kernel -- Add O_FRNONBLOCKING, to allow reads which do not block on disk accesses.
[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
51 #include "hammer.h"
52
53 /*
54  * USERFS VNOPS
55  */
56 /*static int hammer_vop_vnoperate(struct vop_generic_args *);*/
57 static int hammer_vop_fsync(struct vop_fsync_args *);
58 static int hammer_vop_read(struct vop_read_args *);
59 static int hammer_vop_write(struct vop_write_args *);
60 static int hammer_vop_access(struct vop_access_args *);
61 static int hammer_vop_advlock(struct vop_advlock_args *);
62 static int hammer_vop_close(struct vop_close_args *);
63 static int hammer_vop_ncreate(struct vop_ncreate_args *);
64 static int hammer_vop_getattr(struct vop_getattr_args *);
65 static int hammer_vop_nresolve(struct vop_nresolve_args *);
66 static int hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *);
67 static int hammer_vop_nlink(struct vop_nlink_args *);
68 static int hammer_vop_nmkdir(struct vop_nmkdir_args *);
69 static int hammer_vop_nmknod(struct vop_nmknod_args *);
70 static int hammer_vop_open(struct vop_open_args *);
71 static int hammer_vop_print(struct vop_print_args *);
72 static int hammer_vop_readdir(struct vop_readdir_args *);
73 static int hammer_vop_readlink(struct vop_readlink_args *);
74 static int hammer_vop_nremove(struct vop_nremove_args *);
75 static int hammer_vop_nrename(struct vop_nrename_args *);
76 static int hammer_vop_nrmdir(struct vop_nrmdir_args *);
77 static int hammer_vop_markatime(struct vop_markatime_args *);
78 static int hammer_vop_setattr(struct vop_setattr_args *);
79 static int hammer_vop_strategy(struct vop_strategy_args *);
80 static int hammer_vop_bmap(struct vop_bmap_args *ap);
81 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
82 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
83 static int hammer_vop_ioctl(struct vop_ioctl_args *);
84 static int hammer_vop_mountctl(struct vop_mountctl_args *);
85 static int hammer_vop_kqfilter (struct vop_kqfilter_args *);
86
87 static int hammer_vop_fifoclose (struct vop_close_args *);
88 static int hammer_vop_fiforead (struct vop_read_args *);
89 static int hammer_vop_fifowrite (struct vop_write_args *);
90 static int hammer_vop_fifokqfilter (struct vop_kqfilter_args *);
91
92 struct vop_ops hammer_vnode_vops = {
93         .vop_default =          vop_defaultop,
94         .vop_fsync =            hammer_vop_fsync,
95         .vop_getpages =         vop_stdgetpages,
96         .vop_putpages =         vop_stdputpages,
97         .vop_read =             hammer_vop_read,
98         .vop_write =            hammer_vop_write,
99         .vop_access =           hammer_vop_access,
100         .vop_advlock =          hammer_vop_advlock,
101         .vop_close =            hammer_vop_close,
102         .vop_ncreate =          hammer_vop_ncreate,
103         .vop_getattr =          hammer_vop_getattr,
104         .vop_inactive =         hammer_vop_inactive,
105         .vop_reclaim =          hammer_vop_reclaim,
106         .vop_nresolve =         hammer_vop_nresolve,
107         .vop_nlookupdotdot =    hammer_vop_nlookupdotdot,
108         .vop_nlink =            hammer_vop_nlink,
109         .vop_nmkdir =           hammer_vop_nmkdir,
110         .vop_nmknod =           hammer_vop_nmknod,
111         .vop_open =             hammer_vop_open,
112         .vop_pathconf =         vop_stdpathconf,
113         .vop_print =            hammer_vop_print,
114         .vop_readdir =          hammer_vop_readdir,
115         .vop_readlink =         hammer_vop_readlink,
116         .vop_nremove =          hammer_vop_nremove,
117         .vop_nrename =          hammer_vop_nrename,
118         .vop_nrmdir =           hammer_vop_nrmdir,
119         .vop_markatime =        hammer_vop_markatime,
120         .vop_setattr =          hammer_vop_setattr,
121         .vop_bmap =             hammer_vop_bmap,
122         .vop_strategy =         hammer_vop_strategy,
123         .vop_nsymlink =         hammer_vop_nsymlink,
124         .vop_nwhiteout =        hammer_vop_nwhiteout,
125         .vop_ioctl =            hammer_vop_ioctl,
126         .vop_mountctl =         hammer_vop_mountctl,
127         .vop_kqfilter =         hammer_vop_kqfilter
128 };
129
130 struct vop_ops hammer_spec_vops = {
131         .vop_default =          vop_defaultop,
132         .vop_fsync =            hammer_vop_fsync,
133         .vop_read =             vop_stdnoread,
134         .vop_write =            vop_stdnowrite,
135         .vop_access =           hammer_vop_access,
136         .vop_close =            hammer_vop_close,
137         .vop_markatime =        hammer_vop_markatime,
138         .vop_getattr =          hammer_vop_getattr,
139         .vop_inactive =         hammer_vop_inactive,
140         .vop_reclaim =          hammer_vop_reclaim,
141         .vop_setattr =          hammer_vop_setattr
142 };
143
144 struct vop_ops hammer_fifo_vops = {
145         .vop_default =          fifo_vnoperate,
146         .vop_fsync =            hammer_vop_fsync,
147         .vop_read =             hammer_vop_fiforead,
148         .vop_write =            hammer_vop_fifowrite,
149         .vop_access =           hammer_vop_access,
150         .vop_close =            hammer_vop_fifoclose,
151         .vop_markatime =        hammer_vop_markatime,
152         .vop_getattr =          hammer_vop_getattr,
153         .vop_inactive =         hammer_vop_inactive,
154         .vop_reclaim =          hammer_vop_reclaim,
155         .vop_setattr =          hammer_vop_setattr,
156         .vop_kqfilter =         hammer_vop_fifokqfilter
157 };
158
159 static __inline
160 void
161 hammer_knote(struct vnode *vp, int flags)
162 {
163         if (flags)
164                 KNOTE(&vp->v_pollinfo.vpi_kqinfo.ki_note, flags);
165 }
166
167 #ifdef DEBUG_TRUNCATE
168 struct hammer_inode *HammerTruncIp;
169 #endif
170
171 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
172                            struct vnode *dvp, struct ucred *cred,
173                            int flags, int isdir);
174 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
175 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
176
177 #if 0
178 static
179 int
180 hammer_vop_vnoperate(struct vop_generic_args *)
181 {
182         return (VOCALL(&hammer_vnode_vops, ap));
183 }
184 #endif
185
186 /*
187  * hammer_vop_fsync { vp, waitfor }
188  *
189  * fsync() an inode to disk and wait for it to be completely committed
190  * such that the information would not be undone if a crash occured after
191  * return.
192  *
193  * NOTE: HAMMER's fsync()'s are going to remain expensive until we implement
194  *       a REDO log.  A sysctl is provided to relax HAMMER's fsync()
195  *       operation.
196  *
197  *       Ultimately the combination of a REDO log and use of fast storage
198  *       to front-end cluster caches will make fsync fast, but it aint
199  *       here yet.  And, in anycase, we need real transactional
200  *       all-or-nothing features which are not restricted to a single file.
201  */
202 static
203 int
204 hammer_vop_fsync(struct vop_fsync_args *ap)
205 {
206         hammer_inode_t ip = VTOI(ap->a_vp);
207         hammer_mount_t hmp = ip->hmp;
208         int waitfor = ap->a_waitfor;
209         int mode;
210
211         lwkt_gettoken(&hmp->fs_token);
212
213         /*
214          * Fsync rule relaxation (default is either full synchronous flush
215          * or REDO semantics with synchronous flush).
216          */
217         if (ap->a_flags & VOP_FSYNC_SYSCALL) {
218                 switch(hammer_fsync_mode) {
219                 case 0:
220 mode0:
221                         /* no REDO, full synchronous flush */
222                         goto skip;
223                 case 1:
224 mode1:
225                         /* no REDO, full asynchronous flush */
226                         if (waitfor == MNT_WAIT)
227                                 waitfor = MNT_NOWAIT;
228                         goto skip;
229                 case 2:
230                         /* REDO semantics, synchronous flush */
231                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
232                                 goto mode0;
233                         mode = HAMMER_FLUSH_UNDOS_AUTO;
234                         break;
235                 case 3:
236                         /* REDO semantics, relaxed asynchronous flush */
237                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
238                                 goto mode1;
239                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
240                         if (waitfor == MNT_WAIT)
241                                 waitfor = MNT_NOWAIT;
242                         break;
243                 case 4:
244                         /* ignore the fsync() system call */
245                         lwkt_reltoken(&hmp->fs_token);
246                         return(0);
247                 default:
248                         /* we have to do something */
249                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
250                         if (waitfor == MNT_WAIT)
251                                 waitfor = MNT_NOWAIT;
252                         break;
253                 }
254
255                 /*
256                  * Fast fsync only needs to flush the UNDO/REDO fifo if
257                  * HAMMER_INODE_REDO is non-zero and the only modifications
258                  * made to the file are write or write-extends.
259                  */
260                 if ((ip->flags & HAMMER_INODE_REDO) &&
261                     (ip->flags & HAMMER_INODE_MODMASK_NOREDO) == 0
262                 ) {
263                         ++hammer_count_fsyncs;
264                         hammer_flusher_flush_undos(hmp, mode);
265                         ip->redo_count = 0;
266                         lwkt_reltoken(&hmp->fs_token);
267                         return(0);
268                 }
269
270                 /*
271                  * REDO is enabled by fsync(), the idea being we really only
272                  * want to lay down REDO records when programs are using
273                  * fsync() heavily.  The first fsync() on the file starts
274                  * the gravy train going and later fsync()s keep it hot by
275                  * resetting the redo_count.
276                  *
277                  * We weren't running REDOs before now so we have to fall
278                  * through and do a full fsync of what we have.
279                  */
280                 if (hmp->version >= HAMMER_VOL_VERSION_FOUR &&
281                     (hmp->flags & HAMMER_MOUNT_REDO_RECOVERY_RUN) == 0) {
282                         ip->flags |= HAMMER_INODE_REDO;
283                         ip->redo_count = 0;
284                 }
285         }
286 skip:
287
288         /*
289          * Do a full flush sequence.
290          */
291         ++hammer_count_fsyncs;
292         vfsync(ap->a_vp, waitfor, 1, NULL, NULL);
293         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
294         if (waitfor == MNT_WAIT) {
295                 vn_unlock(ap->a_vp);
296                 hammer_wait_inode(ip);
297                 vn_lock(ap->a_vp, LK_EXCLUSIVE | LK_RETRY);
298         }
299         lwkt_reltoken(&hmp->fs_token);
300         return (ip->error);
301 }
302
303 /*
304  * hammer_vop_read { vp, uio, ioflag, cred }
305  *
306  * MPSAFE (for the cache safe does not require fs_token)
307  */
308 static
309 int
310 hammer_vop_read(struct vop_read_args *ap)
311 {
312         struct hammer_transaction trans;
313         hammer_inode_t ip;
314         hammer_mount_t hmp;
315         off_t offset;
316         struct buf *bp;
317         struct uio *uio;
318         int error;
319         int n;
320         int seqcount;
321         int ioseqcount;
322         int blksize;
323         int bigread;
324         int got_fstoken;
325
326         if (ap->a_vp->v_type != VREG)
327                 return (EINVAL);
328         ip = VTOI(ap->a_vp);
329         hmp = ip->hmp;
330         error = 0;
331         uio = ap->a_uio;
332
333         /*
334          * Allow the UIO's size to override the sequential heuristic.
335          */
336         blksize = hammer_blocksize(uio->uio_offset);
337         seqcount = (uio->uio_resid + (BKVASIZE - 1)) / BKVASIZE;
338         ioseqcount = (ap->a_ioflag >> 16);
339         if (seqcount < ioseqcount)
340                 seqcount = ioseqcount;
341
342         /*
343          * If reading or writing a huge amount of data we have to break
344          * atomicy and allow the operation to be interrupted by a signal
345          * or it can DOS the machine.
346          */
347         bigread = (uio->uio_resid > 100 * 1024 * 1024);
348         got_fstoken = 0;
349
350         /*
351          * Access the data typically in HAMMER_BUFSIZE blocks via the
352          * buffer cache, but HAMMER may use a variable block size based
353          * on the offset.
354          *
355          * XXX Temporary hack, delay the start transaction while we remain
356          *     MPSAFE.  NOTE: ino_data.size cannot change while vnode is
357          *     locked-shared.
358          */
359         while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
360                 int64_t base_offset;
361                 int64_t file_limit;
362
363                 blksize = hammer_blocksize(uio->uio_offset);
364                 offset = (int)uio->uio_offset & (blksize - 1);
365                 base_offset = uio->uio_offset - offset;
366
367                 if (bigread && (error = hammer_signal_check(ip->hmp)) != 0)
368                         break;
369
370                 /*
371                  * MPSAFE
372                  */
373                 bp = getcacheblk(ap->a_vp, base_offset, blksize);
374                 if (bp) {
375                         error = 0;
376                         goto skip;
377                 } else {
378                         if (ap->a_ioflag & (IO_NDELAY | IO_NRDELAY))
379                                 return (EWOULDBLOCK);
380                 }
381
382                 /*
383                  * MPUNSAFE
384                  */
385                 if (got_fstoken == 0) {
386                         lwkt_gettoken(&hmp->fs_token);
387                         got_fstoken = 1;
388                         hammer_start_transaction(&trans, ip->hmp);
389                 }
390
391                 if (hammer_cluster_enable) {
392                         /*
393                          * Use file_limit to prevent cluster_read() from
394                          * creating buffers of the wrong block size past
395                          * the demarc.
396                          */
397                         file_limit = ip->ino_data.size;
398                         if (base_offset < HAMMER_XDEMARC &&
399                             file_limit > HAMMER_XDEMARC) {
400                                 file_limit = HAMMER_XDEMARC;
401                         }
402                         error = cluster_read(ap->a_vp,
403                                              file_limit, base_offset,
404                                              blksize, uio->uio_resid,
405                                              seqcount * BKVASIZE, &bp);
406                 } else {
407                         error = bread(ap->a_vp, base_offset, blksize, &bp);
408                 }
409                 if (error) {
410                         brelse(bp);
411                         break;
412                 }
413 skip:
414                 if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IODEBUG)) {
415                         kprintf("doff %016jx read file %016jx@%016jx\n",
416                                 (intmax_t)bp->b_bio2.bio_offset,
417                                 (intmax_t)ip->obj_id,
418                                 (intmax_t)bp->b_loffset);
419                 }
420                 bp->b_flags &= ~B_IODEBUG;
421
422                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
423                 n = blksize - offset;
424                 if (n > uio->uio_resid)
425                         n = uio->uio_resid;
426                 if (n > ip->ino_data.size - uio->uio_offset)
427                         n = (int)(ip->ino_data.size - uio->uio_offset);
428                 if (got_fstoken)
429                         lwkt_reltoken(&hmp->fs_token);
430
431                 /*
432                  * Set B_AGE, data has a lower priority than meta-data.
433                  *
434                  * Use a hold/unlock/drop sequence to run the uiomove
435                  * with the buffer unlocked, avoiding deadlocks against
436                  * read()s on mmap()'d spaces.
437                  */
438                 bp->b_flags |= B_AGE;
439                 bqhold(bp);
440                 bqrelse(bp);
441                 error = uiomove((char *)bp->b_data + offset, n, uio);
442                 bqdrop(bp);
443
444                 if (got_fstoken)
445                         lwkt_gettoken(&hmp->fs_token);
446
447                 if (error)
448                         break;
449                 hammer_stats_file_read += n;
450         }
451
452         /*
453          * XXX only update the atime if we had to get the MP lock.
454          * XXX hack hack hack, fixme.
455          */
456         if (got_fstoken) {
457                 if ((ip->flags & HAMMER_INODE_RO) == 0 &&
458                     (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
459                         ip->ino_data.atime = trans.time;
460                         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
461                 }
462                 hammer_done_transaction(&trans);
463                 lwkt_reltoken(&hmp->fs_token);
464         }
465         return (error);
466 }
467
468 /*
469  * hammer_vop_write { vp, uio, ioflag, cred }
470  */
471 static
472 int
473 hammer_vop_write(struct vop_write_args *ap)
474 {
475         struct hammer_transaction trans;
476         struct hammer_inode *ip;
477         hammer_mount_t hmp;
478         struct uio *uio;
479         int offset;
480         off_t base_offset;
481         struct buf *bp;
482         int kflags;
483         int error;
484         int n;
485         int flags;
486         int seqcount;
487         int bigwrite;
488
489         if (ap->a_vp->v_type != VREG)
490                 return (EINVAL);
491         ip = VTOI(ap->a_vp);
492         hmp = ip->hmp;
493         error = 0;
494         kflags = 0;
495         seqcount = ap->a_ioflag >> 16;
496
497         if (ip->flags & HAMMER_INODE_RO)
498                 return (EROFS);
499
500         /*
501          * Create a transaction to cover the operations we perform.
502          */
503         lwkt_gettoken(&hmp->fs_token);
504         hammer_start_transaction(&trans, hmp);
505         uio = ap->a_uio;
506
507         /*
508          * Check append mode
509          */
510         if (ap->a_ioflag & IO_APPEND)
511                 uio->uio_offset = ip->ino_data.size;
512
513         /*
514          * Check for illegal write offsets.  Valid range is 0...2^63-1.
515          *
516          * NOTE: the base_off assignment is required to work around what
517          * I consider to be a GCC-4 optimization bug.
518          */
519         if (uio->uio_offset < 0) {
520                 hammer_done_transaction(&trans);
521                 lwkt_reltoken(&hmp->fs_token);
522                 return (EFBIG);
523         }
524         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
525         if (uio->uio_resid > 0 && base_offset <= uio->uio_offset) {
526                 hammer_done_transaction(&trans);
527                 lwkt_reltoken(&hmp->fs_token);
528                 return (EFBIG);
529         }
530
531         /*
532          * If reading or writing a huge amount of data we have to break
533          * atomicy and allow the operation to be interrupted by a signal
534          * or it can DOS the machine.
535          *
536          * Preset redo_count so we stop generating REDOs earlier if the
537          * limit is exceeded.
538          */
539         bigwrite = (uio->uio_resid > 100 * 1024 * 1024);
540         if ((ip->flags & HAMMER_INODE_REDO) &&
541             ip->redo_count < hammer_limit_redo) {
542                 ip->redo_count += uio->uio_resid;
543         }
544
545         /*
546          * Access the data typically in HAMMER_BUFSIZE blocks via the
547          * buffer cache, but HAMMER may use a variable block size based
548          * on the offset.
549          */
550         while (uio->uio_resid > 0) {
551                 int fixsize = 0;
552                 int blksize;
553                 int blkmask;
554                 int trivial;
555                 int endofblk;
556                 off_t nsize;
557
558                 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
559                         break;
560                 if (bigwrite && (error = hammer_signal_check(hmp)) != 0)
561                         break;
562
563                 blksize = hammer_blocksize(uio->uio_offset);
564
565                 /*
566                  * Do not allow HAMMER to blow out the buffer cache.  Very
567                  * large UIOs can lockout other processes due to bwillwrite()
568                  * mechanics.
569                  *
570                  * The hammer inode is not locked during these operations.
571                  * The vnode is locked which can interfere with the pageout
572                  * daemon for non-UIO_NOCOPY writes but should not interfere
573                  * with the buffer cache.  Even so, we cannot afford to
574                  * allow the pageout daemon to build up too many dirty buffer
575                  * cache buffers.
576                  *
577                  * Only call this if we aren't being recursively called from
578                  * a virtual disk device (vn), else we may deadlock.
579                  */
580                 if ((ap->a_ioflag & IO_RECURSE) == 0)
581                         bwillwrite(blksize);
582
583                 /*
584                  * Control the number of pending records associated with
585                  * this inode.  If too many have accumulated start a
586                  * flush.  Try to maintain a pipeline with the flusher.
587                  *
588                  * NOTE: It is possible for other sources to grow the
589                  *       records but not necessarily issue another flush,
590                  *       so use a timeout and ensure that a re-flush occurs.
591                  */
592                 if (ip->rsv_recs >= hammer_limit_inode_recs) {
593                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
594                         while (ip->rsv_recs >= hammer_limit_inode_recs * 2) {
595                                 ip->flags |= HAMMER_INODE_RECSW;
596                                 tsleep(&ip->rsv_recs, 0, "hmrwww", hz);
597                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
598                         }
599                 }
600
601 #if 0
602                 /*
603                  * Do not allow HAMMER to blow out system memory by
604                  * accumulating too many records.   Records are so well
605                  * decoupled from the buffer cache that it is possible
606                  * for userland to push data out to the media via
607                  * direct-write, but build up the records queued to the
608                  * backend faster then the backend can flush them out.
609                  * HAMMER has hit its write limit but the frontend has
610                  * no pushback to slow it down.
611                  */
612                 if (hmp->rsv_recs > hammer_limit_recs / 2) {
613                         /*
614                          * Get the inode on the flush list
615                          */
616                         if (ip->rsv_recs >= 64)
617                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
618                         else if (ip->rsv_recs >= 16)
619                                 hammer_flush_inode(ip, 0);
620
621                         /*
622                          * Keep the flusher going if the system keeps
623                          * queueing records.
624                          */
625                         delta = hmp->count_newrecords -
626                                 hmp->last_newrecords;
627                         if (delta < 0 || delta > hammer_limit_recs / 2) {
628                                 hmp->last_newrecords = hmp->count_newrecords;
629                                 hammer_sync_hmp(hmp, MNT_NOWAIT);
630                         }
631
632                         /*
633                          * If we have gotten behind start slowing
634                          * down the writers.
635                          */
636                         delta = (hmp->rsv_recs - hammer_limit_recs) *
637                                 hz / hammer_limit_recs;
638                         if (delta > 0)
639                                 tsleep(&trans, 0, "hmrslo", delta);
640                 }
641 #endif
642
643                 /*
644                  * Calculate the blocksize at the current offset and figure
645                  * out how much we can actually write.
646                  */
647                 blkmask = blksize - 1;
648                 offset = (int)uio->uio_offset & blkmask;
649                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
650                 n = blksize - offset;
651                 if (n > uio->uio_resid) {
652                         n = uio->uio_resid;
653                         endofblk = 0;
654                 } else {
655                         endofblk = 1;
656                 }
657                 nsize = uio->uio_offset + n;
658                 if (nsize > ip->ino_data.size) {
659                         if (uio->uio_offset > ip->ino_data.size)
660                                 trivial = 0;
661                         else
662                                 trivial = 1;
663                         nvextendbuf(ap->a_vp,
664                                     ip->ino_data.size,
665                                     nsize,
666                                     hammer_blocksize(ip->ino_data.size),
667                                     hammer_blocksize(nsize),
668                                     hammer_blockoff(ip->ino_data.size),
669                                     hammer_blockoff(nsize),
670                                     trivial);
671                         fixsize = 1;
672                         kflags |= NOTE_EXTEND;
673                 }
674
675                 if (uio->uio_segflg == UIO_NOCOPY) {
676                         /*
677                          * Issuing a write with the same data backing the
678                          * buffer.  Instantiate the buffer to collect the
679                          * backing vm pages, then read-in any missing bits.
680                          *
681                          * This case is used by vop_stdputpages().
682                          */
683                         bp = getblk(ap->a_vp, base_offset,
684                                     blksize, GETBLK_BHEAVY, 0);
685                         if ((bp->b_flags & B_CACHE) == 0) {
686                                 bqrelse(bp);
687                                 error = bread(ap->a_vp, base_offset,
688                                               blksize, &bp);
689                         }
690                 } else if (offset == 0 && uio->uio_resid >= blksize) {
691                         /*
692                          * Even though we are entirely overwriting the buffer
693                          * we may still have to zero it out to avoid a 
694                          * mmap/write visibility issue.
695                          */
696                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
697                         if ((bp->b_flags & B_CACHE) == 0)
698                                 vfs_bio_clrbuf(bp);
699                 } else if (base_offset >= ip->ino_data.size) {
700                         /*
701                          * If the base offset of the buffer is beyond the
702                          * file EOF, we don't have to issue a read.
703                          */
704                         bp = getblk(ap->a_vp, base_offset,
705                                     blksize, GETBLK_BHEAVY, 0);
706                         vfs_bio_clrbuf(bp);
707                 } else {
708                         /*
709                          * Partial overwrite, read in any missing bits then
710                          * replace the portion being written.
711                          */
712                         error = bread(ap->a_vp, base_offset, blksize, &bp);
713                         if (error == 0)
714                                 bheavy(bp);
715                 }
716                 if (error == 0) {
717                         lwkt_reltoken(&hmp->fs_token);
718                         error = uiomove(bp->b_data + offset, n, uio);
719                         lwkt_gettoken(&hmp->fs_token);
720                 }
721
722                 /*
723                  * Generate REDO records if enabled and redo_count will not
724                  * exceeded the limit.
725                  *
726                  * If redo_count exceeds the limit we stop generating records
727                  * and clear HAMMER_INODE_REDO.  This will cause the next
728                  * fsync() to do a full meta-data sync instead of just an
729                  * UNDO/REDO fifo update.
730                  *
731                  * When clearing HAMMER_INODE_REDO any pre-existing REDOs
732                  * will still be tracked.  The tracks will be terminated
733                  * when the related meta-data (including possible data
734                  * modifications which are not tracked via REDO) is
735                  * flushed.
736                  */
737                 if ((ip->flags & HAMMER_INODE_REDO) && error == 0) {
738                         if (ip->redo_count < hammer_limit_redo) {
739                                 bp->b_flags |= B_VFSFLAG1;
740                                 error = hammer_generate_redo(&trans, ip,
741                                                      base_offset + offset,
742                                                      HAMMER_REDO_WRITE,
743                                                      bp->b_data + offset,
744                                                      (size_t)n);
745                         } else {
746                                 ip->flags &= ~HAMMER_INODE_REDO;
747                         }
748                 }
749
750                 /*
751                  * If we screwed up we have to undo any VM size changes we
752                  * made.
753                  */
754                 if (error) {
755                         brelse(bp);
756                         if (fixsize) {
757                                 nvtruncbuf(ap->a_vp, ip->ino_data.size,
758                                           hammer_blocksize(ip->ino_data.size),
759                                           hammer_blockoff(ip->ino_data.size));
760                         }
761                         break;
762                 }
763                 kflags |= NOTE_WRITE;
764                 hammer_stats_file_write += n;
765                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
766                 if (ip->ino_data.size < uio->uio_offset) {
767                         ip->ino_data.size = uio->uio_offset;
768                         flags = HAMMER_INODE_SDIRTY;
769                 } else {
770                         flags = 0;
771                 }
772                 ip->ino_data.mtime = trans.time;
773                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
774                 hammer_modify_inode(&trans, ip, flags);
775
776                 /*
777                  * Once we dirty the buffer any cached zone-X offset
778                  * becomes invalid.  HAMMER NOTE: no-history mode cannot 
779                  * allow overwriting over the same data sector unless
780                  * we provide UNDOs for the old data, which we don't.
781                  */
782                 bp->b_bio2.bio_offset = NOOFFSET;
783
784                 /*
785                  * Final buffer disposition.
786                  *
787                  * Because meta-data updates are deferred, HAMMER is
788                  * especially sensitive to excessive bdwrite()s because
789                  * the I/O stream is not broken up by disk reads.  So the
790                  * buffer cache simply cannot keep up.
791                  *
792                  * WARNING!  blksize is variable.  cluster_write() is
793                  *           expected to not blow up if it encounters
794                  *           buffers that do not match the passed blksize.
795                  *
796                  * NOTE!  Hammer shouldn't need to bawrite()/cluster_write().
797                  *        The ip->rsv_recs check should burst-flush the data.
798                  *        If we queue it immediately the buf could be left
799                  *        locked on the device queue for a very long time.
800                  *
801                  * NOTE!  To avoid degenerate stalls due to mismatched block
802                  *        sizes we only honor IO_DIRECT on the write which
803                  *        abuts the end of the buffer.  However, we must
804                  *        honor IO_SYNC in case someone is silly enough to
805                  *        configure a HAMMER file as swap, or when HAMMER
806                  *        is serving NFS (for commits).  Ick ick.
807                  */
808                 bp->b_flags |= B_AGE;
809                 if (ap->a_ioflag & IO_SYNC) {
810                         bwrite(bp);
811                 } else if ((ap->a_ioflag & IO_DIRECT) && endofblk) {
812                         bawrite(bp);
813                 } else {
814 #if 0
815                 if (offset + n == blksize) {
816                         if (hammer_cluster_enable == 0 ||
817                             (ap->a_vp->v_mount->mnt_flag & MNT_NOCLUSTERW)) {
818                                 bawrite(bp);
819                         } else {
820                                 cluster_write(bp, ip->ino_data.size,
821                                               blksize, seqcount);
822                         }
823                 } else {
824 #endif
825                         bdwrite(bp);
826                 }
827         }
828         hammer_done_transaction(&trans);
829         hammer_knote(ap->a_vp, kflags);
830         lwkt_reltoken(&hmp->fs_token);
831         return (error);
832 }
833
834 /*
835  * hammer_vop_access { vp, mode, cred }
836  *
837  * MPSAFE - does not require fs_token
838  */
839 static
840 int
841 hammer_vop_access(struct vop_access_args *ap)
842 {
843         struct hammer_inode *ip = VTOI(ap->a_vp);
844         uid_t uid;
845         gid_t gid;
846         int error;
847
848         ++hammer_stats_file_iopsr;
849         uid = hammer_to_unix_xid(&ip->ino_data.uid);
850         gid = hammer_to_unix_xid(&ip->ino_data.gid);
851
852         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
853                                   ip->ino_data.uflags);
854         return (error);
855 }
856
857 /*
858  * hammer_vop_advlock { vp, id, op, fl, flags }
859  *
860  * MPSAFE - does not require fs_token
861  */
862 static
863 int
864 hammer_vop_advlock(struct vop_advlock_args *ap)
865 {
866         hammer_inode_t ip = VTOI(ap->a_vp);
867
868         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
869 }
870
871 /*
872  * hammer_vop_close { vp, fflag }
873  *
874  * We can only sync-on-close for normal closes.  XXX disabled for now.
875  */
876 static
877 int
878 hammer_vop_close(struct vop_close_args *ap)
879 {
880 #if 0
881         struct vnode *vp = ap->a_vp;
882         hammer_inode_t ip = VTOI(vp);
883         int waitfor;
884         if (ip->flags & (HAMMER_INODE_CLOSESYNC|HAMMER_INODE_CLOSEASYNC)) {
885                 if (vn_islocked(vp) == LK_EXCLUSIVE &&
886                     (vp->v_flag & (VINACTIVE|VRECLAIMED)) == 0) {
887                         if (ip->flags & HAMMER_INODE_CLOSESYNC)
888                                 waitfor = MNT_WAIT;
889                         else
890                                 waitfor = MNT_NOWAIT;
891                         ip->flags &= ~(HAMMER_INODE_CLOSESYNC |
892                                        HAMMER_INODE_CLOSEASYNC);
893                         VOP_FSYNC(vp, MNT_NOWAIT, waitfor);
894                 }
895         }
896 #endif
897         return (vop_stdclose(ap));
898 }
899
900 /*
901  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
902  *
903  * The operating system has already ensured that the directory entry
904  * does not exist and done all appropriate namespace locking.
905  */
906 static
907 int
908 hammer_vop_ncreate(struct vop_ncreate_args *ap)
909 {
910         struct hammer_transaction trans;
911         struct hammer_inode *dip;
912         struct hammer_inode *nip;
913         struct nchandle *nch;
914         hammer_mount_t hmp;
915         int error;
916
917         nch = ap->a_nch;
918         dip = VTOI(ap->a_dvp);
919         hmp = dip->hmp;
920
921         if (dip->flags & HAMMER_INODE_RO)
922                 return (EROFS);
923         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
924                 return (error);
925
926         /*
927          * Create a transaction to cover the operations we perform.
928          */
929         lwkt_gettoken(&hmp->fs_token);
930         hammer_start_transaction(&trans, hmp);
931         ++hammer_stats_file_iopsw;
932
933         /*
934          * Create a new filesystem object of the requested type.  The
935          * returned inode will be referenced and shared-locked to prevent
936          * it from being moved to the flusher.
937          */
938         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
939                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
940                                     NULL, &nip);
941         if (error) {
942                 hkprintf("hammer_create_inode error %d\n", error);
943                 hammer_done_transaction(&trans);
944                 *ap->a_vpp = NULL;
945                 lwkt_reltoken(&hmp->fs_token);
946                 return (error);
947         }
948
949         /*
950          * Add the new filesystem object to the directory.  This will also
951          * bump the inode's link count.
952          */
953         error = hammer_ip_add_directory(&trans, dip,
954                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
955                                         nip);
956         if (error)
957                 hkprintf("hammer_ip_add_directory error %d\n", error);
958
959         /*
960          * Finish up.
961          */
962         if (error) {
963                 hammer_rel_inode(nip, 0);
964                 hammer_done_transaction(&trans);
965                 *ap->a_vpp = NULL;
966         } else {
967                 error = hammer_get_vnode(nip, ap->a_vpp);
968                 hammer_done_transaction(&trans);
969                 hammer_rel_inode(nip, 0);
970                 if (error == 0) {
971                         cache_setunresolved(ap->a_nch);
972                         cache_setvp(ap->a_nch, *ap->a_vpp);
973                 }
974                 hammer_knote(ap->a_dvp, NOTE_WRITE);
975         }
976         lwkt_reltoken(&hmp->fs_token);
977         return (error);
978 }
979
980 /*
981  * hammer_vop_getattr { vp, vap }
982  *
983  * Retrieve an inode's attribute information.  When accessing inodes
984  * historically we fake the atime field to ensure consistent results.
985  * The atime field is stored in the B-Tree element and allowed to be
986  * updated without cycling the element.
987  *
988  * MPSAFE - does not require fs_token
989  */
990 static
991 int
992 hammer_vop_getattr(struct vop_getattr_args *ap)
993 {
994         struct hammer_inode *ip = VTOI(ap->a_vp);
995         struct vattr *vap = ap->a_vap;
996
997         /*
998          * We want the fsid to be different when accessing a filesystem
999          * with different as-of's so programs like diff don't think
1000          * the files are the same.
1001          *
1002          * We also want the fsid to be the same when comparing snapshots,
1003          * or when comparing mirrors (which might be backed by different
1004          * physical devices).  HAMMER fsids are based on the PFS's
1005          * shared_uuid field.
1006          *
1007          * XXX there is a chance of collision here.  The va_fsid reported
1008          * by stat is different from the more involved fsid used in the
1009          * mount structure.
1010          */
1011         ++hammer_stats_file_iopsr;
1012         hammer_lock_sh(&ip->lock);
1013         vap->va_fsid = ip->pfsm->fsid_udev ^ (u_int32_t)ip->obj_asof ^
1014                        (u_int32_t)(ip->obj_asof >> 32);
1015
1016         vap->va_fileid = ip->ino_leaf.base.obj_id;
1017         vap->va_mode = ip->ino_data.mode;
1018         vap->va_nlink = ip->ino_data.nlinks;
1019         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1020         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1021         vap->va_rmajor = 0;
1022         vap->va_rminor = 0;
1023         vap->va_size = ip->ino_data.size;
1024
1025         /*
1026          * Special case for @@PFS softlinks.  The actual size of the
1027          * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
1028          * or for MAX_TID is    "@@-1:%05d" == 10 bytes.
1029          */
1030         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
1031             ip->ino_data.size == 10 &&
1032             ip->obj_asof == HAMMER_MAX_TID &&
1033             ip->obj_localization == 0 &&
1034             strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
1035                     if (ip->pfsm->pfsd.mirror_flags & HAMMER_PFSD_SLAVE)
1036                             vap->va_size = 26;
1037                     else
1038                             vap->va_size = 10;
1039         }
1040
1041         /*
1042          * We must provide a consistent atime and mtime for snapshots
1043          * so people can do a 'tar cf - ... | md5' on them and get
1044          * consistent results.
1045          */
1046         if (ip->flags & HAMMER_INODE_RO) {
1047                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
1048                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
1049         } else {
1050                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
1051                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
1052         }
1053         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
1054         vap->va_flags = ip->ino_data.uflags;
1055         vap->va_gen = 1;        /* hammer inums are unique for all time */
1056         vap->va_blocksize = HAMMER_BUFSIZE;
1057         if (ip->ino_data.size >= HAMMER_XDEMARC) {
1058                 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
1059                                 ~HAMMER_XBUFMASK64;
1060         } else if (ip->ino_data.size > HAMMER_BUFSIZE / 2) {
1061                 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
1062                                 ~HAMMER_BUFMASK64;
1063         } else {
1064                 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
1065         }
1066
1067         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
1068         vap->va_filerev = 0;    /* XXX */
1069         vap->va_uid_uuid = ip->ino_data.uid;
1070         vap->va_gid_uuid = ip->ino_data.gid;
1071         vap->va_fsid_uuid = ip->hmp->fsid;
1072         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
1073                           VA_FSID_UUID_VALID;
1074
1075         switch (ip->ino_data.obj_type) {
1076         case HAMMER_OBJTYPE_CDEV:
1077         case HAMMER_OBJTYPE_BDEV:
1078                 vap->va_rmajor = ip->ino_data.rmajor;
1079                 vap->va_rminor = ip->ino_data.rminor;
1080                 break;
1081         default:
1082                 break;
1083         }
1084         hammer_unlock(&ip->lock);
1085         return(0);
1086 }
1087
1088 /*
1089  * hammer_vop_nresolve { nch, dvp, cred }
1090  *
1091  * Locate the requested directory entry.
1092  */
1093 static
1094 int
1095 hammer_vop_nresolve(struct vop_nresolve_args *ap)
1096 {
1097         struct hammer_transaction trans;
1098         struct namecache *ncp;
1099         hammer_mount_t hmp;
1100         hammer_inode_t dip;
1101         hammer_inode_t ip;
1102         hammer_tid_t asof;
1103         struct hammer_cursor cursor;
1104         struct vnode *vp;
1105         int64_t namekey;
1106         int error;
1107         int i;
1108         int nlen;
1109         int flags;
1110         int ispfs;
1111         int64_t obj_id;
1112         u_int32_t localization;
1113         u_int32_t max_iterations;
1114
1115         /*
1116          * Misc initialization, plus handle as-of name extensions.  Look for
1117          * the '@@' extension.  Note that as-of files and directories cannot
1118          * be modified.
1119          */
1120         dip = VTOI(ap->a_dvp);
1121         ncp = ap->a_nch->ncp;
1122         asof = dip->obj_asof;
1123         localization = dip->obj_localization;   /* for code consistency */
1124         nlen = ncp->nc_nlen;
1125         flags = dip->flags & HAMMER_INODE_RO;
1126         ispfs = 0;
1127         hmp = dip->hmp;
1128
1129         lwkt_gettoken(&hmp->fs_token);
1130         hammer_simple_transaction(&trans, hmp);
1131         ++hammer_stats_file_iopsr;
1132
1133         for (i = 0; i < nlen; ++i) {
1134                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
1135                         error = hammer_str_to_tid(ncp->nc_name + i + 2,
1136                                                   &ispfs, &asof, &localization);
1137                         if (error != 0) {
1138                                 i = nlen;
1139                                 break;
1140                         }
1141                         if (asof != HAMMER_MAX_TID)
1142                                 flags |= HAMMER_INODE_RO;
1143                         break;
1144                 }
1145         }
1146         nlen = i;
1147
1148         /*
1149          * If this is a PFS softlink we dive into the PFS
1150          */
1151         if (ispfs && nlen == 0) {
1152                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1153                                       asof, localization,
1154                                       flags, &error);
1155                 if (error == 0) {
1156                         error = hammer_get_vnode(ip, &vp);
1157                         hammer_rel_inode(ip, 0);
1158                 } else {
1159                         vp = NULL;
1160                 }
1161                 if (error == 0) {
1162                         vn_unlock(vp);
1163                         cache_setvp(ap->a_nch, vp);
1164                         vrele(vp);
1165                 }
1166                 goto done;
1167         }
1168
1169         /*
1170          * If there is no path component the time extension is relative to dip.
1171          * e.g. "fubar/@@<snapshot>"
1172          *
1173          * "." is handled by the kernel, but ".@@<snapshot>" is not.
1174          * e.g. "fubar/.@@<snapshot>"
1175          *
1176          * ".." is handled by the kernel.  We do not currently handle
1177          * "..@<snapshot>".
1178          */
1179         if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1180                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
1181                                       asof, dip->obj_localization,
1182                                       flags, &error);
1183                 if (error == 0) {
1184                         error = hammer_get_vnode(ip, &vp);
1185                         hammer_rel_inode(ip, 0);
1186                 } else {
1187                         vp = NULL;
1188                 }
1189                 if (error == 0) {
1190                         vn_unlock(vp);
1191                         cache_setvp(ap->a_nch, vp);
1192                         vrele(vp);
1193                 }
1194                 goto done;
1195         }
1196
1197         /*
1198          * Calculate the namekey and setup the key range for the scan.  This
1199          * works kinda like a chained hash table where the lower 32 bits
1200          * of the namekey synthesize the chain.
1201          *
1202          * The key range is inclusive of both key_beg and key_end.
1203          */
1204         namekey = hammer_directory_namekey(dip, ncp->nc_name, nlen,
1205                                            &max_iterations);
1206
1207         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1208         cursor.key_beg.localization = dip->obj_localization +
1209                                       hammer_dir_localization(dip);
1210         cursor.key_beg.obj_id = dip->obj_id;
1211         cursor.key_beg.key = namekey;
1212         cursor.key_beg.create_tid = 0;
1213         cursor.key_beg.delete_tid = 0;
1214         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1215         cursor.key_beg.obj_type = 0;
1216
1217         cursor.key_end = cursor.key_beg;
1218         cursor.key_end.key += max_iterations;
1219         cursor.asof = asof;
1220         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1221
1222         /*
1223          * Scan all matching records (the chain), locate the one matching
1224          * the requested path component.
1225          *
1226          * The hammer_ip_*() functions merge in-memory records with on-disk
1227          * records for the purposes of the search.
1228          */
1229         obj_id = 0;
1230         localization = HAMMER_DEF_LOCALIZATION;
1231
1232         if (error == 0) {
1233                 error = hammer_ip_first(&cursor);
1234                 while (error == 0) {
1235                         error = hammer_ip_resolve_data(&cursor);
1236                         if (error)
1237                                 break;
1238                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1239                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1240                                 obj_id = cursor.data->entry.obj_id;
1241                                 localization = cursor.data->entry.localization;
1242                                 break;
1243                         }
1244                         error = hammer_ip_next(&cursor);
1245                 }
1246         }
1247         hammer_done_cursor(&cursor);
1248
1249         /*
1250          * Lookup the obj_id.  This should always succeed.  If it does not
1251          * the filesystem may be damaged and we return a dummy inode.
1252          */
1253         if (error == 0) {
1254                 ip = hammer_get_inode(&trans, dip, obj_id,
1255                                       asof, localization,
1256                                       flags, &error);
1257                 if (error == ENOENT) {
1258                         kprintf("HAMMER: WARNING: Missing "
1259                                 "inode for dirent \"%s\"\n"
1260                                 "\tobj_id = %016llx, asof=%016llx, lo=%08x\n",
1261                                 ncp->nc_name,
1262                                 (long long)obj_id, (long long)asof,
1263                                 localization);
1264                         error = 0;
1265                         ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1266                                                     asof, localization,
1267                                                     flags, &error);
1268                 }
1269                 if (error == 0) {
1270                         error = hammer_get_vnode(ip, &vp);
1271                         hammer_rel_inode(ip, 0);
1272                 } else {
1273                         vp = NULL;
1274                 }
1275                 if (error == 0) {
1276                         vn_unlock(vp);
1277                         cache_setvp(ap->a_nch, vp);
1278                         vrele(vp);
1279                 }
1280         } else if (error == ENOENT) {
1281                 cache_setvp(ap->a_nch, NULL);
1282         }
1283 done:
1284         hammer_done_transaction(&trans);
1285         lwkt_reltoken(&hmp->fs_token);
1286         return (error);
1287 }
1288
1289 /*
1290  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1291  *
1292  * Locate the parent directory of a directory vnode.
1293  *
1294  * dvp is referenced but not locked.  *vpp must be returned referenced and
1295  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
1296  * at the root, instead it could indicate that the directory we were in was
1297  * removed.
1298  *
1299  * NOTE: as-of sequences are not linked into the directory structure.  If
1300  * we are at the root with a different asof then the mount point, reload
1301  * the same directory with the mount point's asof.   I'm not sure what this
1302  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
1303  * get confused, but it hasn't been tested.
1304  */
1305 static
1306 int
1307 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1308 {
1309         struct hammer_transaction trans;
1310         struct hammer_inode *dip;
1311         struct hammer_inode *ip;
1312         hammer_mount_t hmp;
1313         int64_t parent_obj_id;
1314         u_int32_t parent_obj_localization;
1315         hammer_tid_t asof;
1316         int error;
1317
1318         dip = VTOI(ap->a_dvp);
1319         asof = dip->obj_asof;
1320         hmp = dip->hmp;
1321
1322         /*
1323          * Whos are parent?  This could be the root of a pseudo-filesystem
1324          * whos parent is in another localization domain.
1325          */
1326         lwkt_gettoken(&hmp->fs_token);
1327         parent_obj_id = dip->ino_data.parent_obj_id;
1328         if (dip->obj_id == HAMMER_OBJID_ROOT)
1329                 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
1330         else
1331                 parent_obj_localization = dip->obj_localization;
1332
1333         if (parent_obj_id == 0) {
1334                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1335                    asof != hmp->asof) {
1336                         parent_obj_id = dip->obj_id;
1337                         asof = hmp->asof;
1338                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1339                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
1340                                   (long long)dip->obj_asof);
1341                 } else {
1342                         *ap->a_vpp = NULL;
1343                         lwkt_reltoken(&hmp->fs_token);
1344                         return ENOENT;
1345                 }
1346         }
1347
1348         hammer_simple_transaction(&trans, hmp);
1349         ++hammer_stats_file_iopsr;
1350
1351         ip = hammer_get_inode(&trans, dip, parent_obj_id,
1352                               asof, parent_obj_localization,
1353                               dip->flags, &error);
1354         if (ip) {
1355                 error = hammer_get_vnode(ip, ap->a_vpp);
1356                 hammer_rel_inode(ip, 0);
1357         } else {
1358                 *ap->a_vpp = NULL;
1359         }
1360         hammer_done_transaction(&trans);
1361         lwkt_reltoken(&hmp->fs_token);
1362         return (error);
1363 }
1364
1365 /*
1366  * hammer_vop_nlink { nch, dvp, vp, cred }
1367  */
1368 static
1369 int
1370 hammer_vop_nlink(struct vop_nlink_args *ap)
1371 {
1372         struct hammer_transaction trans;
1373         struct hammer_inode *dip;
1374         struct hammer_inode *ip;
1375         struct nchandle *nch;
1376         hammer_mount_t hmp;
1377         int error;
1378
1379         if (ap->a_dvp->v_mount != ap->a_vp->v_mount)    
1380                 return(EXDEV);
1381
1382         nch = ap->a_nch;
1383         dip = VTOI(ap->a_dvp);
1384         ip = VTOI(ap->a_vp);
1385         hmp = dip->hmp;
1386
1387         if (dip->obj_localization != ip->obj_localization)
1388                 return(EXDEV);
1389
1390         if (dip->flags & HAMMER_INODE_RO)
1391                 return (EROFS);
1392         if (ip->flags & HAMMER_INODE_RO)
1393                 return (EROFS);
1394         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1395                 return (error);
1396
1397         /*
1398          * Create a transaction to cover the operations we perform.
1399          */
1400         lwkt_gettoken(&hmp->fs_token);
1401         hammer_start_transaction(&trans, hmp);
1402         ++hammer_stats_file_iopsw;
1403
1404         /*
1405          * Add the filesystem object to the directory.  Note that neither
1406          * dip nor ip are referenced or locked, but their vnodes are
1407          * referenced.  This function will bump the inode's link count.
1408          */
1409         error = hammer_ip_add_directory(&trans, dip,
1410                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1411                                         ip);
1412
1413         /*
1414          * Finish up.
1415          */
1416         if (error == 0) {
1417                 cache_setunresolved(nch);
1418                 cache_setvp(nch, ap->a_vp);
1419         }
1420         hammer_done_transaction(&trans);
1421         hammer_knote(ap->a_vp, NOTE_LINK);
1422         hammer_knote(ap->a_dvp, NOTE_WRITE);
1423         lwkt_reltoken(&hmp->fs_token);
1424         return (error);
1425 }
1426
1427 /*
1428  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1429  *
1430  * The operating system has already ensured that the directory entry
1431  * does not exist and done all appropriate namespace locking.
1432  */
1433 static
1434 int
1435 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1436 {
1437         struct hammer_transaction trans;
1438         struct hammer_inode *dip;
1439         struct hammer_inode *nip;
1440         struct nchandle *nch;
1441         hammer_mount_t hmp;
1442         int error;
1443
1444         nch = ap->a_nch;
1445         dip = VTOI(ap->a_dvp);
1446         hmp = dip->hmp;
1447
1448         if (dip->flags & HAMMER_INODE_RO)
1449                 return (EROFS);
1450         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1451                 return (error);
1452
1453         /*
1454          * Create a transaction to cover the operations we perform.
1455          */
1456         lwkt_gettoken(&hmp->fs_token);
1457         hammer_start_transaction(&trans, hmp);
1458         ++hammer_stats_file_iopsw;
1459
1460         /*
1461          * Create a new filesystem object of the requested type.  The
1462          * returned inode will be referenced but not locked.
1463          */
1464         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1465                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1466                                     NULL, &nip);
1467         if (error) {
1468                 hkprintf("hammer_mkdir error %d\n", error);
1469                 hammer_done_transaction(&trans);
1470                 *ap->a_vpp = NULL;
1471                 lwkt_reltoken(&hmp->fs_token);
1472                 return (error);
1473         }
1474         /*
1475          * Add the new filesystem object to the directory.  This will also
1476          * bump the inode's link count.
1477          */
1478         error = hammer_ip_add_directory(&trans, dip,
1479                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1480                                         nip);
1481         if (error)
1482                 hkprintf("hammer_mkdir (add) error %d\n", error);
1483
1484         /*
1485          * Finish up.
1486          */
1487         if (error) {
1488                 hammer_rel_inode(nip, 0);
1489                 *ap->a_vpp = NULL;
1490         } else {
1491                 error = hammer_get_vnode(nip, ap->a_vpp);
1492                 hammer_rel_inode(nip, 0);
1493                 if (error == 0) {
1494                         cache_setunresolved(ap->a_nch);
1495                         cache_setvp(ap->a_nch, *ap->a_vpp);
1496                 }
1497         }
1498         hammer_done_transaction(&trans);
1499         if (error == 0)
1500                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1501         lwkt_reltoken(&hmp->fs_token);
1502         return (error);
1503 }
1504
1505 /*
1506  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1507  *
1508  * The operating system has already ensured that the directory entry
1509  * does not exist and done all appropriate namespace locking.
1510  */
1511 static
1512 int
1513 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1514 {
1515         struct hammer_transaction trans;
1516         struct hammer_inode *dip;
1517         struct hammer_inode *nip;
1518         struct nchandle *nch;
1519         hammer_mount_t hmp;
1520         int error;
1521
1522         nch = ap->a_nch;
1523         dip = VTOI(ap->a_dvp);
1524         hmp = dip->hmp;
1525
1526         if (dip->flags & HAMMER_INODE_RO)
1527                 return (EROFS);
1528         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1529                 return (error);
1530
1531         /*
1532          * Create a transaction to cover the operations we perform.
1533          */
1534         lwkt_gettoken(&hmp->fs_token);
1535         hammer_start_transaction(&trans, hmp);
1536         ++hammer_stats_file_iopsw;
1537
1538         /*
1539          * Create a new filesystem object of the requested type.  The
1540          * returned inode will be referenced but not locked.
1541          *
1542          * If mknod specifies a directory a pseudo-fs is created.
1543          */
1544         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1545                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1546                                     NULL, &nip);
1547         if (error) {
1548                 hammer_done_transaction(&trans);
1549                 *ap->a_vpp = NULL;
1550                 lwkt_reltoken(&hmp->fs_token);
1551                 return (error);
1552         }
1553
1554         /*
1555          * Add the new filesystem object to the directory.  This will also
1556          * bump the inode's link count.
1557          */
1558         error = hammer_ip_add_directory(&trans, dip,
1559                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1560                                         nip);
1561
1562         /*
1563          * Finish up.
1564          */
1565         if (error) {
1566                 hammer_rel_inode(nip, 0);
1567                 *ap->a_vpp = NULL;
1568         } else {
1569                 error = hammer_get_vnode(nip, ap->a_vpp);
1570                 hammer_rel_inode(nip, 0);
1571                 if (error == 0) {
1572                         cache_setunresolved(ap->a_nch);
1573                         cache_setvp(ap->a_nch, *ap->a_vpp);
1574                 }
1575         }
1576         hammer_done_transaction(&trans);
1577         if (error == 0)
1578                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1579         lwkt_reltoken(&hmp->fs_token);
1580         return (error);
1581 }
1582
1583 /*
1584  * hammer_vop_open { vp, mode, cred, fp }
1585  *
1586  * MPSAFE (does not require fs_token)
1587  */
1588 static
1589 int
1590 hammer_vop_open(struct vop_open_args *ap)
1591 {
1592         hammer_inode_t ip;
1593
1594         ++hammer_stats_file_iopsr;
1595         ip = VTOI(ap->a_vp);
1596
1597         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1598                 return (EROFS);
1599         return(vop_stdopen(ap));
1600 }
1601
1602 /*
1603  * hammer_vop_print { vp }
1604  */
1605 static
1606 int
1607 hammer_vop_print(struct vop_print_args *ap)
1608 {
1609         return EOPNOTSUPP;
1610 }
1611
1612 /*
1613  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1614  */
1615 static
1616 int
1617 hammer_vop_readdir(struct vop_readdir_args *ap)
1618 {
1619         struct hammer_transaction trans;
1620         struct hammer_cursor cursor;
1621         struct hammer_inode *ip;
1622         hammer_mount_t hmp;
1623         struct uio *uio;
1624         hammer_base_elm_t base;
1625         int error;
1626         int cookie_index;
1627         int ncookies;
1628         off_t *cookies;
1629         off_t saveoff;
1630         int r;
1631         int dtype;
1632
1633         ++hammer_stats_file_iopsr;
1634         ip = VTOI(ap->a_vp);
1635         uio = ap->a_uio;
1636         saveoff = uio->uio_offset;
1637         hmp = ip->hmp;
1638
1639         if (ap->a_ncookies) {
1640                 ncookies = uio->uio_resid / 16 + 1;
1641                 if (ncookies > 1024)
1642                         ncookies = 1024;
1643                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1644                 cookie_index = 0;
1645         } else {
1646                 ncookies = -1;
1647                 cookies = NULL;
1648                 cookie_index = 0;
1649         }
1650
1651         lwkt_gettoken(&hmp->fs_token);
1652         hammer_simple_transaction(&trans, hmp);
1653
1654         /*
1655          * Handle artificial entries
1656          *
1657          * It should be noted that the minimum value for a directory
1658          * hash key on-media is 0x0000000100000000, so we can use anything
1659          * less then that to represent our 'special' key space.
1660          */
1661         error = 0;
1662         if (saveoff == 0) {
1663                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1664                 if (r)
1665                         goto done;
1666                 if (cookies)
1667                         cookies[cookie_index] = saveoff;
1668                 ++saveoff;
1669                 ++cookie_index;
1670                 if (cookie_index == ncookies)
1671                         goto done;
1672         }
1673         if (saveoff == 1) {
1674                 if (ip->ino_data.parent_obj_id) {
1675                         r = vop_write_dirent(&error, uio,
1676                                              ip->ino_data.parent_obj_id,
1677                                              DT_DIR, 2, "..");
1678                 } else {
1679                         r = vop_write_dirent(&error, uio,
1680                                              ip->obj_id, DT_DIR, 2, "..");
1681                 }
1682                 if (r)
1683                         goto done;
1684                 if (cookies)
1685                         cookies[cookie_index] = saveoff;
1686                 ++saveoff;
1687                 ++cookie_index;
1688                 if (cookie_index == ncookies)
1689                         goto done;
1690         }
1691
1692         /*
1693          * Key range (begin and end inclusive) to scan.  Directory keys
1694          * directly translate to a 64 bit 'seek' position.
1695          */
1696         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1697         cursor.key_beg.localization = ip->obj_localization +
1698                                       hammer_dir_localization(ip);
1699         cursor.key_beg.obj_id = ip->obj_id;
1700         cursor.key_beg.create_tid = 0;
1701         cursor.key_beg.delete_tid = 0;
1702         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1703         cursor.key_beg.obj_type = 0;
1704         cursor.key_beg.key = saveoff;
1705
1706         cursor.key_end = cursor.key_beg;
1707         cursor.key_end.key = HAMMER_MAX_KEY;
1708         cursor.asof = ip->obj_asof;
1709         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1710
1711         error = hammer_ip_first(&cursor);
1712
1713         while (error == 0) {
1714                 error = hammer_ip_resolve_data(&cursor);
1715                 if (error)
1716                         break;
1717                 base = &cursor.leaf->base;
1718                 saveoff = base->key;
1719                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1720
1721                 if (base->obj_id != ip->obj_id)
1722                         panic("readdir: bad record at %p", cursor.node);
1723
1724                 /*
1725                  * Convert pseudo-filesystems into softlinks
1726                  */
1727                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1728                 r = vop_write_dirent(
1729                              &error, uio, cursor.data->entry.obj_id,
1730                              dtype,
1731                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1732                              (void *)cursor.data->entry.name);
1733                 if (r)
1734                         break;
1735                 ++saveoff;
1736                 if (cookies)
1737                         cookies[cookie_index] = base->key;
1738                 ++cookie_index;
1739                 if (cookie_index == ncookies)
1740                         break;
1741                 error = hammer_ip_next(&cursor);
1742         }
1743         hammer_done_cursor(&cursor);
1744
1745 done:
1746         hammer_done_transaction(&trans);
1747
1748         if (ap->a_eofflag)
1749                 *ap->a_eofflag = (error == ENOENT);
1750         uio->uio_offset = saveoff;
1751         if (error && cookie_index == 0) {
1752                 if (error == ENOENT)
1753                         error = 0;
1754                 if (cookies) {
1755                         kfree(cookies, M_TEMP);
1756                         *ap->a_ncookies = 0;
1757                         *ap->a_cookies = NULL;
1758                 }
1759         } else {
1760                 if (error == ENOENT)
1761                         error = 0;
1762                 if (cookies) {
1763                         *ap->a_ncookies = cookie_index;
1764                         *ap->a_cookies = cookies;
1765                 }
1766         }
1767         lwkt_reltoken(&hmp->fs_token);
1768         return(error);
1769 }
1770
1771 /*
1772  * hammer_vop_readlink { vp, uio, cred }
1773  */
1774 static
1775 int
1776 hammer_vop_readlink(struct vop_readlink_args *ap)
1777 {
1778         struct hammer_transaction trans;
1779         struct hammer_cursor cursor;
1780         struct hammer_inode *ip;
1781         hammer_mount_t hmp;
1782         char buf[32];
1783         u_int32_t localization;
1784         hammer_pseudofs_inmem_t pfsm;
1785         int error;
1786
1787         ip = VTOI(ap->a_vp);
1788         hmp = ip->hmp;
1789
1790         lwkt_gettoken(&hmp->fs_token);
1791
1792         /*
1793          * Shortcut if the symlink data was stuffed into ino_data.
1794          *
1795          * Also expand special "@@PFS%05d" softlinks (expansion only
1796          * occurs for non-historical (current) accesses made from the
1797          * primary filesystem).
1798          */
1799         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1800                 char *ptr;
1801                 int bytes;
1802
1803                 ptr = ip->ino_data.ext.symlink;
1804                 bytes = (int)ip->ino_data.size;
1805                 if (bytes == 10 &&
1806                     ip->obj_asof == HAMMER_MAX_TID &&
1807                     ip->obj_localization == 0 &&
1808                     strncmp(ptr, "@@PFS", 5) == 0) {
1809                         hammer_simple_transaction(&trans, hmp);
1810                         bcopy(ptr + 5, buf, 5);
1811                         buf[5] = 0;
1812                         localization = strtoul(buf, NULL, 10) << 16;
1813                         pfsm = hammer_load_pseudofs(&trans, localization,
1814                                                     &error);
1815                         if (error == 0) {
1816                                 if (pfsm->pfsd.mirror_flags &
1817                                     HAMMER_PFSD_SLAVE) {
1818                                         /* vap->va_size == 26 */
1819                                         ksnprintf(buf, sizeof(buf),
1820                                                   "@@0x%016llx:%05d",
1821                                                   (long long)pfsm->pfsd.sync_end_tid,
1822                                                   localization >> 16);
1823                                 } else {
1824                                         /* vap->va_size == 10 */
1825                                         ksnprintf(buf, sizeof(buf),
1826                                                   "@@-1:%05d",
1827                                                   localization >> 16);
1828 #if 0
1829                                         ksnprintf(buf, sizeof(buf),
1830                                                   "@@0x%016llx:%05d",
1831                                                   (long long)HAMMER_MAX_TID,
1832                                                   localization >> 16);
1833 #endif
1834                                 }
1835                                 ptr = buf;
1836                                 bytes = strlen(buf);
1837                         }
1838                         if (pfsm)
1839                                 hammer_rel_pseudofs(hmp, pfsm);
1840                         hammer_done_transaction(&trans);
1841                 }
1842                 error = uiomove(ptr, bytes, ap->a_uio);
1843                 lwkt_reltoken(&hmp->fs_token);
1844                 return(error);
1845         }
1846
1847         /*
1848          * Long version
1849          */
1850         hammer_simple_transaction(&trans, hmp);
1851         ++hammer_stats_file_iopsr;
1852         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1853
1854         /*
1855          * Key range (begin and end inclusive) to scan.  Directory keys
1856          * directly translate to a 64 bit 'seek' position.
1857          */
1858         cursor.key_beg.localization = ip->obj_localization +
1859                                       HAMMER_LOCALIZE_MISC;
1860         cursor.key_beg.obj_id = ip->obj_id;
1861         cursor.key_beg.create_tid = 0;
1862         cursor.key_beg.delete_tid = 0;
1863         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1864         cursor.key_beg.obj_type = 0;
1865         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1866         cursor.asof = ip->obj_asof;
1867         cursor.flags |= HAMMER_CURSOR_ASOF;
1868
1869         error = hammer_ip_lookup(&cursor);
1870         if (error == 0) {
1871                 error = hammer_ip_resolve_data(&cursor);
1872                 if (error == 0) {
1873                         KKASSERT(cursor.leaf->data_len >=
1874                                  HAMMER_SYMLINK_NAME_OFF);
1875                         error = uiomove(cursor.data->symlink.name,
1876                                         cursor.leaf->data_len -
1877                                                 HAMMER_SYMLINK_NAME_OFF,
1878                                         ap->a_uio);
1879                 }
1880         }
1881         hammer_done_cursor(&cursor);
1882         hammer_done_transaction(&trans);
1883         lwkt_reltoken(&hmp->fs_token);
1884         return(error);
1885 }
1886
1887 /*
1888  * hammer_vop_nremove { nch, dvp, cred }
1889  */
1890 static
1891 int
1892 hammer_vop_nremove(struct vop_nremove_args *ap)
1893 {
1894         struct hammer_transaction trans;
1895         struct hammer_inode *dip;
1896         hammer_mount_t hmp;
1897         int error;
1898
1899         dip = VTOI(ap->a_dvp);
1900         hmp = dip->hmp;
1901
1902         if (hammer_nohistory(dip) == 0 &&
1903             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1904                 return (error);
1905         }
1906
1907         lwkt_gettoken(&hmp->fs_token);
1908         hammer_start_transaction(&trans, hmp);
1909         ++hammer_stats_file_iopsw;
1910         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1911         hammer_done_transaction(&trans);
1912         if (error == 0)
1913                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1914         lwkt_reltoken(&hmp->fs_token);
1915         return (error);
1916 }
1917
1918 /*
1919  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1920  */
1921 static
1922 int
1923 hammer_vop_nrename(struct vop_nrename_args *ap)
1924 {
1925         struct hammer_transaction trans;
1926         struct namecache *fncp;
1927         struct namecache *tncp;
1928         struct hammer_inode *fdip;
1929         struct hammer_inode *tdip;
1930         struct hammer_inode *ip;
1931         hammer_mount_t hmp;
1932         struct hammer_cursor cursor;
1933         int64_t namekey;
1934         u_int32_t max_iterations;
1935         int nlen, error;
1936
1937         if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount) 
1938                 return(EXDEV);
1939         if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1940                 return(EXDEV);
1941
1942         fdip = VTOI(ap->a_fdvp);
1943         tdip = VTOI(ap->a_tdvp);
1944         fncp = ap->a_fnch->ncp;
1945         tncp = ap->a_tnch->ncp;
1946         ip = VTOI(fncp->nc_vp);
1947         KKASSERT(ip != NULL);
1948
1949         hmp = ip->hmp;
1950
1951         if (fdip->obj_localization != tdip->obj_localization)
1952                 return(EXDEV);
1953         if (fdip->obj_localization != ip->obj_localization)
1954                 return(EXDEV);
1955
1956         if (fdip->flags & HAMMER_INODE_RO)
1957                 return (EROFS);
1958         if (tdip->flags & HAMMER_INODE_RO)
1959                 return (EROFS);
1960         if (ip->flags & HAMMER_INODE_RO)
1961                 return (EROFS);
1962         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1963                 return (error);
1964
1965         lwkt_gettoken(&hmp->fs_token);
1966         hammer_start_transaction(&trans, hmp);
1967         ++hammer_stats_file_iopsw;
1968
1969         /*
1970          * Remove tncp from the target directory and then link ip as
1971          * tncp. XXX pass trans to dounlink
1972          *
1973          * Force the inode sync-time to match the transaction so it is
1974          * in-sync with the creation of the target directory entry.
1975          */
1976         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
1977                                 ap->a_cred, 0, -1);
1978         if (error == 0 || error == ENOENT) {
1979                 error = hammer_ip_add_directory(&trans, tdip,
1980                                                 tncp->nc_name, tncp->nc_nlen,
1981                                                 ip);
1982                 if (error == 0) {
1983                         ip->ino_data.parent_obj_id = tdip->obj_id;
1984                         ip->ino_data.ctime = trans.time;
1985                         hammer_modify_inode(&trans, ip, HAMMER_INODE_DDIRTY);
1986                 }
1987         }
1988         if (error)
1989                 goto failed; /* XXX */
1990
1991         /*
1992          * Locate the record in the originating directory and remove it.
1993          *
1994          * Calculate the namekey and setup the key range for the scan.  This
1995          * works kinda like a chained hash table where the lower 32 bits
1996          * of the namekey synthesize the chain.
1997          *
1998          * The key range is inclusive of both key_beg and key_end.
1999          */
2000         namekey = hammer_directory_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
2001                                            &max_iterations);
2002 retry:
2003         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
2004         cursor.key_beg.localization = fdip->obj_localization +
2005                                       hammer_dir_localization(fdip);
2006         cursor.key_beg.obj_id = fdip->obj_id;
2007         cursor.key_beg.key = namekey;
2008         cursor.key_beg.create_tid = 0;
2009         cursor.key_beg.delete_tid = 0;
2010         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2011         cursor.key_beg.obj_type = 0;
2012
2013         cursor.key_end = cursor.key_beg;
2014         cursor.key_end.key += max_iterations;
2015         cursor.asof = fdip->obj_asof;
2016         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2017
2018         /*
2019          * Scan all matching records (the chain), locate the one matching
2020          * the requested path component.
2021          *
2022          * The hammer_ip_*() functions merge in-memory records with on-disk
2023          * records for the purposes of the search.
2024          */
2025         error = hammer_ip_first(&cursor);
2026         while (error == 0) {
2027                 if (hammer_ip_resolve_data(&cursor) != 0)
2028                         break;
2029                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2030                 KKASSERT(nlen > 0);
2031                 if (fncp->nc_nlen == nlen &&
2032                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2033                         break;
2034                 }
2035                 error = hammer_ip_next(&cursor);
2036         }
2037
2038         /*
2039          * If all is ok we have to get the inode so we can adjust nlinks.
2040          *
2041          * WARNING: hammer_ip_del_directory() may have to terminate the
2042          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
2043          * twice.
2044          */
2045         if (error == 0)
2046                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
2047
2048         /*
2049          * XXX A deadlock here will break rename's atomicy for the purposes
2050          * of crash recovery.
2051          */
2052         if (error == EDEADLK) {
2053                 hammer_done_cursor(&cursor);
2054                 goto retry;
2055         }
2056
2057         /*
2058          * Cleanup and tell the kernel that the rename succeeded.
2059          *
2060          * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2061          *       without formally acquiring the vp since the vp might
2062          *       have zero refs on it, or in the middle of a reclaim,
2063          *       etc.
2064          */
2065         hammer_done_cursor(&cursor);
2066         if (error == 0) {
2067                 cache_rename(ap->a_fnch, ap->a_tnch);
2068                 hammer_knote(ap->a_fdvp, NOTE_WRITE);
2069                 hammer_knote(ap->a_tdvp, NOTE_WRITE);
2070                 while (ip->vp) {
2071                         struct vnode *vp;
2072
2073                         error = hammer_get_vnode(ip, &vp);
2074                         if (error == 0 && vp) {
2075                                 vn_unlock(vp);
2076                                 hammer_knote(ip->vp, NOTE_RENAME);
2077                                 vrele(vp);
2078                                 break;
2079                         }
2080                         kprintf("Debug: HAMMER ip/vp race2 avoided\n");
2081                 }
2082         }
2083
2084 failed:
2085         hammer_done_transaction(&trans);
2086         lwkt_reltoken(&hmp->fs_token);
2087         return (error);
2088 }
2089
2090 /*
2091  * hammer_vop_nrmdir { nch, dvp, cred }
2092  */
2093 static
2094 int
2095 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
2096 {
2097         struct hammer_transaction trans;
2098         struct hammer_inode *dip;
2099         hammer_mount_t hmp;
2100         int error;
2101
2102         dip = VTOI(ap->a_dvp);
2103         hmp = dip->hmp;
2104
2105         if (hammer_nohistory(dip) == 0 &&
2106             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2107                 return (error);
2108         }
2109
2110         lwkt_gettoken(&hmp->fs_token);
2111         hammer_start_transaction(&trans, hmp);
2112         ++hammer_stats_file_iopsw;
2113         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
2114         hammer_done_transaction(&trans);
2115         if (error == 0)
2116                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
2117         lwkt_reltoken(&hmp->fs_token);
2118         return (error);
2119 }
2120
2121 /*
2122  * hammer_vop_markatime { vp, cred }
2123  */
2124 static
2125 int
2126 hammer_vop_markatime(struct vop_markatime_args *ap)
2127 {
2128         struct hammer_transaction trans;
2129         struct hammer_inode *ip;
2130         hammer_mount_t hmp;
2131
2132         ip = VTOI(ap->a_vp);
2133         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2134                 return (EROFS);
2135         if (ip->flags & HAMMER_INODE_RO)
2136                 return (EROFS);
2137         hmp = ip->hmp;
2138         if (hmp->mp->mnt_flag & MNT_NOATIME)
2139                 return (0);
2140         lwkt_gettoken(&hmp->fs_token);
2141         hammer_start_transaction(&trans, hmp);
2142         ++hammer_stats_file_iopsw;
2143
2144         ip->ino_data.atime = trans.time;
2145         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
2146         hammer_done_transaction(&trans);
2147         hammer_knote(ap->a_vp, NOTE_ATTRIB);
2148         lwkt_reltoken(&hmp->fs_token);
2149         return (0);
2150 }
2151
2152 /*
2153  * hammer_vop_setattr { vp, vap, cred }
2154  */
2155 static
2156 int
2157 hammer_vop_setattr(struct vop_setattr_args *ap)
2158 {
2159         struct hammer_transaction trans;
2160         struct hammer_inode *ip;
2161         struct vattr *vap;
2162         hammer_mount_t hmp;
2163         int modflags;
2164         int error;
2165         int truncating;
2166         int blksize;
2167         int kflags;
2168 #if 0
2169         int64_t aligned_size;
2170 #endif
2171         u_int32_t flags;
2172
2173         vap = ap->a_vap;
2174         ip = ap->a_vp->v_data;
2175         modflags = 0;
2176         kflags = 0;
2177         hmp = ip->hmp;
2178
2179         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2180                 return(EROFS);
2181         if (ip->flags & HAMMER_INODE_RO)
2182                 return (EROFS);
2183         if (hammer_nohistory(ip) == 0 &&
2184             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2185                 return (error);
2186         }
2187
2188         lwkt_gettoken(&hmp->fs_token);
2189         hammer_start_transaction(&trans, hmp);
2190         ++hammer_stats_file_iopsw;
2191         error = 0;
2192
2193         if (vap->va_flags != VNOVAL) {
2194                 flags = ip->ino_data.uflags;
2195                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
2196                                          hammer_to_unix_xid(&ip->ino_data.uid),
2197                                          ap->a_cred);
2198                 if (error == 0) {
2199                         if (ip->ino_data.uflags != flags) {
2200                                 ip->ino_data.uflags = flags;
2201                                 ip->ino_data.ctime = trans.time;
2202                                 modflags |= HAMMER_INODE_DDIRTY;
2203                                 kflags |= NOTE_ATTRIB;
2204                         }
2205                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2206                                 error = 0;
2207                                 goto done;
2208                         }
2209                 }
2210                 goto done;
2211         }
2212         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2213                 error = EPERM;
2214                 goto done;
2215         }
2216         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2217                 mode_t cur_mode = ip->ino_data.mode;
2218                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2219                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2220                 uuid_t uuid_uid;
2221                 uuid_t uuid_gid;
2222
2223                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2224                                          ap->a_cred,
2225                                          &cur_uid, &cur_gid, &cur_mode);
2226                 if (error == 0) {
2227                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
2228                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
2229                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
2230                                  sizeof(uuid_uid)) ||
2231                             bcmp(&uuid_gid, &ip->ino_data.gid,
2232                                  sizeof(uuid_gid)) ||
2233                             ip->ino_data.mode != cur_mode
2234                         ) {
2235                                 ip->ino_data.uid = uuid_uid;
2236                                 ip->ino_data.gid = uuid_gid;
2237                                 ip->ino_data.mode = cur_mode;
2238                                 ip->ino_data.ctime = trans.time;
2239                                 modflags |= HAMMER_INODE_DDIRTY;
2240                         }
2241                         kflags |= NOTE_ATTRIB;
2242                 }
2243         }
2244         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2245                 switch(ap->a_vp->v_type) {
2246                 case VREG:
2247                         if (vap->va_size == ip->ino_data.size)
2248                                 break;
2249
2250                         /*
2251                          * Log the operation if in fast-fsync mode or if
2252                          * there are unterminated redo write records present.
2253                          *
2254                          * The second check is needed so the recovery code
2255                          * properly truncates write redos even if nominal
2256                          * REDO operations is turned off due to excessive
2257                          * writes, because the related records might be
2258                          * destroyed and never lay down a TERM_WRITE.
2259                          */
2260                         if ((ip->flags & HAMMER_INODE_REDO) ||
2261                             (ip->flags & HAMMER_INODE_RDIRTY)) {
2262                                 error = hammer_generate_redo(&trans, ip,
2263                                                              vap->va_size,
2264                                                              HAMMER_REDO_TRUNC,
2265                                                              NULL, 0);
2266                         }
2267                         blksize = hammer_blocksize(vap->va_size);
2268
2269                         /*
2270                          * XXX break atomicy, we can deadlock the backend
2271                          * if we do not release the lock.  Probably not a
2272                          * big deal here.
2273                          */
2274                         if (vap->va_size < ip->ino_data.size) {
2275                                 nvtruncbuf(ap->a_vp, vap->va_size,
2276                                            blksize,
2277                                            hammer_blockoff(vap->va_size));
2278                                 truncating = 1;
2279                                 kflags |= NOTE_WRITE;
2280                         } else {
2281                                 nvextendbuf(ap->a_vp,
2282                                             ip->ino_data.size,
2283                                             vap->va_size,
2284                                             hammer_blocksize(ip->ino_data.size),
2285                                             hammer_blocksize(vap->va_size),
2286                                             hammer_blockoff(ip->ino_data.size),
2287                                             hammer_blockoff(vap->va_size),
2288                                             0);
2289                                 truncating = 0;
2290                                 kflags |= NOTE_WRITE | NOTE_EXTEND;
2291                         }
2292                         ip->ino_data.size = vap->va_size;
2293                         ip->ino_data.mtime = trans.time;
2294                         /* XXX safe to use SDIRTY instead of DDIRTY here? */
2295                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2296
2297                         /*
2298                          * On-media truncation is cached in the inode until
2299                          * the inode is synchronized.  We must immediately
2300                          * handle any frontend records.
2301                          */
2302                         if (truncating) {
2303                                 hammer_ip_frontend_trunc(ip, vap->va_size);
2304 #ifdef DEBUG_TRUNCATE
2305                                 if (HammerTruncIp == NULL)
2306                                         HammerTruncIp = ip;
2307 #endif
2308                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2309                                         ip->flags |= HAMMER_INODE_TRUNCATED;
2310                                         ip->trunc_off = vap->va_size;
2311 #ifdef DEBUG_TRUNCATE
2312                                         if (ip == HammerTruncIp)
2313                                         kprintf("truncate1 %016llx\n",
2314                                                 (long long)ip->trunc_off);
2315 #endif
2316                                 } else if (ip->trunc_off > vap->va_size) {
2317                                         ip->trunc_off = vap->va_size;
2318 #ifdef DEBUG_TRUNCATE
2319                                         if (ip == HammerTruncIp)
2320                                         kprintf("truncate2 %016llx\n",
2321                                                 (long long)ip->trunc_off);
2322 #endif
2323                                 } else {
2324 #ifdef DEBUG_TRUNCATE
2325                                         if (ip == HammerTruncIp)
2326                                         kprintf("truncate3 %016llx (ignored)\n",
2327                                                 (long long)vap->va_size);
2328 #endif
2329                                 }
2330                         }
2331
2332 #if 0
2333                         /*
2334                          * When truncating, nvtruncbuf() may have cleaned out
2335                          * a portion of the last block on-disk in the buffer
2336                          * cache.  We must clean out any frontend records
2337                          * for blocks beyond the new last block.
2338                          */
2339                         aligned_size = (vap->va_size + (blksize - 1)) &
2340                                        ~(int64_t)(blksize - 1);
2341                         if (truncating && vap->va_size < aligned_size) {
2342                                 aligned_size -= blksize;
2343                                 hammer_ip_frontend_trunc(ip, aligned_size);
2344                         }
2345 #endif
2346                         break;
2347                 case VDATABASE:
2348                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2349                                 ip->flags |= HAMMER_INODE_TRUNCATED;
2350                                 ip->trunc_off = vap->va_size;
2351                         } else if (ip->trunc_off > vap->va_size) {
2352                                 ip->trunc_off = vap->va_size;
2353                         }
2354                         hammer_ip_frontend_trunc(ip, vap->va_size);
2355                         ip->ino_data.size = vap->va_size;
2356                         ip->ino_data.mtime = trans.time;
2357                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2358                         kflags |= NOTE_ATTRIB;
2359                         break;
2360                 default:
2361                         error = EINVAL;
2362                         goto done;
2363                 }
2364                 break;
2365         }
2366         if (vap->va_atime.tv_sec != VNOVAL) {
2367                 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2368                 modflags |= HAMMER_INODE_ATIME;
2369                 kflags |= NOTE_ATTRIB;
2370         }
2371         if (vap->va_mtime.tv_sec != VNOVAL) {
2372                 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2373                 modflags |= HAMMER_INODE_MTIME;
2374                 kflags |= NOTE_ATTRIB;
2375         }
2376         if (vap->va_mode != (mode_t)VNOVAL) {
2377                 mode_t   cur_mode = ip->ino_data.mode;
2378                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2379                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2380
2381                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2382                                          cur_uid, cur_gid, &cur_mode);
2383                 if (error == 0 && ip->ino_data.mode != cur_mode) {
2384                         ip->ino_data.mode = cur_mode;
2385                         ip->ino_data.ctime = trans.time;
2386                         modflags |= HAMMER_INODE_DDIRTY;
2387                         kflags |= NOTE_ATTRIB;
2388                 }
2389         }
2390 done:
2391         if (error == 0)
2392                 hammer_modify_inode(&trans, ip, modflags);
2393         hammer_done_transaction(&trans);
2394         hammer_knote(ap->a_vp, kflags);
2395         lwkt_reltoken(&hmp->fs_token);
2396         return (error);
2397 }
2398
2399 /*
2400  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2401  */
2402 static
2403 int
2404 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2405 {
2406         struct hammer_transaction trans;
2407         struct hammer_inode *dip;
2408         struct hammer_inode *nip;
2409         hammer_record_t record;
2410         struct nchandle *nch;
2411         hammer_mount_t hmp;
2412         int error;
2413         int bytes;
2414
2415         ap->a_vap->va_type = VLNK;
2416
2417         nch = ap->a_nch;
2418         dip = VTOI(ap->a_dvp);
2419         hmp = dip->hmp;
2420
2421         if (dip->flags & HAMMER_INODE_RO)
2422                 return (EROFS);
2423         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
2424                 return (error);
2425
2426         /*
2427          * Create a transaction to cover the operations we perform.
2428          */
2429         lwkt_gettoken(&hmp->fs_token);
2430         hammer_start_transaction(&trans, hmp);
2431         ++hammer_stats_file_iopsw;
2432
2433         /*
2434          * Create a new filesystem object of the requested type.  The
2435          * returned inode will be referenced but not locked.
2436          */
2437
2438         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2439                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2440                                     NULL, &nip);
2441         if (error) {
2442                 hammer_done_transaction(&trans);
2443                 *ap->a_vpp = NULL;
2444                 lwkt_reltoken(&hmp->fs_token);
2445                 return (error);
2446         }
2447
2448         /*
2449          * Add a record representing the symlink.  symlink stores the link
2450          * as pure data, not a string, and is no \0 terminated.
2451          */
2452         if (error == 0) {
2453                 bytes = strlen(ap->a_target);
2454
2455                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2456                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2457                 } else {
2458                         record = hammer_alloc_mem_record(nip, bytes);
2459                         record->type = HAMMER_MEM_RECORD_GENERAL;
2460
2461                         record->leaf.base.localization = nip->obj_localization +
2462                                                          HAMMER_LOCALIZE_MISC;
2463                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2464                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2465                         record->leaf.data_len = bytes;
2466                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2467                         bcopy(ap->a_target, record->data->symlink.name, bytes);
2468                         error = hammer_ip_add_record(&trans, record);
2469                 }
2470
2471                 /*
2472                  * Set the file size to the length of the link.
2473                  */
2474                 if (error == 0) {
2475                         nip->ino_data.size = bytes;
2476                         hammer_modify_inode(&trans, nip, HAMMER_INODE_DDIRTY);
2477                 }
2478         }
2479         if (error == 0)
2480                 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
2481                                                 nch->ncp->nc_nlen, nip);
2482
2483         /*
2484          * Finish up.
2485          */
2486         if (error) {
2487                 hammer_rel_inode(nip, 0);
2488                 *ap->a_vpp = NULL;
2489         } else {
2490                 error = hammer_get_vnode(nip, ap->a_vpp);
2491                 hammer_rel_inode(nip, 0);
2492                 if (error == 0) {
2493                         cache_setunresolved(ap->a_nch);
2494                         cache_setvp(ap->a_nch, *ap->a_vpp);
2495                         hammer_knote(ap->a_dvp, NOTE_WRITE);
2496                 }
2497         }
2498         hammer_done_transaction(&trans);
2499         lwkt_reltoken(&hmp->fs_token);
2500         return (error);
2501 }
2502
2503 /*
2504  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2505  */
2506 static
2507 int
2508 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2509 {
2510         struct hammer_transaction trans;
2511         struct hammer_inode *dip;
2512         hammer_mount_t hmp;
2513         int error;
2514
2515         dip = VTOI(ap->a_dvp);
2516         hmp = dip->hmp;
2517
2518         if (hammer_nohistory(dip) == 0 &&
2519             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2520                 return (error);
2521         }
2522
2523         lwkt_gettoken(&hmp->fs_token);
2524         hammer_start_transaction(&trans, hmp);
2525         ++hammer_stats_file_iopsw;
2526         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2527                                 ap->a_cred, ap->a_flags, -1);
2528         hammer_done_transaction(&trans);
2529         lwkt_reltoken(&hmp->fs_token);
2530
2531         return (error);
2532 }
2533
2534 /*
2535  * hammer_vop_ioctl { vp, command, data, fflag, cred }
2536  */
2537 static
2538 int
2539 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2540 {
2541         struct hammer_inode *ip = ap->a_vp->v_data;
2542         hammer_mount_t hmp = ip->hmp;
2543         int error;
2544
2545         ++hammer_stats_file_iopsr;
2546         lwkt_gettoken(&hmp->fs_token);
2547         error = hammer_ioctl(ip, ap->a_command, ap->a_data,
2548                              ap->a_fflag, ap->a_cred);
2549         lwkt_reltoken(&hmp->fs_token);
2550         return (error);
2551 }
2552
2553 static
2554 int
2555 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2556 {
2557         static const struct mountctl_opt extraopt[] = {
2558                 { HMNT_NOHISTORY,       "nohistory" },
2559                 { HMNT_MASTERID,        "master" },
2560                 { 0, NULL}
2561
2562         };
2563         struct hammer_mount *hmp;
2564         struct mount *mp;
2565         int usedbytes;
2566         int error;
2567
2568         error = 0;
2569         usedbytes = 0;
2570         mp = ap->a_head.a_ops->head.vv_mount;
2571         KKASSERT(mp->mnt_data != NULL);
2572         hmp = (struct hammer_mount *)mp->mnt_data;
2573
2574         lwkt_gettoken(&hmp->fs_token);
2575
2576         switch(ap->a_op) {
2577         case MOUNTCTL_SET_EXPORT:
2578                 if (ap->a_ctllen != sizeof(struct export_args))
2579                         error = EINVAL;
2580                 else
2581                         error = hammer_vfs_export(mp, ap->a_op,
2582                                       (const struct export_args *)ap->a_ctl);
2583                 break;
2584         case MOUNTCTL_MOUNTFLAGS:
2585         {
2586                 /*
2587                  * Call standard mountctl VOP function
2588                  * so we get user mount flags.
2589                  */
2590                 error = vop_stdmountctl(ap);
2591                 if (error)
2592                         break;
2593
2594                 usedbytes = *ap->a_res;
2595
2596                 if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2597                         usedbytes += vfs_flagstostr(hmp->hflags, extraopt,
2598                                                     ap->a_buf,
2599                                                     ap->a_buflen - usedbytes,
2600                                                     &error);
2601                 }
2602
2603                 *ap->a_res += usedbytes;
2604                 break;
2605         }
2606         default:
2607                 error = vop_stdmountctl(ap);
2608                 break;
2609         }
2610         lwkt_reltoken(&hmp->fs_token);
2611         return(error);
2612 }
2613
2614 /*
2615  * hammer_vop_strategy { vp, bio }
2616  *
2617  * Strategy call, used for regular file read & write only.  Note that the
2618  * bp may represent a cluster.
2619  *
2620  * To simplify operation and allow better optimizations in the future,
2621  * this code does not make any assumptions with regards to buffer alignment
2622  * or size.
2623  */
2624 static
2625 int
2626 hammer_vop_strategy(struct vop_strategy_args *ap)
2627 {
2628         struct buf *bp;
2629         int error;
2630
2631         bp = ap->a_bio->bio_buf;
2632
2633         switch(bp->b_cmd) {
2634         case BUF_CMD_READ:
2635                 error = hammer_vop_strategy_read(ap);
2636                 break;
2637         case BUF_CMD_WRITE:
2638                 error = hammer_vop_strategy_write(ap);
2639                 break;
2640         default:
2641                 bp->b_error = error = EINVAL;
2642                 bp->b_flags |= B_ERROR;
2643                 biodone(ap->a_bio);
2644                 break;
2645         }
2646
2647         /* hammer_dump_dedup_cache(((hammer_inode_t)ap->a_vp->v_data)->hmp); */
2648
2649         return (error);
2650 }
2651
2652 /*
2653  * Read from a regular file.  Iterate the related records and fill in the
2654  * BIO/BUF.  Gaps are zero-filled.
2655  *
2656  * The support code in hammer_object.c should be used to deal with mixed
2657  * in-memory and on-disk records.
2658  *
2659  * NOTE: Can be called from the cluster code with an oversized buf.
2660  *
2661  * XXX atime update
2662  */
2663 static
2664 int
2665 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2666 {
2667         struct hammer_transaction trans;
2668         struct hammer_inode *ip;
2669         struct hammer_inode *dip;
2670         hammer_mount_t hmp;
2671         struct hammer_cursor cursor;
2672         hammer_base_elm_t base;
2673         hammer_off_t disk_offset;
2674         struct bio *bio;
2675         struct bio *nbio;
2676         struct buf *bp;
2677         int64_t rec_offset;
2678         int64_t ran_end;
2679         int64_t tmp64;
2680         int error;
2681         int boff;
2682         int roff;
2683         int n;
2684         int isdedupable;
2685
2686         bio = ap->a_bio;
2687         bp = bio->bio_buf;
2688         ip = ap->a_vp->v_data;
2689         hmp = ip->hmp;
2690
2691         /*
2692          * The zone-2 disk offset may have been set by the cluster code via
2693          * a BMAP operation, or else should be NOOFFSET.
2694          *
2695          * Checking the high bits for a match against zone-2 should suffice.
2696          *
2697          * In cases where a lot of data duplication is present it may be
2698          * more beneficial to drop through and doubule-buffer through the
2699          * device.
2700          */
2701         nbio = push_bio(bio);
2702         if (hammer_double_buffer == 0 &&
2703             (nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2704             HAMMER_ZONE_LARGE_DATA) {
2705                 lwkt_gettoken(&hmp->fs_token);
2706                 error = hammer_io_direct_read(hmp, nbio, NULL);
2707                 lwkt_reltoken(&hmp->fs_token);
2708                 return (error);
2709         }
2710
2711         /*
2712          * Well, that sucked.  Do it the hard way.  If all the stars are
2713          * aligned we may still be able to issue a direct-read.
2714          */
2715         lwkt_gettoken(&hmp->fs_token);
2716         hammer_simple_transaction(&trans, hmp);
2717         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2718
2719         /*
2720          * Key range (begin and end inclusive) to scan.  Note that the key's
2721          * stored in the actual records represent BASE+LEN, not BASE.  The
2722          * first record containing bio_offset will have a key > bio_offset.
2723          */
2724         cursor.key_beg.localization = ip->obj_localization +
2725                                       HAMMER_LOCALIZE_MISC;
2726         cursor.key_beg.obj_id = ip->obj_id;
2727         cursor.key_beg.create_tid = 0;
2728         cursor.key_beg.delete_tid = 0;
2729         cursor.key_beg.obj_type = 0;
2730         cursor.key_beg.key = bio->bio_offset + 1;
2731         cursor.asof = ip->obj_asof;
2732         cursor.flags |= HAMMER_CURSOR_ASOF;
2733
2734         cursor.key_end = cursor.key_beg;
2735         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2736 #if 0
2737         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2738                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2739                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2740                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2741         } else
2742 #endif
2743         {
2744                 ran_end = bio->bio_offset + bp->b_bufsize;
2745                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2746                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2747                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2748                 if (tmp64 < ran_end)
2749                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2750                 else
2751                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2752         }
2753         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2754
2755         error = hammer_ip_first(&cursor);
2756         boff = 0;
2757
2758         while (error == 0) {
2759                 /*
2760                  * Get the base file offset of the record.  The key for
2761                  * data records is (base + bytes) rather then (base).
2762                  */
2763                 base = &cursor.leaf->base;
2764                 rec_offset = base->key - cursor.leaf->data_len;
2765
2766                 /*
2767                  * Calculate the gap, if any, and zero-fill it.
2768                  *
2769                  * n is the offset of the start of the record verses our
2770                  * current seek offset in the bio.
2771                  */
2772                 n = (int)(rec_offset - (bio->bio_offset + boff));
2773                 if (n > 0) {
2774                         if (n > bp->b_bufsize - boff)
2775                                 n = bp->b_bufsize - boff;
2776                         bzero((char *)bp->b_data + boff, n);
2777                         boff += n;
2778                         n = 0;
2779                 }
2780
2781                 /*
2782                  * Calculate the data offset in the record and the number
2783                  * of bytes we can copy.
2784                  *
2785                  * There are two degenerate cases.  First, boff may already
2786                  * be at bp->b_bufsize.  Secondly, the data offset within
2787                  * the record may exceed the record's size.
2788                  */
2789                 roff = -n;
2790                 rec_offset += roff;
2791                 n = cursor.leaf->data_len - roff;
2792                 if (n <= 0) {
2793                         kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2794                         n = 0;
2795                 } else if (n > bp->b_bufsize - boff) {
2796                         n = bp->b_bufsize - boff;
2797                 }
2798
2799                 /*
2800                  * Deal with cached truncations.  This cool bit of code
2801                  * allows truncate()/ftruncate() to avoid having to sync
2802                  * the file.
2803                  *
2804                  * If the frontend is truncated then all backend records are
2805                  * subject to the frontend's truncation.
2806                  *
2807                  * If the backend is truncated then backend records on-disk
2808                  * (but not in-memory) are subject to the backend's
2809                  * truncation.  In-memory records owned by the backend
2810                  * represent data written after the truncation point on the
2811                  * backend and must not be truncated.
2812                  *
2813                  * Truncate operations deal with frontend buffer cache
2814                  * buffers and frontend-owned in-memory records synchronously.
2815                  */
2816                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2817                         if (hammer_cursor_ondisk(&cursor)/* ||
2818                             cursor.iprec->flush_state == HAMMER_FST_FLUSH*/) {
2819                                 if (ip->trunc_off <= rec_offset)
2820                                         n = 0;
2821                                 else if (ip->trunc_off < rec_offset + n)
2822                                         n = (int)(ip->trunc_off - rec_offset);
2823                         }
2824                 }
2825                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2826                         if (hammer_cursor_ondisk(&cursor)) {
2827                                 if (ip->sync_trunc_off <= rec_offset)
2828                                         n = 0;
2829                                 else if (ip->sync_trunc_off < rec_offset + n)
2830                                         n = (int)(ip->sync_trunc_off - rec_offset);
2831                         }
2832                 }
2833
2834                 /*
2835                  * Try to issue a direct read into our bio if possible,
2836                  * otherwise resolve the element data into a hammer_buffer
2837                  * and copy.
2838                  *
2839                  * The buffer on-disk should be zerod past any real
2840                  * truncation point, but may not be for any synthesized
2841                  * truncation point from above.
2842                  */
2843                 disk_offset = cursor.leaf->data_offset + roff;
2844                 isdedupable = (boff == 0 && n == bp->b_bufsize &&
2845                                hammer_cursor_ondisk(&cursor) &&
2846                                ((int)disk_offset & HAMMER_BUFMASK) == 0);
2847
2848                 if (isdedupable && hammer_double_buffer == 0) {
2849                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2850                                  HAMMER_ZONE_LARGE_DATA);
2851                         nbio->bio_offset = disk_offset;
2852                         error = hammer_io_direct_read(hmp, nbio, cursor.leaf);
2853                         if (hammer_live_dedup && error == 0)
2854                                 hammer_dedup_cache_add(ip, cursor.leaf);
2855                         goto done;
2856                 } else if (n) {
2857                         error = hammer_ip_resolve_data(&cursor);
2858                         if (error == 0) {
2859                                 if (hammer_live_dedup && isdedupable)
2860                                         hammer_dedup_cache_add(ip, cursor.leaf);
2861                                 bcopy((char *)cursor.data + roff,
2862                                       (char *)bp->b_data + boff, n);
2863                         }
2864                 }
2865                 if (error)
2866                         break;
2867
2868                 /*
2869                  * We have to be sure that the only elements added to the
2870                  * dedup cache are those which are already on-media.
2871                  */
2872                 if (hammer_live_dedup && hammer_cursor_ondisk(&cursor))
2873                         hammer_dedup_cache_add(ip, cursor.leaf);
2874
2875                 /*
2876                  * Iterate until we have filled the request.
2877                  */
2878                 boff += n;
2879                 if (boff == bp->b_bufsize)
2880                         break;
2881                 error = hammer_ip_next(&cursor);
2882         }
2883
2884         /*
2885          * There may have been a gap after the last record
2886          */
2887         if (error == ENOENT)
2888                 error = 0;
2889         if (error == 0 && boff != bp->b_bufsize) {
2890                 KKASSERT(boff < bp->b_bufsize);
2891                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2892                 /* boff = bp->b_bufsize; */
2893         }
2894         bp->b_resid = 0;
2895         bp->b_error = error;
2896         if (error)
2897                 bp->b_flags |= B_ERROR;
2898         biodone(ap->a_bio);
2899
2900 done:
2901         /*
2902          * Cache the b-tree node for the last data read in cache[1].
2903          *
2904          * If we hit the file EOF then also cache the node in the
2905          * governing director's cache[3], it will be used to initialize
2906          * the inode's cache[1] for any inodes looked up via the directory.
2907          *
2908          * This doesn't reduce disk accesses since the B-Tree chain is
2909          * likely cached, but it does reduce cpu overhead when looking
2910          * up file offsets for cpdup/tar/cpio style iterations.
2911          */
2912         if (cursor.node)
2913                 hammer_cache_node(&ip->cache[1], cursor.node);
2914         if (ran_end >= ip->ino_data.size) {
2915                 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
2916                                         ip->obj_asof, ip->obj_localization);
2917                 if (dip) {
2918                         hammer_cache_node(&dip->cache[3], cursor.node);
2919                         hammer_rel_inode(dip, 0);
2920                 }
2921         }
2922         hammer_done_cursor(&cursor);
2923         hammer_done_transaction(&trans);
2924         lwkt_reltoken(&hmp->fs_token);
2925         return(error);
2926 }
2927
2928 /*
2929  * BMAP operation - used to support cluster_read() only.
2930  *
2931  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2932  *
2933  * This routine may return EOPNOTSUPP if the opration is not supported for
2934  * the specified offset.  The contents of the pointer arguments do not
2935  * need to be initialized in that case. 
2936  *
2937  * If a disk address is available and properly aligned return 0 with 
2938  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2939  * to the run-length relative to that offset.  Callers may assume that
2940  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2941  * large, so return EOPNOTSUPP if it is not sufficiently large.
2942  */
2943 static
2944 int
2945 hammer_vop_bmap(struct vop_bmap_args *ap)
2946 {
2947         struct hammer_transaction trans;
2948         struct hammer_inode *ip;
2949         hammer_mount_t hmp;
2950         struct hammer_cursor cursor;
2951         hammer_base_elm_t base;
2952         int64_t rec_offset;
2953         int64_t ran_end;
2954         int64_t tmp64;
2955         int64_t base_offset;
2956         int64_t base_disk_offset;
2957         int64_t last_offset;
2958         hammer_off_t last_disk_offset;
2959         hammer_off_t disk_offset;
2960         int     rec_len;
2961         int     error;
2962         int     blksize;
2963
2964         ++hammer_stats_file_iopsr;
2965         ip = ap->a_vp->v_data;
2966         hmp = ip->hmp;
2967
2968         /*
2969          * We can only BMAP regular files.  We can't BMAP database files,
2970          * directories, etc.
2971          */
2972         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
2973                 return(EOPNOTSUPP);
2974
2975         /*
2976          * bmap is typically called with runp/runb both NULL when used
2977          * for writing.  We do not support BMAP for writing atm.
2978          */
2979         if (ap->a_cmd != BUF_CMD_READ)
2980                 return(EOPNOTSUPP);
2981
2982         /*
2983          * Scan the B-Tree to acquire blockmap addresses, then translate
2984          * to raw addresses.
2985          */
2986         lwkt_gettoken(&hmp->fs_token);
2987         hammer_simple_transaction(&trans, hmp);
2988 #if 0
2989         kprintf("bmap_beg %016llx ip->cache %p\n",
2990                 (long long)ap->a_loffset, ip->cache[1]);
2991 #endif
2992         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2993
2994         /*
2995          * Key range (begin and end inclusive) to scan.  Note that the key's
2996          * stored in the actual records represent BASE+LEN, not BASE.  The
2997          * first record containing bio_offset will have a key > bio_offset.
2998          */
2999         cursor.key_beg.localization = ip->obj_localization +
3000                                       HAMMER_LOCALIZE_MISC;
3001         cursor.key_beg.obj_id = ip->obj_id;
3002         cursor.key_beg.create_tid = 0;
3003         cursor.key_beg.delete_tid = 0;
3004         cursor.key_beg.obj_type = 0;
3005         if (ap->a_runb)
3006                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
3007         else
3008                 cursor.key_beg.key = ap->a_loffset + 1;
3009         if (cursor.key_beg.key < 0)
3010                 cursor.key_beg.key = 0;
3011         cursor.asof = ip->obj_asof;
3012         cursor.flags |= HAMMER_CURSOR_ASOF;
3013
3014         cursor.key_end = cursor.key_beg;
3015         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
3016
3017         ran_end = ap->a_loffset + MAXPHYS;
3018         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
3019         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
3020         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
3021         if (tmp64 < ran_end)
3022                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
3023         else
3024                 cursor.key_end.key = ran_end + MAXPHYS + 1;
3025
3026         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
3027
3028         error = hammer_ip_first(&cursor);
3029         base_offset = last_offset = 0;
3030         base_disk_offset = last_disk_offset = 0;
3031
3032         while (error == 0) {
3033                 /*
3034                  * Get the base file offset of the record.  The key for
3035                  * data records is (base + bytes) rather then (base).
3036                  *
3037                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
3038                  * The extra bytes should be zero on-disk and the BMAP op
3039                  * should still be ok.
3040                  */
3041                 base = &cursor.leaf->base;
3042                 rec_offset = base->key - cursor.leaf->data_len;
3043                 rec_len    = cursor.leaf->data_len;
3044
3045                 /*
3046                  * Incorporate any cached truncation.
3047                  *
3048                  * NOTE: Modifications to rec_len based on synthesized
3049                  * truncation points remove the guarantee that any extended
3050                  * data on disk is zero (since the truncations may not have
3051                  * taken place on-media yet).
3052                  */
3053                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
3054                         if (hammer_cursor_ondisk(&cursor) ||
3055                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
3056                                 if (ip->trunc_off <= rec_offset)
3057                                         rec_len = 0;
3058                                 else if (ip->trunc_off < rec_offset + rec_len)
3059                                         rec_len = (int)(ip->trunc_off - rec_offset);
3060                         }
3061                 }
3062                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
3063                         if (hammer_cursor_ondisk(&cursor)) {
3064                                 if (ip->sync_trunc_off <= rec_offset)
3065                                         rec_len = 0;
3066                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
3067                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
3068                         }
3069                 }
3070
3071                 /*
3072                  * Accumulate information.  If we have hit a discontiguous
3073                  * block reset base_offset unless we are already beyond the
3074                  * requested offset.  If we are, that's it, we stop.
3075                  */
3076                 if (error)
3077                         break;
3078                 if (hammer_cursor_ondisk(&cursor)) {
3079                         disk_offset = cursor.leaf->data_offset;
3080                         if (rec_offset != last_offset ||
3081                             disk_offset != last_disk_offset) {
3082                                 if (rec_offset > ap->a_loffset)
3083                                         break;
3084                                 base_offset = rec_offset;
3085                                 base_disk_offset = disk_offset;
3086                         }
3087                         last_offset = rec_offset + rec_len;
3088                         last_disk_offset = disk_offset + rec_len;
3089
3090                         if (hammer_live_dedup)
3091                                 hammer_dedup_cache_add(ip, cursor.leaf);
3092                 }
3093                 
3094                 error = hammer_ip_next(&cursor);
3095         }
3096
3097 #if 0
3098         kprintf("BMAP %016llx:  %016llx - %016llx\n",
3099                 (long long)ap->a_loffset,
3100                 (long long)base_offset,
3101                 (long long)last_offset);
3102         kprintf("BMAP %16s:  %016llx - %016llx\n", "",
3103                 (long long)base_disk_offset,
3104                 (long long)last_disk_offset);
3105 #endif
3106
3107         if (cursor.node) {
3108                 hammer_cache_node(&ip->cache[1], cursor.node);
3109 #if 0
3110                 kprintf("bmap_end2 %016llx ip->cache %p\n",
3111                         (long long)ap->a_loffset, ip->cache[1]);
3112 #endif
3113         }
3114         hammer_done_cursor(&cursor);
3115         hammer_done_transaction(&trans);
3116         lwkt_reltoken(&hmp->fs_token);
3117
3118         /*
3119          * If we couldn't find any records or the records we did find were
3120          * all behind the requested offset, return failure.  A forward
3121          * truncation can leave a hole w/ no on-disk records.
3122          */
3123         if (last_offset == 0 || last_offset < ap->a_loffset)
3124                 return (EOPNOTSUPP);
3125
3126         /*
3127          * Figure out the block size at the requested offset and adjust
3128          * our limits so the cluster_read() does not create inappropriately
3129          * sized buffer cache buffers.
3130          */
3131         blksize = hammer_blocksize(ap->a_loffset);
3132         if (hammer_blocksize(base_offset) != blksize) {
3133                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
3134         }
3135         if (last_offset != ap->a_loffset &&
3136             hammer_blocksize(last_offset - 1) != blksize) {
3137                 last_offset = hammer_blockdemarc(ap->a_loffset,
3138                                                  last_offset - 1);
3139         }
3140
3141         /*
3142          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3143          * from occuring.
3144          */
3145         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
3146
3147         if ((disk_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_LARGE_DATA) {
3148                 /*
3149                  * Only large-data zones can be direct-IOd
3150                  */
3151                 error = EOPNOTSUPP;
3152         } else if ((disk_offset & HAMMER_BUFMASK) ||
3153                    (last_offset - ap->a_loffset) < blksize) {
3154                 /*
3155                  * doffsetp is not aligned or the forward run size does
3156                  * not cover a whole buffer, disallow the direct I/O.
3157                  */
3158                 error = EOPNOTSUPP;
3159         } else {
3160                 /*
3161                  * We're good.
3162                  */
3163                 *ap->a_doffsetp = disk_offset;
3164                 if (ap->a_runb) {
3165                         *ap->a_runb = ap->a_loffset - base_offset;
3166                         KKASSERT(*ap->a_runb >= 0);
3167                 }
3168                 if (ap->a_runp) {
3169                         *ap->a_runp = last_offset - ap->a_loffset;
3170                         KKASSERT(*ap->a_runp >= 0);
3171                 }
3172                 error = 0;
3173         }
3174         return(error);
3175 }
3176
3177 /*
3178  * Write to a regular file.   Because this is a strategy call the OS is
3179  * trying to actually get data onto the media.
3180  */
3181 static
3182 int
3183 hammer_vop_strategy_write(struct vop_strategy_args *ap)
3184 {
3185         hammer_record_t record;
3186         hammer_mount_t hmp;
3187         hammer_inode_t ip;
3188         struct bio *bio;
3189         struct buf *bp;
3190         int blksize;
3191         int bytes;
3192         int error;
3193
3194         bio = ap->a_bio;
3195         bp = bio->bio_buf;
3196         ip = ap->a_vp->v_data;
3197         hmp = ip->hmp;
3198
3199         blksize = hammer_blocksize(bio->bio_offset);
3200         KKASSERT(bp->b_bufsize == blksize);
3201
3202         if (ip->flags & HAMMER_INODE_RO) {
3203                 bp->b_error = EROFS;
3204                 bp->b_flags |= B_ERROR;
3205                 biodone(ap->a_bio);
3206                 return(EROFS);
3207         }
3208
3209         lwkt_gettoken(&hmp->fs_token);
3210
3211         /*
3212          * Interlock with inode destruction (no in-kernel or directory
3213          * topology visibility).  If we queue new IO while trying to
3214          * destroy the inode we can deadlock the vtrunc call in
3215          * hammer_inode_unloadable_check().
3216          *
3217          * Besides, there's no point flushing a bp associated with an
3218          * inode that is being destroyed on-media and has no kernel
3219          * references.
3220          */
3221         if ((ip->flags | ip->sync_flags) &
3222             (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
3223                 bp->b_resid = 0;
3224                 biodone(ap->a_bio);
3225                 lwkt_reltoken(&hmp->fs_token);
3226                 return(0);
3227         }
3228
3229         /*
3230          * Reserve space and issue a direct-write from the front-end. 
3231          * NOTE: The direct_io code will hammer_bread/bcopy smaller
3232          * allocations.
3233          *
3234          * An in-memory record will be installed to reference the storage
3235          * until the flusher can get to it.
3236          *
3237          * Since we own the high level bio the front-end will not try to
3238          * do a direct-read until the write completes.
3239          *
3240          * NOTE: The only time we do not reserve a full-sized buffers
3241          * worth of data is if the file is small.  We do not try to
3242          * allocate a fragment (from the small-data zone) at the end of
3243          * an otherwise large file as this can lead to wildly separated
3244          * data.
3245          */
3246         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
3247         KKASSERT(bio->bio_offset < ip->ino_data.size);
3248         if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
3249                 bytes = bp->b_bufsize;
3250         else
3251                 bytes = ((int)ip->ino_data.size + 15) & ~15;
3252
3253         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
3254                                     bytes, &error);
3255
3256         /*
3257          * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3258          * in hammer_vop_write().  We must flag the record so the proper
3259          * REDO_TERM_WRITE entry is generated during the flush.
3260          */
3261         if (record) {
3262                 if (bp->b_flags & B_VFSFLAG1) {
3263                         record->flags |= HAMMER_RECF_REDO;
3264                         bp->b_flags &= ~B_VFSFLAG1;
3265                 }
3266                 if (record->flags & HAMMER_RECF_DEDUPED) {
3267                         bp->b_resid = 0;
3268                         hammer_ip_replace_bulk(hmp, record);
3269                         biodone(ap->a_bio);
3270                 } else {
3271                         hammer_io_direct_write(hmp, bio, record);
3272                 }
3273                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
3274                         hammer_flush_inode(ip, 0);
3275         } else {
3276                 bp->b_bio2.bio_offset = NOOFFSET;
3277                 bp->b_error = error;
3278                 bp->b_flags |= B_ERROR;
3279                 biodone(ap->a_bio);
3280         }
3281         lwkt_reltoken(&hmp->fs_token);
3282         return(error);
3283 }
3284
3285 /*
3286  * dounlink - disconnect a directory entry
3287  *
3288  * XXX whiteout support not really in yet
3289  */
3290 static int
3291 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
3292                 struct vnode *dvp, struct ucred *cred, 
3293                 int flags, int isdir)
3294 {
3295         struct namecache *ncp;
3296         hammer_inode_t dip;
3297         hammer_inode_t ip;
3298         hammer_mount_t hmp;
3299         struct hammer_cursor cursor;
3300         int64_t namekey;
3301         u_int32_t max_iterations;
3302         int nlen, error;
3303
3304         /*
3305          * Calculate the namekey and setup the key range for the scan.  This
3306          * works kinda like a chained hash table where the lower 32 bits
3307          * of the namekey synthesize the chain.
3308          *
3309          * The key range is inclusive of both key_beg and key_end.
3310          */
3311         dip = VTOI(dvp);
3312         ncp = nch->ncp;
3313         hmp = dip->hmp;
3314
3315         if (dip->flags & HAMMER_INODE_RO)
3316                 return (EROFS);
3317
3318         namekey = hammer_directory_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3319                                            &max_iterations);
3320 retry:
3321         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3322         cursor.key_beg.localization = dip->obj_localization +
3323                                       hammer_dir_localization(dip);
3324         cursor.key_beg.obj_id = dip->obj_id;
3325         cursor.key_beg.key = namekey;
3326         cursor.key_beg.create_tid = 0;
3327         cursor.key_beg.delete_tid = 0;
3328         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3329         cursor.key_beg.obj_type = 0;
3330
3331         cursor.key_end = cursor.key_beg;
3332         cursor.key_end.key += max_iterations;
3333         cursor.asof = dip->obj_asof;
3334         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3335
3336         /*
3337          * Scan all matching records (the chain), locate the one matching
3338          * the requested path component.  info->last_error contains the
3339          * error code on search termination and could be 0, ENOENT, or
3340          * something else.
3341          *
3342          * The hammer_ip_*() functions merge in-memory records with on-disk
3343          * records for the purposes of the search.
3344          */
3345         error = hammer_ip_first(&cursor);
3346
3347         while (error == 0) {
3348                 error = hammer_ip_resolve_data(&cursor);
3349                 if (error)
3350                         break;
3351                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3352                 KKASSERT(nlen > 0);
3353                 if (ncp->nc_nlen == nlen &&
3354                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3355                         break;
3356                 }
3357                 error = hammer_ip_next(&cursor);
3358         }
3359
3360         /*
3361          * If all is ok we have to get the inode so we can adjust nlinks.
3362          * To avoid a deadlock with the flusher we must release the inode
3363          * lock on the directory when acquiring the inode for the entry.
3364          *
3365          * If the target is a directory, it must be empty.
3366          */
3367         if (error == 0) {
3368                 hammer_unlock(&cursor.ip->lock);
3369                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3370                                       hmp->asof,
3371                                       cursor.data->entry.localization,
3372                                       0, &error);
3373                 hammer_lock_sh(&cursor.ip->lock);
3374                 if (error == ENOENT) {
3375                         kprintf("HAMMER: WARNING: Removing "
3376                                 "dirent w/missing inode \"%s\"\n"
3377                                 "\tobj_id = %016llx\n",
3378                                 ncp->nc_name,
3379                                 (long long)cursor.data->entry.obj_id);
3380                         error = 0;
3381                 }
3382
3383                 /*
3384                  * If isdir >= 0 we validate that the entry is or is not a
3385                  * directory.  If isdir < 0 we don't care.
3386                  */
3387                 if (error == 0 && isdir >= 0 && ip) {
3388                         if (isdir &&
3389                             ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3390                                 error = ENOTDIR;
3391                         } else if (isdir == 0 &&
3392                             ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3393                                 error = EISDIR;
3394                         }
3395                 }
3396
3397                 /*
3398                  * If we are trying to remove a directory the directory must
3399                  * be empty.
3400                  *
3401                  * The check directory code can loop and deadlock/retry.  Our
3402                  * own cursor's node locks must be released to avoid a 3-way
3403                  * deadlock with the flusher if the check directory code
3404                  * blocks.
3405                  *
3406                  * If any changes whatsoever have been made to the cursor
3407                  * set EDEADLK and retry.
3408                  *
3409                  * WARNING: See warnings in hammer_unlock_cursor()
3410                  *          function.
3411                  */
3412                 if (error == 0 && ip && ip->ino_data.obj_type ==
3413                                         HAMMER_OBJTYPE_DIRECTORY) {
3414                         hammer_unlock_cursor(&cursor);
3415                         error = hammer_ip_check_directory_empty(trans, ip);
3416                         hammer_lock_cursor(&cursor);
3417                         if (cursor.flags & HAMMER_CURSOR_RETEST) {
3418                                 kprintf("HAMMER: Warning: avoided deadlock "
3419                                         "on rmdir '%s'\n",
3420                                         ncp->nc_name);
3421                                 error = EDEADLK;
3422                         }
3423                 }
3424
3425                 /*
3426                  * Delete the directory entry.
3427                  *
3428                  * WARNING: hammer_ip_del_directory() may have to terminate
3429                  * the cursor to avoid a deadlock.  It is ok to call
3430                  * hammer_done_cursor() twice.
3431                  */
3432                 if (error == 0) {
3433                         error = hammer_ip_del_directory(trans, &cursor,
3434                                                         dip, ip);
3435                 }
3436                 hammer_done_cursor(&cursor);
3437                 if (error == 0) {
3438                         cache_setunresolved(nch);
3439                         cache_setvp(nch, NULL);
3440
3441                         /*
3442                          * NOTE: ip->vp, if non-NULL, cannot be directly
3443                          *       referenced without formally acquiring the
3444                          *       vp since the vp might have zero refs on it,
3445                          *       or in the middle of a reclaim, etc.
3446                          *
3447                          * NOTE: The cache_setunresolved() can rip the vp
3448                          *       out from under us since the vp may not have
3449                          *       any refs, in which case ip->vp will be NULL
3450                          *       from the outset.
3451                          */
3452                         while (ip && ip->vp) {
3453                                 struct vnode *vp;
3454
3455                                 error = hammer_get_vnode(ip, &vp);
3456                                 if (error == 0 && vp) {
3457                                         vn_unlock(vp);
3458                                         hammer_knote(ip->vp, NOTE_DELETE);
3459                                         cache_inval_vp(ip->vp, CINV_DESTROY);
3460                                         vrele(vp);
3461                                         break;
3462                                 }
3463                                 kprintf("Debug: HAMMER ip/vp race1 avoided\n");
3464                         }
3465                 }
3466                 if (ip)
3467                         hammer_rel_inode(ip, 0);
3468         } else {
3469                 hammer_done_cursor(&cursor);
3470         }
3471         if (error == EDEADLK)
3472                 goto retry;
3473
3474         return (error);
3475 }
3476
3477 /************************************************************************
3478  *                          FIFO AND SPECFS OPS                         *
3479  ************************************************************************
3480  *
3481  */
3482 static int
3483 hammer_vop_fifoclose (struct vop_close_args *ap)
3484 {
3485         /* XXX update itimes */
3486         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3487 }
3488
3489 static int
3490 hammer_vop_fiforead (struct vop_read_args *ap)
3491 {
3492         int error;
3493
3494         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3495         /* XXX update access time */
3496         return (error);
3497 }
3498
3499 static int
3500 hammer_vop_fifowrite (struct vop_write_args *ap)
3501 {
3502         int error;
3503
3504         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3505         /* XXX update access time */
3506         return (error);
3507 }
3508
3509 static
3510 int
3511 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3512 {
3513         int error;
3514
3515         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3516         if (error)
3517                 error = hammer_vop_kqfilter(ap);
3518         return(error);
3519 }
3520
3521 /************************************************************************
3522  *                          KQFILTER OPS                                *
3523  ************************************************************************
3524  *
3525  */
3526 static void filt_hammerdetach(struct knote *kn);
3527 static int filt_hammerread(struct knote *kn, long hint);
3528 static int filt_hammerwrite(struct knote *kn, long hint);
3529 static int filt_hammervnode(struct knote *kn, long hint);
3530
3531 static struct filterops hammerread_filtops =
3532         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammerread };
3533 static struct filterops hammerwrite_filtops =
3534         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammerwrite };
3535 static struct filterops hammervnode_filtops =
3536         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammervnode };
3537
3538 static
3539 int
3540 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3541 {
3542         struct vnode *vp = ap->a_vp;
3543         struct knote *kn = ap->a_kn;
3544
3545         switch (kn->kn_filter) {
3546         case EVFILT_READ:
3547                 kn->kn_fop = &hammerread_filtops;
3548                 break;
3549         case EVFILT_WRITE:
3550                 kn->kn_fop = &hammerwrite_filtops;
3551                 break;
3552         case EVFILT_VNODE:
3553                 kn->kn_fop = &hammervnode_filtops;
3554                 break;
3555         default:
3556                 return (EOPNOTSUPP);
3557         }
3558
3559         kn->kn_hook = (caddr_t)vp;
3560
3561         knote_insert(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3562
3563         return(0);
3564 }
3565
3566 static void
3567 filt_hammerdetach(struct knote *kn)
3568 {
3569         struct vnode *vp = (void *)kn->kn_hook;
3570
3571         knote_remove(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3572 }
3573
3574 static int
3575 filt_hammerread(struct knote *kn, long hint)
3576 {
3577         struct vnode *vp = (void *)kn->kn_hook;
3578         hammer_inode_t ip = VTOI(vp);
3579         hammer_mount_t hmp = ip->hmp;
3580         off_t off;
3581
3582         if (hint == NOTE_REVOKE) {
3583                 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3584                 return(1);
3585         }
3586         lwkt_gettoken(&hmp->fs_token);  /* XXX use per-ip-token */
3587         off = ip->ino_data.size - kn->kn_fp->f_offset;
3588         kn->kn_data = (off < INTPTR_MAX) ? off : INTPTR_MAX;
3589         lwkt_reltoken(&hmp->fs_token);
3590         if (kn->kn_sfflags & NOTE_OLDAPI)
3591                 return(1);
3592         return (kn->kn_data != 0);
3593 }
3594
3595 static int
3596 filt_hammerwrite(struct knote *kn, long hint)
3597 {
3598         if (hint == NOTE_REVOKE)
3599                 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3600         kn->kn_data = 0;
3601         return (1);
3602 }
3603
3604 static int
3605 filt_hammervnode(struct knote *kn, long hint)
3606 {
3607         if (kn->kn_sfflags & hint)
3608                 kn->kn_fflags |= hint;
3609         if (hint == NOTE_REVOKE) {
3610                 kn->kn_flags |= EV_EOF;
3611                 return (1);
3612         }
3613         return (kn->kn_fflags != 0);
3614 }
3615