b73f67c0e8143cf44387fb0f401a646a7e804815
[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 1.102 2008/10/16 17:24:16 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 <sys/file.h>
48 #include <vm/vm_extern.h>
49 #include <vfs/fifofs/fifo.h>
50 #include "hammer.h"
51
52 /*
53  * USERFS VNOPS
54  */
55 /*static int hammer_vop_vnoperate(struct vop_generic_args *);*/
56 static int hammer_vop_fsync(struct vop_fsync_args *);
57 static int hammer_vop_read(struct vop_read_args *);
58 static int hammer_vop_write(struct vop_write_args *);
59 static int hammer_vop_access(struct vop_access_args *);
60 static int hammer_vop_advlock(struct vop_advlock_args *);
61 static int hammer_vop_close(struct vop_close_args *);
62 static int hammer_vop_ncreate(struct vop_ncreate_args *);
63 static int hammer_vop_getattr(struct vop_getattr_args *);
64 static int hammer_vop_nresolve(struct vop_nresolve_args *);
65 static int hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *);
66 static int hammer_vop_nlink(struct vop_nlink_args *);
67 static int hammer_vop_nmkdir(struct vop_nmkdir_args *);
68 static int hammer_vop_nmknod(struct vop_nmknod_args *);
69 static int hammer_vop_open(struct vop_open_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_markatime(struct vop_markatime_args *);
77 static int hammer_vop_setattr(struct vop_setattr_args *);
78 static int hammer_vop_strategy(struct vop_strategy_args *);
79 static int hammer_vop_bmap(struct vop_bmap_args *ap);
80 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
81 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
82 static int hammer_vop_ioctl(struct vop_ioctl_args *);
83 static int hammer_vop_mountctl(struct vop_mountctl_args *);
84 static int hammer_vop_kqfilter (struct vop_kqfilter_args *);
85
86 static int hammer_vop_fifoclose (struct vop_close_args *);
87 static int hammer_vop_fiforead (struct vop_read_args *);
88 static int hammer_vop_fifowrite (struct vop_write_args *);
89 static int hammer_vop_fifokqfilter (struct vop_kqfilter_args *);
90
91 static int hammer_vop_specclose (struct vop_close_args *);
92 static int hammer_vop_specread (struct vop_read_args *);
93 static int hammer_vop_specwrite (struct vop_write_args *);
94 static int hammer_vop_specgetattr (struct vop_getattr_args *);
95
96 struct vop_ops hammer_vnode_vops = {
97         .vop_default =          vop_defaultop,
98         .vop_fsync =            hammer_vop_fsync,
99         .vop_getpages =         vop_stdgetpages,
100         .vop_putpages =         vop_stdputpages,
101         .vop_read =             hammer_vop_read,
102         .vop_write =            hammer_vop_write,
103         .vop_access =           hammer_vop_access,
104         .vop_advlock =          hammer_vop_advlock,
105         .vop_close =            hammer_vop_close,
106         .vop_ncreate =          hammer_vop_ncreate,
107         .vop_getattr =          hammer_vop_getattr,
108         .vop_inactive =         hammer_vop_inactive,
109         .vop_reclaim =          hammer_vop_reclaim,
110         .vop_nresolve =         hammer_vop_nresolve,
111         .vop_nlookupdotdot =    hammer_vop_nlookupdotdot,
112         .vop_nlink =            hammer_vop_nlink,
113         .vop_nmkdir =           hammer_vop_nmkdir,
114         .vop_nmknod =           hammer_vop_nmknod,
115         .vop_open =             hammer_vop_open,
116         .vop_pathconf =         vop_stdpathconf,
117         .vop_print =            hammer_vop_print,
118         .vop_readdir =          hammer_vop_readdir,
119         .vop_readlink =         hammer_vop_readlink,
120         .vop_nremove =          hammer_vop_nremove,
121         .vop_nrename =          hammer_vop_nrename,
122         .vop_nrmdir =           hammer_vop_nrmdir,
123         .vop_markatime =        hammer_vop_markatime,
124         .vop_setattr =          hammer_vop_setattr,
125         .vop_bmap =             hammer_vop_bmap,
126         .vop_strategy =         hammer_vop_strategy,
127         .vop_nsymlink =         hammer_vop_nsymlink,
128         .vop_nwhiteout =        hammer_vop_nwhiteout,
129         .vop_ioctl =            hammer_vop_ioctl,
130         .vop_mountctl =         hammer_vop_mountctl,
131         .vop_kqfilter =         hammer_vop_kqfilter
132 };
133
134 struct vop_ops hammer_spec_vops = {
135         .vop_default =          spec_vnoperate,
136         .vop_fsync =            hammer_vop_fsync,
137         .vop_read =             hammer_vop_specread,
138         .vop_write =            hammer_vop_specwrite,
139         .vop_access =           hammer_vop_access,
140         .vop_close =            hammer_vop_specclose,
141         .vop_markatime =        hammer_vop_markatime,
142         .vop_getattr =          hammer_vop_specgetattr,
143         .vop_inactive =         hammer_vop_inactive,
144         .vop_reclaim =          hammer_vop_reclaim,
145         .vop_setattr =          hammer_vop_setattr
146 };
147
148 struct vop_ops hammer_fifo_vops = {
149         .vop_default =          fifo_vnoperate,
150         .vop_fsync =            hammer_vop_fsync,
151         .vop_read =             hammer_vop_fiforead,
152         .vop_write =            hammer_vop_fifowrite,
153         .vop_access =           hammer_vop_access,
154         .vop_close =            hammer_vop_fifoclose,
155         .vop_markatime =        hammer_vop_markatime,
156         .vop_getattr =          hammer_vop_getattr,
157         .vop_inactive =         hammer_vop_inactive,
158         .vop_reclaim =          hammer_vop_reclaim,
159         .vop_setattr =          hammer_vop_setattr,
160         .vop_kqfilter =         hammer_vop_fifokqfilter
161 };
162
163 static __inline
164 void
165 hammer_knote(struct vnode *vp, int flags)
166 {
167         if (flags)
168                 KNOTE(&vp->v_pollinfo.vpi_selinfo.si_note, flags);
169 }
170
171 #ifdef DEBUG_TRUNCATE
172 struct hammer_inode *HammerTruncIp;
173 #endif
174
175 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
176                            struct vnode *dvp, struct ucred *cred,
177                            int flags, int isdir);
178 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
179 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
180
181 #if 0
182 static
183 int
184 hammer_vop_vnoperate(struct vop_generic_args *)
185 {
186         return (VOCALL(&hammer_vnode_vops, ap));
187 }
188 #endif
189
190 /*
191  * hammer_vop_fsync { vp, waitfor }
192  *
193  * fsync() an inode to disk and wait for it to be completely committed
194  * such that the information would not be undone if a crash occured after
195  * return.
196  */
197 static
198 int
199 hammer_vop_fsync(struct vop_fsync_args *ap)
200 {
201         hammer_inode_t ip = VTOI(ap->a_vp);
202
203         ++hammer_count_fsyncs;
204         vfsync(ap->a_vp, ap->a_waitfor, 1, NULL, NULL);
205         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
206         if (ap->a_waitfor == MNT_WAIT) {
207                 vn_unlock(ap->a_vp);
208                 hammer_wait_inode(ip);
209                 vn_lock(ap->a_vp, LK_EXCLUSIVE | LK_RETRY);
210         }
211         return (ip->error);
212 }
213
214 /*
215  * hammer_vop_read { vp, uio, ioflag, cred }
216  */
217 static
218 int
219 hammer_vop_read(struct vop_read_args *ap)
220 {
221         struct hammer_transaction trans;
222         hammer_inode_t ip;
223         off_t offset;
224         struct buf *bp;
225         struct uio *uio;
226         int error;
227         int n;
228         int seqcount;
229         int ioseqcount;
230         int blksize;
231
232         if (ap->a_vp->v_type != VREG)
233                 return (EINVAL);
234         ip = VTOI(ap->a_vp);
235         error = 0;
236         uio = ap->a_uio;
237
238         /*
239          * Allow the UIO's size to override the sequential heuristic.
240          */
241         blksize = hammer_blocksize(uio->uio_offset);
242         seqcount = (uio->uio_resid + (blksize - 1)) / blksize;
243         ioseqcount = ap->a_ioflag >> 16;
244         if (seqcount < ioseqcount)
245                 seqcount = ioseqcount;
246
247         hammer_start_transaction(&trans, ip->hmp);
248
249         /*
250          * Access the data typically in HAMMER_BUFSIZE blocks via the
251          * buffer cache, but HAMMER may use a variable block size based
252          * on the offset.
253          */
254         while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
255                 int64_t base_offset;
256                 int64_t file_limit;
257
258                 blksize = hammer_blocksize(uio->uio_offset);
259                 offset = (int)uio->uio_offset & (blksize - 1);
260                 base_offset = uio->uio_offset - offset;
261
262                 if (hammer_cluster_enable) {
263                         /*
264                          * Use file_limit to prevent cluster_read() from
265                          * creating buffers of the wrong block size past
266                          * the demarc.
267                          */
268                         file_limit = ip->ino_data.size;
269                         if (base_offset < HAMMER_XDEMARC &&
270                             file_limit > HAMMER_XDEMARC) {
271                                 file_limit = HAMMER_XDEMARC;
272                         }
273                         error = cluster_read(ap->a_vp,
274                                              file_limit, base_offset,
275                                              blksize, MAXPHYS,
276                                              seqcount, &bp);
277                 } else {
278                         error = bread(ap->a_vp, base_offset, blksize, &bp);
279                 }
280                 if (error) {
281                         kprintf("error %d\n", error);
282                         brelse(bp);
283                         break;
284                 }
285
286                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
287                 n = blksize - offset;
288                 if (n > uio->uio_resid)
289                         n = uio->uio_resid;
290                 if (n > ip->ino_data.size - uio->uio_offset)
291                         n = (int)(ip->ino_data.size - uio->uio_offset);
292                 error = uiomove((char *)bp->b_data + offset, n, uio);
293
294                 /* data has a lower priority then meta-data */
295                 bp->b_flags |= B_AGE;
296                 bqrelse(bp);
297                 if (error)
298                         break;
299                 hammer_stats_file_read += n;
300         }
301         if ((ip->flags & HAMMER_INODE_RO) == 0 &&
302             (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
303                 ip->ino_data.atime = trans.time;
304                 hammer_modify_inode(ip, HAMMER_INODE_ATIME);
305         }
306         hammer_done_transaction(&trans);
307         return (error);
308 }
309
310 /*
311  * hammer_vop_write { vp, uio, ioflag, cred }
312  */
313 static
314 int
315 hammer_vop_write(struct vop_write_args *ap)
316 {
317         struct hammer_transaction trans;
318         struct hammer_inode *ip;
319         hammer_mount_t hmp;
320         struct uio *uio;
321         int offset;
322         off_t base_offset;
323         struct buf *bp;
324         int kflags;
325         int error;
326         int n;
327         int flags;
328         int delta;
329         int seqcount;
330
331         if (ap->a_vp->v_type != VREG)
332                 return (EINVAL);
333         ip = VTOI(ap->a_vp);
334         hmp = ip->hmp;
335         error = 0;
336         kflags = 0;
337         seqcount = ap->a_ioflag >> 16;
338
339         if (ip->flags & HAMMER_INODE_RO)
340                 return (EROFS);
341
342         /*
343          * Create a transaction to cover the operations we perform.
344          */
345         hammer_start_transaction(&trans, hmp);
346         uio = ap->a_uio;
347
348         /*
349          * Check append mode
350          */
351         if (ap->a_ioflag & IO_APPEND)
352                 uio->uio_offset = ip->ino_data.size;
353
354         /*
355          * Check for illegal write offsets.  Valid range is 0...2^63-1.
356          *
357          * NOTE: the base_off assignment is required to work around what
358          * I consider to be a GCC-4 optimization bug.
359          */
360         if (uio->uio_offset < 0) {
361                 hammer_done_transaction(&trans);
362                 return (EFBIG);
363         }
364         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
365         if (uio->uio_resid > 0 && base_offset <= 0) {
366                 hammer_done_transaction(&trans);
367                 return (EFBIG);
368         }
369
370         /*
371          * Access the data typically in HAMMER_BUFSIZE blocks via the
372          * buffer cache, but HAMMER may use a variable block size based
373          * on the offset.
374          */
375         while (uio->uio_resid > 0) {
376                 int fixsize = 0;
377                 int blksize;
378                 int blkmask;
379
380                 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
381                         break;
382
383                 blksize = hammer_blocksize(uio->uio_offset);
384
385                 /*
386                  * Do not allow HAMMER to blow out the buffer cache.  Very
387                  * large UIOs can lockout other processes due to bwillwrite()
388                  * mechanics.
389                  *
390                  * The hammer inode is not locked during these operations.
391                  * The vnode is locked which can interfere with the pageout
392                  * daemon for non-UIO_NOCOPY writes but should not interfere
393                  * with the buffer cache.  Even so, we cannot afford to
394                  * allow the pageout daemon to build up too many dirty buffer
395                  * cache buffers.
396                  *
397                  * Only call this if we aren't being recursively called from
398                  * a virtual disk device (vn), else we may deadlock.
399                  */
400                 if ((ap->a_ioflag & IO_RECURSE) == 0)
401                         bwillwrite(blksize);
402
403                 /*
404                  * Do not allow HAMMER to blow out system memory by
405                  * accumulating too many records.   Records are so well
406                  * decoupled from the buffer cache that it is possible
407                  * for userland to push data out to the media via
408                  * direct-write, but build up the records queued to the
409                  * backend faster then the backend can flush them out.
410                  * HAMMER has hit its write limit but the frontend has
411                  * no pushback to slow it down.
412                  */
413                 if (hmp->rsv_recs > hammer_limit_recs / 2) {
414                         /*
415                          * Get the inode on the flush list
416                          */
417                         if (ip->rsv_recs >= 64)
418                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
419                         else if (ip->rsv_recs >= 16)
420                                 hammer_flush_inode(ip, 0);
421
422                         /*
423                          * Keep the flusher going if the system keeps
424                          * queueing records.
425                          */
426                         delta = hmp->count_newrecords -
427                                 hmp->last_newrecords;
428                         if (delta < 0 || delta > hammer_limit_recs / 2) {
429                                 hmp->last_newrecords = hmp->count_newrecords;
430                                 hammer_sync_hmp(hmp, MNT_NOWAIT);
431                         }
432
433                         /*
434                          * If we have gotten behind start slowing
435                          * down the writers.
436                          */
437                         delta = (hmp->rsv_recs - hammer_limit_recs) *
438                                 hz / hammer_limit_recs;
439                         if (delta > 0)
440                                 tsleep(&trans, 0, "hmrslo", delta);
441                 }
442
443                 /*
444                  * Calculate the blocksize at the current offset and figure
445                  * out how much we can actually write.
446                  */
447                 blkmask = blksize - 1;
448                 offset = (int)uio->uio_offset & blkmask;
449                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
450                 n = blksize - offset;
451                 if (n > uio->uio_resid)
452                         n = uio->uio_resid;
453                 if (uio->uio_offset + n > ip->ino_data.size) {
454                         vnode_pager_setsize(ap->a_vp, uio->uio_offset + n);
455                         fixsize = 1;
456                         kflags |= NOTE_EXTEND;
457                 }
458
459                 if (uio->uio_segflg == UIO_NOCOPY) {
460                         /*
461                          * Issuing a write with the same data backing the
462                          * buffer.  Instantiate the buffer to collect the
463                          * backing vm pages, then read-in any missing bits.
464                          *
465                          * This case is used by vop_stdputpages().
466                          */
467                         bp = getblk(ap->a_vp, base_offset,
468                                     blksize, GETBLK_BHEAVY, 0);
469                         if ((bp->b_flags & B_CACHE) == 0) {
470                                 bqrelse(bp);
471                                 error = bread(ap->a_vp, base_offset,
472                                               blksize, &bp);
473                         }
474                 } else if (offset == 0 && uio->uio_resid >= blksize) {
475                         /*
476                          * Even though we are entirely overwriting the buffer
477                          * we may still have to zero it out to avoid a 
478                          * mmap/write visibility issue.
479                          */
480                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
481                         if ((bp->b_flags & B_CACHE) == 0)
482                                 vfs_bio_clrbuf(bp);
483                 } else if (base_offset >= ip->ino_data.size) {
484                         /*
485                          * If the base offset of the buffer is beyond the
486                          * file EOF, we don't have to issue a read.
487                          */
488                         bp = getblk(ap->a_vp, base_offset,
489                                     blksize, GETBLK_BHEAVY, 0);
490                         vfs_bio_clrbuf(bp);
491                 } else {
492                         /*
493                          * Partial overwrite, read in any missing bits then
494                          * replace the portion being written.
495                          */
496                         error = bread(ap->a_vp, base_offset, blksize, &bp);
497                         if (error == 0)
498                                 bheavy(bp);
499                 }
500                 if (error == 0) {
501                         error = uiomove((char *)bp->b_data + offset,
502                                         n, uio);
503                 }
504
505                 /*
506                  * If we screwed up we have to undo any VM size changes we
507                  * made.
508                  */
509                 if (error) {
510                         brelse(bp);
511                         if (fixsize) {
512                                 vtruncbuf(ap->a_vp, ip->ino_data.size,
513                                           hammer_blocksize(ip->ino_data.size));
514                         }
515                         break;
516                 }
517                 kflags |= NOTE_WRITE;
518                 hammer_stats_file_write += n;
519                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
520                 if (ip->ino_data.size < uio->uio_offset) {
521                         ip->ino_data.size = uio->uio_offset;
522                         flags = HAMMER_INODE_DDIRTY;
523                         vnode_pager_setsize(ap->a_vp, ip->ino_data.size);
524                 } else {
525                         flags = 0;
526                 }
527                 ip->ino_data.mtime = trans.time;
528                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
529                 hammer_modify_inode(ip, flags);
530
531                 /*
532                  * Once we dirty the buffer any cached zone-X offset
533                  * becomes invalid.  HAMMER NOTE: no-history mode cannot 
534                  * allow overwriting over the same data sector unless
535                  * we provide UNDOs for the old data, which we don't.
536                  */
537                 bp->b_bio2.bio_offset = NOOFFSET;
538
539                 /*
540                  * Final buffer disposition.
541                  */
542                 bp->b_flags |= B_AGE;
543                 if (ap->a_ioflag & IO_SYNC) {
544                         bwrite(bp);
545                 } else if (ap->a_ioflag & IO_DIRECT) {
546                         bawrite(bp);
547                 } else {
548                         bdwrite(bp);
549                 }
550         }
551         hammer_done_transaction(&trans);
552         hammer_knote(ap->a_vp, kflags);
553         return (error);
554 }
555
556 /*
557  * hammer_vop_access { vp, mode, cred }
558  */
559 static
560 int
561 hammer_vop_access(struct vop_access_args *ap)
562 {
563         struct hammer_inode *ip = VTOI(ap->a_vp);
564         uid_t uid;
565         gid_t gid;
566         int error;
567
568         ++hammer_stats_file_iopsr;
569         uid = hammer_to_unix_xid(&ip->ino_data.uid);
570         gid = hammer_to_unix_xid(&ip->ino_data.gid);
571
572         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
573                                   ip->ino_data.uflags);
574         return (error);
575 }
576
577 /*
578  * hammer_vop_advlock { vp, id, op, fl, flags }
579  */
580 static
581 int
582 hammer_vop_advlock(struct vop_advlock_args *ap)
583 {
584         hammer_inode_t ip = VTOI(ap->a_vp);
585
586         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
587 }
588
589 /*
590  * hammer_vop_close { vp, fflag }
591  */
592 static
593 int
594 hammer_vop_close(struct vop_close_args *ap)
595 {
596         /*hammer_inode_t ip = VTOI(ap->a_vp);*/
597         return (vop_stdclose(ap));
598 }
599
600 /*
601  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
602  *
603  * The operating system has already ensured that the directory entry
604  * does not exist and done all appropriate namespace locking.
605  */
606 static
607 int
608 hammer_vop_ncreate(struct vop_ncreate_args *ap)
609 {
610         struct hammer_transaction trans;
611         struct hammer_inode *dip;
612         struct hammer_inode *nip;
613         struct nchandle *nch;
614         int error;
615
616         nch = ap->a_nch;
617         dip = VTOI(ap->a_dvp);
618
619         if (dip->flags & HAMMER_INODE_RO)
620                 return (EROFS);
621         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
622                 return (error);
623
624         /*
625          * Create a transaction to cover the operations we perform.
626          */
627         hammer_start_transaction(&trans, dip->hmp);
628         ++hammer_stats_file_iopsw;
629
630         /*
631          * Create a new filesystem object of the requested type.  The
632          * returned inode will be referenced and shared-locked to prevent
633          * it from being moved to the flusher.
634          */
635         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
636                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
637                                     NULL, &nip);
638         if (error) {
639                 hkprintf("hammer_create_inode error %d\n", error);
640                 hammer_done_transaction(&trans);
641                 *ap->a_vpp = NULL;
642                 return (error);
643         }
644
645         /*
646          * Add the new filesystem object to the directory.  This will also
647          * bump the inode's link count.
648          */
649         error = hammer_ip_add_directory(&trans, dip,
650                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
651                                         nip);
652         if (error)
653                 hkprintf("hammer_ip_add_directory error %d\n", error);
654
655         /*
656          * Finish up.
657          */
658         if (error) {
659                 hammer_rel_inode(nip, 0);
660                 hammer_done_transaction(&trans);
661                 *ap->a_vpp = NULL;
662         } else {
663                 error = hammer_get_vnode(nip, ap->a_vpp);
664                 hammer_done_transaction(&trans);
665                 hammer_rel_inode(nip, 0);
666                 if (error == 0) {
667                         cache_setunresolved(ap->a_nch);
668                         cache_setvp(ap->a_nch, *ap->a_vpp);
669                 }
670                 hammer_knote(ap->a_dvp, NOTE_WRITE);
671         }
672         return (error);
673 }
674
675 /*
676  * hammer_vop_getattr { vp, vap }
677  *
678  * Retrieve an inode's attribute information.  When accessing inodes
679  * historically we fake the atime field to ensure consistent results.
680  * The atime field is stored in the B-Tree element and allowed to be
681  * updated without cycling the element.
682  */
683 static
684 int
685 hammer_vop_getattr(struct vop_getattr_args *ap)
686 {
687         struct hammer_inode *ip = VTOI(ap->a_vp);
688         struct vattr *vap = ap->a_vap;
689
690         /*
691          * We want the fsid to be different when accessing a filesystem
692          * with different as-of's so programs like diff don't think
693          * the files are the same.
694          *
695          * We also want the fsid to be the same when comparing snapshots,
696          * or when comparing mirrors (which might be backed by different
697          * physical devices).  HAMMER fsids are based on the PFS's
698          * shared_uuid field.
699          *
700          * XXX there is a chance of collision here.  The va_fsid reported
701          * by stat is different from the more involved fsid used in the
702          * mount structure.
703          */
704         ++hammer_stats_file_iopsr;
705         vap->va_fsid = ip->pfsm->fsid_udev ^ (u_int32_t)ip->obj_asof ^
706                        (u_int32_t)(ip->obj_asof >> 32);
707
708         vap->va_fileid = ip->ino_leaf.base.obj_id;
709         vap->va_mode = ip->ino_data.mode;
710         vap->va_nlink = ip->ino_data.nlinks;
711         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
712         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
713         vap->va_rmajor = 0;
714         vap->va_rminor = 0;
715         vap->va_size = ip->ino_data.size;
716
717         /*
718          * Special case for @@PFS softlinks.  The actual size of the
719          * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
720          * or for MAX_TID is    "@@-1:%05d" == 10 bytes.
721          */
722         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
723             ip->ino_data.size == 10 &&
724             ip->obj_asof == HAMMER_MAX_TID &&
725             ip->obj_localization == 0 &&
726             strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
727                     if (ip->pfsm->pfsd.mirror_flags & HAMMER_PFSD_SLAVE)
728                             vap->va_size = 26;
729                     else
730                             vap->va_size = 10;
731         }
732
733         /*
734          * We must provide a consistent atime and mtime for snapshots
735          * so people can do a 'tar cf - ... | md5' on them and get
736          * consistent results.
737          */
738         if (ip->flags & HAMMER_INODE_RO) {
739                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
740                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
741         } else {
742                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
743                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
744         }
745         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
746         vap->va_flags = ip->ino_data.uflags;
747         vap->va_gen = 1;        /* hammer inums are unique for all time */
748         vap->va_blocksize = HAMMER_BUFSIZE;
749         if (ip->ino_data.size >= HAMMER_XDEMARC) {
750                 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
751                                 ~HAMMER_XBUFMASK64;
752         } else if (ip->ino_data.size > HAMMER_BUFSIZE / 2) {
753                 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
754                                 ~HAMMER_BUFMASK64;
755         } else {
756                 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
757         }
758
759         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
760         vap->va_filerev = 0;    /* XXX */
761         /* mtime uniquely identifies any adjustments made to the file XXX */
762         vap->va_fsmid = ip->ino_data.mtime;
763         vap->va_uid_uuid = ip->ino_data.uid;
764         vap->va_gid_uuid = ip->ino_data.gid;
765         vap->va_fsid_uuid = ip->hmp->fsid;
766         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
767                           VA_FSID_UUID_VALID;
768
769         switch (ip->ino_data.obj_type) {
770         case HAMMER_OBJTYPE_CDEV:
771         case HAMMER_OBJTYPE_BDEV:
772                 vap->va_rmajor = ip->ino_data.rmajor;
773                 vap->va_rminor = ip->ino_data.rminor;
774                 break;
775         default:
776                 break;
777         }
778         return(0);
779 }
780
781 /*
782  * hammer_vop_nresolve { nch, dvp, cred }
783  *
784  * Locate the requested directory entry.
785  */
786 static
787 int
788 hammer_vop_nresolve(struct vop_nresolve_args *ap)
789 {
790         struct hammer_transaction trans;
791         struct namecache *ncp;
792         hammer_inode_t dip;
793         hammer_inode_t ip;
794         hammer_tid_t asof;
795         struct hammer_cursor cursor;
796         struct vnode *vp;
797         int64_t namekey;
798         int error;
799         int i;
800         int nlen;
801         int flags;
802         int ispfs;
803         int64_t obj_id;
804         u_int32_t localization;
805         u_int32_t max_iterations;
806
807         /*
808          * Misc initialization, plus handle as-of name extensions.  Look for
809          * the '@@' extension.  Note that as-of files and directories cannot
810          * be modified.
811          */
812         dip = VTOI(ap->a_dvp);
813         ncp = ap->a_nch->ncp;
814         asof = dip->obj_asof;
815         localization = dip->obj_localization;   /* for code consistency */
816         nlen = ncp->nc_nlen;
817         flags = dip->flags & HAMMER_INODE_RO;
818         ispfs = 0;
819
820         hammer_simple_transaction(&trans, dip->hmp);
821         ++hammer_stats_file_iopsr;
822
823         for (i = 0; i < nlen; ++i) {
824                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
825                         error = hammer_str_to_tid(ncp->nc_name + i + 2,
826                                                   &ispfs, &asof, &localization);
827                         if (error != 0) {
828                                 i = nlen;
829                                 break;
830                         }
831                         if (asof != HAMMER_MAX_TID)
832                                 flags |= HAMMER_INODE_RO;
833                         break;
834                 }
835         }
836         nlen = i;
837
838         /*
839          * If this is a PFS softlink we dive into the PFS
840          */
841         if (ispfs && nlen == 0) {
842                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
843                                       asof, localization,
844                                       flags, &error);
845                 if (error == 0) {
846                         error = hammer_get_vnode(ip, &vp);
847                         hammer_rel_inode(ip, 0);
848                 } else {
849                         vp = NULL;
850                 }
851                 if (error == 0) {
852                         vn_unlock(vp);
853                         cache_setvp(ap->a_nch, vp);
854                         vrele(vp);
855                 }
856                 goto done;
857         }
858
859         /*
860          * If there is no path component the time extension is relative to dip.
861          * e.g. "fubar/@@<snapshot>"
862          *
863          * "." is handled by the kernel, but ".@@<snapshot>" is not.
864          * e.g. "fubar/.@@<snapshot>"
865          *
866          * ".." is handled by the kernel.  We do not currently handle
867          * "..@<snapshot>".
868          */
869         if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
870                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
871                                       asof, dip->obj_localization,
872                                       flags, &error);
873                 if (error == 0) {
874                         error = hammer_get_vnode(ip, &vp);
875                         hammer_rel_inode(ip, 0);
876                 } else {
877                         vp = NULL;
878                 }
879                 if (error == 0) {
880                         vn_unlock(vp);
881                         cache_setvp(ap->a_nch, vp);
882                         vrele(vp);
883                 }
884                 goto done;
885         }
886
887         /*
888          * Calculate the namekey and setup the key range for the scan.  This
889          * works kinda like a chained hash table where the lower 32 bits
890          * of the namekey synthesize the chain.
891          *
892          * The key range is inclusive of both key_beg and key_end.
893          */
894         namekey = hammer_directory_namekey(dip, ncp->nc_name, nlen,
895                                            &max_iterations);
896
897         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
898         cursor.key_beg.localization = dip->obj_localization +
899                                       hammer_dir_localization(dip);
900         cursor.key_beg.obj_id = dip->obj_id;
901         cursor.key_beg.key = namekey;
902         cursor.key_beg.create_tid = 0;
903         cursor.key_beg.delete_tid = 0;
904         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
905         cursor.key_beg.obj_type = 0;
906
907         cursor.key_end = cursor.key_beg;
908         cursor.key_end.key += max_iterations;
909         cursor.asof = asof;
910         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
911
912         /*
913          * Scan all matching records (the chain), locate the one matching
914          * the requested path component.
915          *
916          * The hammer_ip_*() functions merge in-memory records with on-disk
917          * records for the purposes of the search.
918          */
919         obj_id = 0;
920         localization = HAMMER_DEF_LOCALIZATION;
921
922         if (error == 0) {
923                 error = hammer_ip_first(&cursor);
924                 while (error == 0) {
925                         error = hammer_ip_resolve_data(&cursor);
926                         if (error)
927                                 break;
928                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
929                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
930                                 obj_id = cursor.data->entry.obj_id;
931                                 localization = cursor.data->entry.localization;
932                                 break;
933                         }
934                         error = hammer_ip_next(&cursor);
935                 }
936         }
937         hammer_done_cursor(&cursor);
938
939         /*
940          * Lookup the obj_id.  This should always succeed.  If it does not
941          * the filesystem may be damaged and we return a dummy inode.
942          */
943         if (error == 0) {
944                 ip = hammer_get_inode(&trans, dip, obj_id,
945                                       asof, localization,
946                                       flags, &error);
947                 if (error == ENOENT) {
948                         kprintf("HAMMER: WARNING: Missing "
949                                 "inode for dirent \"%s\"\n"
950                                 "\tobj_id = %016llx\n",
951                                 ncp->nc_name, (long long)obj_id);
952                         error = 0;
953                         ip = hammer_get_dummy_inode(&trans, dip, obj_id,
954                                                     asof, localization,
955                                                     flags, &error);
956                 }
957                 if (error == 0) {
958                         error = hammer_get_vnode(ip, &vp);
959                         hammer_rel_inode(ip, 0);
960                 } else {
961                         vp = NULL;
962                 }
963                 if (error == 0) {
964                         vn_unlock(vp);
965                         cache_setvp(ap->a_nch, vp);
966                         vrele(vp);
967                 }
968         } else if (error == ENOENT) {
969                 cache_setvp(ap->a_nch, NULL);
970         }
971 done:
972         hammer_done_transaction(&trans);
973         return (error);
974 }
975
976 /*
977  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
978  *
979  * Locate the parent directory of a directory vnode.
980  *
981  * dvp is referenced but not locked.  *vpp must be returned referenced and
982  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
983  * at the root, instead it could indicate that the directory we were in was
984  * removed.
985  *
986  * NOTE: as-of sequences are not linked into the directory structure.  If
987  * we are at the root with a different asof then the mount point, reload
988  * the same directory with the mount point's asof.   I'm not sure what this
989  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
990  * get confused, but it hasn't been tested.
991  */
992 static
993 int
994 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
995 {
996         struct hammer_transaction trans;
997         struct hammer_inode *dip;
998         struct hammer_inode *ip;
999         int64_t parent_obj_id;
1000         u_int32_t parent_obj_localization;
1001         hammer_tid_t asof;
1002         int error;
1003
1004         dip = VTOI(ap->a_dvp);
1005         asof = dip->obj_asof;
1006
1007         /*
1008          * Whos are parent?  This could be the root of a pseudo-filesystem
1009          * whos parent is in another localization domain.
1010          */
1011         parent_obj_id = dip->ino_data.parent_obj_id;
1012         if (dip->obj_id == HAMMER_OBJID_ROOT)
1013                 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
1014         else
1015                 parent_obj_localization = dip->obj_localization;
1016
1017         if (parent_obj_id == 0) {
1018                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1019                    asof != dip->hmp->asof) {
1020                         parent_obj_id = dip->obj_id;
1021                         asof = dip->hmp->asof;
1022                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1023                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
1024                                   (long long)dip->obj_asof);
1025                 } else {
1026                         *ap->a_vpp = NULL;
1027                         return ENOENT;
1028                 }
1029         }
1030
1031         hammer_simple_transaction(&trans, dip->hmp);
1032         ++hammer_stats_file_iopsr;
1033
1034         ip = hammer_get_inode(&trans, dip, parent_obj_id,
1035                               asof, parent_obj_localization,
1036                               dip->flags, &error);
1037         if (ip) {
1038                 error = hammer_get_vnode(ip, ap->a_vpp);
1039                 hammer_rel_inode(ip, 0);
1040         } else {
1041                 *ap->a_vpp = NULL;
1042         }
1043         hammer_done_transaction(&trans);
1044         return (error);
1045 }
1046
1047 /*
1048  * hammer_vop_nlink { nch, dvp, vp, cred }
1049  */
1050 static
1051 int
1052 hammer_vop_nlink(struct vop_nlink_args *ap)
1053 {
1054         struct hammer_transaction trans;
1055         struct hammer_inode *dip;
1056         struct hammer_inode *ip;
1057         struct nchandle *nch;
1058         int error;
1059
1060         if (ap->a_dvp->v_mount != ap->a_vp->v_mount)    
1061                 return(EXDEV);
1062
1063         nch = ap->a_nch;
1064         dip = VTOI(ap->a_dvp);
1065         ip = VTOI(ap->a_vp);
1066
1067         if (dip->obj_localization != ip->obj_localization)
1068                 return(EXDEV);
1069
1070         if (dip->flags & HAMMER_INODE_RO)
1071                 return (EROFS);
1072         if (ip->flags & HAMMER_INODE_RO)
1073                 return (EROFS);
1074         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1075                 return (error);
1076
1077         /*
1078          * Create a transaction to cover the operations we perform.
1079          */
1080         hammer_start_transaction(&trans, dip->hmp);
1081         ++hammer_stats_file_iopsw;
1082
1083         /*
1084          * Add the filesystem object to the directory.  Note that neither
1085          * dip nor ip are referenced or locked, but their vnodes are
1086          * referenced.  This function will bump the inode's link count.
1087          */
1088         error = hammer_ip_add_directory(&trans, dip,
1089                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1090                                         ip);
1091
1092         /*
1093          * Finish up.
1094          */
1095         if (error == 0) {
1096                 cache_setunresolved(nch);
1097                 cache_setvp(nch, ap->a_vp);
1098         }
1099         hammer_done_transaction(&trans);
1100         hammer_knote(ap->a_vp, NOTE_LINK);
1101         hammer_knote(ap->a_dvp, NOTE_WRITE);
1102         return (error);
1103 }
1104
1105 /*
1106  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1107  *
1108  * The operating system has already ensured that the directory entry
1109  * does not exist and done all appropriate namespace locking.
1110  */
1111 static
1112 int
1113 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1114 {
1115         struct hammer_transaction trans;
1116         struct hammer_inode *dip;
1117         struct hammer_inode *nip;
1118         struct nchandle *nch;
1119         int error;
1120
1121         nch = ap->a_nch;
1122         dip = VTOI(ap->a_dvp);
1123
1124         if (dip->flags & HAMMER_INODE_RO)
1125                 return (EROFS);
1126         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1127                 return (error);
1128
1129         /*
1130          * Create a transaction to cover the operations we perform.
1131          */
1132         hammer_start_transaction(&trans, dip->hmp);
1133         ++hammer_stats_file_iopsw;
1134
1135         /*
1136          * Create a new filesystem object of the requested type.  The
1137          * returned inode will be referenced but not locked.
1138          */
1139         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1140                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1141                                     NULL, &nip);
1142         if (error) {
1143                 hkprintf("hammer_mkdir error %d\n", error);
1144                 hammer_done_transaction(&trans);
1145                 *ap->a_vpp = NULL;
1146                 return (error);
1147         }
1148         /*
1149          * Add the new filesystem object to the directory.  This will also
1150          * bump the inode's link count.
1151          */
1152         error = hammer_ip_add_directory(&trans, dip,
1153                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1154                                         nip);
1155         if (error)
1156                 hkprintf("hammer_mkdir (add) error %d\n", error);
1157
1158         /*
1159          * Finish up.
1160          */
1161         if (error) {
1162                 hammer_rel_inode(nip, 0);
1163                 *ap->a_vpp = NULL;
1164         } else {
1165                 error = hammer_get_vnode(nip, ap->a_vpp);
1166                 hammer_rel_inode(nip, 0);
1167                 if (error == 0) {
1168                         cache_setunresolved(ap->a_nch);
1169                         cache_setvp(ap->a_nch, *ap->a_vpp);
1170                 }
1171         }
1172         hammer_done_transaction(&trans);
1173         if (error == 0)
1174                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1175         return (error);
1176 }
1177
1178 /*
1179  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1180  *
1181  * The operating system has already ensured that the directory entry
1182  * does not exist and done all appropriate namespace locking.
1183  */
1184 static
1185 int
1186 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1187 {
1188         struct hammer_transaction trans;
1189         struct hammer_inode *dip;
1190         struct hammer_inode *nip;
1191         struct nchandle *nch;
1192         int error;
1193
1194         nch = ap->a_nch;
1195         dip = VTOI(ap->a_dvp);
1196
1197         if (dip->flags & HAMMER_INODE_RO)
1198                 return (EROFS);
1199         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1200                 return (error);
1201
1202         /*
1203          * Create a transaction to cover the operations we perform.
1204          */
1205         hammer_start_transaction(&trans, dip->hmp);
1206         ++hammer_stats_file_iopsw;
1207
1208         /*
1209          * Create a new filesystem object of the requested type.  The
1210          * returned inode will be referenced but not locked.
1211          *
1212          * If mknod specifies a directory a pseudo-fs is created.
1213          */
1214         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1215                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1216                                     NULL, &nip);
1217         if (error) {
1218                 hammer_done_transaction(&trans);
1219                 *ap->a_vpp = NULL;
1220                 return (error);
1221         }
1222
1223         /*
1224          * Add the new filesystem object to the directory.  This will also
1225          * bump the inode's link count.
1226          */
1227         error = hammer_ip_add_directory(&trans, dip,
1228                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1229                                         nip);
1230
1231         /*
1232          * Finish up.
1233          */
1234         if (error) {
1235                 hammer_rel_inode(nip, 0);
1236                 *ap->a_vpp = NULL;
1237         } else {
1238                 error = hammer_get_vnode(nip, ap->a_vpp);
1239                 hammer_rel_inode(nip, 0);
1240                 if (error == 0) {
1241                         cache_setunresolved(ap->a_nch);
1242                         cache_setvp(ap->a_nch, *ap->a_vpp);
1243                 }
1244         }
1245         hammer_done_transaction(&trans);
1246         if (error == 0)
1247                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1248         return (error);
1249 }
1250
1251 /*
1252  * hammer_vop_open { vp, mode, cred, fp }
1253  */
1254 static
1255 int
1256 hammer_vop_open(struct vop_open_args *ap)
1257 {
1258         hammer_inode_t ip;
1259
1260         ++hammer_stats_file_iopsr;
1261         ip = VTOI(ap->a_vp);
1262
1263         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1264                 return (EROFS);
1265         return(vop_stdopen(ap));
1266 }
1267
1268 /*
1269  * hammer_vop_print { vp }
1270  */
1271 static
1272 int
1273 hammer_vop_print(struct vop_print_args *ap)
1274 {
1275         return EOPNOTSUPP;
1276 }
1277
1278 /*
1279  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1280  */
1281 static
1282 int
1283 hammer_vop_readdir(struct vop_readdir_args *ap)
1284 {
1285         struct hammer_transaction trans;
1286         struct hammer_cursor cursor;
1287         struct hammer_inode *ip;
1288         struct uio *uio;
1289         hammer_base_elm_t base;
1290         int error;
1291         int cookie_index;
1292         int ncookies;
1293         off_t *cookies;
1294         off_t saveoff;
1295         int r;
1296         int dtype;
1297
1298         ++hammer_stats_file_iopsr;
1299         ip = VTOI(ap->a_vp);
1300         uio = ap->a_uio;
1301         saveoff = uio->uio_offset;
1302
1303         if (ap->a_ncookies) {
1304                 ncookies = uio->uio_resid / 16 + 1;
1305                 if (ncookies > 1024)
1306                         ncookies = 1024;
1307                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1308                 cookie_index = 0;
1309         } else {
1310                 ncookies = -1;
1311                 cookies = NULL;
1312                 cookie_index = 0;
1313         }
1314
1315         hammer_simple_transaction(&trans, ip->hmp);
1316
1317         /*
1318          * Handle artificial entries
1319          *
1320          * It should be noted that the minimum value for a directory
1321          * hash key on-media is 0x0000000100000000, so we can use anything
1322          * less then that to represent our 'special' key space.
1323          */
1324         error = 0;
1325         if (saveoff == 0) {
1326                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1327                 if (r)
1328                         goto done;
1329                 if (cookies)
1330                         cookies[cookie_index] = saveoff;
1331                 ++saveoff;
1332                 ++cookie_index;
1333                 if (cookie_index == ncookies)
1334                         goto done;
1335         }
1336         if (saveoff == 1) {
1337                 if (ip->ino_data.parent_obj_id) {
1338                         r = vop_write_dirent(&error, uio,
1339                                              ip->ino_data.parent_obj_id,
1340                                              DT_DIR, 2, "..");
1341                 } else {
1342                         r = vop_write_dirent(&error, uio,
1343                                              ip->obj_id, DT_DIR, 2, "..");
1344                 }
1345                 if (r)
1346                         goto done;
1347                 if (cookies)
1348                         cookies[cookie_index] = saveoff;
1349                 ++saveoff;
1350                 ++cookie_index;
1351                 if (cookie_index == ncookies)
1352                         goto done;
1353         }
1354
1355         /*
1356          * Key range (begin and end inclusive) to scan.  Directory keys
1357          * directly translate to a 64 bit 'seek' position.
1358          */
1359         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1360         cursor.key_beg.localization = ip->obj_localization +
1361                                       hammer_dir_localization(ip);
1362         cursor.key_beg.obj_id = ip->obj_id;
1363         cursor.key_beg.create_tid = 0;
1364         cursor.key_beg.delete_tid = 0;
1365         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1366         cursor.key_beg.obj_type = 0;
1367         cursor.key_beg.key = saveoff;
1368
1369         cursor.key_end = cursor.key_beg;
1370         cursor.key_end.key = HAMMER_MAX_KEY;
1371         cursor.asof = ip->obj_asof;
1372         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1373
1374         error = hammer_ip_first(&cursor);
1375
1376         while (error == 0) {
1377                 error = hammer_ip_resolve_data(&cursor);
1378                 if (error)
1379                         break;
1380                 base = &cursor.leaf->base;
1381                 saveoff = base->key;
1382                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1383
1384                 if (base->obj_id != ip->obj_id)
1385                         panic("readdir: bad record at %p", cursor.node);
1386
1387                 /*
1388                  * Convert pseudo-filesystems into softlinks
1389                  */
1390                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1391                 r = vop_write_dirent(
1392                              &error, uio, cursor.data->entry.obj_id,
1393                              dtype,
1394                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1395                              (void *)cursor.data->entry.name);
1396                 if (r)
1397                         break;
1398                 ++saveoff;
1399                 if (cookies)
1400                         cookies[cookie_index] = base->key;
1401                 ++cookie_index;
1402                 if (cookie_index == ncookies)
1403                         break;
1404                 error = hammer_ip_next(&cursor);
1405         }
1406         hammer_done_cursor(&cursor);
1407
1408 done:
1409         hammer_done_transaction(&trans);
1410
1411         if (ap->a_eofflag)
1412                 *ap->a_eofflag = (error == ENOENT);
1413         uio->uio_offset = saveoff;
1414         if (error && cookie_index == 0) {
1415                 if (error == ENOENT)
1416                         error = 0;
1417                 if (cookies) {
1418                         kfree(cookies, M_TEMP);
1419                         *ap->a_ncookies = 0;
1420                         *ap->a_cookies = NULL;
1421                 }
1422         } else {
1423                 if (error == ENOENT)
1424                         error = 0;
1425                 if (cookies) {
1426                         *ap->a_ncookies = cookie_index;
1427                         *ap->a_cookies = cookies;
1428                 }
1429         }
1430         return(error);
1431 }
1432
1433 /*
1434  * hammer_vop_readlink { vp, uio, cred }
1435  */
1436 static
1437 int
1438 hammer_vop_readlink(struct vop_readlink_args *ap)
1439 {
1440         struct hammer_transaction trans;
1441         struct hammer_cursor cursor;
1442         struct hammer_inode *ip;
1443         char buf[32];
1444         u_int32_t localization;
1445         hammer_pseudofs_inmem_t pfsm;
1446         int error;
1447
1448         ip = VTOI(ap->a_vp);
1449
1450         /*
1451          * Shortcut if the symlink data was stuffed into ino_data.
1452          *
1453          * Also expand special "@@PFS%05d" softlinks (expansion only
1454          * occurs for non-historical (current) accesses made from the
1455          * primary filesystem).
1456          */
1457         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1458                 char *ptr;
1459                 int bytes;
1460
1461                 ptr = ip->ino_data.ext.symlink;
1462                 bytes = (int)ip->ino_data.size;
1463                 if (bytes == 10 &&
1464                     ip->obj_asof == HAMMER_MAX_TID &&
1465                     ip->obj_localization == 0 &&
1466                     strncmp(ptr, "@@PFS", 5) == 0) {
1467                         hammer_simple_transaction(&trans, ip->hmp);
1468                         bcopy(ptr + 5, buf, 5);
1469                         buf[5] = 0;
1470                         localization = strtoul(buf, NULL, 10) << 16;
1471                         pfsm = hammer_load_pseudofs(&trans, localization,
1472                                                     &error);
1473                         if (error == 0) {
1474                                 if (pfsm->pfsd.mirror_flags &
1475                                     HAMMER_PFSD_SLAVE) {
1476                                         /* vap->va_size == 26 */
1477                                         ksnprintf(buf, sizeof(buf),
1478                                                   "@@0x%016llx:%05d",
1479                                                   (long long)pfsm->pfsd.sync_end_tid,
1480                                                   localization >> 16);
1481                                 } else {
1482                                         /* vap->va_size == 10 */
1483                                         ksnprintf(buf, sizeof(buf),
1484                                                   "@@-1:%05d",
1485                                                   localization >> 16);
1486 #if 0
1487                                         ksnprintf(buf, sizeof(buf),
1488                                                   "@@0x%016llx:%05d",
1489                                                   (long long)HAMMER_MAX_TID,
1490                                                   localization >> 16);
1491 #endif
1492                                 }
1493                                 ptr = buf;
1494                                 bytes = strlen(buf);
1495                         }
1496                         if (pfsm)
1497                                 hammer_rel_pseudofs(trans.hmp, pfsm);
1498                         hammer_done_transaction(&trans);
1499                 }
1500                 error = uiomove(ptr, bytes, ap->a_uio);
1501                 return(error);
1502         }
1503
1504         /*
1505          * Long version
1506          */
1507         hammer_simple_transaction(&trans, ip->hmp);
1508         ++hammer_stats_file_iopsr;
1509         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1510
1511         /*
1512          * Key range (begin and end inclusive) to scan.  Directory keys
1513          * directly translate to a 64 bit 'seek' position.
1514          */
1515         cursor.key_beg.localization = ip->obj_localization +
1516                                       HAMMER_LOCALIZE_MISC;
1517         cursor.key_beg.obj_id = ip->obj_id;
1518         cursor.key_beg.create_tid = 0;
1519         cursor.key_beg.delete_tid = 0;
1520         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1521         cursor.key_beg.obj_type = 0;
1522         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1523         cursor.asof = ip->obj_asof;
1524         cursor.flags |= HAMMER_CURSOR_ASOF;
1525
1526         error = hammer_ip_lookup(&cursor);
1527         if (error == 0) {
1528                 error = hammer_ip_resolve_data(&cursor);
1529                 if (error == 0) {
1530                         KKASSERT(cursor.leaf->data_len >=
1531                                  HAMMER_SYMLINK_NAME_OFF);
1532                         error = uiomove(cursor.data->symlink.name,
1533                                         cursor.leaf->data_len -
1534                                                 HAMMER_SYMLINK_NAME_OFF,
1535                                         ap->a_uio);
1536                 }
1537         }
1538         hammer_done_cursor(&cursor);
1539         hammer_done_transaction(&trans);
1540         return(error);
1541 }
1542
1543 /*
1544  * hammer_vop_nremove { nch, dvp, cred }
1545  */
1546 static
1547 int
1548 hammer_vop_nremove(struct vop_nremove_args *ap)
1549 {
1550         struct hammer_transaction trans;
1551         struct hammer_inode *dip;
1552         int error;
1553
1554         dip = VTOI(ap->a_dvp);
1555
1556         if (hammer_nohistory(dip) == 0 &&
1557             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1558                 return (error);
1559         }
1560
1561         hammer_start_transaction(&trans, dip->hmp);
1562         ++hammer_stats_file_iopsw;
1563         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1564         hammer_done_transaction(&trans);
1565         if (error == 0)
1566                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1567         return (error);
1568 }
1569
1570 /*
1571  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1572  */
1573 static
1574 int
1575 hammer_vop_nrename(struct vop_nrename_args *ap)
1576 {
1577         struct hammer_transaction trans;
1578         struct namecache *fncp;
1579         struct namecache *tncp;
1580         struct hammer_inode *fdip;
1581         struct hammer_inode *tdip;
1582         struct hammer_inode *ip;
1583         struct hammer_cursor cursor;
1584         int64_t namekey;
1585         u_int32_t max_iterations;
1586         int nlen, error;
1587
1588         if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount) 
1589                 return(EXDEV);
1590         if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1591                 return(EXDEV);
1592
1593         fdip = VTOI(ap->a_fdvp);
1594         tdip = VTOI(ap->a_tdvp);
1595         fncp = ap->a_fnch->ncp;
1596         tncp = ap->a_tnch->ncp;
1597         ip = VTOI(fncp->nc_vp);
1598         KKASSERT(ip != NULL);
1599
1600         if (fdip->obj_localization != tdip->obj_localization)
1601                 return(EXDEV);
1602         if (fdip->obj_localization != ip->obj_localization)
1603                 return(EXDEV);
1604
1605         if (fdip->flags & HAMMER_INODE_RO)
1606                 return (EROFS);
1607         if (tdip->flags & HAMMER_INODE_RO)
1608                 return (EROFS);
1609         if (ip->flags & HAMMER_INODE_RO)
1610                 return (EROFS);
1611         if ((error = hammer_checkspace(fdip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1612                 return (error);
1613
1614         hammer_start_transaction(&trans, fdip->hmp);
1615         ++hammer_stats_file_iopsw;
1616
1617         /*
1618          * Remove tncp from the target directory and then link ip as
1619          * tncp. XXX pass trans to dounlink
1620          *
1621          * Force the inode sync-time to match the transaction so it is
1622          * in-sync with the creation of the target directory entry.
1623          */
1624         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
1625                                 ap->a_cred, 0, -1);
1626         if (error == 0 || error == ENOENT) {
1627                 error = hammer_ip_add_directory(&trans, tdip,
1628                                                 tncp->nc_name, tncp->nc_nlen,
1629                                                 ip);
1630                 if (error == 0) {
1631                         ip->ino_data.parent_obj_id = tdip->obj_id;
1632                         ip->ino_data.ctime = trans.time;
1633                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
1634                 }
1635         }
1636         if (error)
1637                 goto failed; /* XXX */
1638
1639         /*
1640          * Locate the record in the originating directory and remove it.
1641          *
1642          * Calculate the namekey and setup the key range for the scan.  This
1643          * works kinda like a chained hash table where the lower 32 bits
1644          * of the namekey synthesize the chain.
1645          *
1646          * The key range is inclusive of both key_beg and key_end.
1647          */
1648         namekey = hammer_directory_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
1649                                            &max_iterations);
1650 retry:
1651         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
1652         cursor.key_beg.localization = fdip->obj_localization +
1653                                       hammer_dir_localization(fdip);
1654         cursor.key_beg.obj_id = fdip->obj_id;
1655         cursor.key_beg.key = namekey;
1656         cursor.key_beg.create_tid = 0;
1657         cursor.key_beg.delete_tid = 0;
1658         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1659         cursor.key_beg.obj_type = 0;
1660
1661         cursor.key_end = cursor.key_beg;
1662         cursor.key_end.key += max_iterations;
1663         cursor.asof = fdip->obj_asof;
1664         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1665
1666         /*
1667          * Scan all matching records (the chain), locate the one matching
1668          * the requested path component.
1669          *
1670          * The hammer_ip_*() functions merge in-memory records with on-disk
1671          * records for the purposes of the search.
1672          */
1673         error = hammer_ip_first(&cursor);
1674         while (error == 0) {
1675                 if (hammer_ip_resolve_data(&cursor) != 0)
1676                         break;
1677                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
1678                 KKASSERT(nlen > 0);
1679                 if (fncp->nc_nlen == nlen &&
1680                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1681                         break;
1682                 }
1683                 error = hammer_ip_next(&cursor);
1684         }
1685
1686         /*
1687          * If all is ok we have to get the inode so we can adjust nlinks.
1688          *
1689          * WARNING: hammer_ip_del_directory() may have to terminate the
1690          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
1691          * twice.
1692          */
1693         if (error == 0)
1694                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
1695
1696         /*
1697          * XXX A deadlock here will break rename's atomicy for the purposes
1698          * of crash recovery.
1699          */
1700         if (error == EDEADLK) {
1701                 hammer_done_cursor(&cursor);
1702                 goto retry;
1703         }
1704
1705         /*
1706          * Cleanup and tell the kernel that the rename succeeded.
1707          */
1708         hammer_done_cursor(&cursor);
1709         if (error == 0) {
1710                 cache_rename(ap->a_fnch, ap->a_tnch);
1711                 hammer_knote(ap->a_fdvp, NOTE_WRITE);
1712                 hammer_knote(ap->a_tdvp, NOTE_WRITE);
1713                 if (ip->vp)
1714                         hammer_knote(ip->vp, NOTE_RENAME);
1715         }
1716
1717 failed:
1718         hammer_done_transaction(&trans);
1719         return (error);
1720 }
1721
1722 /*
1723  * hammer_vop_nrmdir { nch, dvp, cred }
1724  */
1725 static
1726 int
1727 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
1728 {
1729         struct hammer_transaction trans;
1730         struct hammer_inode *dip;
1731         int error;
1732
1733         dip = VTOI(ap->a_dvp);
1734
1735         if (hammer_nohistory(dip) == 0 &&
1736             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1737                 return (error);
1738         }
1739
1740         hammer_start_transaction(&trans, dip->hmp);
1741         ++hammer_stats_file_iopsw;
1742         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
1743         hammer_done_transaction(&trans);
1744         if (error == 0)
1745                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1746         return (error);
1747 }
1748
1749 /*
1750  * hammer_vop_markatime { vp, cred }
1751  */
1752 static
1753 int
1754 hammer_vop_markatime(struct vop_markatime_args *ap)
1755 {
1756         struct hammer_transaction trans;
1757         struct hammer_inode *ip;
1758
1759         ip = VTOI(ap->a_vp);
1760         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
1761                 return (EROFS);
1762         if (ip->flags & HAMMER_INODE_RO)
1763                 return (EROFS);
1764         if (ip->hmp->mp->mnt_flag & MNT_NOATIME)
1765                 return (0);
1766         hammer_start_transaction(&trans, ip->hmp);
1767         ++hammer_stats_file_iopsw;
1768
1769         ip->ino_data.atime = trans.time;
1770         hammer_modify_inode(ip, HAMMER_INODE_ATIME);
1771         hammer_done_transaction(&trans);
1772         hammer_knote(ap->a_vp, NOTE_ATTRIB);
1773         return (0);
1774 }
1775
1776 /*
1777  * hammer_vop_setattr { vp, vap, cred }
1778  */
1779 static
1780 int
1781 hammer_vop_setattr(struct vop_setattr_args *ap)
1782 {
1783         struct hammer_transaction trans;
1784         struct vattr *vap;
1785         struct hammer_inode *ip;
1786         int modflags;
1787         int error;
1788         int truncating;
1789         int blksize;
1790         int kflags;
1791         int64_t aligned_size;
1792         u_int32_t flags;
1793
1794         vap = ap->a_vap;
1795         ip = ap->a_vp->v_data;
1796         modflags = 0;
1797         kflags = 0;
1798
1799         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
1800                 return(EROFS);
1801         if (ip->flags & HAMMER_INODE_RO)
1802                 return (EROFS);
1803         if (hammer_nohistory(ip) == 0 &&
1804             (error = hammer_checkspace(ip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1805                 return (error);
1806         }
1807
1808         hammer_start_transaction(&trans, ip->hmp);
1809         ++hammer_stats_file_iopsw;
1810         error = 0;
1811
1812         if (vap->va_flags != VNOVAL) {
1813                 flags = ip->ino_data.uflags;
1814                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
1815                                          hammer_to_unix_xid(&ip->ino_data.uid),
1816                                          ap->a_cred);
1817                 if (error == 0) {
1818                         if (ip->ino_data.uflags != flags) {
1819                                 ip->ino_data.uflags = flags;
1820                                 ip->ino_data.ctime = trans.time;
1821                                 modflags |= HAMMER_INODE_DDIRTY;
1822                                 kflags |= NOTE_ATTRIB;
1823                         }
1824                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
1825                                 error = 0;
1826                                 goto done;
1827                         }
1828                 }
1829                 goto done;
1830         }
1831         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
1832                 error = EPERM;
1833                 goto done;
1834         }
1835         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
1836                 mode_t cur_mode = ip->ino_data.mode;
1837                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1838                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1839                 uuid_t uuid_uid;
1840                 uuid_t uuid_gid;
1841
1842                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
1843                                          ap->a_cred,
1844                                          &cur_uid, &cur_gid, &cur_mode);
1845                 if (error == 0) {
1846                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
1847                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
1848                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
1849                                  sizeof(uuid_uid)) ||
1850                             bcmp(&uuid_gid, &ip->ino_data.gid,
1851                                  sizeof(uuid_gid)) ||
1852                             ip->ino_data.mode != cur_mode
1853                         ) {
1854                                 ip->ino_data.uid = uuid_uid;
1855                                 ip->ino_data.gid = uuid_gid;
1856                                 ip->ino_data.mode = cur_mode;
1857                                 ip->ino_data.ctime = trans.time;
1858                                 modflags |= HAMMER_INODE_DDIRTY;
1859                         }
1860                         kflags |= NOTE_ATTRIB;
1861                 }
1862         }
1863         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
1864                 switch(ap->a_vp->v_type) {
1865                 case VREG:
1866                         if (vap->va_size == ip->ino_data.size)
1867                                 break;
1868                         /*
1869                          * XXX break atomicy, we can deadlock the backend
1870                          * if we do not release the lock.  Probably not a
1871                          * big deal here.
1872                          */
1873                         blksize = hammer_blocksize(vap->va_size);
1874                         if (vap->va_size < ip->ino_data.size) {
1875                                 vtruncbuf(ap->a_vp, vap->va_size, blksize);
1876                                 truncating = 1;
1877                                 kflags |= NOTE_WRITE;
1878                         } else {
1879                                 vnode_pager_setsize(ap->a_vp, vap->va_size);
1880                                 truncating = 0;
1881                                 kflags |= NOTE_WRITE | NOTE_EXTEND;
1882                         }
1883                         ip->ino_data.size = vap->va_size;
1884                         ip->ino_data.mtime = trans.time;
1885                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
1886
1887                         /*
1888                          * on-media truncation is cached in the inode until
1889                          * the inode is synchronized.
1890                          */
1891                         if (truncating) {
1892                                 hammer_ip_frontend_trunc(ip, vap->va_size);
1893 #ifdef DEBUG_TRUNCATE
1894                                 if (HammerTruncIp == NULL)
1895                                         HammerTruncIp = ip;
1896 #endif
1897                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
1898                                         ip->flags |= HAMMER_INODE_TRUNCATED;
1899                                         ip->trunc_off = vap->va_size;
1900 #ifdef DEBUG_TRUNCATE
1901                                         if (ip == HammerTruncIp)
1902                                         kprintf("truncate1 %016llx\n",
1903                                                 (long long)ip->trunc_off);
1904 #endif
1905                                 } else if (ip->trunc_off > vap->va_size) {
1906                                         ip->trunc_off = vap->va_size;
1907 #ifdef DEBUG_TRUNCATE
1908                                         if (ip == HammerTruncIp)
1909                                         kprintf("truncate2 %016llx\n",
1910                                                 (long long)ip->trunc_off);
1911 #endif
1912                                 } else {
1913 #ifdef DEBUG_TRUNCATE
1914                                         if (ip == HammerTruncIp)
1915                                         kprintf("truncate3 %016llx (ignored)\n",
1916                                                 (long long)vap->va_size);
1917 #endif
1918                                 }
1919                         }
1920
1921                         /*
1922                          * If truncating we have to clean out a portion of
1923                          * the last block on-disk.  We do this in the
1924                          * front-end buffer cache.
1925                          */
1926                         aligned_size = (vap->va_size + (blksize - 1)) &
1927                                        ~(int64_t)(blksize - 1);
1928                         if (truncating && vap->va_size < aligned_size) {
1929                                 struct buf *bp;
1930                                 int offset;
1931
1932                                 aligned_size -= blksize;
1933
1934                                 offset = (int)vap->va_size & (blksize - 1);
1935                                 error = bread(ap->a_vp, aligned_size,
1936                                               blksize, &bp);
1937                                 hammer_ip_frontend_trunc(ip, aligned_size);
1938                                 if (error == 0) {
1939                                         bzero(bp->b_data + offset,
1940                                               blksize - offset);
1941                                         /* must de-cache direct-io offset */
1942                                         bp->b_bio2.bio_offset = NOOFFSET;
1943                                         bdwrite(bp);
1944                                 } else {
1945                                         kprintf("ERROR %d\n", error);
1946                                         brelse(bp);
1947                                 }
1948                         }
1949                         break;
1950                 case VDATABASE:
1951                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
1952                                 ip->flags |= HAMMER_INODE_TRUNCATED;
1953                                 ip->trunc_off = vap->va_size;
1954                         } else if (ip->trunc_off > vap->va_size) {
1955                                 ip->trunc_off = vap->va_size;
1956                         }
1957                         hammer_ip_frontend_trunc(ip, vap->va_size);
1958                         ip->ino_data.size = vap->va_size;
1959                         ip->ino_data.mtime = trans.time;
1960                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
1961                         kflags |= NOTE_ATTRIB;
1962                         break;
1963                 default:
1964                         error = EINVAL;
1965                         goto done;
1966                 }
1967                 break;
1968         }
1969         if (vap->va_atime.tv_sec != VNOVAL) {
1970                 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
1971                 modflags |= HAMMER_INODE_ATIME;
1972                 kflags |= NOTE_ATTRIB;
1973         }
1974         if (vap->va_mtime.tv_sec != VNOVAL) {
1975                 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
1976                 modflags |= HAMMER_INODE_MTIME;
1977                 kflags |= NOTE_ATTRIB;
1978         }
1979         if (vap->va_mode != (mode_t)VNOVAL) {
1980                 mode_t   cur_mode = ip->ino_data.mode;
1981                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1982                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1983
1984                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
1985                                          cur_uid, cur_gid, &cur_mode);
1986                 if (error == 0 && ip->ino_data.mode != cur_mode) {
1987                         ip->ino_data.mode = cur_mode;
1988                         ip->ino_data.ctime = trans.time;
1989                         modflags |= HAMMER_INODE_DDIRTY;
1990                         kflags |= NOTE_ATTRIB;
1991                 }
1992         }
1993 done:
1994         if (error == 0)
1995                 hammer_modify_inode(ip, modflags);
1996         hammer_done_transaction(&trans);
1997         hammer_knote(ap->a_vp, kflags);
1998         return (error);
1999 }
2000
2001 /*
2002  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2003  */
2004 static
2005 int
2006 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2007 {
2008         struct hammer_transaction trans;
2009         struct hammer_inode *dip;
2010         struct hammer_inode *nip;
2011         struct nchandle *nch;
2012         hammer_record_t record;
2013         int error;
2014         int bytes;
2015
2016         ap->a_vap->va_type = VLNK;
2017
2018         nch = ap->a_nch;
2019         dip = VTOI(ap->a_dvp);
2020
2021         if (dip->flags & HAMMER_INODE_RO)
2022                 return (EROFS);
2023         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
2024                 return (error);
2025
2026         /*
2027          * Create a transaction to cover the operations we perform.
2028          */
2029         hammer_start_transaction(&trans, dip->hmp);
2030         ++hammer_stats_file_iopsw;
2031
2032         /*
2033          * Create a new filesystem object of the requested type.  The
2034          * returned inode will be referenced but not locked.
2035          */
2036
2037         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2038                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2039                                     NULL, &nip);
2040         if (error) {
2041                 hammer_done_transaction(&trans);
2042                 *ap->a_vpp = NULL;
2043                 return (error);
2044         }
2045
2046         /*
2047          * Add a record representing the symlink.  symlink stores the link
2048          * as pure data, not a string, and is no \0 terminated.
2049          */
2050         if (error == 0) {
2051                 bytes = strlen(ap->a_target);
2052
2053                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2054                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2055                 } else {
2056                         record = hammer_alloc_mem_record(nip, bytes);
2057                         record->type = HAMMER_MEM_RECORD_GENERAL;
2058
2059                         record->leaf.base.localization = nip->obj_localization +
2060                                                          HAMMER_LOCALIZE_MISC;
2061                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2062                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2063                         record->leaf.data_len = bytes;
2064                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2065                         bcopy(ap->a_target, record->data->symlink.name, bytes);
2066                         error = hammer_ip_add_record(&trans, record);
2067                 }
2068
2069                 /*
2070                  * Set the file size to the length of the link.
2071                  */
2072                 if (error == 0) {
2073                         nip->ino_data.size = bytes;
2074                         hammer_modify_inode(nip, HAMMER_INODE_DDIRTY);
2075                 }
2076         }
2077         if (error == 0)
2078                 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
2079                                                 nch->ncp->nc_nlen, nip);
2080
2081         /*
2082          * Finish up.
2083          */
2084         if (error) {
2085                 hammer_rel_inode(nip, 0);
2086                 *ap->a_vpp = NULL;
2087         } else {
2088                 error = hammer_get_vnode(nip, ap->a_vpp);
2089                 hammer_rel_inode(nip, 0);
2090                 if (error == 0) {
2091                         cache_setunresolved(ap->a_nch);
2092                         cache_setvp(ap->a_nch, *ap->a_vpp);
2093                         hammer_knote(ap->a_dvp, NOTE_WRITE);
2094                 }
2095         }
2096         hammer_done_transaction(&trans);
2097         return (error);
2098 }
2099
2100 /*
2101  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2102  */
2103 static
2104 int
2105 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2106 {
2107         struct hammer_transaction trans;
2108         struct hammer_inode *dip;
2109         int error;
2110
2111         dip = VTOI(ap->a_dvp);
2112
2113         if (hammer_nohistory(dip) == 0 &&
2114             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2115                 return (error);
2116         }
2117
2118         hammer_start_transaction(&trans, dip->hmp);
2119         ++hammer_stats_file_iopsw;
2120         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2121                                 ap->a_cred, ap->a_flags, -1);
2122         hammer_done_transaction(&trans);
2123
2124         return (error);
2125 }
2126
2127 /*
2128  * hammer_vop_ioctl { vp, command, data, fflag, cred }
2129  */
2130 static
2131 int
2132 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2133 {
2134         struct hammer_inode *ip = ap->a_vp->v_data;
2135
2136         ++hammer_stats_file_iopsr;
2137         return(hammer_ioctl(ip, ap->a_command, ap->a_data,
2138                             ap->a_fflag, ap->a_cred));
2139 }
2140
2141 static
2142 int
2143 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2144 {
2145         struct mount *mp;
2146         int error;
2147
2148         mp = ap->a_head.a_ops->head.vv_mount;
2149
2150         switch(ap->a_op) {
2151         case MOUNTCTL_SET_EXPORT:
2152                 if (ap->a_ctllen != sizeof(struct export_args))
2153                         error = EINVAL;
2154                 else
2155                         error = hammer_vfs_export(mp, ap->a_op,
2156                                       (const struct export_args *)ap->a_ctl);
2157                 break;
2158         default:
2159                 error = journal_mountctl(ap);
2160                 break;
2161         }
2162         return(error);
2163 }
2164
2165 /*
2166  * hammer_vop_strategy { vp, bio }
2167  *
2168  * Strategy call, used for regular file read & write only.  Note that the
2169  * bp may represent a cluster.
2170  *
2171  * To simplify operation and allow better optimizations in the future,
2172  * this code does not make any assumptions with regards to buffer alignment
2173  * or size.
2174  */
2175 static
2176 int
2177 hammer_vop_strategy(struct vop_strategy_args *ap)
2178 {
2179         struct buf *bp;
2180         int error;
2181
2182         bp = ap->a_bio->bio_buf;
2183
2184         switch(bp->b_cmd) {
2185         case BUF_CMD_READ:
2186                 error = hammer_vop_strategy_read(ap);
2187                 break;
2188         case BUF_CMD_WRITE:
2189                 error = hammer_vop_strategy_write(ap);
2190                 break;
2191         default:
2192                 bp->b_error = error = EINVAL;
2193                 bp->b_flags |= B_ERROR;
2194                 biodone(ap->a_bio);
2195                 break;
2196         }
2197         return (error);
2198 }
2199
2200 /*
2201  * Read from a regular file.  Iterate the related records and fill in the
2202  * BIO/BUF.  Gaps are zero-filled.
2203  *
2204  * The support code in hammer_object.c should be used to deal with mixed
2205  * in-memory and on-disk records.
2206  *
2207  * NOTE: Can be called from the cluster code with an oversized buf.
2208  *
2209  * XXX atime update
2210  */
2211 static
2212 int
2213 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2214 {
2215         struct hammer_transaction trans;
2216         struct hammer_inode *ip;
2217         struct hammer_inode *dip;
2218         struct hammer_cursor cursor;
2219         hammer_base_elm_t base;
2220         hammer_off_t disk_offset;
2221         struct bio *bio;
2222         struct bio *nbio;
2223         struct buf *bp;
2224         int64_t rec_offset;
2225         int64_t ran_end;
2226         int64_t tmp64;
2227         int error;
2228         int boff;
2229         int roff;
2230         int n;
2231
2232         bio = ap->a_bio;
2233         bp = bio->bio_buf;
2234         ip = ap->a_vp->v_data;
2235
2236         /*
2237          * The zone-2 disk offset may have been set by the cluster code via
2238          * a BMAP operation, or else should be NOOFFSET.
2239          *
2240          * Checking the high bits for a match against zone-2 should suffice.
2241          */
2242         nbio = push_bio(bio);
2243         if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2244             HAMMER_ZONE_LARGE_DATA) {
2245                 error = hammer_io_direct_read(ip->hmp, nbio, NULL);
2246                 return (error);
2247         }
2248
2249         /*
2250          * Well, that sucked.  Do it the hard way.  If all the stars are
2251          * aligned we may still be able to issue a direct-read.
2252          */
2253         hammer_simple_transaction(&trans, ip->hmp);
2254         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2255
2256         /*
2257          * Key range (begin and end inclusive) to scan.  Note that the key's
2258          * stored in the actual records represent BASE+LEN, not BASE.  The
2259          * first record containing bio_offset will have a key > bio_offset.
2260          */
2261         cursor.key_beg.localization = ip->obj_localization +
2262                                       HAMMER_LOCALIZE_MISC;
2263         cursor.key_beg.obj_id = ip->obj_id;
2264         cursor.key_beg.create_tid = 0;
2265         cursor.key_beg.delete_tid = 0;
2266         cursor.key_beg.obj_type = 0;
2267         cursor.key_beg.key = bio->bio_offset + 1;
2268         cursor.asof = ip->obj_asof;
2269         cursor.flags |= HAMMER_CURSOR_ASOF;
2270
2271         cursor.key_end = cursor.key_beg;
2272         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2273 #if 0
2274         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2275                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2276                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2277                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2278         } else
2279 #endif
2280         {
2281                 ran_end = bio->bio_offset + bp->b_bufsize;
2282                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2283                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2284                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2285                 if (tmp64 < ran_end)
2286                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2287                 else
2288                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2289         }
2290         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2291
2292         error = hammer_ip_first(&cursor);
2293         boff = 0;
2294
2295         while (error == 0) {
2296                 /*
2297                  * Get the base file offset of the record.  The key for
2298                  * data records is (base + bytes) rather then (base).
2299                  */
2300                 base = &cursor.leaf->base;
2301                 rec_offset = base->key - cursor.leaf->data_len;
2302
2303                 /*
2304                  * Calculate the gap, if any, and zero-fill it.
2305                  *
2306                  * n is the offset of the start of the record verses our
2307                  * current seek offset in the bio.
2308                  */
2309                 n = (int)(rec_offset - (bio->bio_offset + boff));
2310                 if (n > 0) {
2311                         if (n > bp->b_bufsize - boff)
2312                                 n = bp->b_bufsize - boff;
2313                         bzero((char *)bp->b_data + boff, n);
2314                         boff += n;
2315                         n = 0;
2316                 }
2317
2318                 /*
2319                  * Calculate the data offset in the record and the number
2320                  * of bytes we can copy.
2321                  *
2322                  * There are two degenerate cases.  First, boff may already
2323                  * be at bp->b_bufsize.  Secondly, the data offset within
2324                  * the record may exceed the record's size.
2325                  */
2326                 roff = -n;
2327                 rec_offset += roff;
2328                 n = cursor.leaf->data_len - roff;
2329                 if (n <= 0) {
2330                         kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2331                         n = 0;
2332                 } else if (n > bp->b_bufsize - boff) {
2333                         n = bp->b_bufsize - boff;
2334                 }
2335
2336                 /*
2337                  * Deal with cached truncations.  This cool bit of code
2338                  * allows truncate()/ftruncate() to avoid having to sync
2339                  * the file.
2340                  *
2341                  * If the frontend is truncated then all backend records are
2342                  * subject to the frontend's truncation.
2343                  *
2344                  * If the backend is truncated then backend records on-disk
2345                  * (but not in-memory) are subject to the backend's
2346                  * truncation.  In-memory records owned by the backend
2347                  * represent data written after the truncation point on the
2348                  * backend and must not be truncated.
2349                  *
2350                  * Truncate operations deal with frontend buffer cache
2351                  * buffers and frontend-owned in-memory records synchronously.
2352                  */
2353                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2354                         if (hammer_cursor_ondisk(&cursor) ||
2355                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2356                                 if (ip->trunc_off <= rec_offset)
2357                                         n = 0;
2358                                 else if (ip->trunc_off < rec_offset + n)
2359                                         n = (int)(ip->trunc_off - rec_offset);
2360                         }
2361                 }
2362                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2363                         if (hammer_cursor_ondisk(&cursor)) {
2364                                 if (ip->sync_trunc_off <= rec_offset)
2365                                         n = 0;
2366                                 else if (ip->sync_trunc_off < rec_offset + n)
2367                                         n = (int)(ip->sync_trunc_off - rec_offset);
2368                         }
2369                 }
2370
2371                 /*
2372                  * Try to issue a direct read into our bio if possible,
2373                  * otherwise resolve the element data into a hammer_buffer
2374                  * and copy.
2375                  *
2376                  * The buffer on-disk should be zerod past any real
2377                  * truncation point, but may not be for any synthesized
2378                  * truncation point from above.
2379                  */
2380                 disk_offset = cursor.leaf->data_offset + roff;
2381                 if (boff == 0 && n == bp->b_bufsize &&
2382                     hammer_cursor_ondisk(&cursor) &&
2383                     (disk_offset & HAMMER_BUFMASK) == 0) {
2384                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2385                                  HAMMER_ZONE_LARGE_DATA);
2386                         nbio->bio_offset = disk_offset;
2387                         error = hammer_io_direct_read(trans.hmp, nbio,
2388                                                       cursor.leaf);
2389                         goto done;
2390                 } else if (n) {
2391                         error = hammer_ip_resolve_data(&cursor);
2392                         if (error == 0) {
2393                                 bcopy((char *)cursor.data + roff,
2394                                       (char *)bp->b_data + boff, n);
2395                         }
2396                 }
2397                 if (error)
2398                         break;
2399
2400                 /*
2401                  * Iterate until we have filled the request.
2402                  */
2403                 boff += n;
2404                 if (boff == bp->b_bufsize)
2405                         break;
2406                 error = hammer_ip_next(&cursor);
2407         }
2408
2409         /*
2410          * There may have been a gap after the last record
2411          */
2412         if (error == ENOENT)
2413                 error = 0;
2414         if (error == 0 && boff != bp->b_bufsize) {
2415                 KKASSERT(boff < bp->b_bufsize);
2416                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2417                 /* boff = bp->b_bufsize; */
2418         }
2419         bp->b_resid = 0;
2420         bp->b_error = error;
2421         if (error)
2422                 bp->b_flags |= B_ERROR;
2423         biodone(ap->a_bio);
2424
2425 done:
2426         /*
2427          * Cache the b-tree node for the last data read in cache[1].
2428          *
2429          * If we hit the file EOF then also cache the node in the
2430          * governing director's cache[3], it will be used to initialize
2431          * the inode's cache[1] for any inodes looked up via the directory.
2432          *
2433          * This doesn't reduce disk accesses since the B-Tree chain is
2434          * likely cached, but it does reduce cpu overhead when looking
2435          * up file offsets for cpdup/tar/cpio style iterations.
2436          */
2437         if (cursor.node)
2438                 hammer_cache_node(&ip->cache[1], cursor.node);
2439         if (ran_end >= ip->ino_data.size) {
2440                 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
2441                                         ip->obj_asof, ip->obj_localization);
2442                 if (dip) {
2443                         hammer_cache_node(&dip->cache[3], cursor.node);
2444                         hammer_rel_inode(dip, 0);
2445                 }
2446         }
2447         hammer_done_cursor(&cursor);
2448         hammer_done_transaction(&trans);
2449         return(error);
2450 }
2451
2452 /*
2453  * BMAP operation - used to support cluster_read() only.
2454  *
2455  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2456  *
2457  * This routine may return EOPNOTSUPP if the opration is not supported for
2458  * the specified offset.  The contents of the pointer arguments do not
2459  * need to be initialized in that case. 
2460  *
2461  * If a disk address is available and properly aligned return 0 with 
2462  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2463  * to the run-length relative to that offset.  Callers may assume that
2464  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2465  * large, so return EOPNOTSUPP if it is not sufficiently large.
2466  */
2467 static
2468 int
2469 hammer_vop_bmap(struct vop_bmap_args *ap)
2470 {
2471         struct hammer_transaction trans;
2472         struct hammer_inode *ip;
2473         struct hammer_cursor cursor;
2474         hammer_base_elm_t base;
2475         int64_t rec_offset;
2476         int64_t ran_end;
2477         int64_t tmp64;
2478         int64_t base_offset;
2479         int64_t base_disk_offset;
2480         int64_t last_offset;
2481         hammer_off_t last_disk_offset;
2482         hammer_off_t disk_offset;
2483         int     rec_len;
2484         int     error;
2485         int     blksize;
2486
2487         ++hammer_stats_file_iopsr;
2488         ip = ap->a_vp->v_data;
2489
2490         /*
2491          * We can only BMAP regular files.  We can't BMAP database files,
2492          * directories, etc.
2493          */
2494         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
2495                 return(EOPNOTSUPP);
2496
2497         /*
2498          * bmap is typically called with runp/runb both NULL when used
2499          * for writing.  We do not support BMAP for writing atm.
2500          */
2501         if (ap->a_cmd != BUF_CMD_READ)
2502                 return(EOPNOTSUPP);
2503
2504         /*
2505          * Scan the B-Tree to acquire blockmap addresses, then translate
2506          * to raw addresses.
2507          */
2508         hammer_simple_transaction(&trans, ip->hmp);
2509 #if 0
2510         kprintf("bmap_beg %016llx ip->cache %p\n",
2511                 (long long)ap->a_loffset, ip->cache[1]);
2512 #endif
2513         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2514
2515         /*
2516          * Key range (begin and end inclusive) to scan.  Note that the key's
2517          * stored in the actual records represent BASE+LEN, not BASE.  The
2518          * first record containing bio_offset will have a key > bio_offset.
2519          */
2520         cursor.key_beg.localization = ip->obj_localization +
2521                                       HAMMER_LOCALIZE_MISC;
2522         cursor.key_beg.obj_id = ip->obj_id;
2523         cursor.key_beg.create_tid = 0;
2524         cursor.key_beg.delete_tid = 0;
2525         cursor.key_beg.obj_type = 0;
2526         if (ap->a_runb)
2527                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
2528         else
2529                 cursor.key_beg.key = ap->a_loffset + 1;
2530         if (cursor.key_beg.key < 0)
2531                 cursor.key_beg.key = 0;
2532         cursor.asof = ip->obj_asof;
2533         cursor.flags |= HAMMER_CURSOR_ASOF;
2534
2535         cursor.key_end = cursor.key_beg;
2536         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2537
2538         ran_end = ap->a_loffset + MAXPHYS;
2539         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2540         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2541         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2542         if (tmp64 < ran_end)
2543                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2544         else
2545                 cursor.key_end.key = ran_end + MAXPHYS + 1;
2546
2547         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2548
2549         error = hammer_ip_first(&cursor);
2550         base_offset = last_offset = 0;
2551         base_disk_offset = last_disk_offset = 0;
2552
2553         while (error == 0) {
2554                 /*
2555                  * Get the base file offset of the record.  The key for
2556                  * data records is (base + bytes) rather then (base).
2557                  *
2558                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
2559                  * The extra bytes should be zero on-disk and the BMAP op
2560                  * should still be ok.
2561                  */
2562                 base = &cursor.leaf->base;
2563                 rec_offset = base->key - cursor.leaf->data_len;
2564                 rec_len    = cursor.leaf->data_len;
2565
2566                 /*
2567                  * Incorporate any cached truncation.
2568                  *
2569                  * NOTE: Modifications to rec_len based on synthesized
2570                  * truncation points remove the guarantee that any extended
2571                  * data on disk is zero (since the truncations may not have
2572                  * taken place on-media yet).
2573                  */
2574                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2575                         if (hammer_cursor_ondisk(&cursor) ||
2576                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2577                                 if (ip->trunc_off <= rec_offset)
2578                                         rec_len = 0;
2579                                 else if (ip->trunc_off < rec_offset + rec_len)
2580                                         rec_len = (int)(ip->trunc_off - rec_offset);
2581                         }
2582                 }
2583                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2584                         if (hammer_cursor_ondisk(&cursor)) {
2585                                 if (ip->sync_trunc_off <= rec_offset)
2586                                         rec_len = 0;
2587                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
2588                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
2589                         }
2590                 }
2591
2592                 /*
2593                  * Accumulate information.  If we have hit a discontiguous
2594                  * block reset base_offset unless we are already beyond the
2595                  * requested offset.  If we are, that's it, we stop.
2596                  */
2597                 if (error)
2598                         break;
2599                 if (hammer_cursor_ondisk(&cursor)) {
2600                         disk_offset = cursor.leaf->data_offset;
2601                         if (rec_offset != last_offset ||
2602                             disk_offset != last_disk_offset) {
2603                                 if (rec_offset > ap->a_loffset)
2604                                         break;
2605                                 base_offset = rec_offset;
2606                                 base_disk_offset = disk_offset;
2607                         }
2608                         last_offset = rec_offset + rec_len;
2609                         last_disk_offset = disk_offset + rec_len;
2610                 }
2611                 error = hammer_ip_next(&cursor);
2612         }
2613
2614 #if 0
2615         kprintf("BMAP %016llx:  %016llx - %016llx\n",
2616                 (long long)ap->a_loffset,
2617                 (long long)base_offset,
2618                 (long long)last_offset);
2619         kprintf("BMAP %16s:  %016llx - %016llx\n", "",
2620                 (long long)base_disk_offset,
2621                 (long long)last_disk_offset);
2622 #endif
2623
2624         if (cursor.node) {
2625                 hammer_cache_node(&ip->cache[1], cursor.node);
2626 #if 0
2627                 kprintf("bmap_end2 %016llx ip->cache %p\n",
2628                         (long long)ap->a_loffset, ip->cache[1]);
2629 #endif
2630         }
2631         hammer_done_cursor(&cursor);
2632         hammer_done_transaction(&trans);
2633
2634         /*
2635          * If we couldn't find any records or the records we did find were
2636          * all behind the requested offset, return failure.  A forward
2637          * truncation can leave a hole w/ no on-disk records.
2638          */
2639         if (last_offset == 0 || last_offset < ap->a_loffset)
2640                 return (EOPNOTSUPP);
2641
2642         /*
2643          * Figure out the block size at the requested offset and adjust
2644          * our limits so the cluster_read() does not create inappropriately
2645          * sized buffer cache buffers.
2646          */
2647         blksize = hammer_blocksize(ap->a_loffset);
2648         if (hammer_blocksize(base_offset) != blksize) {
2649                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
2650         }
2651         if (last_offset != ap->a_loffset &&
2652             hammer_blocksize(last_offset - 1) != blksize) {
2653                 last_offset = hammer_blockdemarc(ap->a_loffset,
2654                                                  last_offset - 1);
2655         }
2656
2657         /*
2658          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
2659          * from occuring.
2660          */
2661         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
2662
2663         if ((disk_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_LARGE_DATA) {
2664                 /*
2665                  * Only large-data zones can be direct-IOd
2666                  */
2667                 error = EOPNOTSUPP;
2668         } else if ((disk_offset & HAMMER_BUFMASK) ||
2669                    (last_offset - ap->a_loffset) < blksize) {
2670                 /*
2671                  * doffsetp is not aligned or the forward run size does
2672                  * not cover a whole buffer, disallow the direct I/O.
2673                  */
2674                 error = EOPNOTSUPP;
2675         } else {
2676                 /*
2677                  * We're good.
2678                  */
2679                 *ap->a_doffsetp = disk_offset;
2680                 if (ap->a_runb) {
2681                         *ap->a_runb = ap->a_loffset - base_offset;
2682                         KKASSERT(*ap->a_runb >= 0);
2683                 }
2684                 if (ap->a_runp) {
2685                         *ap->a_runp = last_offset - ap->a_loffset;
2686                         KKASSERT(*ap->a_runp >= 0);
2687                 }
2688                 error = 0;
2689         }
2690         return(error);
2691 }
2692
2693 /*
2694  * Write to a regular file.   Because this is a strategy call the OS is
2695  * trying to actually get data onto the media.
2696  */
2697 static
2698 int
2699 hammer_vop_strategy_write(struct vop_strategy_args *ap)
2700 {
2701         hammer_record_t record;
2702         hammer_mount_t hmp;
2703         hammer_inode_t ip;
2704         struct bio *bio;
2705         struct buf *bp;
2706         int blksize;
2707         int bytes;
2708         int error;
2709
2710         bio = ap->a_bio;
2711         bp = bio->bio_buf;
2712         ip = ap->a_vp->v_data;
2713         hmp = ip->hmp;
2714
2715         blksize = hammer_blocksize(bio->bio_offset);
2716         KKASSERT(bp->b_bufsize == blksize);
2717
2718         if (ip->flags & HAMMER_INODE_RO) {
2719                 bp->b_error = EROFS;
2720                 bp->b_flags |= B_ERROR;
2721                 biodone(ap->a_bio);
2722                 return(EROFS);
2723         }
2724
2725         /*
2726          * Interlock with inode destruction (no in-kernel or directory
2727          * topology visibility).  If we queue new IO while trying to
2728          * destroy the inode we can deadlock the vtrunc call in
2729          * hammer_inode_unloadable_check().
2730          *
2731          * Besides, there's no point flushing a bp associated with an
2732          * inode that is being destroyed on-media and has no kernel
2733          * references.
2734          */
2735         if ((ip->flags | ip->sync_flags) &
2736             (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
2737                 bp->b_resid = 0;
2738                 biodone(ap->a_bio);
2739                 return(0);
2740         }
2741
2742         /*
2743          * Reserve space and issue a direct-write from the front-end. 
2744          * NOTE: The direct_io code will hammer_bread/bcopy smaller
2745          * allocations.
2746          *
2747          * An in-memory record will be installed to reference the storage
2748          * until the flusher can get to it.
2749          *
2750          * Since we own the high level bio the front-end will not try to
2751          * do a direct-read until the write completes.
2752          *
2753          * NOTE: The only time we do not reserve a full-sized buffers
2754          * worth of data is if the file is small.  We do not try to
2755          * allocate a fragment (from the small-data zone) at the end of
2756          * an otherwise large file as this can lead to wildly separated
2757          * data.
2758          */
2759         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
2760         KKASSERT(bio->bio_offset < ip->ino_data.size);
2761         if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
2762                 bytes = bp->b_bufsize;
2763         else
2764                 bytes = ((int)ip->ino_data.size + 15) & ~15;
2765
2766         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
2767                                     bytes, &error);
2768         if (record) {
2769                 hammer_io_direct_write(hmp, record, bio);
2770                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
2771                         hammer_flush_inode(ip, 0);
2772         } else {
2773                 bp->b_bio2.bio_offset = NOOFFSET;
2774                 bp->b_error = error;
2775                 bp->b_flags |= B_ERROR;
2776                 biodone(ap->a_bio);
2777         }
2778         return(error);
2779 }
2780
2781 /*
2782  * dounlink - disconnect a directory entry
2783  *
2784  * XXX whiteout support not really in yet
2785  */
2786 static int
2787 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
2788                 struct vnode *dvp, struct ucred *cred, 
2789                 int flags, int isdir)
2790 {
2791         struct namecache *ncp;
2792         hammer_inode_t dip;
2793         hammer_inode_t ip;
2794         struct hammer_cursor cursor;
2795         int64_t namekey;
2796         u_int32_t max_iterations;
2797         int nlen, error;
2798
2799         /*
2800          * Calculate the namekey and setup the key range for the scan.  This
2801          * works kinda like a chained hash table where the lower 32 bits
2802          * of the namekey synthesize the chain.
2803          *
2804          * The key range is inclusive of both key_beg and key_end.
2805          */
2806         dip = VTOI(dvp);
2807         ncp = nch->ncp;
2808
2809         if (dip->flags & HAMMER_INODE_RO)
2810                 return (EROFS);
2811
2812         namekey = hammer_directory_namekey(dip, ncp->nc_name, ncp->nc_nlen,
2813                                            &max_iterations);
2814 retry:
2815         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
2816         cursor.key_beg.localization = dip->obj_localization +
2817                                       hammer_dir_localization(dip);
2818         cursor.key_beg.obj_id = dip->obj_id;
2819         cursor.key_beg.key = namekey;
2820         cursor.key_beg.create_tid = 0;
2821         cursor.key_beg.delete_tid = 0;
2822         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2823         cursor.key_beg.obj_type = 0;
2824
2825         cursor.key_end = cursor.key_beg;
2826         cursor.key_end.key += max_iterations;
2827         cursor.asof = dip->obj_asof;
2828         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2829
2830         /*
2831          * Scan all matching records (the chain), locate the one matching
2832          * the requested path component.  info->last_error contains the
2833          * error code on search termination and could be 0, ENOENT, or
2834          * something else.
2835          *
2836          * The hammer_ip_*() functions merge in-memory records with on-disk
2837          * records for the purposes of the search.
2838          */
2839         error = hammer_ip_first(&cursor);
2840
2841         while (error == 0) {
2842                 error = hammer_ip_resolve_data(&cursor);
2843                 if (error)
2844                         break;
2845                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2846                 KKASSERT(nlen > 0);
2847                 if (ncp->nc_nlen == nlen &&
2848                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2849                         break;
2850                 }
2851                 error = hammer_ip_next(&cursor);
2852         }
2853
2854         /*
2855          * If all is ok we have to get the inode so we can adjust nlinks.
2856          * To avoid a deadlock with the flusher we must release the inode
2857          * lock on the directory when acquiring the inode for the entry.
2858          *
2859          * If the target is a directory, it must be empty.
2860          */
2861         if (error == 0) {
2862                 hammer_unlock(&cursor.ip->lock);
2863                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
2864                                       dip->hmp->asof,
2865                                       cursor.data->entry.localization,
2866                                       0, &error);
2867                 hammer_lock_sh(&cursor.ip->lock);
2868                 if (error == ENOENT) {
2869                         kprintf("HAMMER: WARNING: Removing "
2870                                 "dirent w/missing inode \"%s\"\n"
2871                                 "\tobj_id = %016llx\n",
2872                                 ncp->nc_name,
2873                                 (long long)cursor.data->entry.obj_id);
2874                         error = 0;
2875                 }
2876
2877                 /*
2878                  * If isdir >= 0 we validate that the entry is or is not a
2879                  * directory.  If isdir < 0 we don't care.
2880                  */
2881                 if (error == 0 && isdir >= 0 && ip) {
2882                         if (isdir &&
2883                             ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
2884                                 error = ENOTDIR;
2885                         } else if (isdir == 0 &&
2886                             ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
2887                                 error = EISDIR;
2888                         }
2889                 }
2890
2891                 /*
2892                  * If we are trying to remove a directory the directory must
2893                  * be empty.
2894                  *
2895                  * The check directory code can loop and deadlock/retry.  Our
2896                  * own cursor's node locks must be released to avoid a 3-way
2897                  * deadlock with the flusher if the check directory code
2898                  * blocks.
2899                  *
2900                  * If any changes whatsoever have been made to the cursor
2901                  * set EDEADLK and retry.
2902                  */
2903                 if (error == 0 && ip && ip->ino_data.obj_type ==
2904                                         HAMMER_OBJTYPE_DIRECTORY) {
2905                         hammer_unlock_cursor(&cursor);
2906                         error = hammer_ip_check_directory_empty(trans, ip);
2907                         hammer_lock_cursor(&cursor);
2908                         if (cursor.flags & HAMMER_CURSOR_RETEST) {
2909                                 kprintf("HAMMER: Warning: avoided deadlock "
2910                                         "on rmdir '%s'\n",
2911                                         ncp->nc_name);
2912                                 error = EDEADLK;
2913                         }
2914                 }
2915
2916                 /*
2917                  * Delete the directory entry.
2918                  *
2919                  * WARNING: hammer_ip_del_directory() may have to terminate
2920                  * the cursor to avoid a deadlock.  It is ok to call
2921                  * hammer_done_cursor() twice.
2922                  */
2923                 if (error == 0) {
2924                         error = hammer_ip_del_directory(trans, &cursor,
2925                                                         dip, ip);
2926                 }
2927                 hammer_done_cursor(&cursor);
2928                 if (error == 0) {
2929                         cache_setunresolved(nch);
2930                         cache_setvp(nch, NULL);
2931                         /* XXX locking */
2932                         if (ip && ip->vp) {
2933                                 hammer_knote(ip->vp, NOTE_DELETE);
2934                                 cache_inval_vp(ip->vp, CINV_DESTROY);
2935                         }
2936                 }
2937                 if (ip)
2938                         hammer_rel_inode(ip, 0);
2939         } else {
2940                 hammer_done_cursor(&cursor);
2941         }
2942         if (error == EDEADLK)
2943                 goto retry;
2944
2945         return (error);
2946 }
2947
2948 /************************************************************************
2949  *                          FIFO AND SPECFS OPS                         *
2950  ************************************************************************
2951  *
2952  */
2953
2954 static int
2955 hammer_vop_fifoclose (struct vop_close_args *ap)
2956 {
2957         /* XXX update itimes */
2958         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
2959 }
2960
2961 static int
2962 hammer_vop_fiforead (struct vop_read_args *ap)
2963 {
2964         int error;
2965
2966         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2967         /* XXX update access time */
2968         return (error);
2969 }
2970
2971 static int
2972 hammer_vop_fifowrite (struct vop_write_args *ap)
2973 {
2974         int error;
2975
2976         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2977         /* XXX update access time */
2978         return (error);
2979 }
2980
2981 static
2982 int
2983 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
2984 {
2985         int error;
2986
2987         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2988         if (error)
2989                 error = hammer_vop_kqfilter(ap);
2990         return(error);
2991 }
2992
2993 static int
2994 hammer_vop_specclose (struct vop_close_args *ap)
2995 {
2996         /* XXX update itimes */
2997         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2998 }
2999
3000 static int
3001 hammer_vop_specread (struct vop_read_args *ap)
3002 {
3003         /* XXX update access time */
3004         return (VOCALL(&spec_vnode_vops, &ap->a_head));
3005 }
3006
3007 static int
3008 hammer_vop_specwrite (struct vop_write_args *ap)
3009 {
3010         /* XXX update last change time */
3011         return (VOCALL(&spec_vnode_vops, &ap->a_head));
3012 }
3013
3014 /*
3015  * SPECFS's getattr will override fields as necessary, but does not fill
3016  *          stuff in from scratch.
3017  */
3018 static
3019 int
3020 hammer_vop_specgetattr (struct vop_getattr_args *ap)
3021 {
3022         int error;
3023
3024         error = hammer_vop_getattr(ap);
3025         if (error == 0)
3026                 VOCALL(&spec_vnode_vops, &ap->a_head);
3027         return (error);
3028 }
3029
3030
3031 /************************************************************************
3032  *                          KQFILTER OPS                                *
3033  ************************************************************************
3034  *
3035  */
3036 static void filt_hammerdetach(struct knote *kn);
3037 static int filt_hammerread(struct knote *kn, long hint);
3038 static int filt_hammerwrite(struct knote *kn, long hint);
3039 static int filt_hammervnode(struct knote *kn, long hint);
3040
3041 static struct filterops hammerread_filtops =
3042         { 1, NULL, filt_hammerdetach, filt_hammerread };
3043 static struct filterops hammerwrite_filtops =
3044         { 1, NULL, filt_hammerdetach, filt_hammerwrite };
3045 static struct filterops hammervnode_filtops =
3046         { 1, NULL, filt_hammerdetach, filt_hammervnode };
3047
3048 static
3049 int
3050 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3051 {
3052         struct vnode *vp = ap->a_vp;
3053         struct knote *kn = ap->a_kn;
3054         lwkt_tokref vlock;
3055
3056         switch (kn->kn_filter) {
3057         case EVFILT_READ:
3058                 kn->kn_fop = &hammerread_filtops;
3059                 break;
3060         case EVFILT_WRITE:
3061                 kn->kn_fop = &hammerwrite_filtops;
3062                 break;
3063         case EVFILT_VNODE:
3064                 kn->kn_fop = &hammervnode_filtops;
3065                 break;
3066         default:
3067                 return (1);
3068         }
3069
3070         kn->kn_hook = (caddr_t)vp;
3071
3072         lwkt_gettoken(&vlock, &vp->v_token);
3073         SLIST_INSERT_HEAD(&vp->v_pollinfo.vpi_selinfo.si_note, kn, kn_selnext);
3074         lwkt_reltoken(&vlock);
3075
3076         return(0);
3077 }
3078
3079 static void
3080 filt_hammerdetach(struct knote *kn)
3081 {
3082         struct vnode *vp = (void *)kn->kn_hook;
3083         lwkt_tokref vlock;
3084
3085         lwkt_gettoken(&vlock, &vp->v_token);
3086         SLIST_REMOVE(&vp->v_pollinfo.vpi_selinfo.si_note,
3087                      kn, knote, kn_selnext);
3088         lwkt_reltoken(&vlock);
3089 }
3090
3091 static int
3092 filt_hammerread(struct knote *kn, long hint)
3093 {
3094         struct vnode *vp = (void *)kn->kn_hook;
3095         hammer_inode_t ip = VTOI(vp);
3096
3097         if (hint == NOTE_REVOKE) {
3098                 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3099                 return(1);
3100         }
3101         kn->kn_data = ip->ino_data.size - kn->kn_fp->f_offset;
3102         return (kn->kn_data != 0);
3103 }
3104
3105 static int
3106 filt_hammerwrite(struct knote *kn, long hint)
3107 {
3108         if (hint == NOTE_REVOKE)
3109                 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3110         kn->kn_data = 0;
3111         return (1);
3112 }
3113
3114 static int
3115 filt_hammervnode(struct knote *kn, long hint)
3116 {
3117         if (kn->kn_sfflags & hint)
3118                 kn->kn_fflags |= hint;
3119         if (hint == NOTE_REVOKE) {
3120                 kn->kn_flags |= EV_EOF;
3121                 return (1);
3122         }
3123         return (kn->kn_fflags != 0);
3124 }
3125