Pre-2.0 release: Sync with HAMMER 64 - NFS and cross-device link fixes.
[dragonfly.git] / sys / vfs / hammer / hammer_vnops.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_vnops.c,v 1.91.2.2 2008/07/19 04:51:09 dillon Exp $
35  */
36
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
40 #include <sys/fcntl.h>
41 #include <sys/namecache.h>
42 #include <sys/vnode.h>
43 #include <sys/lockf.h>
44 #include <sys/event.h>
45 #include <sys/stat.h>
46 #include <sys/dirent.h>
47 #include <vm/vm_extern.h>
48 #include <vfs/fifofs/fifo.h>
49 #include "hammer.h"
50
51 /*
52  * USERFS VNOPS
53  */
54 /*static int hammer_vop_vnoperate(struct vop_generic_args *);*/
55 static int hammer_vop_fsync(struct vop_fsync_args *);
56 static int hammer_vop_read(struct vop_read_args *);
57 static int hammer_vop_write(struct vop_write_args *);
58 static int hammer_vop_access(struct vop_access_args *);
59 static int hammer_vop_advlock(struct vop_advlock_args *);
60 static int hammer_vop_close(struct vop_close_args *);
61 static int hammer_vop_ncreate(struct vop_ncreate_args *);
62 static int hammer_vop_getattr(struct vop_getattr_args *);
63 static int hammer_vop_nresolve(struct vop_nresolve_args *);
64 static int hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *);
65 static int hammer_vop_nlink(struct vop_nlink_args *);
66 static int hammer_vop_nmkdir(struct vop_nmkdir_args *);
67 static int hammer_vop_nmknod(struct vop_nmknod_args *);
68 static int hammer_vop_open(struct vop_open_args *);
69 static int hammer_vop_pathconf(struct vop_pathconf_args *);
70 static int hammer_vop_print(struct vop_print_args *);
71 static int hammer_vop_readdir(struct vop_readdir_args *);
72 static int hammer_vop_readlink(struct vop_readlink_args *);
73 static int hammer_vop_nremove(struct vop_nremove_args *);
74 static int hammer_vop_nrename(struct vop_nrename_args *);
75 static int hammer_vop_nrmdir(struct vop_nrmdir_args *);
76 static int hammer_vop_setattr(struct vop_setattr_args *);
77 static int hammer_vop_strategy(struct vop_strategy_args *);
78 static int hammer_vop_bmap(struct vop_bmap_args *ap);
79 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
80 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
81 static int hammer_vop_ioctl(struct vop_ioctl_args *);
82 static int hammer_vop_mountctl(struct vop_mountctl_args *);
83
84 static int hammer_vop_fifoclose (struct vop_close_args *);
85 static int hammer_vop_fiforead (struct vop_read_args *);
86 static int hammer_vop_fifowrite (struct vop_write_args *);
87
88 static int hammer_vop_specclose (struct vop_close_args *);
89 static int hammer_vop_specread (struct vop_read_args *);
90 static int hammer_vop_specwrite (struct vop_write_args *);
91
92 struct vop_ops hammer_vnode_vops = {
93         .vop_default =          vop_defaultop,
94         .vop_fsync =            hammer_vop_fsync,
95         .vop_getpages =         vop_stdgetpages,
96         .vop_putpages =         vop_stdputpages,
97         .vop_read =             hammer_vop_read,
98         .vop_write =            hammer_vop_write,
99         .vop_access =           hammer_vop_access,
100         .vop_advlock =          hammer_vop_advlock,
101         .vop_close =            hammer_vop_close,
102         .vop_ncreate =          hammer_vop_ncreate,
103         .vop_getattr =          hammer_vop_getattr,
104         .vop_inactive =         hammer_vop_inactive,
105         .vop_reclaim =          hammer_vop_reclaim,
106         .vop_nresolve =         hammer_vop_nresolve,
107         .vop_nlookupdotdot =    hammer_vop_nlookupdotdot,
108         .vop_nlink =            hammer_vop_nlink,
109         .vop_nmkdir =           hammer_vop_nmkdir,
110         .vop_nmknod =           hammer_vop_nmknod,
111         .vop_open =             hammer_vop_open,
112         .vop_pathconf =         hammer_vop_pathconf,
113         .vop_print =            hammer_vop_print,
114         .vop_readdir =          hammer_vop_readdir,
115         .vop_readlink =         hammer_vop_readlink,
116         .vop_nremove =          hammer_vop_nremove,
117         .vop_nrename =          hammer_vop_nrename,
118         .vop_nrmdir =           hammer_vop_nrmdir,
119         .vop_setattr =          hammer_vop_setattr,
120         .vop_bmap =             hammer_vop_bmap,
121         .vop_strategy =         hammer_vop_strategy,
122         .vop_nsymlink =         hammer_vop_nsymlink,
123         .vop_nwhiteout =        hammer_vop_nwhiteout,
124         .vop_ioctl =            hammer_vop_ioctl,
125         .vop_mountctl =         hammer_vop_mountctl
126 };
127
128 struct vop_ops hammer_spec_vops = {
129         .vop_default =          spec_vnoperate,
130         .vop_fsync =            hammer_vop_fsync,
131         .vop_read =             hammer_vop_specread,
132         .vop_write =            hammer_vop_specwrite,
133         .vop_access =           hammer_vop_access,
134         .vop_close =            hammer_vop_specclose,
135         .vop_getattr =          hammer_vop_getattr,
136         .vop_inactive =         hammer_vop_inactive,
137         .vop_reclaim =          hammer_vop_reclaim,
138         .vop_setattr =          hammer_vop_setattr
139 };
140
141 struct vop_ops hammer_fifo_vops = {
142         .vop_default =          fifo_vnoperate,
143         .vop_fsync =            hammer_vop_fsync,
144         .vop_read =             hammer_vop_fiforead,
145         .vop_write =            hammer_vop_fifowrite,
146         .vop_access =           hammer_vop_access,
147         .vop_close =            hammer_vop_fifoclose,
148         .vop_getattr =          hammer_vop_getattr,
149         .vop_inactive =         hammer_vop_inactive,
150         .vop_reclaim =          hammer_vop_reclaim,
151         .vop_setattr =          hammer_vop_setattr
152 };
153
154 #ifdef DEBUG_TRUNCATE
155 struct hammer_inode *HammerTruncIp;
156 #endif
157
158 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
159                            struct vnode *dvp, struct ucred *cred, int flags);
160 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
161 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
162
163 #if 0
164 static
165 int
166 hammer_vop_vnoperate(struct vop_generic_args *)
167 {
168         return (VOCALL(&hammer_vnode_vops, ap));
169 }
170 #endif
171
172 /*
173  * hammer_vop_fsync { vp, waitfor }
174  *
175  * fsync() an inode to disk and wait for it to be completely committed
176  * such that the information would not be undone if a crash occured after
177  * return.
178  */
179 static
180 int
181 hammer_vop_fsync(struct vop_fsync_args *ap)
182 {
183         hammer_inode_t ip = VTOI(ap->a_vp);
184
185         ++hammer_count_fsyncs;
186         vfsync(ap->a_vp, ap->a_waitfor, 1, NULL, NULL);
187         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
188         if (ap->a_waitfor == MNT_WAIT)
189                 hammer_wait_inode(ip);
190         return (ip->error);
191 }
192
193 /*
194  * hammer_vop_read { vp, uio, ioflag, cred }
195  */
196 static
197 int
198 hammer_vop_read(struct vop_read_args *ap)
199 {
200         struct hammer_transaction trans;
201         hammer_inode_t ip;
202         off_t offset;
203         struct buf *bp;
204         struct uio *uio;
205         int error;
206         int n;
207         int seqcount;
208         int ioseqcount;
209         int blksize;
210
211         if (ap->a_vp->v_type != VREG)
212                 return (EINVAL);
213         ip = VTOI(ap->a_vp);
214         error = 0;
215         uio = ap->a_uio;
216
217         /*
218          * Allow the UIO's size to override the sequential heuristic.
219          */
220         blksize = hammer_blocksize(uio->uio_offset);
221         seqcount = (uio->uio_resid + (blksize - 1)) / blksize;
222         ioseqcount = ap->a_ioflag >> 16;
223         if (seqcount < ioseqcount)
224                 seqcount = ioseqcount;
225
226         hammer_start_transaction(&trans, ip->hmp);
227
228         /*
229          * Access the data typically in HAMMER_BUFSIZE blocks via the
230          * buffer cache, but HAMMER may use a variable block size based
231          * on the offset.
232          */
233         while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
234                 int64_t base_offset;
235                 int64_t file_limit;
236
237                 blksize = hammer_blocksize(uio->uio_offset);
238                 offset = (int)uio->uio_offset & (blksize - 1);
239                 base_offset = uio->uio_offset - offset;
240
241                 if (hammer_cluster_enable) {
242                         /*
243                          * Use file_limit to prevent cluster_read() from
244                          * creating buffers of the wrong block size past
245                          * the demarc.
246                          */
247                         file_limit = ip->ino_data.size;
248                         if (base_offset < HAMMER_XDEMARC &&
249                             file_limit > HAMMER_XDEMARC) {
250                                 file_limit = HAMMER_XDEMARC;
251                         }
252                         error = cluster_read(ap->a_vp,
253                                              file_limit, base_offset,
254                                              blksize, MAXPHYS,
255                                              seqcount, &bp);
256                 } else {
257                         error = bread(ap->a_vp, base_offset, blksize, &bp);
258                 }
259                 if (error) {
260                         kprintf("error %d\n", error);
261                         brelse(bp);
262                         break;
263                 }
264
265                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
266                 n = blksize - offset;
267                 if (n > uio->uio_resid)
268                         n = uio->uio_resid;
269                 if (n > ip->ino_data.size - uio->uio_offset)
270                         n = (int)(ip->ino_data.size - uio->uio_offset);
271                 error = uiomove((char *)bp->b_data + offset, n, uio);
272
273                 /* data has a lower priority then meta-data */
274                 bp->b_flags |= B_AGE;
275                 bqrelse(bp);
276                 if (error)
277                         break;
278                 hammer_stats_file_read += n;
279         }
280         if ((ip->flags & HAMMER_INODE_RO) == 0 &&
281             (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
282                 ip->ino_data.atime = trans.time;
283                 hammer_modify_inode(ip, HAMMER_INODE_ATIME);
284         }
285         hammer_done_transaction(&trans);
286         return (error);
287 }
288
289 /*
290  * hammer_vop_write { vp, uio, ioflag, cred }
291  */
292 static
293 int
294 hammer_vop_write(struct vop_write_args *ap)
295 {
296         struct hammer_transaction trans;
297         struct hammer_inode *ip;
298         hammer_mount_t hmp;
299         struct uio *uio;
300         int offset;
301         off_t base_offset;
302         struct buf *bp;
303         int error;
304         int n;
305         int flags;
306         int delta;
307         int seqcount;
308
309         if (ap->a_vp->v_type != VREG)
310                 return (EINVAL);
311         ip = VTOI(ap->a_vp);
312         hmp = ip->hmp;
313         error = 0;
314         seqcount = ap->a_ioflag >> 16;
315
316         if (ip->flags & HAMMER_INODE_RO)
317                 return (EROFS);
318
319         /*
320          * Create a transaction to cover the operations we perform.
321          */
322         hammer_start_transaction(&trans, hmp);
323         uio = ap->a_uio;
324
325         /*
326          * Check append mode
327          */
328         if (ap->a_ioflag & IO_APPEND)
329                 uio->uio_offset = ip->ino_data.size;
330
331         /*
332          * Check for illegal write offsets.  Valid range is 0...2^63-1.
333          *
334          * NOTE: the base_off assignment is required to work around what
335          * I consider to be a GCC-4 optimization bug.
336          */
337         if (uio->uio_offset < 0) {
338                 hammer_done_transaction(&trans);
339                 return (EFBIG);
340         }
341         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
342         if (uio->uio_resid > 0 && base_offset <= 0) {
343                 hammer_done_transaction(&trans);
344                 return (EFBIG);
345         }
346
347         /*
348          * Access the data typically in HAMMER_BUFSIZE blocks via the
349          * buffer cache, but HAMMER may use a variable block size based
350          * on the offset.
351          */
352         while (uio->uio_resid > 0) {
353                 int fixsize = 0;
354                 int blksize;
355                 int blkmask;
356
357                 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
358                         break;
359
360                 blksize = hammer_blocksize(uio->uio_offset);
361
362                 /*
363                  * Do not allow HAMMER to blow out the buffer cache.  Very
364                  * large UIOs can lockout other processes due to bwillwrite()
365                  * mechanics.
366                  *
367                  * The hammer inode is not locked during these operations.
368                  * The vnode is locked which can interfere with the pageout
369                  * daemon for non-UIO_NOCOPY writes but should not interfere
370                  * with the buffer cache.  Even so, we cannot afford to
371                  * allow the pageout daemon to build up too many dirty buffer
372                  * cache buffers.
373                  */
374                 /*if (((int)uio->uio_offset & (blksize - 1)) == 0)*/
375                 bwillwrite(blksize);
376
377                 /*
378                  * Do not allow HAMMER to blow out system memory by
379                  * accumulating too many records.   Records are so well
380                  * decoupled from the buffer cache that it is possible
381                  * for userland to push data out to the media via
382                  * direct-write, but build up the records queued to the
383                  * backend faster then the backend can flush them out.
384                  * HAMMER has hit its write limit but the frontend has
385                  * no pushback to slow it down.
386                  */
387                 if (hmp->rsv_recs > hammer_limit_recs / 2) {
388                         /*
389                          * Get the inode on the flush list
390                          */
391                         if (ip->rsv_recs >= 64)
392                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
393                         else if (ip->rsv_recs >= 16)
394                                 hammer_flush_inode(ip, 0);
395
396                         /*
397                          * Keep the flusher going if the system keeps
398                          * queueing records.
399                          */
400                         delta = hmp->count_newrecords -
401                                 hmp->last_newrecords;
402                         if (delta < 0 || delta > hammer_limit_recs / 2) {
403                                 hmp->last_newrecords = hmp->count_newrecords;
404                                 hammer_sync_hmp(hmp, MNT_NOWAIT);
405                         }
406
407                         /*
408                          * If we have gotten behind start slowing
409                          * down the writers.
410                          */
411                         delta = (hmp->rsv_recs - hammer_limit_recs) *
412                                 hz / hammer_limit_recs;
413                         if (delta > 0)
414                                 tsleep(&trans, 0, "hmrslo", delta);
415                 }
416
417                 /*
418                  * Calculate the blocksize at the current offset and figure
419                  * out how much we can actually write.
420                  */
421                 blkmask = blksize - 1;
422                 offset = (int)uio->uio_offset & blkmask;
423                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
424                 n = blksize - offset;
425                 if (n > uio->uio_resid)
426                         n = uio->uio_resid;
427                 if (uio->uio_offset + n > ip->ino_data.size) {
428                         vnode_pager_setsize(ap->a_vp, uio->uio_offset + n);
429                         fixsize = 1;
430                 }
431
432                 if (uio->uio_segflg == UIO_NOCOPY) {
433                         /*
434                          * Issuing a write with the same data backing the
435                          * buffer.  Instantiate the buffer to collect the
436                          * backing vm pages, then read-in any missing bits.
437                          *
438                          * This case is used by vop_stdputpages().
439                          */
440                         bp = getblk(ap->a_vp, base_offset,
441                                     blksize, GETBLK_BHEAVY, 0);
442                         if ((bp->b_flags & B_CACHE) == 0) {
443                                 bqrelse(bp);
444                                 error = bread(ap->a_vp, base_offset,
445                                               blksize, &bp);
446                         }
447                 } else if (offset == 0 && uio->uio_resid >= blksize) {
448                         /*
449                          * Even though we are entirely overwriting the buffer
450                          * we may still have to zero it out to avoid a 
451                          * mmap/write visibility issue.
452                          */
453                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
454                         if ((bp->b_flags & B_CACHE) == 0)
455                                 vfs_bio_clrbuf(bp);
456                 } else if (base_offset >= ip->ino_data.size) {
457                         /*
458                          * If the base offset of the buffer is beyond the
459                          * file EOF, we don't have to issue a read.
460                          */
461                         bp = getblk(ap->a_vp, base_offset,
462                                     blksize, GETBLK_BHEAVY, 0);
463                         vfs_bio_clrbuf(bp);
464                 } else {
465                         /*
466                          * Partial overwrite, read in any missing bits then
467                          * replace the portion being written.
468                          */
469                         error = bread(ap->a_vp, base_offset, blksize, &bp);
470                         if (error == 0)
471                                 bheavy(bp);
472                 }
473                 if (error == 0) {
474                         error = uiomove((char *)bp->b_data + offset,
475                                         n, uio);
476                 }
477
478                 /*
479                  * If we screwed up we have to undo any VM size changes we
480                  * made.
481                  */
482                 if (error) {
483                         brelse(bp);
484                         if (fixsize) {
485                                 vtruncbuf(ap->a_vp, ip->ino_data.size,
486                                           hammer_blocksize(ip->ino_data.size));
487                         }
488                         break;
489                 }
490                 hammer_stats_file_write += n;
491                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
492                 if (ip->ino_data.size < uio->uio_offset) {
493                         ip->ino_data.size = uio->uio_offset;
494                         flags = HAMMER_INODE_DDIRTY;
495                         vnode_pager_setsize(ap->a_vp, ip->ino_data.size);
496                 } else {
497                         flags = 0;
498                 }
499                 ip->ino_data.mtime = trans.time;
500                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
501                 hammer_modify_inode(ip, flags);
502
503                 /*
504                  * Once we dirty the buffer any cached zone-X offset
505                  * becomes invalid.  HAMMER NOTE: no-history mode cannot 
506                  * allow overwriting over the same data sector unless
507                  * we provide UNDOs for the old data, which we don't.
508                  */
509                 bp->b_bio2.bio_offset = NOOFFSET;
510
511                 /*
512                  * Final buffer disposition.
513                  */
514                 bp->b_flags |= B_AGE;
515                 if (ap->a_ioflag & IO_SYNC) {
516                         bwrite(bp);
517                 } else if (ap->a_ioflag & IO_DIRECT) {
518                         bawrite(bp);
519                 } else {
520                         bdwrite(bp);
521                 }
522         }
523         hammer_done_transaction(&trans);
524         return (error);
525 }
526
527 /*
528  * hammer_vop_access { vp, mode, cred }
529  */
530 static
531 int
532 hammer_vop_access(struct vop_access_args *ap)
533 {
534         struct hammer_inode *ip = VTOI(ap->a_vp);
535         uid_t uid;
536         gid_t gid;
537         int error;
538
539         ++hammer_stats_file_iopsr;
540         uid = hammer_to_unix_xid(&ip->ino_data.uid);
541         gid = hammer_to_unix_xid(&ip->ino_data.gid);
542
543         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
544                                   ip->ino_data.uflags);
545         return (error);
546 }
547
548 /*
549  * hammer_vop_advlock { vp, id, op, fl, flags }
550  */
551 static
552 int
553 hammer_vop_advlock(struct vop_advlock_args *ap)
554 {
555         hammer_inode_t ip = VTOI(ap->a_vp);
556
557         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
558 }
559
560 /*
561  * hammer_vop_close { vp, fflag }
562  */
563 static
564 int
565 hammer_vop_close(struct vop_close_args *ap)
566 {
567         hammer_inode_t ip = VTOI(ap->a_vp);
568
569         if ((ip->flags | ip->sync_flags) & HAMMER_INODE_MODMASK)
570                 hammer_inode_waitreclaims(ip->hmp);
571         return (vop_stdclose(ap));
572 }
573
574 /*
575  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
576  *
577  * The operating system has already ensured that the directory entry
578  * does not exist and done all appropriate namespace locking.
579  */
580 static
581 int
582 hammer_vop_ncreate(struct vop_ncreate_args *ap)
583 {
584         struct hammer_transaction trans;
585         struct hammer_inode *dip;
586         struct hammer_inode *nip;
587         struct nchandle *nch;
588         int error;
589
590         nch = ap->a_nch;
591         dip = VTOI(ap->a_dvp);
592
593         if (dip->flags & HAMMER_INODE_RO)
594                 return (EROFS);
595         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
596                 return (error);
597
598         /*
599          * Create a transaction to cover the operations we perform.
600          */
601         hammer_start_transaction(&trans, dip->hmp);
602         ++hammer_stats_file_iopsw;
603
604         /*
605          * Create a new filesystem object of the requested type.  The
606          * returned inode will be referenced and shared-locked to prevent
607          * it from being moved to the flusher.
608          */
609
610         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
611                                     dip, NULL, &nip);
612         if (error) {
613                 hkprintf("hammer_create_inode error %d\n", error);
614                 hammer_done_transaction(&trans);
615                 *ap->a_vpp = NULL;
616                 return (error);
617         }
618
619         /*
620          * Add the new filesystem object to the directory.  This will also
621          * bump the inode's link count.
622          */
623         error = hammer_ip_add_directory(&trans, dip,
624                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
625                                         nip);
626         if (error)
627                 hkprintf("hammer_ip_add_directory error %d\n", error);
628
629         /*
630          * Finish up.
631          */
632         if (error) {
633                 hammer_rel_inode(nip, 0);
634                 hammer_done_transaction(&trans);
635                 *ap->a_vpp = NULL;
636         } else {
637                 error = hammer_get_vnode(nip, ap->a_vpp);
638                 hammer_done_transaction(&trans);
639                 hammer_rel_inode(nip, 0);
640                 if (error == 0) {
641                         cache_setunresolved(ap->a_nch);
642                         cache_setvp(ap->a_nch, *ap->a_vpp);
643                 }
644         }
645         return (error);
646 }
647
648 /*
649  * hammer_vop_getattr { vp, vap }
650  *
651  * Retrieve an inode's attribute information.  When accessing inodes
652  * historically we fake the atime field to ensure consistent results.
653  * The atime field is stored in the B-Tree element and allowed to be
654  * updated without cycling the element.
655  */
656 static
657 int
658 hammer_vop_getattr(struct vop_getattr_args *ap)
659 {
660         struct hammer_inode *ip = VTOI(ap->a_vp);
661         struct vattr *vap = ap->a_vap;
662
663         /*
664          * We want the fsid to be different when accessing a filesystem
665          * with different as-of's so programs like diff don't think
666          * the files are the same.
667          *
668          * We also want the fsid to be the same when comparing snapshots,
669          * or when comparing mirrors (which might be backed by different
670          * physical devices).  HAMMER fsids are based on the PFS's
671          * shared_uuid field.
672          *
673          * XXX there is a chance of collision here.  The va_fsid reported
674          * by stat is different from the more involved fsid used in the
675          * mount structure.
676          */
677         ++hammer_stats_file_iopsr;
678         vap->va_fsid = ip->pfsm->fsid_udev ^ (u_int32_t)ip->obj_asof ^
679                        (u_int32_t)(ip->obj_asof >> 32);
680
681         vap->va_fileid = ip->ino_leaf.base.obj_id;
682         vap->va_mode = ip->ino_data.mode;
683         vap->va_nlink = ip->ino_data.nlinks;
684         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
685         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
686         vap->va_rmajor = 0;
687         vap->va_rminor = 0;
688         vap->va_size = ip->ino_data.size;
689
690         /*
691          * Special case for @@PFS softlinks.  The actual size of the
692          * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
693          */
694         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
695             ip->ino_data.size == 10 &&
696             ip->obj_asof == HAMMER_MAX_TID &&
697             ip->obj_localization == 0 &&
698             strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
699                     vap->va_size = 26;
700         }
701
702         /*
703          * We must provide a consistent atime and mtime for snapshots
704          * so people can do a 'tar cf - ... | md5' on them and get
705          * consistent results.
706          */
707         if (ip->flags & HAMMER_INODE_RO) {
708                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
709                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
710         } else {
711                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
712                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
713         }
714         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
715         vap->va_flags = ip->ino_data.uflags;
716         vap->va_gen = 1;        /* hammer inums are unique for all time */
717         vap->va_blocksize = HAMMER_BUFSIZE;
718         if (ip->ino_data.size >= HAMMER_XDEMARC) {
719                 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
720                                 ~HAMMER_XBUFMASK64;
721         } else if (ip->ino_data.size > HAMMER_BUFSIZE / 2) {
722                 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
723                                 ~HAMMER_BUFMASK64;
724         } else {
725                 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
726         }
727         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
728         vap->va_filerev = 0;    /* XXX */
729         /* mtime uniquely identifies any adjustments made to the file XXX */
730         vap->va_fsmid = ip->ino_data.mtime;
731         vap->va_uid_uuid = ip->ino_data.uid;
732         vap->va_gid_uuid = ip->ino_data.gid;
733         vap->va_fsid_uuid = ip->hmp->fsid;
734         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
735                           VA_FSID_UUID_VALID;
736
737         switch (ip->ino_data.obj_type) {
738         case HAMMER_OBJTYPE_CDEV:
739         case HAMMER_OBJTYPE_BDEV:
740                 vap->va_rmajor = ip->ino_data.rmajor;
741                 vap->va_rminor = ip->ino_data.rminor;
742                 break;
743         default:
744                 break;
745         }
746         return(0);
747 }
748
749 /*
750  * hammer_vop_nresolve { nch, dvp, cred }
751  *
752  * Locate the requested directory entry.
753  */
754 static
755 int
756 hammer_vop_nresolve(struct vop_nresolve_args *ap)
757 {
758         struct hammer_transaction trans;
759         struct namecache *ncp;
760         hammer_inode_t dip;
761         hammer_inode_t ip;
762         hammer_tid_t asof;
763         struct hammer_cursor cursor;
764         struct vnode *vp;
765         int64_t namekey;
766         int error;
767         int i;
768         int nlen;
769         int flags;
770         int ispfs;
771         int64_t obj_id;
772         u_int32_t localization;
773
774         /*
775          * Misc initialization, plus handle as-of name extensions.  Look for
776          * the '@@' extension.  Note that as-of files and directories cannot
777          * be modified.
778          */
779         dip = VTOI(ap->a_dvp);
780         ncp = ap->a_nch->ncp;
781         asof = dip->obj_asof;
782         nlen = ncp->nc_nlen;
783         flags = dip->flags & HAMMER_INODE_RO;
784         ispfs = 0;
785
786         hammer_simple_transaction(&trans, dip->hmp);
787         ++hammer_stats_file_iopsr;
788
789         for (i = 0; i < nlen; ++i) {
790                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
791                         asof = hammer_str_to_tid(ncp->nc_name + i + 2,
792                                                  &ispfs, &localization);
793                         if (asof != HAMMER_MAX_TID)
794                                 flags |= HAMMER_INODE_RO;
795                         break;
796                 }
797         }
798         nlen = i;
799
800         /*
801          * If this is a PFS softlink we dive into the PFS
802          */
803         if (ispfs && nlen == 0) {
804                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
805                                       asof, localization,
806                                       flags, &error);
807                 if (error == 0) {
808                         error = hammer_get_vnode(ip, &vp);
809                         hammer_rel_inode(ip, 0);
810                 } else {
811                         vp = NULL;
812                 }
813                 if (error == 0) {
814                         vn_unlock(vp);
815                         cache_setvp(ap->a_nch, vp);
816                         vrele(vp);
817                 }
818                 goto done;
819         }
820
821         /*
822          * If there is no path component the time extension is relative to
823          * dip.
824          */
825         if (nlen == 0) {
826                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
827                                       asof, dip->obj_localization,
828                                       flags, &error);
829                 if (error == 0) {
830                         error = hammer_get_vnode(ip, &vp);
831                         hammer_rel_inode(ip, 0);
832                 } else {
833                         vp = NULL;
834                 }
835                 if (error == 0) {
836                         vn_unlock(vp);
837                         cache_setvp(ap->a_nch, vp);
838                         vrele(vp);
839                 }
840                 goto done;
841         }
842
843         /*
844          * Calculate the namekey and setup the key range for the scan.  This
845          * works kinda like a chained hash table where the lower 32 bits
846          * of the namekey synthesize the chain.
847          *
848          * The key range is inclusive of both key_beg and key_end.
849          */
850         namekey = hammer_directory_namekey(ncp->nc_name, nlen);
851
852         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
853         cursor.key_beg.localization = dip->obj_localization +
854                                       HAMMER_LOCALIZE_MISC;
855         cursor.key_beg.obj_id = dip->obj_id;
856         cursor.key_beg.key = namekey;
857         cursor.key_beg.create_tid = 0;
858         cursor.key_beg.delete_tid = 0;
859         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
860         cursor.key_beg.obj_type = 0;
861
862         cursor.key_end = cursor.key_beg;
863         cursor.key_end.key |= 0xFFFFFFFFULL;
864         cursor.asof = asof;
865         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
866
867         /*
868          * Scan all matching records (the chain), locate the one matching
869          * the requested path component.
870          *
871          * The hammer_ip_*() functions merge in-memory records with on-disk
872          * records for the purposes of the search.
873          */
874         obj_id = 0;
875         localization = HAMMER_DEF_LOCALIZATION;
876
877         if (error == 0) {
878                 error = hammer_ip_first(&cursor);
879                 while (error == 0) {
880                         error = hammer_ip_resolve_data(&cursor);
881                         if (error)
882                                 break;
883                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
884                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
885                                 obj_id = cursor.data->entry.obj_id;
886                                 localization = cursor.data->entry.localization;
887                                 break;
888                         }
889                         error = hammer_ip_next(&cursor);
890                 }
891         }
892         hammer_done_cursor(&cursor);
893         if (error == 0) {
894                 ip = hammer_get_inode(&trans, dip, obj_id,
895                                       asof, localization,
896                                       flags, &error);
897                 if (error == 0) {
898                         error = hammer_get_vnode(ip, &vp);
899                         hammer_rel_inode(ip, 0);
900                 } else {
901                         vp = NULL;
902                 }
903                 if (error == 0) {
904                         vn_unlock(vp);
905                         cache_setvp(ap->a_nch, vp);
906                         vrele(vp);
907                 }
908         } else if (error == ENOENT) {
909                 cache_setvp(ap->a_nch, NULL);
910         }
911 done:
912         hammer_done_transaction(&trans);
913         return (error);
914 }
915
916 /*
917  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
918  *
919  * Locate the parent directory of a directory vnode.
920  *
921  * dvp is referenced but not locked.  *vpp must be returned referenced and
922  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
923  * at the root, instead it could indicate that the directory we were in was
924  * removed.
925  *
926  * NOTE: as-of sequences are not linked into the directory structure.  If
927  * we are at the root with a different asof then the mount point, reload
928  * the same directory with the mount point's asof.   I'm not sure what this
929  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
930  * get confused, but it hasn't been tested.
931  */
932 static
933 int
934 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
935 {
936         struct hammer_transaction trans;
937         struct hammer_inode *dip;
938         struct hammer_inode *ip;
939         int64_t parent_obj_id;
940         u_int32_t parent_obj_localization;
941         hammer_tid_t asof;
942         int error;
943
944         dip = VTOI(ap->a_dvp);
945         asof = dip->obj_asof;
946
947         /*
948          * Whos are parent?  This could be the root of a pseudo-filesystem
949          * whos parent is in another localization domain.
950          */
951         parent_obj_id = dip->ino_data.parent_obj_id;
952         if (dip->obj_id == HAMMER_OBJID_ROOT)
953                 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
954         else
955                 parent_obj_localization = dip->obj_localization;
956
957         if (parent_obj_id == 0) {
958                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
959                    asof != dip->hmp->asof) {
960                         parent_obj_id = dip->obj_id;
961                         asof = dip->hmp->asof;
962                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
963                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
964                                    dip->obj_asof);
965                 } else {
966                         *ap->a_vpp = NULL;
967                         return ENOENT;
968                 }
969         }
970
971         hammer_simple_transaction(&trans, dip->hmp);
972         ++hammer_stats_file_iopsr;
973
974         ip = hammer_get_inode(&trans, dip, parent_obj_id,
975                               asof, parent_obj_localization,
976                               dip->flags, &error);
977         if (ip) {
978                 error = hammer_get_vnode(ip, ap->a_vpp);
979                 hammer_rel_inode(ip, 0);
980         } else {
981                 *ap->a_vpp = NULL;
982         }
983         hammer_done_transaction(&trans);
984         return (error);
985 }
986
987 /*
988  * hammer_vop_nlink { nch, dvp, vp, cred }
989  */
990 static
991 int
992 hammer_vop_nlink(struct vop_nlink_args *ap)
993 {
994         struct hammer_transaction trans;
995         struct hammer_inode *dip;
996         struct hammer_inode *ip;
997         struct nchandle *nch;
998         int error;
999
1000         if (ap->a_dvp->v_mount != ap->a_vp->v_mount)    
1001                 return(EXDEV);
1002
1003         nch = ap->a_nch;
1004         dip = VTOI(ap->a_dvp);
1005         ip = VTOI(ap->a_vp);
1006
1007         if (dip->obj_localization != ip->obj_localization)
1008                 return(EXDEV);
1009
1010         if (dip->flags & HAMMER_INODE_RO)
1011                 return (EROFS);
1012         if (ip->flags & HAMMER_INODE_RO)
1013                 return (EROFS);
1014         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1015                 return (error);
1016
1017         /*
1018          * Create a transaction to cover the operations we perform.
1019          */
1020         hammer_start_transaction(&trans, dip->hmp);
1021         ++hammer_stats_file_iopsw;
1022
1023         /*
1024          * Add the filesystem object to the directory.  Note that neither
1025          * dip nor ip are referenced or locked, but their vnodes are
1026          * referenced.  This function will bump the inode's link count.
1027          */
1028         error = hammer_ip_add_directory(&trans, dip,
1029                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1030                                         ip);
1031
1032         /*
1033          * Finish up.
1034          */
1035         if (error == 0) {
1036                 cache_setunresolved(nch);
1037                 cache_setvp(nch, ap->a_vp);
1038         }
1039         hammer_done_transaction(&trans);
1040         return (error);
1041 }
1042
1043 /*
1044  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1045  *
1046  * The operating system has already ensured that the directory entry
1047  * does not exist and done all appropriate namespace locking.
1048  */
1049 static
1050 int
1051 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1052 {
1053         struct hammer_transaction trans;
1054         struct hammer_inode *dip;
1055         struct hammer_inode *nip;
1056         struct nchandle *nch;
1057         int error;
1058
1059         nch = ap->a_nch;
1060         dip = VTOI(ap->a_dvp);
1061
1062         if (dip->flags & HAMMER_INODE_RO)
1063                 return (EROFS);
1064         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1065                 return (error);
1066
1067         /*
1068          * Create a transaction to cover the operations we perform.
1069          */
1070         hammer_start_transaction(&trans, dip->hmp);
1071         ++hammer_stats_file_iopsw;
1072
1073         /*
1074          * Create a new filesystem object of the requested type.  The
1075          * returned inode will be referenced but not locked.
1076          */
1077         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1078                                     dip, NULL, &nip);
1079         if (error) {
1080                 hkprintf("hammer_mkdir error %d\n", error);
1081                 hammer_done_transaction(&trans);
1082                 *ap->a_vpp = NULL;
1083                 return (error);
1084         }
1085         /*
1086          * Add the new filesystem object to the directory.  This will also
1087          * bump the inode's link count.
1088          */
1089         error = hammer_ip_add_directory(&trans, dip,
1090                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1091                                         nip);
1092         if (error)
1093                 hkprintf("hammer_mkdir (add) error %d\n", error);
1094
1095         /*
1096          * Finish up.
1097          */
1098         if (error) {
1099                 hammer_rel_inode(nip, 0);
1100                 *ap->a_vpp = NULL;
1101         } else {
1102                 error = hammer_get_vnode(nip, ap->a_vpp);
1103                 hammer_rel_inode(nip, 0);
1104                 if (error == 0) {
1105                         cache_setunresolved(ap->a_nch);
1106                         cache_setvp(ap->a_nch, *ap->a_vpp);
1107                 }
1108         }
1109         hammer_done_transaction(&trans);
1110         return (error);
1111 }
1112
1113 /*
1114  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1115  *
1116  * The operating system has already ensured that the directory entry
1117  * does not exist and done all appropriate namespace locking.
1118  */
1119 static
1120 int
1121 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1122 {
1123         struct hammer_transaction trans;
1124         struct hammer_inode *dip;
1125         struct hammer_inode *nip;
1126         struct nchandle *nch;
1127         int error;
1128
1129         nch = ap->a_nch;
1130         dip = VTOI(ap->a_dvp);
1131
1132         if (dip->flags & HAMMER_INODE_RO)
1133                 return (EROFS);
1134         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1135                 return (error);
1136
1137         /*
1138          * Create a transaction to cover the operations we perform.
1139          */
1140         hammer_start_transaction(&trans, dip->hmp);
1141         ++hammer_stats_file_iopsw;
1142
1143         /*
1144          * Create a new filesystem object of the requested type.  The
1145          * returned inode will be referenced but not locked.
1146          *
1147          * If mknod specifies a directory a pseudo-fs is created.
1148          */
1149         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1150                                     dip, NULL, &nip);
1151         if (error) {
1152                 hammer_done_transaction(&trans);
1153                 *ap->a_vpp = NULL;
1154                 return (error);
1155         }
1156
1157         /*
1158          * Add the new filesystem object to the directory.  This will also
1159          * bump the inode's link count.
1160          */
1161         error = hammer_ip_add_directory(&trans, dip,
1162                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1163                                         nip);
1164
1165         /*
1166          * Finish up.
1167          */
1168         if (error) {
1169                 hammer_rel_inode(nip, 0);
1170                 *ap->a_vpp = NULL;
1171         } else {
1172                 error = hammer_get_vnode(nip, ap->a_vpp);
1173                 hammer_rel_inode(nip, 0);
1174                 if (error == 0) {
1175                         cache_setunresolved(ap->a_nch);
1176                         cache_setvp(ap->a_nch, *ap->a_vpp);
1177                 }
1178         }
1179         hammer_done_transaction(&trans);
1180         return (error);
1181 }
1182
1183 /*
1184  * hammer_vop_open { vp, mode, cred, fp }
1185  */
1186 static
1187 int
1188 hammer_vop_open(struct vop_open_args *ap)
1189 {
1190         hammer_inode_t ip;
1191
1192         ++hammer_stats_file_iopsr;
1193         ip = VTOI(ap->a_vp);
1194
1195         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1196                 return (EROFS);
1197         return(vop_stdopen(ap));
1198 }
1199
1200 /*
1201  * hammer_vop_pathconf { vp, name, retval }
1202  */
1203 static
1204 int
1205 hammer_vop_pathconf(struct vop_pathconf_args *ap)
1206 {
1207         return EOPNOTSUPP;
1208 }
1209
1210 /*
1211  * hammer_vop_print { vp }
1212  */
1213 static
1214 int
1215 hammer_vop_print(struct vop_print_args *ap)
1216 {
1217         return EOPNOTSUPP;
1218 }
1219
1220 /*
1221  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1222  */
1223 static
1224 int
1225 hammer_vop_readdir(struct vop_readdir_args *ap)
1226 {
1227         struct hammer_transaction trans;
1228         struct hammer_cursor cursor;
1229         struct hammer_inode *ip;
1230         struct uio *uio;
1231         hammer_base_elm_t base;
1232         int error;
1233         int cookie_index;
1234         int ncookies;
1235         off_t *cookies;
1236         off_t saveoff;
1237         int r;
1238         int dtype;
1239
1240         ++hammer_stats_file_iopsr;
1241         ip = VTOI(ap->a_vp);
1242         uio = ap->a_uio;
1243         saveoff = uio->uio_offset;
1244
1245         if (ap->a_ncookies) {
1246                 ncookies = uio->uio_resid / 16 + 1;
1247                 if (ncookies > 1024)
1248                         ncookies = 1024;
1249                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1250                 cookie_index = 0;
1251         } else {
1252                 ncookies = -1;
1253                 cookies = NULL;
1254                 cookie_index = 0;
1255         }
1256
1257         hammer_simple_transaction(&trans, ip->hmp);
1258
1259         /*
1260          * Handle artificial entries
1261          */
1262         error = 0;
1263         if (saveoff == 0) {
1264                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1265                 if (r)
1266                         goto done;
1267                 if (cookies)
1268                         cookies[cookie_index] = saveoff;
1269                 ++saveoff;
1270                 ++cookie_index;
1271                 if (cookie_index == ncookies)
1272                         goto done;
1273         }
1274         if (saveoff == 1) {
1275                 if (ip->ino_data.parent_obj_id) {
1276                         r = vop_write_dirent(&error, uio,
1277                                              ip->ino_data.parent_obj_id,
1278                                              DT_DIR, 2, "..");
1279                 } else {
1280                         r = vop_write_dirent(&error, uio,
1281                                              ip->obj_id, DT_DIR, 2, "..");
1282                 }
1283                 if (r)
1284                         goto done;
1285                 if (cookies)
1286                         cookies[cookie_index] = saveoff;
1287                 ++saveoff;
1288                 ++cookie_index;
1289                 if (cookie_index == ncookies)
1290                         goto done;
1291         }
1292
1293         /*
1294          * Key range (begin and end inclusive) to scan.  Directory keys
1295          * directly translate to a 64 bit 'seek' position.
1296          */
1297         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1298         cursor.key_beg.localization = ip->obj_localization +
1299                                       HAMMER_LOCALIZE_MISC;
1300         cursor.key_beg.obj_id = ip->obj_id;
1301         cursor.key_beg.create_tid = 0;
1302         cursor.key_beg.delete_tid = 0;
1303         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1304         cursor.key_beg.obj_type = 0;
1305         cursor.key_beg.key = saveoff;
1306
1307         cursor.key_end = cursor.key_beg;
1308         cursor.key_end.key = HAMMER_MAX_KEY;
1309         cursor.asof = ip->obj_asof;
1310         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1311
1312         error = hammer_ip_first(&cursor);
1313
1314         while (error == 0) {
1315                 error = hammer_ip_resolve_data(&cursor);
1316                 if (error)
1317                         break;
1318                 base = &cursor.leaf->base;
1319                 saveoff = base->key;
1320                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1321
1322                 if (base->obj_id != ip->obj_id)
1323                         panic("readdir: bad record at %p", cursor.node);
1324
1325                 /*
1326                  * Convert pseudo-filesystems into softlinks
1327                  */
1328                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1329                 r = vop_write_dirent(
1330                              &error, uio, cursor.data->entry.obj_id,
1331                              dtype,
1332                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1333                              (void *)cursor.data->entry.name);
1334                 if (r)
1335                         break;
1336                 ++saveoff;
1337                 if (cookies)
1338                         cookies[cookie_index] = base->key;
1339                 ++cookie_index;
1340                 if (cookie_index == ncookies)
1341                         break;
1342                 error = hammer_ip_next(&cursor);
1343         }
1344         hammer_done_cursor(&cursor);
1345
1346 done:
1347         hammer_done_transaction(&trans);
1348
1349         if (ap->a_eofflag)
1350                 *ap->a_eofflag = (error == ENOENT);
1351         uio->uio_offset = saveoff;
1352         if (error && cookie_index == 0) {
1353                 if (error == ENOENT)
1354                         error = 0;
1355                 if (cookies) {
1356                         kfree(cookies, M_TEMP);
1357                         *ap->a_ncookies = 0;
1358                         *ap->a_cookies = NULL;
1359                 }
1360         } else {
1361                 if (error == ENOENT)
1362                         error = 0;
1363                 if (cookies) {
1364                         *ap->a_ncookies = cookie_index;
1365                         *ap->a_cookies = cookies;
1366                 }
1367         }
1368         return(error);
1369 }
1370
1371 /*
1372  * hammer_vop_readlink { vp, uio, cred }
1373  */
1374 static
1375 int
1376 hammer_vop_readlink(struct vop_readlink_args *ap)
1377 {
1378         struct hammer_transaction trans;
1379         struct hammer_cursor cursor;
1380         struct hammer_inode *ip;
1381         char buf[32];
1382         u_int32_t localization;
1383         hammer_pseudofs_inmem_t pfsm;
1384         int error;
1385
1386         ip = VTOI(ap->a_vp);
1387
1388         /*
1389          * Shortcut if the symlink data was stuffed into ino_data.
1390          *
1391          * Also expand special "@@PFS%05d" softlinks (expansion only
1392          * occurs for non-historical (current) accesses made from the
1393          * primary filesystem).
1394          */
1395         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1396                 char *ptr;
1397                 int bytes;
1398
1399                 ptr = ip->ino_data.ext.symlink;
1400                 bytes = (int)ip->ino_data.size;
1401                 if (bytes == 10 &&
1402                     ip->obj_asof == HAMMER_MAX_TID &&
1403                     ip->obj_localization == 0 &&
1404                     strncmp(ptr, "@@PFS", 5) == 0) {
1405                         hammer_simple_transaction(&trans, ip->hmp);
1406                         bcopy(ptr + 5, buf, 5);
1407                         buf[5] = 0;
1408                         localization = strtoul(buf, NULL, 10) << 16;
1409                         pfsm = hammer_load_pseudofs(&trans, localization,
1410                                                     &error);
1411                         if (error == 0) {
1412                                 if (pfsm->pfsd.mirror_flags &
1413                                     HAMMER_PFSD_SLAVE) {
1414                                         ksnprintf(buf, sizeof(buf),
1415                                                   "@@0x%016llx:%05d",
1416                                                   pfsm->pfsd.sync_end_tid,
1417                                                   localization >> 16);
1418                                 } else {
1419                                         ksnprintf(buf, sizeof(buf),
1420                                                   "@@0x%016llx:%05d",
1421                                                   HAMMER_MAX_TID,
1422                                                   localization >> 16);
1423                                 }
1424                                 ptr = buf;
1425                                 bytes = strlen(buf);
1426                         }
1427                         if (pfsm)
1428                                 hammer_rel_pseudofs(trans.hmp, pfsm);
1429                         hammer_done_transaction(&trans);
1430                 }
1431                 error = uiomove(ptr, bytes, ap->a_uio);
1432                 return(error);
1433         }
1434
1435         /*
1436          * Long version
1437          */
1438         hammer_simple_transaction(&trans, ip->hmp);
1439         ++hammer_stats_file_iopsr;
1440         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1441
1442         /*
1443          * Key range (begin and end inclusive) to scan.  Directory keys
1444          * directly translate to a 64 bit 'seek' position.
1445          */
1446         cursor.key_beg.localization = ip->obj_localization +
1447                                       HAMMER_LOCALIZE_MISC;
1448         cursor.key_beg.obj_id = ip->obj_id;
1449         cursor.key_beg.create_tid = 0;
1450         cursor.key_beg.delete_tid = 0;
1451         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1452         cursor.key_beg.obj_type = 0;
1453         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1454         cursor.asof = ip->obj_asof;
1455         cursor.flags |= HAMMER_CURSOR_ASOF;
1456
1457         error = hammer_ip_lookup(&cursor);
1458         if (error == 0) {
1459                 error = hammer_ip_resolve_data(&cursor);
1460                 if (error == 0) {
1461                         KKASSERT(cursor.leaf->data_len >=
1462                                  HAMMER_SYMLINK_NAME_OFF);
1463                         error = uiomove(cursor.data->symlink.name,
1464                                         cursor.leaf->data_len -
1465                                                 HAMMER_SYMLINK_NAME_OFF,
1466                                         ap->a_uio);
1467                 }
1468         }
1469         hammer_done_cursor(&cursor);
1470         hammer_done_transaction(&trans);
1471         return(error);
1472 }
1473
1474 /*
1475  * hammer_vop_nremove { nch, dvp, cred }
1476  */
1477 static
1478 int
1479 hammer_vop_nremove(struct vop_nremove_args *ap)
1480 {
1481         struct hammer_transaction trans;
1482         struct hammer_inode *dip;
1483         int error;
1484
1485         dip = VTOI(ap->a_dvp);
1486
1487         if (hammer_nohistory(dip) == 0 &&
1488             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1489                 return (error);
1490         }
1491
1492         hammer_start_transaction(&trans, dip->hmp);
1493         ++hammer_stats_file_iopsw;
1494         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0);
1495         hammer_done_transaction(&trans);
1496
1497         return (error);
1498 }
1499
1500 /*
1501  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1502  */
1503 static
1504 int
1505 hammer_vop_nrename(struct vop_nrename_args *ap)
1506 {
1507         struct hammer_transaction trans;
1508         struct namecache *fncp;
1509         struct namecache *tncp;
1510         struct hammer_inode *fdip;
1511         struct hammer_inode *tdip;
1512         struct hammer_inode *ip;
1513         struct hammer_cursor cursor;
1514         int64_t namekey;
1515         int nlen, error;
1516
1517         if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount) 
1518                 return(EXDEV);
1519         if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1520                 return(EXDEV);
1521
1522         fdip = VTOI(ap->a_fdvp);
1523         tdip = VTOI(ap->a_tdvp);
1524         fncp = ap->a_fnch->ncp;
1525         tncp = ap->a_tnch->ncp;
1526         ip = VTOI(fncp->nc_vp);
1527         KKASSERT(ip != NULL);
1528
1529         if (fdip->obj_localization != tdip->obj_localization)
1530                 return(EXDEV);
1531         if (fdip->obj_localization != ip->obj_localization)
1532                 return(EXDEV);
1533
1534         if (fdip->flags & HAMMER_INODE_RO)
1535                 return (EROFS);
1536         if (tdip->flags & HAMMER_INODE_RO)
1537                 return (EROFS);
1538         if (ip->flags & HAMMER_INODE_RO)
1539                 return (EROFS);
1540         if ((error = hammer_checkspace(fdip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1541                 return (error);
1542
1543         hammer_start_transaction(&trans, fdip->hmp);
1544         ++hammer_stats_file_iopsw;
1545
1546         /*
1547          * Remove tncp from the target directory and then link ip as
1548          * tncp. XXX pass trans to dounlink
1549          *
1550          * Force the inode sync-time to match the transaction so it is
1551          * in-sync with the creation of the target directory entry.
1552          */
1553         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp, ap->a_cred, 0);
1554         if (error == 0 || error == ENOENT) {
1555                 error = hammer_ip_add_directory(&trans, tdip,
1556                                                 tncp->nc_name, tncp->nc_nlen,
1557                                                 ip);
1558                 if (error == 0) {
1559                         ip->ino_data.parent_obj_id = tdip->obj_id;
1560                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
1561                 }
1562         }
1563         if (error)
1564                 goto failed; /* XXX */
1565
1566         /*
1567          * Locate the record in the originating directory and remove it.
1568          *
1569          * Calculate the namekey and setup the key range for the scan.  This
1570          * works kinda like a chained hash table where the lower 32 bits
1571          * of the namekey synthesize the chain.
1572          *
1573          * The key range is inclusive of both key_beg and key_end.
1574          */
1575         namekey = hammer_directory_namekey(fncp->nc_name, fncp->nc_nlen);
1576 retry:
1577         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
1578         cursor.key_beg.localization = fdip->obj_localization +
1579                                       HAMMER_LOCALIZE_MISC;
1580         cursor.key_beg.obj_id = fdip->obj_id;
1581         cursor.key_beg.key = namekey;
1582         cursor.key_beg.create_tid = 0;
1583         cursor.key_beg.delete_tid = 0;
1584         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1585         cursor.key_beg.obj_type = 0;
1586
1587         cursor.key_end = cursor.key_beg;
1588         cursor.key_end.key |= 0xFFFFFFFFULL;
1589         cursor.asof = fdip->obj_asof;
1590         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1591
1592         /*
1593          * Scan all matching records (the chain), locate the one matching
1594          * the requested path component.
1595          *
1596          * The hammer_ip_*() functions merge in-memory records with on-disk
1597          * records for the purposes of the search.
1598          */
1599         error = hammer_ip_first(&cursor);
1600         while (error == 0) {
1601                 if (hammer_ip_resolve_data(&cursor) != 0)
1602                         break;
1603                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
1604                 KKASSERT(nlen > 0);
1605                 if (fncp->nc_nlen == nlen &&
1606                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1607                         break;
1608                 }
1609                 error = hammer_ip_next(&cursor);
1610         }
1611
1612         /*
1613          * If all is ok we have to get the inode so we can adjust nlinks.
1614          *
1615          * WARNING: hammer_ip_del_directory() may have to terminate the
1616          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
1617          * twice.
1618          */
1619         if (error == 0)
1620                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
1621
1622         /*
1623          * XXX A deadlock here will break rename's atomicy for the purposes
1624          * of crash recovery.
1625          */
1626         if (error == EDEADLK) {
1627                 hammer_done_cursor(&cursor);
1628                 goto retry;
1629         }
1630
1631         /*
1632          * Cleanup and tell the kernel that the rename succeeded.
1633          */
1634         hammer_done_cursor(&cursor);
1635         if (error == 0)
1636                 cache_rename(ap->a_fnch, ap->a_tnch);
1637
1638 failed:
1639         hammer_done_transaction(&trans);
1640         return (error);
1641 }
1642
1643 /*
1644  * hammer_vop_nrmdir { nch, dvp, cred }
1645  */
1646 static
1647 int
1648 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
1649 {
1650         struct hammer_transaction trans;
1651         struct hammer_inode *dip;
1652         int error;
1653
1654         dip = VTOI(ap->a_dvp);
1655
1656         if (hammer_nohistory(dip) == 0 &&
1657             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1658                 return (error);
1659         }
1660
1661         hammer_start_transaction(&trans, dip->hmp);
1662         ++hammer_stats_file_iopsw;
1663         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0);
1664         hammer_done_transaction(&trans);
1665
1666         return (error);
1667 }
1668
1669 /*
1670  * hammer_vop_setattr { vp, vap, cred }
1671  */
1672 static
1673 int
1674 hammer_vop_setattr(struct vop_setattr_args *ap)
1675 {
1676         struct hammer_transaction trans;
1677         struct vattr *vap;
1678         struct hammer_inode *ip;
1679         int modflags;
1680         int error;
1681         int truncating;
1682         int blksize;
1683         int64_t aligned_size;
1684         u_int32_t flags;
1685
1686         vap = ap->a_vap;
1687         ip = ap->a_vp->v_data;
1688         modflags = 0;
1689
1690         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
1691                 return(EROFS);
1692         if (ip->flags & HAMMER_INODE_RO)
1693                 return (EROFS);
1694         if (hammer_nohistory(ip) == 0 &&
1695             (error = hammer_checkspace(ip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1696                 return (error);
1697         }
1698
1699         hammer_start_transaction(&trans, ip->hmp);
1700         ++hammer_stats_file_iopsw;
1701         error = 0;
1702
1703         if (vap->va_flags != VNOVAL) {
1704                 flags = ip->ino_data.uflags;
1705                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
1706                                          hammer_to_unix_xid(&ip->ino_data.uid),
1707                                          ap->a_cred);
1708                 if (error == 0) {
1709                         if (ip->ino_data.uflags != flags) {
1710                                 ip->ino_data.uflags = flags;
1711                                 modflags |= HAMMER_INODE_DDIRTY;
1712                         }
1713                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
1714                                 error = 0;
1715                                 goto done;
1716                         }
1717                 }
1718                 goto done;
1719         }
1720         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
1721                 error = EPERM;
1722                 goto done;
1723         }
1724         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
1725                 mode_t cur_mode = ip->ino_data.mode;
1726                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1727                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1728                 uuid_t uuid_uid;
1729                 uuid_t uuid_gid;
1730
1731                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
1732                                          ap->a_cred,
1733                                          &cur_uid, &cur_gid, &cur_mode);
1734                 if (error == 0) {
1735                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
1736                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
1737                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
1738                                  sizeof(uuid_uid)) ||
1739                             bcmp(&uuid_gid, &ip->ino_data.gid,
1740                                  sizeof(uuid_gid)) ||
1741                             ip->ino_data.mode != cur_mode
1742                         ) {
1743                                 ip->ino_data.uid = uuid_uid;
1744                                 ip->ino_data.gid = uuid_gid;
1745                                 ip->ino_data.mode = cur_mode;
1746                         }
1747                         modflags |= HAMMER_INODE_DDIRTY;
1748                 }
1749         }
1750         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
1751                 switch(ap->a_vp->v_type) {
1752                 case VREG:
1753                         if (vap->va_size == ip->ino_data.size)
1754                                 break;
1755                         /*
1756                          * XXX break atomicy, we can deadlock the backend
1757                          * if we do not release the lock.  Probably not a
1758                          * big deal here.
1759                          */
1760                         blksize = hammer_blocksize(vap->va_size);
1761                         if (vap->va_size < ip->ino_data.size) {
1762                                 vtruncbuf(ap->a_vp, vap->va_size, blksize);
1763                                 truncating = 1;
1764                         } else {
1765                                 vnode_pager_setsize(ap->a_vp, vap->va_size);
1766                                 truncating = 0;
1767                         }
1768                         ip->ino_data.size = vap->va_size;
1769                         modflags |= HAMMER_INODE_DDIRTY;
1770
1771                         /*
1772                          * on-media truncation is cached in the inode until
1773                          * the inode is synchronized.
1774                          */
1775                         if (truncating) {
1776                                 hammer_ip_frontend_trunc(ip, vap->va_size);
1777 #ifdef DEBUG_TRUNCATE
1778                                 if (HammerTruncIp == NULL)
1779                                         HammerTruncIp = ip;
1780 #endif
1781                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
1782                                         ip->flags |= HAMMER_INODE_TRUNCATED;
1783                                         ip->trunc_off = vap->va_size;
1784 #ifdef DEBUG_TRUNCATE
1785                                         if (ip == HammerTruncIp)
1786                                         kprintf("truncate1 %016llx\n", ip->trunc_off);
1787 #endif
1788                                 } else if (ip->trunc_off > vap->va_size) {
1789                                         ip->trunc_off = vap->va_size;
1790 #ifdef DEBUG_TRUNCATE
1791                                         if (ip == HammerTruncIp)
1792                                         kprintf("truncate2 %016llx\n", ip->trunc_off);
1793 #endif
1794                                 } else {
1795 #ifdef DEBUG_TRUNCATE
1796                                         if (ip == HammerTruncIp)
1797                                         kprintf("truncate3 %016llx (ignored)\n", vap->va_size);
1798 #endif
1799                                 }
1800                         }
1801
1802                         /*
1803                          * If truncating we have to clean out a portion of
1804                          * the last block on-disk.  We do this in the
1805                          * front-end buffer cache.
1806                          */
1807                         aligned_size = (vap->va_size + (blksize - 1)) &
1808                                        ~(int64_t)(blksize - 1);
1809                         if (truncating && vap->va_size < aligned_size) {
1810                                 struct buf *bp;
1811                                 int offset;
1812
1813                                 aligned_size -= blksize;
1814
1815                                 offset = (int)vap->va_size & (blksize - 1);
1816                                 error = bread(ap->a_vp, aligned_size,
1817                                               blksize, &bp);
1818                                 hammer_ip_frontend_trunc(ip, aligned_size);
1819                                 if (error == 0) {
1820                                         bzero(bp->b_data + offset,
1821                                               blksize - offset);
1822                                         /* must de-cache direct-io offset */
1823                                         bp->b_bio2.bio_offset = NOOFFSET;
1824                                         bdwrite(bp);
1825                                 } else {
1826                                         kprintf("ERROR %d\n", error);
1827                                         brelse(bp);
1828                                 }
1829                         }
1830                         break;
1831                 case VDATABASE:
1832                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
1833                                 ip->flags |= HAMMER_INODE_TRUNCATED;
1834                                 ip->trunc_off = vap->va_size;
1835                         } else if (ip->trunc_off > vap->va_size) {
1836                                 ip->trunc_off = vap->va_size;
1837                         }
1838                         hammer_ip_frontend_trunc(ip, vap->va_size);
1839                         ip->ino_data.size = vap->va_size;
1840                         modflags |= HAMMER_INODE_DDIRTY;
1841                         break;
1842                 default:
1843                         error = EINVAL;
1844                         goto done;
1845                 }
1846                 break;
1847         }
1848         if (vap->va_atime.tv_sec != VNOVAL) {
1849                 ip->ino_data.atime =
1850                         hammer_timespec_to_time(&vap->va_atime);
1851                 modflags |= HAMMER_INODE_ATIME;
1852         }
1853         if (vap->va_mtime.tv_sec != VNOVAL) {
1854                 ip->ino_data.mtime =
1855                         hammer_timespec_to_time(&vap->va_mtime);
1856                 modflags |= HAMMER_INODE_MTIME;
1857         }
1858         if (vap->va_mode != (mode_t)VNOVAL) {
1859                 mode_t   cur_mode = ip->ino_data.mode;
1860                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1861                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1862
1863                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
1864                                          cur_uid, cur_gid, &cur_mode);
1865                 if (error == 0 && ip->ino_data.mode != cur_mode) {
1866                         ip->ino_data.mode = cur_mode;
1867                         modflags |= HAMMER_INODE_DDIRTY;
1868                 }
1869         }
1870 done:
1871         if (error == 0)
1872                 hammer_modify_inode(ip, modflags);
1873         hammer_done_transaction(&trans);
1874         return (error);
1875 }
1876
1877 /*
1878  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
1879  */
1880 static
1881 int
1882 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
1883 {
1884         struct hammer_transaction trans;
1885         struct hammer_inode *dip;
1886         struct hammer_inode *nip;
1887         struct nchandle *nch;
1888         hammer_record_t record;
1889         int error;
1890         int bytes;
1891
1892         ap->a_vap->va_type = VLNK;
1893
1894         nch = ap->a_nch;
1895         dip = VTOI(ap->a_dvp);
1896
1897         if (dip->flags & HAMMER_INODE_RO)
1898                 return (EROFS);
1899         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1900                 return (error);
1901
1902         /*
1903          * Create a transaction to cover the operations we perform.
1904          */
1905         hammer_start_transaction(&trans, dip->hmp);
1906         ++hammer_stats_file_iopsw;
1907
1908         /*
1909          * Create a new filesystem object of the requested type.  The
1910          * returned inode will be referenced but not locked.
1911          */
1912
1913         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1914                                     dip, NULL, &nip);
1915         if (error) {
1916                 hammer_done_transaction(&trans);
1917                 *ap->a_vpp = NULL;
1918                 return (error);
1919         }
1920
1921         /*
1922          * Add a record representing the symlink.  symlink stores the link
1923          * as pure data, not a string, and is no \0 terminated.
1924          */
1925         if (error == 0) {
1926                 bytes = strlen(ap->a_target);
1927
1928                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
1929                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
1930                 } else {
1931                         record = hammer_alloc_mem_record(nip, bytes);
1932                         record->type = HAMMER_MEM_RECORD_GENERAL;
1933
1934                         record->leaf.base.localization = nip->obj_localization +
1935                                                          HAMMER_LOCALIZE_MISC;
1936                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
1937                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
1938                         record->leaf.data_len = bytes;
1939                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
1940                         bcopy(ap->a_target, record->data->symlink.name, bytes);
1941                         error = hammer_ip_add_record(&trans, record);
1942                 }
1943
1944                 /*
1945                  * Set the file size to the length of the link.
1946                  */
1947                 if (error == 0) {
1948                         nip->ino_data.size = bytes;
1949                         hammer_modify_inode(nip, HAMMER_INODE_DDIRTY);
1950                 }
1951         }
1952         if (error == 0)
1953                 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
1954                                                 nch->ncp->nc_nlen, nip);
1955
1956         /*
1957          * Finish up.
1958          */
1959         if (error) {
1960                 hammer_rel_inode(nip, 0);
1961                 *ap->a_vpp = NULL;
1962         } else {
1963                 error = hammer_get_vnode(nip, ap->a_vpp);
1964                 hammer_rel_inode(nip, 0);
1965                 if (error == 0) {
1966                         cache_setunresolved(ap->a_nch);
1967                         cache_setvp(ap->a_nch, *ap->a_vpp);
1968                 }
1969         }
1970         hammer_done_transaction(&trans);
1971         return (error);
1972 }
1973
1974 /*
1975  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
1976  */
1977 static
1978 int
1979 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
1980 {
1981         struct hammer_transaction trans;
1982         struct hammer_inode *dip;
1983         int error;
1984
1985         dip = VTOI(ap->a_dvp);
1986
1987         if (hammer_nohistory(dip) == 0 &&
1988             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0) {
1989                 return (error);
1990         }
1991
1992         hammer_start_transaction(&trans, dip->hmp);
1993         ++hammer_stats_file_iopsw;
1994         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
1995                                 ap->a_cred, ap->a_flags);
1996         hammer_done_transaction(&trans);
1997
1998         return (error);
1999 }
2000
2001 /*
2002  * hammer_vop_ioctl { vp, command, data, fflag, cred }
2003  */
2004 static
2005 int
2006 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2007 {
2008         struct hammer_inode *ip = ap->a_vp->v_data;
2009
2010         ++hammer_stats_file_iopsr;
2011         return(hammer_ioctl(ip, ap->a_command, ap->a_data,
2012                             ap->a_fflag, ap->a_cred));
2013 }
2014
2015 static
2016 int
2017 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2018 {
2019         struct mount *mp;
2020         int error;
2021
2022         mp = ap->a_head.a_ops->head.vv_mount;
2023
2024         switch(ap->a_op) {
2025         case MOUNTCTL_SET_EXPORT:
2026                 if (ap->a_ctllen != sizeof(struct export_args))
2027                         error = EINVAL;
2028                 error = hammer_vfs_export(mp, ap->a_op,
2029                                       (const struct export_args *)ap->a_ctl);
2030                 break;
2031         default:
2032                 error = journal_mountctl(ap);
2033                 break;
2034         }
2035         return(error);
2036 }
2037
2038 /*
2039  * hammer_vop_strategy { vp, bio }
2040  *
2041  * Strategy call, used for regular file read & write only.  Note that the
2042  * bp may represent a cluster.
2043  *
2044  * To simplify operation and allow better optimizations in the future,
2045  * this code does not make any assumptions with regards to buffer alignment
2046  * or size.
2047  */
2048 static
2049 int
2050 hammer_vop_strategy(struct vop_strategy_args *ap)
2051 {
2052         struct buf *bp;
2053         int error;
2054
2055         bp = ap->a_bio->bio_buf;
2056
2057         switch(bp->b_cmd) {
2058         case BUF_CMD_READ:
2059                 error = hammer_vop_strategy_read(ap);
2060                 break;
2061         case BUF_CMD_WRITE:
2062                 error = hammer_vop_strategy_write(ap);
2063                 break;
2064         default:
2065                 bp->b_error = error = EINVAL;
2066                 bp->b_flags |= B_ERROR;
2067                 biodone(ap->a_bio);
2068                 break;
2069         }
2070         return (error);
2071 }
2072
2073 /*
2074  * Read from a regular file.  Iterate the related records and fill in the
2075  * BIO/BUF.  Gaps are zero-filled.
2076  *
2077  * The support code in hammer_object.c should be used to deal with mixed
2078  * in-memory and on-disk records.
2079  *
2080  * NOTE: Can be called from the cluster code with an oversized buf.
2081  *
2082  * XXX atime update
2083  */
2084 static
2085 int
2086 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2087 {
2088         struct hammer_transaction trans;
2089         struct hammer_inode *ip;
2090         struct hammer_cursor cursor;
2091         hammer_base_elm_t base;
2092         hammer_off_t disk_offset;
2093         struct bio *bio;
2094         struct bio *nbio;
2095         struct buf *bp;
2096         int64_t rec_offset;
2097         int64_t ran_end;
2098         int64_t tmp64;
2099         int error;
2100         int boff;
2101         int roff;
2102         int n;
2103
2104         bio = ap->a_bio;
2105         bp = bio->bio_buf;
2106         ip = ap->a_vp->v_data;
2107
2108         /*
2109          * The zone-2 disk offset may have been set by the cluster code via
2110          * a BMAP operation, or else should be NOOFFSET.
2111          *
2112          * Checking the high bits for a match against zone-2 should suffice.
2113          */
2114         nbio = push_bio(bio);
2115         if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2116             HAMMER_ZONE_LARGE_DATA) {
2117                 error = hammer_io_direct_read(ip->hmp, nbio, NULL);
2118                 return (error);
2119         }
2120
2121         /*
2122          * Well, that sucked.  Do it the hard way.  If all the stars are
2123          * aligned we may still be able to issue a direct-read.
2124          */
2125         hammer_simple_transaction(&trans, ip->hmp);
2126         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2127
2128         /*
2129          * Key range (begin and end inclusive) to scan.  Note that the key's
2130          * stored in the actual records represent BASE+LEN, not BASE.  The
2131          * first record containing bio_offset will have a key > bio_offset.
2132          */
2133         cursor.key_beg.localization = ip->obj_localization +
2134                                       HAMMER_LOCALIZE_MISC;
2135         cursor.key_beg.obj_id = ip->obj_id;
2136         cursor.key_beg.create_tid = 0;
2137         cursor.key_beg.delete_tid = 0;
2138         cursor.key_beg.obj_type = 0;
2139         cursor.key_beg.key = bio->bio_offset + 1;
2140         cursor.asof = ip->obj_asof;
2141         cursor.flags |= HAMMER_CURSOR_ASOF;
2142
2143         cursor.key_end = cursor.key_beg;
2144         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2145 #if 0
2146         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2147                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2148                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2149                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2150         } else
2151 #endif
2152         {
2153                 ran_end = bio->bio_offset + bp->b_bufsize;
2154                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2155                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2156                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2157                 if (tmp64 < ran_end)
2158                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2159                 else
2160                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2161         }
2162         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2163
2164         error = hammer_ip_first(&cursor);
2165         boff = 0;
2166
2167         while (error == 0) {
2168                 /*
2169                  * Get the base file offset of the record.  The key for
2170                  * data records is (base + bytes) rather then (base).
2171                  */
2172                 base = &cursor.leaf->base;
2173                 rec_offset = base->key - cursor.leaf->data_len;
2174
2175                 /*
2176                  * Calculate the gap, if any, and zero-fill it.
2177                  *
2178                  * n is the offset of the start of the record verses our
2179                  * current seek offset in the bio.
2180                  */
2181                 n = (int)(rec_offset - (bio->bio_offset + boff));
2182                 if (n > 0) {
2183                         if (n > bp->b_bufsize - boff)
2184                                 n = bp->b_bufsize - boff;
2185                         bzero((char *)bp->b_data + boff, n);
2186                         boff += n;
2187                         n = 0;
2188                 }
2189
2190                 /*
2191                  * Calculate the data offset in the record and the number
2192                  * of bytes we can copy.
2193                  *
2194                  * There are two degenerate cases.  First, boff may already
2195                  * be at bp->b_bufsize.  Secondly, the data offset within
2196                  * the record may exceed the record's size.
2197                  */
2198                 roff = -n;
2199                 rec_offset += roff;
2200                 n = cursor.leaf->data_len - roff;
2201                 if (n <= 0) {
2202                         kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2203                         n = 0;
2204                 } else if (n > bp->b_bufsize - boff) {
2205                         n = bp->b_bufsize - boff;
2206                 }
2207
2208                 /*
2209                  * Deal with cached truncations.  This cool bit of code
2210                  * allows truncate()/ftruncate() to avoid having to sync
2211                  * the file.
2212                  *
2213                  * If the frontend is truncated then all backend records are
2214                  * subject to the frontend's truncation.
2215                  *
2216                  * If the backend is truncated then backend records on-disk
2217                  * (but not in-memory) are subject to the backend's
2218                  * truncation.  In-memory records owned by the backend
2219                  * represent data written after the truncation point on the
2220                  * backend and must not be truncated.
2221                  *
2222                  * Truncate operations deal with frontend buffer cache
2223                  * buffers and frontend-owned in-memory records synchronously.
2224                  */
2225                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2226                         if (hammer_cursor_ondisk(&cursor) ||
2227                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2228                                 if (ip->trunc_off <= rec_offset)
2229                                         n = 0;
2230                                 else if (ip->trunc_off < rec_offset + n)
2231                                         n = (int)(ip->trunc_off - rec_offset);
2232                         }
2233                 }
2234                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2235                         if (hammer_cursor_ondisk(&cursor)) {
2236                                 if (ip->sync_trunc_off <= rec_offset)
2237                                         n = 0;
2238                                 else if (ip->sync_trunc_off < rec_offset + n)
2239                                         n = (int)(ip->sync_trunc_off - rec_offset);
2240                         }
2241                 }
2242
2243                 /*
2244                  * Try to issue a direct read into our bio if possible,
2245                  * otherwise resolve the element data into a hammer_buffer
2246                  * and copy.
2247                  *
2248                  * The buffer on-disk should be zerod past any real
2249                  * truncation point, but may not be for any synthesized
2250                  * truncation point from above.
2251                  */
2252                 disk_offset = cursor.leaf->data_offset + roff;
2253                 if (boff == 0 && n == bp->b_bufsize &&
2254                     hammer_cursor_ondisk(&cursor) &&
2255                     (disk_offset & HAMMER_BUFMASK) == 0) {
2256                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2257                                  HAMMER_ZONE_LARGE_DATA);
2258                         nbio->bio_offset = disk_offset;
2259                         error = hammer_io_direct_read(trans.hmp, nbio,
2260                                                       cursor.leaf);
2261                         goto done;
2262                 } else if (n) {
2263                         error = hammer_ip_resolve_data(&cursor);
2264                         if (error == 0) {
2265                                 bcopy((char *)cursor.data + roff,
2266                                       (char *)bp->b_data + boff, n);
2267                         }
2268                 }
2269                 if (error)
2270                         break;
2271
2272                 /*
2273                  * Iterate until we have filled the request.
2274                  */
2275                 boff += n;
2276                 if (boff == bp->b_bufsize)
2277                         break;
2278                 error = hammer_ip_next(&cursor);
2279         }
2280
2281         /*
2282          * There may have been a gap after the last record
2283          */
2284         if (error == ENOENT)
2285                 error = 0;
2286         if (error == 0 && boff != bp->b_bufsize) {
2287                 KKASSERT(boff < bp->b_bufsize);
2288                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2289                 /* boff = bp->b_bufsize; */
2290         }
2291         bp->b_resid = 0;
2292         bp->b_error = error;
2293         if (error)
2294                 bp->b_flags |= B_ERROR;
2295         biodone(ap->a_bio);
2296
2297 done:
2298         if (cursor.node)
2299                 hammer_cache_node(&ip->cache[1], cursor.node);
2300         hammer_done_cursor(&cursor);
2301         hammer_done_transaction(&trans);
2302         return(error);
2303 }
2304
2305 /*
2306  * BMAP operation - used to support cluster_read() only.
2307  *
2308  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2309  *
2310  * This routine may return EOPNOTSUPP if the opration is not supported for
2311  * the specified offset.  The contents of the pointer arguments do not
2312  * need to be initialized in that case. 
2313  *
2314  * If a disk address is available and properly aligned return 0 with 
2315  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2316  * to the run-length relative to that offset.  Callers may assume that
2317  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2318  * large, so return EOPNOTSUPP if it is not sufficiently large.
2319  */
2320 static
2321 int
2322 hammer_vop_bmap(struct vop_bmap_args *ap)
2323 {
2324         struct hammer_transaction trans;
2325         struct hammer_inode *ip;
2326         struct hammer_cursor cursor;
2327         hammer_base_elm_t base;
2328         int64_t rec_offset;
2329         int64_t ran_end;
2330         int64_t tmp64;
2331         int64_t base_offset;
2332         int64_t base_disk_offset;
2333         int64_t last_offset;
2334         hammer_off_t last_disk_offset;
2335         hammer_off_t disk_offset;
2336         int     rec_len;
2337         int     error;
2338         int     blksize;
2339
2340         ++hammer_stats_file_iopsr;
2341         ip = ap->a_vp->v_data;
2342
2343         /*
2344          * We can only BMAP regular files.  We can't BMAP database files,
2345          * directories, etc.
2346          */
2347         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
2348                 return(EOPNOTSUPP);
2349
2350         /*
2351          * bmap is typically called with runp/runb both NULL when used
2352          * for writing.  We do not support BMAP for writing atm.
2353          */
2354         if (ap->a_cmd != BUF_CMD_READ)
2355                 return(EOPNOTSUPP);
2356
2357         /*
2358          * Scan the B-Tree to acquire blockmap addresses, then translate
2359          * to raw addresses.
2360          */
2361         hammer_simple_transaction(&trans, ip->hmp);
2362 #if 0
2363         kprintf("bmap_beg %016llx ip->cache %p\n", ap->a_loffset, ip->cache[1]);
2364 #endif
2365         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2366
2367         /*
2368          * Key range (begin and end inclusive) to scan.  Note that the key's
2369          * stored in the actual records represent BASE+LEN, not BASE.  The
2370          * first record containing bio_offset will have a key > bio_offset.
2371          */
2372         cursor.key_beg.localization = ip->obj_localization +
2373                                       HAMMER_LOCALIZE_MISC;
2374         cursor.key_beg.obj_id = ip->obj_id;
2375         cursor.key_beg.create_tid = 0;
2376         cursor.key_beg.delete_tid = 0;
2377         cursor.key_beg.obj_type = 0;
2378         if (ap->a_runb)
2379                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
2380         else
2381                 cursor.key_beg.key = ap->a_loffset + 1;
2382         if (cursor.key_beg.key < 0)
2383                 cursor.key_beg.key = 0;
2384         cursor.asof = ip->obj_asof;
2385         cursor.flags |= HAMMER_CURSOR_ASOF;
2386
2387         cursor.key_end = cursor.key_beg;
2388         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2389
2390         ran_end = ap->a_loffset + MAXPHYS;
2391         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2392         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2393         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2394         if (tmp64 < ran_end)
2395                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2396         else
2397                 cursor.key_end.key = ran_end + MAXPHYS + 1;
2398
2399         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2400
2401         error = hammer_ip_first(&cursor);
2402         base_offset = last_offset = 0;
2403         base_disk_offset = last_disk_offset = 0;
2404
2405         while (error == 0) {
2406                 /*
2407                  * Get the base file offset of the record.  The key for
2408                  * data records is (base + bytes) rather then (base).
2409                  *
2410                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
2411                  * The extra bytes should be zero on-disk and the BMAP op
2412                  * should still be ok.
2413                  */
2414                 base = &cursor.leaf->base;
2415                 rec_offset = base->key - cursor.leaf->data_len;
2416                 rec_len    = cursor.leaf->data_len;
2417
2418                 /*
2419                  * Incorporate any cached truncation.
2420                  *
2421                  * NOTE: Modifications to rec_len based on synthesized
2422                  * truncation points remove the guarantee that any extended
2423                  * data on disk is zero (since the truncations may not have
2424                  * taken place on-media yet).
2425                  */
2426                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2427                         if (hammer_cursor_ondisk(&cursor) ||
2428                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2429                                 if (ip->trunc_off <= rec_offset)
2430                                         rec_len = 0;
2431                                 else if (ip->trunc_off < rec_offset + rec_len)
2432                                         rec_len = (int)(ip->trunc_off - rec_offset);
2433                         }
2434                 }
2435                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2436                         if (hammer_cursor_ondisk(&cursor)) {
2437                                 if (ip->sync_trunc_off <= rec_offset)
2438                                         rec_len = 0;
2439                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
2440                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
2441                         }
2442                 }
2443
2444                 /*
2445                  * Accumulate information.  If we have hit a discontiguous
2446                  * block reset base_offset unless we are already beyond the
2447                  * requested offset.  If we are, that's it, we stop.
2448                  */
2449                 if (error)
2450                         break;
2451                 if (hammer_cursor_ondisk(&cursor)) {
2452                         disk_offset = cursor.leaf->data_offset;
2453                         if (rec_offset != last_offset ||
2454                             disk_offset != last_disk_offset) {
2455                                 if (rec_offset > ap->a_loffset)
2456                                         break;
2457                                 base_offset = rec_offset;
2458                                 base_disk_offset = disk_offset;
2459                         }
2460                         last_offset = rec_offset + rec_len;
2461                         last_disk_offset = disk_offset + rec_len;
2462                 }
2463                 error = hammer_ip_next(&cursor);
2464         }
2465
2466 #if 0
2467         kprintf("BMAP %016llx:  %016llx - %016llx\n",
2468                 ap->a_loffset, base_offset, last_offset);
2469         kprintf("BMAP %16s:  %016llx - %016llx\n",
2470                 "", base_disk_offset, last_disk_offset);
2471 #endif
2472
2473         if (cursor.node) {
2474                 hammer_cache_node(&ip->cache[1], cursor.node);
2475 #if 0
2476                 kprintf("bmap_end2 %016llx ip->cache %p\n", ap->a_loffset, ip->cache[1]);
2477 #endif
2478         }
2479         hammer_done_cursor(&cursor);
2480         hammer_done_transaction(&trans);
2481
2482         /*
2483          * If we couldn't find any records or the records we did find were
2484          * all behind the requested offset, return failure.  A forward
2485          * truncation can leave a hole w/ no on-disk records.
2486          */
2487         if (last_offset == 0 || last_offset < ap->a_loffset)
2488                 return (EOPNOTSUPP);
2489
2490         /*
2491          * Figure out the block size at the requested offset and adjust
2492          * our limits so the cluster_read() does not create inappropriately
2493          * sized buffer cache buffers.
2494          */
2495         blksize = hammer_blocksize(ap->a_loffset);
2496         if (hammer_blocksize(base_offset) != blksize) {
2497                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
2498         }
2499         if (last_offset != ap->a_loffset &&
2500             hammer_blocksize(last_offset - 1) != blksize) {
2501                 last_offset = hammer_blockdemarc(ap->a_loffset,
2502                                                  last_offset - 1);
2503         }
2504
2505         /*
2506          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
2507          * from occuring.
2508          */
2509         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
2510
2511         if ((disk_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_LARGE_DATA) {
2512                 /*
2513                  * Only large-data zones can be direct-IOd
2514                  */
2515                 error = EOPNOTSUPP;
2516         } else if ((disk_offset & HAMMER_BUFMASK) ||
2517                    (last_offset - ap->a_loffset) < blksize) {
2518                 /*
2519                  * doffsetp is not aligned or the forward run size does
2520                  * not cover a whole buffer, disallow the direct I/O.
2521                  */
2522                 error = EOPNOTSUPP;
2523         } else {
2524                 /*
2525                  * We're good.
2526                  */
2527                 *ap->a_doffsetp = disk_offset;
2528                 if (ap->a_runb) {
2529                         *ap->a_runb = ap->a_loffset - base_offset;
2530                         KKASSERT(*ap->a_runb >= 0);
2531                 }
2532                 if (ap->a_runp) {
2533                         *ap->a_runp = last_offset - ap->a_loffset;
2534                         KKASSERT(*ap->a_runp >= 0);
2535                 }
2536                 error = 0;
2537         }
2538         return(error);
2539 }
2540
2541 /*
2542  * Write to a regular file.   Because this is a strategy call the OS is
2543  * trying to actually get data onto the media.
2544  */
2545 static
2546 int
2547 hammer_vop_strategy_write(struct vop_strategy_args *ap)
2548 {
2549         hammer_record_t record;
2550         hammer_mount_t hmp;
2551         hammer_inode_t ip;
2552         struct bio *bio;
2553         struct buf *bp;
2554         int blksize;
2555         int bytes;
2556         int error;
2557
2558         bio = ap->a_bio;
2559         bp = bio->bio_buf;
2560         ip = ap->a_vp->v_data;
2561         hmp = ip->hmp;
2562
2563         blksize = hammer_blocksize(bio->bio_offset);
2564         KKASSERT(bp->b_bufsize == blksize);
2565
2566         if (ip->flags & HAMMER_INODE_RO) {
2567                 bp->b_error = EROFS;
2568                 bp->b_flags |= B_ERROR;
2569                 biodone(ap->a_bio);
2570                 return(EROFS);
2571         }
2572
2573         /*
2574          * Interlock with inode destruction (no in-kernel or directory
2575          * topology visibility).  If we queue new IO while trying to
2576          * destroy the inode we can deadlock the vtrunc call in
2577          * hammer_inode_unloadable_check().
2578          */
2579         if (ip->flags & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
2580                 bp->b_resid = 0;
2581                 biodone(ap->a_bio);
2582                 return(0);
2583         }
2584
2585         /*
2586          * Reserve space and issue a direct-write from the front-end. 
2587          * NOTE: The direct_io code will hammer_bread/bcopy smaller
2588          * allocations.
2589          *
2590          * An in-memory record will be installed to reference the storage
2591          * until the flusher can get to it.
2592          *
2593          * Since we own the high level bio the front-end will not try to
2594          * do a direct-read until the write completes.
2595          *
2596          * NOTE: The only time we do not reserve a full-sized buffers
2597          * worth of data is if the file is small.  We do not try to
2598          * allocate a fragment (from the small-data zone) at the end of
2599          * an otherwise large file as this can lead to wildly separated
2600          * data.
2601          */
2602         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
2603         KKASSERT(bio->bio_offset < ip->ino_data.size);
2604         if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
2605                 bytes = bp->b_bufsize;
2606         else
2607                 bytes = ((int)ip->ino_data.size + 15) & ~15;
2608
2609         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
2610                                     bytes, &error);
2611         if (record) {
2612                 hammer_io_direct_write(hmp, record, bio);
2613                 hammer_rel_mem_record(record);
2614                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
2615                         hammer_flush_inode(ip, 0);
2616         } else {
2617                 bp->b_bio2.bio_offset = NOOFFSET;
2618                 bp->b_error = error;
2619                 bp->b_flags |= B_ERROR;
2620                 biodone(ap->a_bio);
2621         }
2622         return(error);
2623 }
2624
2625 /*
2626  * dounlink - disconnect a directory entry
2627  *
2628  * XXX whiteout support not really in yet
2629  */
2630 static int
2631 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
2632                 struct vnode *dvp, struct ucred *cred, int flags)
2633 {
2634         struct namecache *ncp;
2635         hammer_inode_t dip;
2636         hammer_inode_t ip;
2637         struct hammer_cursor cursor;
2638         int64_t namekey;
2639         int nlen, error;
2640
2641         /*
2642          * Calculate the namekey and setup the key range for the scan.  This
2643          * works kinda like a chained hash table where the lower 32 bits
2644          * of the namekey synthesize the chain.
2645          *
2646          * The key range is inclusive of both key_beg and key_end.
2647          */
2648         dip = VTOI(dvp);
2649         ncp = nch->ncp;
2650
2651         if (dip->flags & HAMMER_INODE_RO)
2652                 return (EROFS);
2653
2654         namekey = hammer_directory_namekey(ncp->nc_name, ncp->nc_nlen);
2655 retry:
2656         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
2657         cursor.key_beg.localization = dip->obj_localization +
2658                                       HAMMER_LOCALIZE_MISC;
2659         cursor.key_beg.obj_id = dip->obj_id;
2660         cursor.key_beg.key = namekey;
2661         cursor.key_beg.create_tid = 0;
2662         cursor.key_beg.delete_tid = 0;
2663         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2664         cursor.key_beg.obj_type = 0;
2665
2666         cursor.key_end = cursor.key_beg;
2667         cursor.key_end.key |= 0xFFFFFFFFULL;
2668         cursor.asof = dip->obj_asof;
2669         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2670
2671         /*
2672          * Scan all matching records (the chain), locate the one matching
2673          * the requested path component.  info->last_error contains the
2674          * error code on search termination and could be 0, ENOENT, or
2675          * something else.
2676          *
2677          * The hammer_ip_*() functions merge in-memory records with on-disk
2678          * records for the purposes of the search.
2679          */
2680         error = hammer_ip_first(&cursor);
2681
2682         while (error == 0) {
2683                 error = hammer_ip_resolve_data(&cursor);
2684                 if (error)
2685                         break;
2686                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2687                 KKASSERT(nlen > 0);
2688                 if (ncp->nc_nlen == nlen &&
2689                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2690                         break;
2691                 }
2692                 error = hammer_ip_next(&cursor);
2693         }
2694
2695         /*
2696          * If all is ok we have to get the inode so we can adjust nlinks.
2697          * To avoid a deadlock with the flusher we must release the inode
2698          * lock on the directory when acquiring the inode for the entry.
2699          *
2700          * If the target is a directory, it must be empty.
2701          */
2702         if (error == 0) {
2703                 hammer_unlock(&cursor.ip->lock);
2704                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
2705                                       dip->hmp->asof,
2706                                       cursor.data->entry.localization,
2707                                       0, &error);
2708                 hammer_lock_sh(&cursor.ip->lock);
2709                 if (error == ENOENT) {
2710                         kprintf("obj_id %016llx\n", cursor.data->entry.obj_id);
2711                         Debugger("ENOENT unlinking object that should exist");
2712                 }
2713
2714                 /*
2715                  * If we are trying to remove a directory the directory must
2716                  * be empty.
2717                  *
2718                  * WARNING: hammer_ip_check_directory_empty() may have to
2719                  * terminate the cursor to avoid a deadlock.  It is ok to
2720                  * call hammer_done_cursor() twice.
2721                  */
2722                 if (error == 0 && ip->ino_data.obj_type ==
2723                                   HAMMER_OBJTYPE_DIRECTORY) {
2724                         error = hammer_ip_check_directory_empty(trans, ip);
2725                 }
2726
2727                 /*
2728                  * Delete the directory entry.
2729                  *
2730                  * WARNING: hammer_ip_del_directory() may have to terminate
2731                  * the cursor to avoid a deadlock.  It is ok to call
2732                  * hammer_done_cursor() twice.
2733                  */
2734                 if (error == 0) {
2735                         error = hammer_ip_del_directory(trans, &cursor,
2736                                                         dip, ip);
2737                 }
2738                 hammer_done_cursor(&cursor);
2739                 if (error == 0) {
2740                         cache_setunresolved(nch);
2741                         cache_setvp(nch, NULL);
2742                         /* XXX locking */
2743                         if (ip->vp)
2744                                 cache_inval_vp(ip->vp, CINV_DESTROY);
2745                 }
2746                 if (ip)
2747                         hammer_rel_inode(ip, 0);
2748         } else {
2749                 hammer_done_cursor(&cursor);
2750         }
2751         hammer_inode_waitreclaims(dip->hmp);
2752         if (error == EDEADLK)
2753                 goto retry;
2754
2755         return (error);
2756 }
2757
2758 /************************************************************************
2759  *                          FIFO AND SPECFS OPS                         *
2760  ************************************************************************
2761  *
2762  */
2763
2764 static int
2765 hammer_vop_fifoclose (struct vop_close_args *ap)
2766 {
2767         /* XXX update itimes */
2768         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
2769 }
2770
2771 static int
2772 hammer_vop_fiforead (struct vop_read_args *ap)
2773 {
2774         int error;
2775
2776         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2777         /* XXX update access time */
2778         return (error);
2779 }
2780
2781 static int
2782 hammer_vop_fifowrite (struct vop_write_args *ap)
2783 {
2784         int error;
2785
2786         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2787         /* XXX update access time */
2788         return (error);
2789 }
2790
2791 static int
2792 hammer_vop_specclose (struct vop_close_args *ap)
2793 {
2794         /* XXX update itimes */
2795         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2796 }
2797
2798 static int
2799 hammer_vop_specread (struct vop_read_args *ap)
2800 {
2801         /* XXX update access time */
2802         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2803 }
2804
2805 static int
2806 hammer_vop_specwrite (struct vop_write_args *ap)
2807 {
2808         /* XXX update last change time */
2809         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2810 }
2811