip->sync_ino_leaf.base.create_tid = trans->tid;
2872                 ip->sync_ino_leaf.create_ts = trans->time32;
2873                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2874                 break;
2875         }
2876
2877         /*
2878          * If RDIRTY or DDIRTY is set, write out a new record.  If the inode
2879          * is already on-disk the old record is marked as deleted.
2880          *
2881          * If DELETED is set hammer_update_inode() will delete the existing
2882          * record without writing out a new one.
2883          *
2884          * If *ONLY* the ITIMES flag is set we can update the record in-place.
2885          */
2886         if (ip->flags & HAMMER_INODE_DELETED) {
2887                 error = hammer_update_inode(&cursor, ip);
2888         } else 
2889         if ((ip->sync_flags & HAMMER_INODE_DDIRTY) == 0 &&
2890             (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
2891                 error = hammer_update_itimes(&cursor, ip);
2892         } else
2893         if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
2894                 error = hammer_update_inode(&cursor, ip);
2895         }
2896 done:
2897         if (error) {
2898                 hammer_critical_error(ip->hmp, ip, error,
2899                                       "while syncing inode");
2900         }
2901         hammer_done_cursor(&cursor);
2902         return(error);
2903 }
2904
2905 /*
2906  * This routine is called when the OS is no longer actively referencing
2907  * the inode (but might still be keeping it cached), or when releasing
2908  * the last reference to an inode.
2909  *
2910  * At this point if the inode's nlinks count is zero we want to destroy
2911  * it, which may mean destroying it on-media too.
2912  */
2913 void
2914 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
2915 {
2916         struct vnode *vp;
2917
2918         /*
2919          * Set the DELETING flag when the link count drops to 0 and the
2920          * OS no longer has any opens on the inode.
2921          *
2922          * The backend will clear DELETING (a mod flag) and set DELETED
2923          * (a state flag) when it is actually able to perform the
2924          * operation.
2925          *
2926          * Don't reflag the deletion if the flusher is currently syncing
2927          * one that was already flagged.  A previously set DELETING flag
2928          * may bounce around flags and sync_flags until the operation is
2929          * completely done.
2930          */
2931         if (ip->ino_data.nlinks == 0 &&
2932             ((ip->flags | ip->sync_flags) & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
2933                 ip->flags |= HAMMER_INODE_DELETING;
2934                 ip->flags |= HAMMER_INODE_TRUNCATED;
2935                 ip->trunc_off = 0;
2936                 vp = NULL;
2937                 if (getvp) {
2938                         if (hammer_get_vnode(ip, &vp) != 0)
2939                                 return;
2940                 }
2941
2942                 /*
2943                  * Final cleanup
2944                  */
2945                 if (ip->vp) {
2946                         vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
2947                         vnode_pager_setsize(ip->vp, 0);
2948                 }
2949                 if (getvp) {
2950                         vput(vp);
2951                 }
2952         }
2953 }
2954
2955 /*
2956  * After potentially resolving a dependancy the inode is tested
2957  * to determine whether it needs to be reflushed.
2958  */
2959 void
2960 hammer_test_inode(hammer_inode_t ip)
2961 {
2962         if (ip->flags & HAMMER_INODE_REFLUSH) {
2963                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2964                 hammer_ref(&ip->lock);
2965                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2966                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
2967                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2968                 } else {
2969                         hammer_flush_inode(ip, 0);
2970                 }
2971                 hammer_rel_inode(ip, 0);
2972         }
2973 }
2974
2975 /*
2976  * Clear the RECLAIM flag on an inode.  This occurs when the inode is
2977  * reassociated with a vp or just before it gets freed.
2978  *
2979  * Pipeline wakeups to threads blocked due to an excessive number of
2980  * detached inodes.  The reclaim count generates a bit of negative
2981  * feedback.
2982  */
2983 static void
2984 hammer_inode_wakereclaims(hammer_inode_t ip, int dowake)
2985 {
2986         struct hammer_reclaim *reclaim;
2987         hammer_mount_t hmp = ip->hmp;
2988
2989         if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
2990                 return;
2991
2992         --hammer_count_reclaiming;
2993         --hmp->inode_reclaims;
2994         ip->flags &= ~HAMMER_INODE_RECLAIM;
2995
2996         if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT || dowake) {
2997                 reclaim = TAILQ_FIRST(&hmp->reclaim_list);
2998                 if (reclaim && reclaim->count > 0 && --reclaim->count == 0) {
2999                         TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
3000                         wakeup(reclaim);
3001                 }
3002         }
3003 }
3004
3005 /*
3006  * Setup our reclaim pipeline.  We only let so many detached (and dirty)
3007  * inodes build up before we start blocking.
3008  *
3009  * When we block we don't care *which* inode has finished reclaiming,
3010  * as lone as one does.  This is somewhat heuristical... we also put a
3011  * cap on how long we are willing to wait.
3012  */
3013 void
3014 hammer_inode_waitreclaims(hammer_mount_t hmp)
3015 {
3016         struct hammer_reclaim reclaim;
3017         int delay;
3018
3019         if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT)
3020                 return;
3021         delay = (hmp->inode_reclaims - HAMMER_RECLAIM_WAIT) * hz /
3022                 (HAMMER_RECLAIM_WAIT * 3) + 1;
3023         if (delay > 0) {
3024                 reclaim.count = 2;
3025                 TAILQ_INSERT_TAIL(&hmp->reclaim_list, &reclaim, entry);
3026                 tsleep(&reclaim, 0, "hmrrcm", delay);
3027                 if (reclaim.count > 0)
3028                         TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
3029         }
3030 }
3031
3032 /*
3033  * A larger then normal backlog of inodes is sitting in the flusher,
3034  * enforce a general slowdown to let it catch up.  This routine is only
3035  * called on completion of a non-flusher-related transaction which
3036  * performed B-Tree node I/O.
3037  *
3038  * It is possible for the flusher to stall in a continuous load.
3039  * blogbench -i1000 -o seems to do a good job generating this sort of load.
3040  * If the flusher is unable to catch up the inode count can bloat until
3041  * we run out of kvm.
3042  *
3043  * This is a bit of a hack.
3044  */
3045 void
3046 hammer_inode_waithard(hammer_mount_t hmp)
3047 {
3048         /*
3049          * Hysteresis.
3050          */
3051         if (hmp->flags & HAMMER_MOUNT_FLUSH_RECOVERY) {
3052                 if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT / 2 &&
3053                     hmp->count_iqueued < hmp->count_inodes / 20) {
3054                         hmp->flags &= ~HAMMER_MOUNT_FLUSH_RECOVERY;
3055                         return;
3056                 }
3057         } else {
3058                 if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT ||
3059                     hmp->count_iqueued < hmp->count_inodes / 10) {
3060                         return;
3061                 }
3062                 hmp->flags |= HAMMER_MOUNT_FLUSH_RECOVERY;
3063         }
3064
3065         /*
3066          * Block for one flush cycle.
3067          */
3068         hammer_flusher_wait_next(hmp);
3069 }
3070