AMD64 - Fix many compile-time warnings. int/ptr type mismatches, %llx, etc.
[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",
841                       (long long)ip->obj_id);
842                 /* not reached */
843                 hammer_free_inode(ip);
844         }
845         *ipp = ip;
846         return(error);
847 }
848
849 /*
850  * Final cleanup / freeing of an inode structure
851  */
852 static void
853 hammer_free_inode(hammer_inode_t ip)
854 {
855         struct hammer_mount *hmp;
856
857         hmp = ip->hmp;
858         KKASSERT(ip->lock.refs == 1);
859         hammer_uncache_node(&ip->cache[0]);
860         hammer_uncache_node(&ip->cache[1]);
861         hammer_uncache_node(&ip->cache[2]);
862         hammer_uncache_node(&ip->cache[3]);
863         hammer_inode_wakereclaims(ip, 1);
864         if (ip->objid_cache)
865                 hammer_clear_objid(ip);
866         --hammer_count_inodes;
867         --hmp->count_inodes;
868         if (ip->pfsm) {
869                 hammer_rel_pseudofs(hmp, ip->pfsm);
870                 ip->pfsm = NULL;
871         }
872         kfree(ip, hmp->m_inodes);
873         ip = NULL;
874 }
875
876 /*
877  * Retrieve pseudo-fs data.  NULL will never be returned.
878  *
879  * If an error occurs *errorp will be set and a default template is returned,
880  * otherwise *errorp is set to 0.  Typically when an error occurs it will
881  * be ENOENT.
882  */
883 hammer_pseudofs_inmem_t
884 hammer_load_pseudofs(hammer_transaction_t trans,
885                      u_int32_t localization, int *errorp)
886 {
887         hammer_mount_t hmp = trans->hmp;
888         hammer_inode_t ip;
889         hammer_pseudofs_inmem_t pfsm;
890         struct hammer_cursor cursor;
891         int bytes;
892
893 retry:
894         pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
895         if (pfsm) {
896                 hammer_ref(&pfsm->lock);
897                 *errorp = 0;
898                 return(pfsm);
899         }
900
901         /*
902          * PFS records are stored in the root inode (not the PFS root inode,
903          * but the real root).  Avoid an infinite recursion if loading
904          * the PFS for the real root.
905          */
906         if (localization) {
907                 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
908                                       HAMMER_MAX_TID,
909                                       HAMMER_DEF_LOCALIZATION, 0, errorp);
910         } else {
911                 ip = NULL;
912         }
913
914         pfsm = kmalloc(sizeof(*pfsm), hmp->m_misc, M_WAITOK | M_ZERO);
915         pfsm->localization = localization;
916         pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
917         pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
918
919         hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
920         cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION +
921                                       HAMMER_LOCALIZE_MISC;
922         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
923         cursor.key_beg.create_tid = 0;
924         cursor.key_beg.delete_tid = 0;
925         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
926         cursor.key_beg.obj_type = 0;
927         cursor.key_beg.key = localization;
928         cursor.asof = HAMMER_MAX_TID;
929         cursor.flags |= HAMMER_CURSOR_ASOF;
930
931         if (ip)
932                 *errorp = hammer_ip_lookup(&cursor);
933         else
934                 *errorp = hammer_btree_lookup(&cursor);
935         if (*errorp == 0) {
936                 *errorp = hammer_ip_resolve_data(&cursor);
937                 if (*errorp == 0) {
938                         if (cursor.data->pfsd.mirror_flags &
939                             HAMMER_PFSD_DELETED) {
940                                 *errorp = ENOENT;
941                         } else {
942                                 bytes = cursor.leaf->data_len;
943                                 if (bytes > sizeof(pfsm->pfsd))
944                                         bytes = sizeof(pfsm->pfsd);
945                                 bcopy(cursor.data, &pfsm->pfsd, bytes);
946                         }
947                 }
948         }
949         hammer_done_cursor(&cursor);
950
951         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
952         hammer_ref(&pfsm->lock);
953         if (ip)
954                 hammer_rel_inode(ip, 0);
955         if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
956                 kfree(pfsm, hmp->m_misc);
957                 goto retry;
958         }
959         return(pfsm);
960 }
961
962 /*
963  * Store pseudo-fs data.  The backend will automatically delete any prior
964  * on-disk pseudo-fs data but we have to delete in-memory versions.
965  */
966 int
967 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
968 {
969         struct hammer_cursor cursor;
970         hammer_record_t record;
971         hammer_inode_t ip;
972         int error;
973
974         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
975                               HAMMER_DEF_LOCALIZATION, 0, &error);
976 retry:
977         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
978         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
979         cursor.key_beg.localization = ip->obj_localization +
980                                       HAMMER_LOCALIZE_MISC;
981         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
982         cursor.key_beg.create_tid = 0;
983         cursor.key_beg.delete_tid = 0;
984         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
985         cursor.key_beg.obj_type = 0;
986         cursor.key_beg.key = pfsm->localization;
987         cursor.asof = HAMMER_MAX_TID;
988         cursor.flags |= HAMMER_CURSOR_ASOF;
989
990         /*
991          * Replace any in-memory version of the record.
992          */
993         error = hammer_ip_lookup(&cursor);
994         if (error == 0 && hammer_cursor_inmem(&cursor)) {
995                 record = cursor.iprec;
996                 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
997                         KKASSERT(cursor.deadlk_rec == NULL);
998                         hammer_ref(&record->lock);
999                         cursor.deadlk_rec = record;
1000                         error = EDEADLK;
1001                 } else {
1002                         record->flags |= HAMMER_RECF_DELETED_FE;
1003                         error = 0;
1004                 }
1005         }
1006
1007         /*
1008          * Allocate replacement general record.  The backend flush will
1009          * delete any on-disk version of the record.
1010          */
1011         if (error == 0 || error == ENOENT) {
1012                 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
1013                 record->type = HAMMER_MEM_RECORD_GENERAL;
1014
1015                 record->leaf.base.localization = ip->obj_localization +
1016                                                  HAMMER_LOCALIZE_MISC;
1017                 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
1018                 record->leaf.base.key = pfsm->localization;
1019                 record->leaf.data_len = sizeof(pfsm->pfsd);
1020                 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
1021                 error = hammer_ip_add_record(trans, record);
1022         }
1023         hammer_done_cursor(&cursor);
1024         if (error == EDEADLK)
1025                 goto retry;
1026         hammer_rel_inode(ip, 0);
1027         return(error);
1028 }
1029
1030 /*
1031  * Create a root directory for a PFS if one does not alredy exist.
1032  *
1033  * The PFS root stands alone so we must also bump the nlinks count
1034  * to prevent it from being destroyed on release.
1035  */
1036 int
1037 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1038                        hammer_pseudofs_inmem_t pfsm)
1039 {
1040         hammer_inode_t ip;
1041         struct vattr vap;
1042         int error;
1043
1044         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1045                               pfsm->localization, 0, &error);
1046         if (ip == NULL) {
1047                 vattr_null(&vap);
1048                 vap.va_mode = 0755;
1049                 vap.va_type = VDIR;
1050                 error = hammer_create_inode(trans, &vap, cred,
1051                                             NULL, NULL, 0,
1052                                             pfsm, &ip);
1053                 if (error == 0) {
1054                         ++ip->ino_data.nlinks;
1055                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
1056                 }
1057         }
1058         if (ip)
1059                 hammer_rel_inode(ip, 0);
1060         return(error);
1061 }
1062
1063 /*
1064  * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
1065  * if we are unable to disassociate all the inodes.
1066  */
1067 static
1068 int
1069 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
1070 {
1071         int res;
1072
1073         hammer_ref(&ip->lock);
1074         if (ip->lock.refs == 2 && ip->vp)
1075                 vclean_unlocked(ip->vp);
1076         if (ip->lock.refs == 1 && ip->vp == NULL)
1077                 res = 0;
1078         else
1079                 res = -1;       /* stop, someone is using the inode */
1080         hammer_rel_inode(ip, 0);
1081         return(res);
1082 }
1083
1084 int
1085 hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
1086 {
1087         int res;
1088         int try;
1089
1090         for (try = res = 0; try < 4; ++try) {
1091                 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
1092                                            hammer_inode_pfs_cmp,
1093                                            hammer_unload_pseudofs_callback,
1094                                            &localization);
1095                 if (res == 0 && try > 1)
1096                         break;
1097                 hammer_flusher_sync(trans->hmp);
1098         }
1099         if (res != 0)
1100                 res = ENOTEMPTY;
1101         return(res);
1102 }
1103
1104
1105 /*
1106  * Release a reference on a PFS
1107  */
1108 void
1109 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
1110 {
1111         hammer_unref(&pfsm->lock);
1112         if (pfsm->lock.refs == 0) {
1113                 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
1114                 kfree(pfsm, hmp->m_misc);
1115         }
1116 }
1117
1118 /*
1119  * Called by hammer_sync_inode().
1120  */
1121 static int
1122 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
1123 {
1124         hammer_transaction_t trans = cursor->trans;
1125         hammer_record_t record;
1126         int error;
1127         int redirty;
1128
1129 retry:
1130         error = 0;
1131
1132         /*
1133          * If the inode has a presence on-disk then locate it and mark
1134          * it deleted, setting DELONDISK.
1135          *
1136          * The record may or may not be physically deleted, depending on
1137          * the retention policy.
1138          */
1139         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
1140             HAMMER_INODE_ONDISK) {
1141                 hammer_normalize_cursor(cursor);
1142                 cursor->key_beg.localization = ip->obj_localization + 
1143                                                HAMMER_LOCALIZE_INODE;
1144                 cursor->key_beg.obj_id = ip->obj_id;
1145                 cursor->key_beg.key = 0;
1146                 cursor->key_beg.create_tid = 0;
1147                 cursor->key_beg.delete_tid = 0;
1148                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1149                 cursor->key_beg.obj_type = 0;
1150                 cursor->asof = ip->obj_asof;
1151                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1152                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
1153                 cursor->flags |= HAMMER_CURSOR_BACKEND;
1154
1155                 error = hammer_btree_lookup(cursor);
1156                 if (hammer_debug_inode)
1157                         kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
1158
1159                 if (error == 0) {
1160                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1161                         if (hammer_debug_inode)
1162                                 kprintf(" error %d\n", error);
1163                         if (error == 0) {
1164                                 ip->flags |= HAMMER_INODE_DELONDISK;
1165                         }
1166                         if (cursor->node)
1167                                 hammer_cache_node(&ip->cache[0], cursor->node);
1168                 }
1169                 if (error == EDEADLK) {
1170                         hammer_done_cursor(cursor);
1171                         error = hammer_init_cursor(trans, cursor,
1172                                                    &ip->cache[0], ip);
1173                         if (hammer_debug_inode)
1174                                 kprintf("IPDED %p %d\n", ip, error);
1175                         if (error == 0)
1176                                 goto retry;
1177                 }
1178         }
1179
1180         /*
1181          * Ok, write out the initial record or a new record (after deleting
1182          * the old one), unless the DELETED flag is set.  This routine will
1183          * clear DELONDISK if it writes out a record.
1184          *
1185          * Update our inode statistics if this is the first application of
1186          * the inode on-disk.
1187          */
1188         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
1189                 /*
1190                  * Generate a record and write it to the media.  We clean-up
1191                  * the state before releasing so we do not have to set-up
1192                  * a flush_group.
1193                  */
1194                 record = hammer_alloc_mem_record(ip, 0);
1195                 record->type = HAMMER_MEM_RECORD_INODE;
1196                 record->flush_state = HAMMER_FST_FLUSH;
1197                 record->leaf = ip->sync_ino_leaf;
1198                 record->leaf.base.create_tid = trans->tid;
1199                 record->leaf.data_len = sizeof(ip->sync_ino_data);
1200                 record->leaf.create_ts = trans->time32;
1201                 record->data = (void *)&ip->sync_ino_data;
1202                 record->flags |= HAMMER_RECF_INTERLOCK_BE;
1203
1204                 /*
1205                  * If this flag is set we cannot sync the new file size
1206                  * because we haven't finished related truncations.  The
1207                  * inode will be flushed in another flush group to finish
1208                  * the job.
1209                  */
1210                 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1211                     ip->sync_ino_data.size != ip->ino_data.size) {
1212                         redirty = 1;
1213                         ip->sync_ino_data.size = ip->ino_data.size;
1214                 } else {
1215                         redirty = 0;
1216                 }
1217
1218                 for (;;) {
1219                         error = hammer_ip_sync_record_cursor(cursor, record);
1220                         if (hammer_debug_inode)
1221                                 kprintf("GENREC %p rec %08x %d\n",      
1222                                         ip, record->flags, error);
1223                         if (error != EDEADLK)
1224                                 break;
1225                         hammer_done_cursor(cursor);
1226                         error = hammer_init_cursor(trans, cursor,
1227                                                    &ip->cache[0], ip);
1228                         if (hammer_debug_inode)
1229                                 kprintf("GENREC reinit %d\n", error);
1230                         if (error)
1231                                 break;
1232                 }
1233
1234                 /*
1235                  * Note:  The record was never on the inode's record tree
1236                  * so just wave our hands importantly and destroy it.
1237                  */
1238                 record->flags |= HAMMER_RECF_COMMITTED;
1239                 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1240                 record->flush_state = HAMMER_FST_IDLE;
1241                 ++ip->rec_generation;
1242                 hammer_rel_mem_record(record);
1243
1244                 /*
1245                  * Finish up.
1246                  */
1247                 if (error == 0) {
1248                         if (hammer_debug_inode)
1249                                 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1250                         ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1251                                             HAMMER_INODE_ATIME |
1252                                             HAMMER_INODE_MTIME);
1253                         ip->flags &= ~HAMMER_INODE_DELONDISK;
1254                         if (redirty)
1255                                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1256
1257                         /*
1258                          * Root volume count of inodes
1259                          */
1260                         hammer_sync_lock_sh(trans);
1261                         if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1262                                 hammer_modify_volume_field(trans,
1263                                                            trans->rootvol,
1264                                                            vol0_stat_inodes);
1265                                 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1266                                 hammer_modify_volume_done(trans->rootvol);
1267                                 ip->flags |= HAMMER_INODE_ONDISK;
1268                                 if (hammer_debug_inode)
1269                                         kprintf("NOWONDISK %p\n", ip);
1270                         }
1271                         hammer_sync_unlock(trans);
1272                 }
1273         }
1274
1275         /*
1276          * If the inode has been destroyed, clean out any left-over flags
1277          * that may have been set by the frontend.
1278          */
1279         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) { 
1280                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1281                                     HAMMER_INODE_ATIME |
1282                                     HAMMER_INODE_MTIME);
1283         }
1284         return(error);
1285 }
1286
1287 /*
1288  * Update only the itimes fields.
1289  *
1290  * ATIME can be updated without generating any UNDO.  MTIME is updated
1291  * with UNDO so it is guaranteed to be synchronized properly in case of
1292  * a crash.
1293  *
1294  * Neither field is included in the B-Tree leaf element's CRC, which is how
1295  * we can get away with updating ATIME the way we do.
1296  */
1297 static int
1298 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1299 {
1300         hammer_transaction_t trans = cursor->trans;
1301         int error;
1302
1303 retry:
1304         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1305             HAMMER_INODE_ONDISK) {
1306                 return(0);
1307         }
1308
1309         hammer_normalize_cursor(cursor);
1310         cursor->key_beg.localization = ip->obj_localization + 
1311                                        HAMMER_LOCALIZE_INODE;
1312         cursor->key_beg.obj_id = ip->obj_id;
1313         cursor->key_beg.key = 0;
1314         cursor->key_beg.create_tid = 0;
1315         cursor->key_beg.delete_tid = 0;
1316         cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1317         cursor->key_beg.obj_type = 0;
1318         cursor->asof = ip->obj_asof;
1319         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1320         cursor->flags |= HAMMER_CURSOR_ASOF;
1321         cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1322         cursor->flags |= HAMMER_CURSOR_GET_DATA;
1323         cursor->flags |= HAMMER_CURSOR_BACKEND;
1324
1325         error = hammer_btree_lookup(cursor);
1326         if (error == 0) {
1327                 hammer_cache_node(&ip->cache[0], cursor->node);
1328                 if (ip->sync_flags & HAMMER_INODE_MTIME) {
1329                         /*
1330                          * Updating MTIME requires an UNDO.  Just cover
1331                          * both atime and mtime.
1332                          */
1333                         hammer_sync_lock_sh(trans);
1334                         hammer_modify_buffer(trans, cursor->data_buffer,
1335                                      HAMMER_ITIMES_BASE(&cursor->data->inode),
1336                                      HAMMER_ITIMES_BYTES);
1337                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1338                         cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1339                         hammer_modify_buffer_done(cursor->data_buffer);
1340                         hammer_sync_unlock(trans);
1341                 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1342                         /*
1343                          * Updating atime only can be done in-place with
1344                          * no UNDO.
1345                          */
1346                         hammer_sync_lock_sh(trans);
1347                         hammer_modify_buffer(trans, cursor->data_buffer,
1348                                              NULL, 0);
1349                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1350                         hammer_modify_buffer_done(cursor->data_buffer);
1351                         hammer_sync_unlock(trans);
1352                 }
1353                 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1354         }
1355         if (error == EDEADLK) {
1356                 hammer_done_cursor(cursor);
1357                 error = hammer_init_cursor(trans, cursor,
1358                                            &ip->cache[0], ip);
1359                 if (error == 0)
1360                         goto retry;
1361         }
1362         return(error);
1363 }
1364
1365 /*
1366  * Release a reference on an inode, flush as requested.
1367  *
1368  * On the last reference we queue the inode to the flusher for its final
1369  * disposition.
1370  */
1371 void
1372 hammer_rel_inode(struct hammer_inode *ip, int flush)
1373 {
1374         /*hammer_mount_t hmp = ip->hmp;*/
1375
1376         /*
1377          * Handle disposition when dropping the last ref.
1378          */
1379         for (;;) {
1380                 if (ip->lock.refs == 1) {
1381                         /*
1382                          * Determine whether on-disk action is needed for
1383                          * the inode's final disposition.
1384                          */
1385                         KKASSERT(ip->vp == NULL);
1386                         hammer_inode_unloadable_check(ip, 0);
1387                         if (ip->flags & HAMMER_INODE_MODMASK) {
1388                                 hammer_flush_inode(ip, 0);
1389                         } else if (ip->lock.refs == 1) {
1390                                 hammer_unload_inode(ip);
1391                                 break;
1392                         }
1393                 } else {
1394                         if (flush)
1395                                 hammer_flush_inode(ip, 0);
1396
1397                         /*
1398                          * The inode still has multiple refs, try to drop
1399                          * one ref.
1400                          */
1401                         KKASSERT(ip->lock.refs >= 1);
1402                         if (ip->lock.refs > 1) {
1403                                 hammer_unref(&ip->lock);
1404                                 break;
1405                         }
1406                 }
1407         }
1408 }
1409
1410 /*
1411  * Unload and destroy the specified inode.  Must be called with one remaining
1412  * reference.  The reference is disposed of.
1413  *
1414  * The inode must be completely clean.
1415  */
1416 static int
1417 hammer_unload_inode(struct hammer_inode *ip)
1418 {
1419         hammer_mount_t hmp = ip->hmp;
1420
1421         KASSERT(ip->lock.refs == 1,
1422                 ("hammer_unload_inode: %d refs\n", ip->lock.refs));
1423         KKASSERT(ip->vp == NULL);
1424         KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1425         KKASSERT(ip->cursor_ip_refs == 0);
1426         KKASSERT(ip->lock.lockcount == 0);
1427         KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1428
1429         KKASSERT(RB_EMPTY(&ip->rec_tree));
1430         KKASSERT(TAILQ_EMPTY(&ip->target_list));
1431
1432         RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1433
1434         hammer_free_inode(ip);
1435         return(0);
1436 }
1437
1438 /*
1439  * Called during unmounting if a critical error occured.  The in-memory
1440  * inode and all related structures are destroyed.
1441  *
1442  * If a critical error did not occur the unmount code calls the standard
1443  * release and asserts that the inode is gone.
1444  */
1445 int
1446 hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
1447 {
1448         hammer_record_t rec;
1449
1450         /*
1451          * Get rid of the inodes in-memory records, regardless of their
1452          * state, and clear the mod-mask.
1453          */
1454         while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1455                 TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1456                 rec->target_ip = NULL;
1457                 if (rec->flush_state == HAMMER_FST_SETUP)
1458                         rec->flush_state = HAMMER_FST_IDLE;
1459         }
1460         while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1461                 if (rec->flush_state == HAMMER_FST_FLUSH)
1462                         --rec->flush_group->refs;
1463                 else
1464                         hammer_ref(&rec->lock);
1465                 KKASSERT(rec->lock.refs == 1);
1466                 rec->flush_state = HAMMER_FST_IDLE;
1467                 rec->flush_group = NULL;
1468                 rec->flags |= HAMMER_RECF_DELETED_FE; /* wave hands */
1469                 rec->flags |= HAMMER_RECF_DELETED_BE; /* wave hands */
1470                 ++ip->rec_generation;
1471                 hammer_rel_mem_record(rec);
1472         }
1473         ip->flags &= ~HAMMER_INODE_MODMASK;
1474         ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1475         KKASSERT(ip->vp == NULL);
1476
1477         /*
1478          * Remove the inode from any flush group, force it idle.  FLUSH
1479          * and SETUP states have an inode ref.
1480          */
1481         switch(ip->flush_state) {
1482         case HAMMER_FST_FLUSH:
1483                 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
1484                 --ip->flush_group->refs;
1485                 ip->flush_group = NULL;
1486                 /* fall through */
1487         case HAMMER_FST_SETUP:
1488                 hammer_unref(&ip->lock);
1489                 ip->flush_state = HAMMER_FST_IDLE;
1490                 /* fall through */
1491         case HAMMER_FST_IDLE:
1492                 break;
1493         }
1494
1495         /*
1496          * There shouldn't be any associated vnode.  The unload needs at
1497          * least one ref, if we do have a vp steal its ip ref.
1498          */
1499         if (ip->vp) {
1500                 kprintf("hammer_destroy_inode_callback: Unexpected "
1501                         "vnode association ip %p vp %p\n", ip, ip->vp);
1502                 ip->vp->v_data = NULL;
1503                 ip->vp = NULL;
1504         } else {
1505                 hammer_ref(&ip->lock);
1506         }
1507         hammer_unload_inode(ip);
1508         return(0);
1509 }
1510
1511 /*
1512  * Called on mount -u when switching from RW to RO or vise-versa.  Adjust
1513  * the read-only flag for cached inodes.
1514  *
1515  * This routine is called from a RB_SCAN().
1516  */
1517 int
1518 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1519 {
1520         hammer_mount_t hmp = ip->hmp;
1521
1522         if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1523                 ip->flags |= HAMMER_INODE_RO;
1524         else
1525                 ip->flags &= ~HAMMER_INODE_RO;
1526         return(0);
1527 }
1528
1529 /*
1530  * A transaction has modified an inode, requiring updates as specified by
1531  * the passed flags.
1532  *
1533  * HAMMER_INODE_DDIRTY: Inode data has been updated
1534  * HAMMER_INODE_XDIRTY: Dirty in-memory records
1535  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
1536  * HAMMER_INODE_DELETED: Inode record/data must be deleted
1537  * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1538  */
1539 void
1540 hammer_modify_inode(hammer_inode_t ip, int flags)
1541 {
1542         /* 
1543          * ronly of 0 or 2 does not trigger assertion.
1544          * 2 is a special error state 
1545          */
1546         KKASSERT(ip->hmp->ronly != 1 ||
1547                   (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY | 
1548                             HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1549                             HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1550         if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1551                 ip->flags |= HAMMER_INODE_RSV_INODES;
1552                 ++ip->hmp->rsv_inodes;
1553         }
1554
1555         ip->flags |= flags;
1556 }
1557
1558 /*
1559  * Request that an inode be flushed.  This whole mess cannot block and may
1560  * recurse (if not synchronous).  Once requested HAMMER will attempt to
1561  * actively flush the inode until the flush can be done.
1562  *
1563  * The inode may already be flushing, or may be in a setup state.  We can
1564  * place the inode in a flushing state if it is currently idle and flag it
1565  * to reflush if it is currently flushing.
1566  *
1567  * Upon return if the inode could not be flushed due to a setup
1568  * dependancy, then it will be automatically flushed when the dependancy
1569  * is satisfied.
1570  */
1571 void
1572 hammer_flush_inode(hammer_inode_t ip, int flags)
1573 {
1574         hammer_mount_t hmp;
1575         hammer_flush_group_t flg;
1576         int good;
1577
1578         /*
1579          * next_flush_group is the first flush group we can place the inode
1580          * in.  It may be NULL.  If it becomes full we append a new flush
1581          * group and make that the next_flush_group.
1582          */
1583         hmp = ip->hmp;
1584         while ((flg = hmp->next_flush_group) != NULL) {
1585                 KKASSERT(flg->running == 0);
1586                 if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit)
1587                         break;
1588                 hmp->next_flush_group = TAILQ_NEXT(flg, flush_entry);
1589                 hammer_flusher_async(ip->hmp, flg);
1590         }
1591         if (flg == NULL) {
1592                 flg = kmalloc(sizeof(*flg), hmp->m_misc, M_WAITOK|M_ZERO);
1593                 hmp->next_flush_group = flg;
1594                 TAILQ_INIT(&flg->flush_list);
1595                 TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1596         }
1597
1598         /*
1599          * Trivial 'nothing to flush' case.  If the inode is in a SETUP
1600          * state we have to put it back into an IDLE state so we can
1601          * drop the extra ref.
1602          *
1603          * If we have a parent dependancy we must still fall through
1604          * so we can run it.
1605          */
1606         if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1607                 if (ip->flush_state == HAMMER_FST_SETUP &&
1608                     TAILQ_EMPTY(&ip->target_list)) {
1609                         ip->flush_state = HAMMER_FST_IDLE;
1610                         hammer_rel_inode(ip, 0);
1611                 }
1612                 if (ip->flush_state == HAMMER_FST_IDLE)
1613                         return;
1614         }
1615
1616         /*
1617          * Our flush action will depend on the current state.
1618          */
1619         switch(ip->flush_state) {
1620         case HAMMER_FST_IDLE:
1621                 /*
1622                  * We have no dependancies and can flush immediately.  Some
1623                  * our children may not be flushable so we have to re-test
1624                  * with that additional knowledge.
1625                  */
1626                 hammer_flush_inode_core(ip, flg, flags);
1627                 break;
1628         case HAMMER_FST_SETUP:
1629                 /*
1630                  * Recurse upwards through dependancies via target_list
1631                  * and start their flusher actions going if possible.
1632                  *
1633                  * 'good' is our connectivity.  -1 means we have none and
1634                  * can't flush, 0 means there weren't any dependancies, and
1635                  * 1 means we have good connectivity.
1636                  */
1637                 good = hammer_setup_parent_inodes(ip, 0, flg);
1638
1639                 if (good >= 0) {
1640                         /*
1641                          * We can continue if good >= 0.  Determine how 
1642                          * many records under our inode can be flushed (and
1643                          * mark them).
1644                          */
1645                         hammer_flush_inode_core(ip, flg, flags);
1646                 } else {
1647                         /*
1648                          * Parent has no connectivity, tell it to flush
1649                          * us as soon as it does.
1650                          *
1651                          * The REFLUSH flag is also needed to trigger
1652                          * dependancy wakeups.
1653                          */
1654                         ip->flags |= HAMMER_INODE_CONN_DOWN |
1655                                      HAMMER_INODE_REFLUSH;
1656                         if (flags & HAMMER_FLUSH_SIGNAL) {
1657                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1658                                 hammer_flusher_async(ip->hmp, flg);
1659                         }
1660                 }
1661                 break;
1662         case HAMMER_FST_FLUSH:
1663                 /*
1664                  * We are already flushing, flag the inode to reflush
1665                  * if needed after it completes its current flush.
1666                  *
1667                  * The REFLUSH flag is also needed to trigger
1668                  * dependancy wakeups.
1669                  */
1670                 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1671                         ip->flags |= HAMMER_INODE_REFLUSH;
1672                 if (flags & HAMMER_FLUSH_SIGNAL) {
1673                         ip->flags |= HAMMER_INODE_RESIGNAL;
1674                         hammer_flusher_async(ip->hmp, flg);
1675                 }
1676                 break;
1677         }
1678 }
1679
1680 /*
1681  * Scan ip->target_list, which is a list of records owned by PARENTS to our
1682  * ip which reference our ip.
1683  *
1684  * XXX This is a huge mess of recursive code, but not one bit of it blocks
1685  *     so for now do not ref/deref the structures.  Note that if we use the
1686  *     ref/rel code later, the rel CAN block.
1687  */
1688 static int
1689 hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
1690                            hammer_flush_group_t flg)
1691 {
1692         hammer_record_t depend;
1693         int good;
1694         int r;
1695
1696         /*
1697          * If we hit our recursion limit and we have parent dependencies
1698          * We cannot continue.  Returning < 0 will cause us to be flagged
1699          * for reflush.  Returning -2 cuts off additional dependency checks
1700          * because they are likely to also hit the depth limit.
1701          *
1702          * We cannot return < 0 if there are no dependencies or there might
1703          * not be anything to wakeup (ip).
1704          */
1705         if (depth == 20 && TAILQ_FIRST(&ip->target_list)) {
1706                 kprintf("HAMMER Warning: depth limit reached on "
1707                         "setup recursion, inode %p %016llx\n",
1708                         ip, (long long)ip->obj_id);
1709                 return(-2);
1710         }
1711
1712         /*
1713          * Scan dependencies
1714          */
1715         good = 0;
1716         TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1717                 r = hammer_setup_parent_inodes_helper(depend, depth, flg);
1718                 KKASSERT(depend->target_ip == ip);
1719                 if (r < 0 && good == 0)
1720                         good = -1;
1721                 if (r > 0)
1722                         good = 1;
1723
1724                 /*
1725                  * If we failed due to the recursion depth limit then stop
1726                  * now.
1727                  */
1728                 if (r == -2)
1729                         break;
1730         }
1731         return(good);
1732 }
1733
1734 /*
1735  * This helper function takes a record representing the dependancy between
1736  * the parent inode and child inode.
1737  *
1738  * record->ip           = parent inode
1739  * record->target_ip    = child inode
1740  * 
1741  * We are asked to recurse upwards and convert the record from SETUP
1742  * to FLUSH if possible.
1743  *
1744  * Return 1 if the record gives us connectivity
1745  *
1746  * Return 0 if the record is not relevant 
1747  *
1748  * Return -1 if we can't resolve the dependancy and there is no connectivity.
1749  */
1750 static int
1751 hammer_setup_parent_inodes_helper(hammer_record_t record, int depth,
1752                                   hammer_flush_group_t flg)
1753 {
1754         hammer_mount_t hmp;
1755         hammer_inode_t pip;
1756         int good;
1757
1758         KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1759         pip = record->ip;
1760         hmp = pip->hmp;
1761
1762         /*
1763          * If the record is already flushing, is it in our flush group?
1764          *
1765          * If it is in our flush group but it is a general record or a 
1766          * delete-on-disk, it does not improve our connectivity (return 0),
1767          * and if the target inode is not trying to destroy itself we can't
1768          * allow the operation yet anyway (the second return -1).
1769          */
1770         if (record->flush_state == HAMMER_FST_FLUSH) {
1771                 /*
1772                  * If not in our flush group ask the parent to reflush
1773                  * us as soon as possible.
1774                  */
1775                 if (record->flush_group != flg) {
1776                         pip->flags |= HAMMER_INODE_REFLUSH;
1777                         record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1778                         return(-1);
1779                 }
1780
1781                 /*
1782                  * If in our flush group everything is already set up,
1783                  * just return whether the record will improve our
1784                  * visibility or not.
1785                  */
1786                 if (record->type == HAMMER_MEM_RECORD_ADD)
1787                         return(1);
1788                 return(0);
1789         }
1790
1791         /*
1792          * It must be a setup record.  Try to resolve the setup dependancies
1793          * by recursing upwards so we can place ip on the flush list.
1794          *
1795          * Limit ourselves to 20 levels of recursion to avoid blowing out
1796          * the kernel stack.  If we hit the recursion limit we can't flush
1797          * until the parent flushes.  The parent will flush independantly
1798          * on its own and ultimately a deep recursion will be resolved.
1799          */
1800         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1801
1802         good = hammer_setup_parent_inodes(pip, depth + 1, flg);
1803
1804         /*
1805          * If good < 0 the parent has no connectivity and we cannot safely
1806          * flush the directory entry, which also means we can't flush our
1807          * ip.  Flag us for downward recursion once the parent's
1808          * connectivity is resolved.  Flag the parent for [re]flush or it
1809          * may not check for downward recursions.
1810          */
1811         if (good < 0) {
1812                 pip->flags |= HAMMER_INODE_REFLUSH;
1813                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1814                 return(good);
1815         }
1816
1817         /*
1818          * We are go, place the parent inode in a flushing state so we can
1819          * place its record in a flushing state.  Note that the parent
1820          * may already be flushing.  The record must be in the same flush
1821          * group as the parent.
1822          */
1823         if (pip->flush_state != HAMMER_FST_FLUSH)
1824                 hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
1825         KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
1826         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1827
1828 #if 0
1829         if (record->type == HAMMER_MEM_RECORD_DEL &&
1830             (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
1831                 /*
1832                  * Regardless of flushing state we cannot sync this path if the
1833                  * record represents a delete-on-disk but the target inode
1834                  * is not ready to sync its own deletion.
1835                  *
1836                  * XXX need to count effective nlinks to determine whether
1837                  * the flush is ok, otherwise removing a hardlink will
1838                  * just leave the DEL record to rot.
1839                  */
1840                 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
1841                 return(-1);
1842         } else
1843 #endif
1844         if (pip->flush_group == flg) {
1845                 /*
1846                  * Because we have not calculated nlinks yet we can just
1847                  * set records to the flush state if the parent is in
1848                  * the same flush group as we are.
1849                  */
1850                 record->flush_state = HAMMER_FST_FLUSH;
1851                 record->flush_group = flg;
1852                 ++record->flush_group->refs;
1853                 hammer_ref(&record->lock);
1854
1855                 /*
1856                  * A general directory-add contributes to our visibility.
1857                  *
1858                  * Otherwise it is probably a directory-delete or 
1859                  * delete-on-disk record and does not contribute to our
1860                  * visbility (but we can still flush it).
1861                  */
1862                 if (record->type == HAMMER_MEM_RECORD_ADD)
1863                         return(1);
1864                 return(0);
1865         } else {
1866                 /*
1867                  * If the parent is not in our flush group we cannot
1868                  * flush this record yet, there is no visibility.
1869                  * We tell the parent to reflush and mark ourselves
1870                  * so the parent knows it should flush us too.
1871                  */
1872                 pip->flags |= HAMMER_INODE_REFLUSH;
1873                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1874                 return(-1);
1875         }
1876 }
1877
1878 /*
1879  * This is the core routine placing an inode into the FST_FLUSH state.
1880  */
1881 static void
1882 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
1883 {
1884         int go_count;
1885
1886         /*
1887          * Set flush state and prevent the flusher from cycling into
1888          * the next flush group.  Do not place the ip on the list yet.
1889          * Inodes not in the idle state get an extra reference.
1890          */
1891         KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
1892         if (ip->flush_state == HAMMER_FST_IDLE)
1893                 hammer_ref(&ip->lock);
1894         ip->flush_state = HAMMER_FST_FLUSH;
1895         ip->flush_group = flg;
1896         ++ip->hmp->flusher.group_lock;
1897         ++ip->hmp->count_iqueued;
1898         ++hammer_count_iqueued;
1899         ++flg->total_count;
1900
1901         /*
1902          * If the flush group reaches the autoflush limit we want to signal
1903          * the flusher.  This is particularly important for remove()s.
1904          */
1905         if (flg->total_count == hammer_autoflush)
1906                 flags |= HAMMER_FLUSH_SIGNAL;
1907
1908         /*
1909          * We need to be able to vfsync/truncate from the backend.
1910          */
1911         KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
1912         if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
1913                 ip->flags |= HAMMER_INODE_VHELD;
1914                 vref(ip->vp);
1915         }
1916
1917         /*
1918          * Figure out how many in-memory records we can actually flush
1919          * (not including inode meta-data, buffers, etc).
1920          */
1921         KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
1922         if (flags & HAMMER_FLUSH_RECURSION) {
1923                 /*
1924                  * If this is a upwards recursion we do not want to
1925                  * recurse down again!
1926                  */
1927                 go_count = 1;
1928 #if 0
1929         } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
1930                 /*
1931                  * No new records are added if we must complete a flush
1932                  * from a previous cycle, but we do have to move the records
1933                  * from the previous cycle to the current one.
1934                  */
1935 #if 0
1936                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1937                                    hammer_syncgrp_child_callback, NULL);
1938 #endif
1939                 go_count = 1;
1940 #endif
1941         } else {
1942                 /*
1943                  * Normal flush, scan records and bring them into the flush.
1944                  * Directory adds and deletes are usually skipped (they are
1945                  * grouped with the related inode rather then with the
1946                  * directory).
1947                  *
1948                  * go_count can be negative, which means the scan aborted
1949                  * due to the flush group being over-full and we should
1950                  * flush what we have.
1951                  */
1952                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
1953                                    hammer_setup_child_callback, NULL);
1954         }
1955
1956         /*
1957          * This is a more involved test that includes go_count.  If we
1958          * can't flush, flag the inode and return.  If go_count is 0 we
1959          * were are unable to flush any records in our rec_tree and
1960          * must ignore the XDIRTY flag.
1961          */
1962         if (go_count == 0) {
1963                 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
1964                         --ip->hmp->count_iqueued;
1965                         --hammer_count_iqueued;
1966
1967                         --flg->total_count;
1968                         ip->flush_state = HAMMER_FST_SETUP;
1969                         ip->flush_group = NULL;
1970                         if (ip->flags & HAMMER_INODE_VHELD) {
1971                                 ip->flags &= ~HAMMER_INODE_VHELD;
1972                                 vrele(ip->vp);
1973                         }
1974
1975                         /*
1976                          * REFLUSH is needed to trigger dependancy wakeups
1977                          * when an inode is in SETUP.
1978                          */
1979                         ip->flags |= HAMMER_INODE_REFLUSH;
1980                         if (flags & HAMMER_FLUSH_SIGNAL) {
1981                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1982                                 hammer_flusher_async(ip->hmp, flg);
1983                         }
1984                         if (--ip->hmp->flusher.group_lock == 0)
1985                                 wakeup(&ip->hmp->flusher.group_lock);
1986                         return;
1987                 }
1988         }
1989
1990         /*
1991          * Snapshot the state of the inode for the backend flusher.
1992          *
1993          * We continue to retain save_trunc_off even when all truncations
1994          * have been resolved as an optimization to determine if we can
1995          * skip the B-Tree lookup for overwrite deletions.
1996          *
1997          * NOTE: The DELETING flag is a mod flag, but it is also sticky,
1998          * and stays in ip->flags.  Once set, it stays set until the
1999          * inode is destroyed.
2000          */
2001         if (ip->flags & HAMMER_INODE_TRUNCATED) {
2002                 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
2003                 ip->sync_trunc_off = ip->trunc_off;
2004                 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
2005                 ip->flags &= ~HAMMER_INODE_TRUNCATED;
2006                 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
2007
2008                 /*
2009                  * The save_trunc_off used to cache whether the B-Tree
2010                  * holds any records past that point is not used until
2011                  * after the truncation has succeeded, so we can safely
2012                  * set it now.
2013                  */
2014                 if (ip->save_trunc_off > ip->sync_trunc_off)
2015                         ip->save_trunc_off = ip->sync_trunc_off;
2016         }
2017         ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
2018                            ~HAMMER_INODE_TRUNCATED);
2019         ip->sync_ino_leaf = ip->ino_leaf;
2020         ip->sync_ino_data = ip->ino_data;
2021         ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
2022 #ifdef DEBUG_TRUNCATE
2023         if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
2024                 kprintf("truncateS %016llx\n", ip->sync_trunc_off);
2025 #endif
2026
2027         /*
2028          * The flusher list inherits our inode and reference.
2029          */
2030         KKASSERT(flg->running == 0);
2031         TAILQ_INSERT_TAIL(&flg->flush_list, ip, flush_entry);
2032         if (--ip->hmp->flusher.group_lock == 0)
2033                 wakeup(&ip->hmp->flusher.group_lock);
2034
2035         if (flags & HAMMER_FLUSH_SIGNAL) {
2036                 hammer_flusher_async(ip->hmp, flg);
2037         }
2038 }
2039
2040 /*
2041  * Callback for scan of ip->rec_tree.  Try to include each record in our
2042  * flush.  ip->flush_group has been set but the inode has not yet been
2043  * moved into a flushing state.
2044  *
2045  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
2046  * both inodes.
2047  *
2048  * We return 1 for any record placed or found in FST_FLUSH, which prevents
2049  * the caller from shortcutting the flush.
2050  */
2051 static int
2052 hammer_setup_child_callback(hammer_record_t rec, void *data)
2053 {
2054         hammer_flush_group_t flg;
2055         hammer_inode_t target_ip;
2056         hammer_inode_t ip;
2057         int r;
2058
2059         /*
2060          * Records deleted or committed by the backend are ignored.
2061          * Note that the flush detects deleted frontend records at
2062          * multiple points to deal with races.  This is just the first
2063          * line of defense.  The only time HAMMER_RECF_DELETED_FE cannot
2064          * be set is when HAMMER_RECF_INTERLOCK_BE is set, because it
2065          * messes up link-count calculations.
2066          *
2067          * NOTE: Don't get confused between record deletion and, say,
2068          * directory entry deletion.  The deletion of a directory entry
2069          * which is on-media has nothing to do with the record deletion
2070          * flags.
2071          */
2072         if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
2073                           HAMMER_RECF_COMMITTED)) {
2074                 if (rec->flush_state == HAMMER_FST_FLUSH) {
2075                         KKASSERT(rec->flush_group == rec->ip->flush_group);
2076                         r = 1;
2077                 } else {
2078                         r = 0;
2079                 }
2080                 return(r);
2081         }
2082
2083         /*
2084          * If the record is in an idle state it has no dependancies and
2085          * can be flushed.
2086          */
2087         ip = rec->ip;
2088         flg = ip->flush_group;
2089         r = 0;
2090
2091         switch(rec->flush_state) {
2092         case HAMMER_FST_IDLE:
2093                 /*
2094                  * The record has no setup dependancy, we can flush it.
2095                  */
2096                 KKASSERT(rec->target_ip == NULL);
2097                 rec->flush_state = HAMMER_FST_FLUSH;
2098                 rec->flush_group = flg;
2099                 ++flg->refs;
2100                 hammer_ref(&rec->lock);
2101                 r = 1;
2102                 break;
2103         case HAMMER_FST_SETUP:
2104                 /*
2105                  * The record has a setup dependancy.  These are typically
2106                  * directory entry adds and deletes.  Such entries will be
2107                  * flushed when their inodes are flushed so we do not
2108                  * usually have to add them to the flush here.  However,
2109                  * if the target_ip has set HAMMER_INODE_CONN_DOWN then
2110                  * it is asking us to flush this record (and it).
2111                  */
2112                 target_ip = rec->target_ip;
2113                 KKASSERT(target_ip != NULL);
2114                 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
2115
2116                 /*
2117                  * If the target IP is already flushing in our group
2118                  * we could associate the record, but target_ip has
2119                  * already synced ino_data to sync_ino_data and we
2120                  * would also have to adjust nlinks.   Plus there are
2121                  * ordering issues for adds and deletes.
2122                  *
2123                  * Reflush downward if this is an ADD, and upward if
2124                  * this is a DEL.
2125                  */
2126                 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
2127                         if (rec->flush_state == HAMMER_MEM_RECORD_ADD)
2128                                 ip->flags |= HAMMER_INODE_REFLUSH;
2129                         else
2130                                 target_ip->flags |= HAMMER_INODE_REFLUSH;
2131                         break;
2132                 } 
2133
2134                 /*
2135                  * Target IP is not yet flushing.  This can get complex
2136                  * because we have to be careful about the recursion.
2137                  *
2138                  * Directories create an issue for us in that if a flush
2139                  * of a directory is requested the expectation is to flush
2140                  * any pending directory entries, but this will cause the
2141                  * related inodes to recursively flush as well.  We can't
2142                  * really defer the operation so just get as many as we
2143                  * can and
2144                  */
2145 #if 0
2146                 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
2147                     (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
2148                         /*
2149                          * We aren't reclaiming and the target ip was not
2150                          * previously prevented from flushing due to this
2151                          * record dependancy.  Do not flush this record.
2152                          */
2153                         /*r = 0;*/
2154                 } else
2155 #endif
2156                 if (flg->total_count + flg->refs >
2157                            ip->hmp->undo_rec_limit) {
2158                         /*
2159                          * Our flush group is over-full and we risk blowing
2160                          * out the UNDO FIFO.  Stop the scan, flush what we
2161                          * have, then reflush the directory.
2162                          *
2163                          * The directory may be forced through multiple
2164                          * flush groups before it can be completely
2165                          * flushed.
2166                          */
2167                         ip->flags |= HAMMER_INODE_RESIGNAL |
2168                                      HAMMER_INODE_REFLUSH;
2169                         r = -1;
2170                 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
2171                         /*
2172                          * If the target IP is not flushing we can force
2173                          * it to flush, even if it is unable to write out
2174                          * any of its own records we have at least one in
2175                          * hand that we CAN deal with.
2176                          */
2177                         rec->flush_state = HAMMER_FST_FLUSH;
2178                         rec->flush_group = flg;
2179                         ++flg->refs;
2180                         hammer_ref(&rec->lock);
2181                         hammer_flush_inode_core(target_ip, flg,
2182                                                 HAMMER_FLUSH_RECURSION);
2183                         r = 1;
2184                 } else {
2185                         /*
2186                          * General or delete-on-disk record.
2187                          *
2188                          * XXX this needs help.  If a delete-on-disk we could
2189                          * disconnect the target.  If the target has its own
2190                          * dependancies they really need to be flushed.
2191                          *
2192                          * XXX
2193                          */
2194                         rec->flush_state = HAMMER_FST_FLUSH;
2195                         rec->flush_group = flg;
2196                         ++flg->refs;
2197                         hammer_ref(&rec->lock);
2198                         hammer_flush_inode_core(target_ip, flg,
2199                                                 HAMMER_FLUSH_RECURSION);
2200                         r = 1;
2201                 }
2202                 break;
2203         case HAMMER_FST_FLUSH:
2204                 /* 
2205                  * The flush_group should already match.
2206                  */
2207                 KKASSERT(rec->flush_group == flg);
2208                 r = 1;
2209                 break;
2210         }
2211         return(r);
2212 }
2213
2214 #if 0
2215 /*
2216  * This version just moves records already in a flush state to the new
2217  * flush group and that is it.
2218  */
2219 static int
2220 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
2221 {
2222         hammer_inode_t ip = rec->ip;
2223
2224         switch(rec->flush_state) {
2225         case HAMMER_FST_FLUSH:
2226                 KKASSERT(rec->flush_group == ip->flush_group);
2227                 break;
2228         default:
2229                 break;
2230         }
2231         return(0);
2232 }
2233 #endif
2234
2235 /*
2236  * Wait for a previously queued flush to complete.
2237  *
2238  * If a critical error occured we don't try to wait.
2239  */
2240 void
2241 hammer_wait_inode(hammer_inode_t ip)
2242 {
2243         hammer_flush_group_t flg;
2244
2245         flg = NULL;
2246         if ((ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2247                 while (ip->flush_state != HAMMER_FST_IDLE &&
2248                        (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
2249                         if (ip->flush_state == HAMMER_FST_SETUP)
2250                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2251                         if (ip->flush_state != HAMMER_FST_IDLE) {
2252                                 ip->flags |= HAMMER_INODE_FLUSHW;
2253                                 tsleep(&ip->flags, 0, "hmrwin", 0);
2254                         }
2255                 }
2256         }
2257 }
2258
2259 /*
2260  * Called by the backend code when a flush has been completed.
2261  * The inode has already been removed from the flush list.
2262  *
2263  * A pipelined flush can occur, in which case we must re-enter the
2264  * inode on the list and re-copy its fields.
2265  */
2266 void
2267 hammer_flush_inode_done(hammer_inode_t ip, int error)
2268 {
2269         hammer_mount_t hmp;
2270         int dorel;
2271
2272         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2273
2274         hmp = ip->hmp;
2275
2276         /*
2277          * Auto-reflush if the backend could not completely flush
2278          * the inode.  This fixes a case where a deferred buffer flush
2279          * could cause fsync to return early.
2280          */
2281         if (ip->sync_flags & HAMMER_INODE_MODMASK)
2282                 ip->flags |= HAMMER_INODE_REFLUSH;
2283
2284         /*
2285          * Merge left-over flags back into the frontend and fix the state.
2286          * Incomplete truncations are retained by the backend.
2287          */
2288         ip->error = error;
2289         ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2290         ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2291
2292         /*
2293          * The backend may have adjusted nlinks, so if the adjusted nlinks
2294          * does not match the fronttend set the frontend's RDIRTY flag again.
2295          */
2296         if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2297                 ip->flags |= HAMMER_INODE_DDIRTY;
2298
2299         /*
2300          * Fix up the dirty buffer status.
2301          */
2302         if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2303                 ip->flags |= HAMMER_INODE_BUFS;
2304         }
2305
2306         /*
2307          * Re-set the XDIRTY flag if some of the inode's in-memory records
2308          * could not be flushed.
2309          */
2310         KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2311                   (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2312                  (!RB_EMPTY(&ip->rec_tree) &&
2313                   (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2314
2315         /*
2316          * Do not lose track of inodes which no longer have vnode
2317          * assocations, otherwise they may never get flushed again.
2318          *
2319          * The reflush flag can be set superfluously, causing extra pain
2320          * for no reason.  If the inode is no longer modified it no longer
2321          * needs to be flushed.
2322          */
2323         if (ip->flags & HAMMER_INODE_MODMASK) {
2324                 if (ip->vp == NULL)
2325                         ip->flags |= HAMMER_INODE_REFLUSH;
2326         } else {
2327                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2328         }
2329
2330         /*
2331          * Adjust the flush state.
2332          */
2333         if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2334                 /*
2335                  * We were unable to flush out all our records, leave the
2336                  * inode in a flush state and in the current flush group.
2337                  * The flush group will be re-run.
2338                  *
2339                  * This occurs if the UNDO block gets too full or there is
2340                  * too much dirty meta-data and allows the flusher to
2341                  * finalize the UNDO block and then re-flush.
2342                  */
2343                 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2344                 dorel = 0;
2345         } else {
2346                 /*
2347                  * Remove from the flush_group
2348                  */
2349                 TAILQ_REMOVE(&ip->flush_group->flush_list, ip, flush_entry);
2350                 ip->flush_group = NULL;
2351
2352                 /*
2353                  * Clean up the vnode ref and tracking counts.
2354                  */
2355                 if (ip->flags & HAMMER_INODE_VHELD) {
2356                         ip->flags &= ~HAMMER_INODE_VHELD;
2357                         vrele(ip->vp);
2358                 }
2359                 --hmp->count_iqueued;
2360                 --hammer_count_iqueued;
2361
2362                 /*
2363                  * And adjust the state.
2364                  */
2365                 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2366                         ip->flush_state = HAMMER_FST_IDLE;
2367                         dorel = 1;
2368                 } else {
2369                         ip->flush_state = HAMMER_FST_SETUP;
2370                         dorel = 0;
2371                 }
2372
2373                 /*
2374                  * If the frontend is waiting for a flush to complete,
2375                  * wake it up.
2376                  */
2377                 if (ip->flags & HAMMER_INODE_FLUSHW) {
2378                         ip->flags &= ~HAMMER_INODE_FLUSHW;
2379                         wakeup(&ip->flags);
2380                 }
2381
2382                 /*
2383                  * If the frontend made more changes and requested another
2384                  * flush, then try to get it running.
2385                  *
2386                  * Reflushes are aborted when the inode is errored out.
2387                  */
2388                 if (ip->flags & HAMMER_INODE_REFLUSH) {
2389                         ip->flags &= ~HAMMER_INODE_REFLUSH;
2390                         if (ip->flags & HAMMER_INODE_RESIGNAL) {
2391                                 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2392                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2393                         } else {
2394                                 hammer_flush_inode(ip, 0);
2395                         }
2396                 }
2397         }
2398
2399         /*
2400          * If we have no parent dependancies we can clear CONN_DOWN
2401          */
2402         if (TAILQ_EMPTY(&ip->target_list))
2403                 ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2404
2405         /*
2406          * If the inode is now clean drop the space reservation.
2407          */
2408         if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2409             (ip->flags & HAMMER_INODE_RSV_INODES)) {
2410                 ip->flags &= ~HAMMER_INODE_RSV_INODES;
2411                 --hmp->rsv_inodes;
2412         }
2413
2414         if (dorel)
2415                 hammer_rel_inode(ip, 0);
2416 }
2417
2418 /*
2419  * Called from hammer_sync_inode() to synchronize in-memory records
2420  * to the media.
2421  */
2422 static int
2423 hammer_sync_record_callback(hammer_record_t record, void *data)
2424 {
2425         hammer_cursor_t cursor = data;
2426         hammer_transaction_t trans = cursor->trans;
2427         hammer_mount_t hmp = trans->hmp;
2428         int error;
2429
2430         /*
2431          * Skip records that do not belong to the current flush.
2432          */
2433         ++hammer_stats_record_iterations;
2434         if (record->flush_state != HAMMER_FST_FLUSH)
2435                 return(0);
2436
2437 #if 1
2438         if (record->flush_group != record->ip->flush_group) {
2439                 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2440                 Debugger("blah2");
2441                 return(0);
2442         }
2443 #endif
2444         KKASSERT(record->flush_group == record->ip->flush_group);
2445
2446         /*
2447          * Interlock the record using the BE flag.  Once BE is set the
2448          * frontend cannot change the state of FE.
2449          *
2450          * NOTE: If FE is set prior to us setting BE we still sync the
2451          * record out, but the flush completion code converts it to 
2452          * a delete-on-disk record instead of destroying it.
2453          */
2454         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2455         record->flags |= HAMMER_RECF_INTERLOCK_BE;
2456
2457         /*
2458          * The backend has already disposed of the record.
2459          */
2460         if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
2461                 error = 0;
2462                 goto done;
2463         }
2464
2465         /*
2466          * If the whole inode is being deleting all on-disk records will
2467          * be deleted very soon, we can't sync any new records to disk
2468          * because they will be deleted in the same transaction they were
2469          * created in (delete_tid == create_tid), which will assert.
2470          *
2471          * XXX There may be a case with RECORD_ADD with DELETED_FE set
2472          * that we currently panic on.
2473          */
2474         if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2475                 switch(record->type) {
2476                 case HAMMER_MEM_RECORD_DATA:
2477                         /*
2478                          * We don't have to do anything, if the record was
2479                          * committed the space will have been accounted for
2480                          * in the blockmap.
2481                          */
2482                         /* fall through */
2483                 case HAMMER_MEM_RECORD_GENERAL:
2484                         /*
2485                          * Set deleted-by-backend flag.  Do not set the
2486                          * backend committed flag, because we are throwing
2487                          * the record away.
2488                          */
2489                         record->flags |= HAMMER_RECF_DELETED_BE;
2490                         ++record->ip->rec_generation;
2491                         error = 0;
2492                         goto done;
2493                 case HAMMER_MEM_RECORD_ADD:
2494                         panic("hammer_sync_record_callback: illegal add "
2495                               "during inode deletion record %p", record);
2496                         break; /* NOT REACHED */
2497                 case HAMMER_MEM_RECORD_INODE:
2498                         panic("hammer_sync_record_callback: attempt to "
2499                               "sync inode record %p?", record);
2500                         break; /* NOT REACHED */
2501                 case HAMMER_MEM_RECORD_DEL:
2502                         /* 
2503                          * Follow through and issue the on-disk deletion
2504                          */
2505                         break;
2506                 }
2507         }
2508
2509         /*
2510          * If DELETED_FE is set special handling is needed for directory
2511          * entries.  Dependant pieces related to the directory entry may
2512          * have already been synced to disk.  If this occurs we have to
2513          * sync the directory entry and then change the in-memory record
2514          * from an ADD to a DELETE to cover the fact that it's been
2515          * deleted by the frontend.
2516          *
2517          * A directory delete covering record (MEM_RECORD_DEL) can never
2518          * be deleted by the frontend.
2519          *
2520          * Any other record type (aka DATA) can be deleted by the frontend.
2521          * XXX At the moment the flusher must skip it because there may
2522          * be another data record in the flush group for the same block,
2523          * meaning that some frontend data changes can leak into the backend's
2524          * synchronization point.
2525          */
2526         if (record->flags & HAMMER_RECF_DELETED_FE) {
2527                 if (record->type == HAMMER_MEM_RECORD_ADD) {
2528                         /*
2529                          * Convert a front-end deleted directory-add to
2530                          * a directory-delete entry later.
2531                          */
2532                         record->flags |= HAMMER_RECF_CONVERT_DELETE;
2533                 } else {
2534                         /*
2535                          * Dispose of the record (race case).  Mark as
2536                          * deleted by backend (and not committed).
2537                          */
2538                         KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2539                         record->flags |= HAMMER_RECF_DELETED_BE;
2540                         ++record->ip->rec_generation;
2541                         error = 0;
2542                         goto done;
2543                 }
2544         }
2545
2546         /*
2547          * Assign the create_tid for new records.  Deletions already
2548          * have the record's entire key properly set up.
2549          */
2550         if (record->type != HAMMER_MEM_RECORD_DEL) {
2551                 record->leaf.base.create_tid = trans->tid;
2552                 record->leaf.create_ts = trans->time32;
2553         }
2554         for (;;) {
2555                 error = hammer_ip_sync_record_cursor(cursor, record);
2556                 if (error != EDEADLK)
2557                         break;
2558                 hammer_done_cursor(cursor);
2559                 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2560                                            record->ip);
2561                 if (error)
2562                         break;
2563         }
2564         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2565
2566         if (error)
2567                 error = -error;
2568 done:
2569         hammer_flush_record_done(record, error);
2570
2571         /*
2572          * Do partial finalization if we have built up too many dirty
2573          * buffers.  Otherwise a buffer cache deadlock can occur when
2574          * doing things like creating tens of thousands of tiny files.
2575          *
2576          * We must release our cursor lock to avoid a 3-way deadlock
2577          * due to the exclusive sync lock the finalizer must get.
2578          */
2579         if (hammer_flusher_meta_limit(hmp)) {
2580                 hammer_unlock_cursor(cursor);
2581                 hammer_flusher_finalize(trans, 0);
2582                 hammer_lock_cursor(cursor);
2583         }
2584
2585         return(error);
2586 }
2587
2588 /*
2589  * Backend function called by the flusher to sync an inode to media.
2590  */
2591 int
2592 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2593 {
2594         struct hammer_cursor cursor;
2595         hammer_node_t tmp_node;
2596         hammer_record_t depend;
2597         hammer_record_t next;
2598         int error, tmp_error;
2599         u_int64_t nlinks;
2600
2601         if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2602                 return(0);
2603
2604         error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2605         if (error)
2606                 goto done;
2607
2608         /*
2609          * Any directory records referencing this inode which are not in
2610          * our current flush group must adjust our nlink count for the
2611          * purposes of synchronization to disk.
2612          *
2613          * Records which are in our flush group can be unlinked from our
2614          * inode now, potentially allowing the inode to be physically
2615          * deleted.
2616          *
2617          * This cannot block.
2618          */
2619         nlinks = ip->ino_data.nlinks;
2620         next = TAILQ_FIRST(&ip->target_list);
2621         while ((depend = next) != NULL) {
2622                 next = TAILQ_NEXT(depend, target_entry);
2623                 if (depend->flush_state == HAMMER_FST_FLUSH &&
2624                     depend->flush_group == ip->flush_group) {
2625                         /*
2626                          * If this is an ADD that was deleted by the frontend
2627                          * the frontend nlinks count will have already been
2628                          * decremented, but the backend is going to sync its
2629                          * directory entry and must account for it.  The
2630                          * record will be converted to a delete-on-disk when
2631                          * it gets synced.
2632                          *
2633                          * If the ADD was not deleted by the frontend we
2634                          * can remove the dependancy from our target_list.
2635                          */
2636                         if (depend->flags & HAMMER_RECF_DELETED_FE) {
2637                                 ++nlinks;
2638                         } else {
2639                                 TAILQ_REMOVE(&ip->target_list, depend,
2640                                              target_entry);
2641                                 depend->target_ip = NULL;
2642                         }
2643                 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2644                         /*
2645                          * Not part of our flush group and not deleted by
2646                          * the front-end, adjust the link count synced to
2647                          * the media (undo what the frontend did when it
2648                          * queued the record).
2649                          */
2650                         KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2651                         switch(depend->type) {
2652                         case HAMMER_MEM_RECORD_ADD:
2653                                 --nlinks;
2654                                 break;
2655                         case HAMMER_MEM_RECORD_DEL:
2656                                 ++nlinks;
2657                                 break;
2658                         default:
2659                                 break;
2660                         }
2661                 }
2662         }
2663
2664         /*
2665          * Set dirty if we had to modify the link count.
2666          */
2667         if (ip->sync_ino_data.nlinks != nlinks) {
2668                 KKASSERT((int64_t)nlinks >= 0);
2669                 ip->sync_ino_data.nlinks = nlinks;
2670                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2671         }
2672
2673         /*
2674          * If there is a trunction queued destroy any data past the (aligned)
2675          * truncation point.  Userland will have dealt with the buffer
2676          * containing the truncation point for us.
2677          *
2678          * We don't flush pending frontend data buffers until after we've
2679          * dealt with the truncation.
2680          */
2681         if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2682                 /*
2683                  * Interlock trunc_off.  The VOP front-end may continue to
2684                  * make adjustments to it while we are blocked.
2685                  */
2686                 off_t trunc_off;
2687                 off_t aligned_trunc_off;
2688                 int blkmask;
2689
2690                 trunc_off = ip->sync_trunc_off;
2691                 blkmask = hammer_blocksize(trunc_off) - 1;
2692                 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
2693
2694                 /*
2695                  * Delete any whole blocks on-media.  The front-end has
2696                  * already cleaned out any partial block and made it
2697                  * pending.  The front-end may have updated trunc_off
2698                  * while we were blocked so we only use sync_trunc_off.
2699                  *
2700                  * This operation can blow out the buffer cache, EWOULDBLOCK
2701                  * means we were unable to complete the deletion.  The
2702                  * deletion will update sync_trunc_off in that case.
2703                  */
2704                 error = hammer_ip_delete_range(&cursor, ip,
2705                                                 aligned_trunc_off,
2706                                                 0x7FFFFFFFFFFFFFFFLL, 2);
2707                 if (error == EWOULDBLOCK) {
2708                         ip->flags |= HAMMER_INODE_WOULDBLOCK;
2709                         error = 0;
2710                         goto defer_buffer_flush;
2711                 }
2712
2713                 if (error)
2714                         goto done;
2715
2716                 /*
2717                  * Clear the truncation flag on the backend after we have
2718                  * complete the deletions.  Backend data is now good again
2719                  * (including new records we are about to sync, below).
2720                  *
2721                  * Leave sync_trunc_off intact.  As we write additional
2722                  * records the backend will update sync_trunc_off.  This
2723                  * tells the backend whether it can skip the overwrite
2724                  * test.  This should work properly even when the backend
2725                  * writes full blocks where the truncation point straddles
2726                  * the block because the comparison is against the base
2727                  * offset of the record.
2728                  */
2729                 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2730                 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2731         } else {
2732                 error = 0;
2733         }
2734
2735         /*
2736          * Now sync related records.  These will typically be directory
2737          * entries, records tracking direct-writes, or delete-on-disk records.
2738          */
2739         if (error == 0) {
2740                 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2741                                     hammer_sync_record_callback, &cursor);
2742                 if (tmp_error < 0)
2743                         tmp_error = -error;
2744                 if (tmp_error)
2745                         error = tmp_error;
2746         }
2747         hammer_cache_node(&ip->cache[1], cursor.node);
2748
2749         /*
2750          * Re-seek for inode update, assuming our cache hasn't been ripped
2751          * out from under us.
2752          */
2753         if (error == 0) {
2754                 tmp_node = hammer_ref_node_safe(trans, &ip->cache[0], &error);
2755                 if (tmp_node) {
2756                         hammer_cursor_downgrade(&cursor);
2757                         hammer_lock_sh(&tmp_node->lock);
2758                         if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
2759                                 hammer_cursor_seek(&cursor, tmp_node, 0);
2760                         hammer_unlock(&tmp_node->lock);
2761                         hammer_rel_node(tmp_node);
2762                 }
2763                 error = 0;
2764         }
2765
2766         /*
2767          * If we are deleting the inode the frontend had better not have
2768          * any active references on elements making up the inode.
2769          *
2770          * The call to hammer_ip_delete_clean() cleans up auxillary records
2771          * but not DB or DATA records.  Those must have already been deleted
2772          * by the normal truncation mechanic.
2773          */
2774         if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
2775                 RB_EMPTY(&ip->rec_tree)  &&
2776             (ip->sync_flags & HAMMER_INODE_DELETING) &&
2777             (ip->flags & HAMMER_INODE_DELETED) == 0) {
2778                 int count1 = 0;
2779
2780                 error = hammer_ip_delete_clean(&cursor, ip, &count1);
2781                 if (error == 0) {
2782                         ip->flags |= HAMMER_INODE_DELETED;
2783                         ip->sync_flags &= ~HAMMER_INODE_DELETING;
2784                         ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2785                         KKASSERT(RB_EMPTY(&ip->rec_tree));
2786
2787                         /*
2788                          * Set delete_tid in both the frontend and backend
2789                          * copy of the inode record.  The DELETED flag handles
2790                          * this, do not set RDIRTY.
2791                          */
2792                         ip->ino_leaf.base.delete_tid = trans->tid;
2793                         ip->sync_ino_leaf.base.delete_tid = trans->tid;
2794                         ip->ino_leaf.delete_ts = trans->time32;
2795                         ip->sync_ino_leaf.delete_ts = trans->time32;
2796
2797
2798                         /*
2799                          * Adjust the inode count in the volume header
2800                          */
2801                         hammer_sync_lock_sh(trans);
2802                         if (ip->flags & HAMMER_INODE_ONDISK) {
2803                                 hammer_modify_volume_field(trans,
2804                                                            trans->rootvol,
2805                                                            vol0_stat_inodes);
2806                                 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
2807                                 hammer_modify_volume_done(trans->rootvol);
2808                         }
2809                         hammer_sync_unlock(trans);
2810                 }
2811         }
2812
2813         if (error)
2814                 goto done;
2815         ip->sync_flags &= ~HAMMER_INODE_BUFS;
2816
2817 defer_buffer_flush:
2818         /*
2819          * Now update the inode's on-disk inode-data and/or on-disk record.
2820          * DELETED and ONDISK are managed only in ip->flags.
2821          *
2822          * In the case of a defered buffer flush we still update the on-disk
2823          * inode to satisfy visibility requirements if there happen to be
2824          * directory dependancies.
2825          */
2826         switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
2827         case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
2828                 /*
2829                  * If deleted and on-disk, don't set any additional flags.
2830                  * the delete flag takes care of things.
2831                  *
2832                  * Clear flags which may have been set by the frontend.
2833                  */
2834                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2835                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2836                                     HAMMER_INODE_DELETING);
2837                 break;
2838         case HAMMER_INODE_DELETED:
2839                 /*
2840                  * Take care of the case where a deleted inode was never
2841                  * flushed to the disk in the first place.
2842                  *
2843                  * Clear flags which may have been set by the frontend.
2844                  */
2845                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
2846                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
2847                                     HAMMER_INODE_DELETING);
2848                 while (RB_ROOT(&ip->rec_tree)) {
2849                         hammer_record_t record = RB_ROOT(&ip->rec_tree);
2850                         hammer_ref(&record->lock);
2851                         KKASSERT(record->lock.refs == 1);
2852                         record->flags |= HAMMER_RECF_DELETED_BE;
2853                         ++record->ip->rec_generation;
2854                         hammer_rel_mem_record(record);
2855                 }
2856                 break;
2857         case HAMMER_INODE_ONDISK:
2858                 /*
2859                  * If already on-disk, do not set any additional flags.
2860                  */
2861                 break;
2862         default:
2863                 /*
2864                  * If not on-disk and not deleted, set DDIRTY to force
2865                  * an initial record to be written.
2866                  *
2867                  * Also set the create_tid in both the frontend and backend
2868                  * copy of the inode record.
2869                  */
2870                 ip->ino_leaf.base.create_tid = trans->tid;
2871                 ip->ino_leaf.create_ts = trans->time32;
2872                 ip->sync_ino_leaf.base.create_tid = trans->tid;
2873                 ip->sync_ino_leaf.create_ts = trans->time32;
2874                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2875                 break;
2876         }
2877
2878         /*
2879          * If RDIRTY or DDIRTY is set, write out a new record.  If the inode
2880          * is already on-disk the old record is marked as deleted.
2881          *
2882          * If DELETED is set hammer_update_inode() will delete the existing
2883          * record without writing out a new one.
2884          *
2885          * If *ONLY* the ITIMES flag is set we can update the record in-place.
2886          */
2887         if (ip->flags & HAMMER_INODE_DELETED) {
2888                 error = hammer_update_inode(&cursor, ip);
2889         } else 
2890         if ((ip->sync_flags & HAMMER_INODE_DDIRTY) == 0 &&
2891             (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
2892                 error = hammer_update_itimes(&cursor, ip);
2893         } else
2894         if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
2895                 error = hammer_update_inode(&cursor, ip);
2896         }
2897 done:
2898         if (error) {
2899                 hammer_critical_error(ip->hmp, ip, error,
2900                                       "while syncing inode");
2901         }
2902         hammer_done_cursor(&cursor);
2903         return(error);
2904 }
2905
2906 /*
2907  * This routine is called when the OS is no longer actively referencing
2908  * the inode (but might still be keeping it cached), or when releasing
2909  * the last reference to an inode.
2910  *
2911  * At this point if the inode's nlinks count is zero we want to destroy
2912  * it, which may mean destroying it on-media too.
2913  */
2914 void
2915 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
2916 {
2917         struct vnode *vp;
2918
2919         /*
2920          * Set the DELETING flag when the link count drops to 0 and the
2921          * OS no longer has any opens on the inode.
2922          *
2923          * The backend will clear DELETING (a mod flag) and set DELETED
2924          * (a state flag) when it is actually able to perform the
2925          * operation.
2926          *
2927          * Don't reflag the deletion if the flusher is currently syncing
2928          * one that was already flagged.  A previously set DELETING flag
2929          * may bounce around flags and sync_flags until the operation is
2930          * completely done.
2931          */
2932         if (ip->ino_data.nlinks == 0 &&
2933             ((ip->flags | ip->sync_flags) & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
2934                 ip->flags |= HAMMER_INODE_DELETING;
2935                 ip->flags |= HAMMER_INODE_TRUNCATED;
2936                 ip->trunc_off = 0;
2937                 vp = NULL;
2938                 if (getvp) {
2939                         if (hammer_get_vnode(ip, &vp) != 0)
2940                                 return;
2941                 }
2942
2943                 /*
2944                  * Final cleanup
2945                  */
2946                 if (ip->vp) {
2947                         vtruncbuf(ip->vp, 0, HAMMER_BUFSIZE);
2948                         vnode_pager_setsize(ip->vp, 0);
2949                 }
2950                 if (getvp) {
2951                         vput(vp);
2952                 }
2953         }
2954 }
2955
2956 /*
2957  * After potentially resolving a dependancy the inode is tested
2958  * to determine whether it needs to be reflushed.
2959  */
2960 void
2961 hammer_test_inode(hammer_inode_t ip)
2962 {
2963         if (ip->flags & HAMMER_INODE_REFLUSH) {
2964                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2965                 hammer_ref(&ip->lock);
2966                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2967                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
2968                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2969                 } else {
2970                         hammer_flush_inode(ip, 0);
2971                 }
2972                 hammer_rel_inode(ip, 0);
2973         }
2974 }
2975
2976 /*
2977  * Clear the RECLAIM flag on an inode.  This occurs when the inode is
2978  * reassociated with a vp or just before it gets freed.
2979  *
2980  * Pipeline wakeups to threads blocked due to an excessive number of
2981  * detached inodes.  The reclaim count generates a bit of negative
2982  * feedback.
2983  */
2984 static void
2985 hammer_inode_wakereclaims(hammer_inode_t ip, int dowake)
2986 {
2987         struct hammer_reclaim *reclaim;
2988         hammer_mount_t hmp = ip->hmp;
2989
2990         if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
2991                 return;
2992
2993         --hammer_count_reclaiming;
2994         --hmp->inode_reclaims;
2995         ip->flags &= ~HAMMER_INODE_RECLAIM;
2996
2997         if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT || dowake) {
2998                 reclaim = TAILQ_FIRST(&hmp->reclaim_list);
2999                 if (reclaim && reclaim->count > 0 && --reclaim->count == 0) {
3000                         TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
3001                         wakeup(reclaim);
3002                 }
3003         }
3004 }
3005
3006 /*
3007  * Setup our reclaim pipeline.  We only let so many detached (and dirty)
3008  * inodes build up before we start blocking.
3009  *
3010  * When we block we don't care *which* inode has finished reclaiming,
3011  * as lone as one does.  This is somewhat heuristical... we also put a
3012  * cap on how long we are willing to wait.
3013  */
3014 void
3015 hammer_inode_waitreclaims(hammer_mount_t hmp)
3016 {
3017         struct hammer_reclaim reclaim;
3018         int delay;
3019
3020         if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT)
3021                 return;
3022         delay = (hmp->inode_reclaims - HAMMER_RECLAIM_WAIT) * hz /
3023                 (HAMMER_RECLAIM_WAIT * 3) + 1;
3024         if (delay > 0) {
3025                 reclaim.count = 2;
3026                 TAILQ_INSERT_TAIL(&hmp->reclaim_list, &reclaim, entry);
3027                 tsleep(&reclaim, 0, "hmrrcm", delay);
3028                 if (reclaim.count > 0)
3029                         TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
3030         }
3031 }
3032
3033 /*
3034  * A larger then normal backlog of inodes is sitting in the flusher,
3035  * enforce a general slowdown to let it catch up.  This routine is only
3036  * called on completion of a non-flusher-related transaction which
3037  * performed B-Tree node I/O.
3038  *
3039  * It is possible for the flusher to stall in a continuous load.
3040  * blogbench -i1000 -o seems to do a good job generating this sort of load.
3041  * If the flusher is unable to catch up the inode count can bloat until
3042  * we run out of kvm.
3043  *
3044  * This is a bit of a hack.
3045  */
3046 void
3047 hammer_inode_waithard(hammer_mount_t hmp)
3048 {
3049         /*
3050          * Hysteresis.
3051          */
3052         if (hmp->flags & HAMMER_MOUNT_FLUSH_RECOVERY) {
3053                 if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT / 2 &&
3054                     hmp->count_iqueued < hmp->count_inodes / 20) {
3055                         hmp->flags &= ~HAMMER_MOUNT_FLUSH_RECOVERY;
3056                         return;
3057                 }
3058         } else {
3059                 if (hmp->inode_reclaims < HAMMER_RECLAIM_WAIT ||
3060                     hmp->count_iqueued < hmp->count_inodes / 10) {
3061                         return;
3062                 }
3063                 hmp->flags |= HAMMER_MOUNT_FLUSH_RECOVERY;
3064         }
3065
3066         /*
3067          * Block for one flush cycle.
3068          */
3069         hammer_flusher_wait_next(hmp);
3070 }
3071