HAMMER VFS - Only set B_CLUSTEROK on 64K buffers
[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 <vm/swap_pager.h>
50 #include <vfs/fifofs/fifo.h>
51
52 #include "hammer.h"
53
54 /*
55  * USERFS VNOPS
56  */
57 /*static int hammer_vop_vnoperate(struct vop_generic_args *);*/
58 static int hammer_vop_fsync(struct vop_fsync_args *);
59 static int hammer_vop_read(struct vop_read_args *);
60 static int hammer_vop_write(struct vop_write_args *);
61 static int hammer_vop_access(struct vop_access_args *);
62 static int hammer_vop_advlock(struct vop_advlock_args *);
63 static int hammer_vop_close(struct vop_close_args *);
64 static int hammer_vop_ncreate(struct vop_ncreate_args *);
65 static int hammer_vop_getattr(struct vop_getattr_args *);
66 static int hammer_vop_nresolve(struct vop_nresolve_args *);
67 static int hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *);
68 static int hammer_vop_nlink(struct vop_nlink_args *);
69 static int hammer_vop_nmkdir(struct vop_nmkdir_args *);
70 static int hammer_vop_nmknod(struct vop_nmknod_args *);
71 static int hammer_vop_open(struct vop_open_args *);
72 static int hammer_vop_print(struct vop_print_args *);
73 static int hammer_vop_readdir(struct vop_readdir_args *);
74 static int hammer_vop_readlink(struct vop_readlink_args *);
75 static int hammer_vop_nremove(struct vop_nremove_args *);
76 static int hammer_vop_nrename(struct vop_nrename_args *);
77 static int hammer_vop_nrmdir(struct vop_nrmdir_args *);
78 static int hammer_vop_markatime(struct vop_markatime_args *);
79 static int hammer_vop_setattr(struct vop_setattr_args *);
80 static int hammer_vop_strategy(struct vop_strategy_args *);
81 static int hammer_vop_bmap(struct vop_bmap_args *ap);
82 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
83 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
84 static int hammer_vop_ioctl(struct vop_ioctl_args *);
85 static int hammer_vop_mountctl(struct vop_mountctl_args *);
86 static int hammer_vop_kqfilter (struct vop_kqfilter_args *);
87
88 static int hammer_vop_fifoclose (struct vop_close_args *);
89 static int hammer_vop_fiforead (struct vop_read_args *);
90 static int hammer_vop_fifowrite (struct vop_write_args *);
91 static int hammer_vop_fifokqfilter (struct vop_kqfilter_args *);
92
93 struct vop_ops hammer_vnode_vops = {
94         .vop_default =          vop_defaultop,
95         .vop_fsync =            hammer_vop_fsync,
96         .vop_getpages =         vop_stdgetpages,
97         .vop_putpages =         vop_stdputpages,
98         .vop_read =             hammer_vop_read,
99         .vop_write =            hammer_vop_write,
100         .vop_access =           hammer_vop_access,
101         .vop_advlock =          hammer_vop_advlock,
102         .vop_close =            hammer_vop_close,
103         .vop_ncreate =          hammer_vop_ncreate,
104         .vop_getattr =          hammer_vop_getattr,
105         .vop_inactive =         hammer_vop_inactive,
106         .vop_reclaim =          hammer_vop_reclaim,
107         .vop_nresolve =         hammer_vop_nresolve,
108         .vop_nlookupdotdot =    hammer_vop_nlookupdotdot,
109         .vop_nlink =            hammer_vop_nlink,
110         .vop_nmkdir =           hammer_vop_nmkdir,
111         .vop_nmknod =           hammer_vop_nmknod,
112         .vop_open =             hammer_vop_open,
113         .vop_pathconf =         vop_stdpathconf,
114         .vop_print =            hammer_vop_print,
115         .vop_readdir =          hammer_vop_readdir,
116         .vop_readlink =         hammer_vop_readlink,
117         .vop_nremove =          hammer_vop_nremove,
118         .vop_nrename =          hammer_vop_nrename,
119         .vop_nrmdir =           hammer_vop_nrmdir,
120         .vop_markatime =        hammer_vop_markatime,
121         .vop_setattr =          hammer_vop_setattr,
122         .vop_bmap =             hammer_vop_bmap,
123         .vop_strategy =         hammer_vop_strategy,
124         .vop_nsymlink =         hammer_vop_nsymlink,
125         .vop_nwhiteout =        hammer_vop_nwhiteout,
126         .vop_ioctl =            hammer_vop_ioctl,
127         .vop_mountctl =         hammer_vop_mountctl,
128         .vop_kqfilter =         hammer_vop_kqfilter
129 };
130
131 struct vop_ops hammer_spec_vops = {
132         .vop_default =          vop_defaultop,
133         .vop_fsync =            hammer_vop_fsync,
134         .vop_read =             vop_stdnoread,
135         .vop_write =            vop_stdnowrite,
136         .vop_access =           hammer_vop_access,
137         .vop_close =            hammer_vop_close,
138         .vop_markatime =        hammer_vop_markatime,
139         .vop_getattr =          hammer_vop_getattr,
140         .vop_inactive =         hammer_vop_inactive,
141         .vop_reclaim =          hammer_vop_reclaim,
142         .vop_setattr =          hammer_vop_setattr
143 };
144
145 struct vop_ops hammer_fifo_vops = {
146         .vop_default =          fifo_vnoperate,
147         .vop_fsync =            hammer_vop_fsync,
148         .vop_read =             hammer_vop_fiforead,
149         .vop_write =            hammer_vop_fifowrite,
150         .vop_access =           hammer_vop_access,
151         .vop_close =            hammer_vop_fifoclose,
152         .vop_markatime =        hammer_vop_markatime,
153         .vop_getattr =          hammer_vop_getattr,
154         .vop_inactive =         hammer_vop_inactive,
155         .vop_reclaim =          hammer_vop_reclaim,
156         .vop_setattr =          hammer_vop_setattr,
157         .vop_kqfilter =         hammer_vop_fifokqfilter
158 };
159
160 static __inline
161 void
162 hammer_knote(struct vnode *vp, int flags)
163 {
164         if (flags)
165                 KNOTE(&vp->v_pollinfo.vpi_kqinfo.ki_note, flags);
166 }
167
168 #ifdef DEBUG_TRUNCATE
169 struct hammer_inode *HammerTruncIp;
170 #endif
171
172 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
173                            struct vnode *dvp, struct ucred *cred,
174                            int flags, int isdir);
175 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
176 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
177
178 #if 0
179 static
180 int
181 hammer_vop_vnoperate(struct vop_generic_args *)
182 {
183         return (VOCALL(&hammer_vnode_vops, ap));
184 }
185 #endif
186
187 /*
188  * hammer_vop_fsync { vp, waitfor }
189  *
190  * fsync() an inode to disk and wait for it to be completely committed
191  * such that the information would not be undone if a crash occured after
192  * return.
193  *
194  * NOTE: HAMMER's fsync()'s are going to remain expensive until we implement
195  *       a REDO log.  A sysctl is provided to relax HAMMER's fsync()
196  *       operation.
197  *
198  *       Ultimately the combination of a REDO log and use of fast storage
199  *       to front-end cluster caches will make fsync fast, but it aint
200  *       here yet.  And, in anycase, we need real transactional
201  *       all-or-nothing features which are not restricted to a single file.
202  */
203 static
204 int
205 hammer_vop_fsync(struct vop_fsync_args *ap)
206 {
207         hammer_inode_t ip = VTOI(ap->a_vp);
208         hammer_mount_t hmp = ip->hmp;
209         int waitfor = ap->a_waitfor;
210         int mode;
211
212         lwkt_gettoken(&hmp->fs_token);
213
214         /*
215          * Fsync rule relaxation (default is either full synchronous flush
216          * or REDO semantics with synchronous flush).
217          */
218         if (ap->a_flags & VOP_FSYNC_SYSCALL) {
219                 switch(hammer_fsync_mode) {
220                 case 0:
221 mode0:
222                         /* no REDO, full synchronous flush */
223                         goto skip;
224                 case 1:
225 mode1:
226                         /* no REDO, full asynchronous flush */
227                         if (waitfor == MNT_WAIT)
228                                 waitfor = MNT_NOWAIT;
229                         goto skip;
230                 case 2:
231                         /* REDO semantics, synchronous flush */
232                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
233                                 goto mode0;
234                         mode = HAMMER_FLUSH_UNDOS_AUTO;
235                         break;
236                 case 3:
237                         /* REDO semantics, relaxed asynchronous flush */
238                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
239                                 goto mode1;
240                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
241                         if (waitfor == MNT_WAIT)
242                                 waitfor = MNT_NOWAIT;
243                         break;
244                 case 4:
245                         /* ignore the fsync() system call */
246                         lwkt_reltoken(&hmp->fs_token);
247                         return(0);
248                 default:
249                         /* we have to do something */
250                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
251                         if (waitfor == MNT_WAIT)
252                                 waitfor = MNT_NOWAIT;
253                         break;
254                 }
255
256                 /*
257                  * Fast fsync only needs to flush the UNDO/REDO fifo if
258                  * HAMMER_INODE_REDO is non-zero and the only modifications
259                  * made to the file are write or write-extends.
260                  */
261                 if ((ip->flags & HAMMER_INODE_REDO) &&
262                     (ip->flags & HAMMER_INODE_MODMASK_NOREDO) == 0
263                 ) {
264                         ++hammer_count_fsyncs;
265                         hammer_flusher_flush_undos(hmp, mode);
266                         ip->redo_count = 0;
267                         lwkt_reltoken(&hmp->fs_token);
268                         return(0);
269                 }
270
271                 /*
272                  * REDO is enabled by fsync(), the idea being we really only
273                  * want to lay down REDO records when programs are using
274                  * fsync() heavily.  The first fsync() on the file starts
275                  * the gravy train going and later fsync()s keep it hot by
276                  * resetting the redo_count.
277                  *
278                  * We weren't running REDOs before now so we have to fall
279                  * through and do a full fsync of what we have.
280                  */
281                 if (hmp->version >= HAMMER_VOL_VERSION_FOUR &&
282                     (hmp->flags & HAMMER_MOUNT_REDO_RECOVERY_RUN) == 0) {
283                         ip->flags |= HAMMER_INODE_REDO;
284                         ip->redo_count = 0;
285                 }
286         }
287 skip:
288
289         /*
290          * Do a full flush sequence.
291          *
292          * Attempt to release the vnode while waiting for the inode to
293          * finish flushing.  This can really mess up inactive->reclaim
294          * sequences so only do it if the vnode is active.
295          */
296         ++hammer_count_fsyncs;
297         vfsync(ap->a_vp, waitfor, 1, NULL, NULL);
298         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
299         if (waitfor == MNT_WAIT) {
300                 if ((ap->a_vp->v_flag & VINACTIVE) == 0)
301                         vn_unlock(ap->a_vp);
302                 hammer_wait_inode(ip);
303                 if ((ap->a_vp->v_flag & VINACTIVE) == 0)
304                         vn_lock(ap->a_vp, LK_EXCLUSIVE | LK_RETRY);
305         }
306         lwkt_reltoken(&hmp->fs_token);
307         return (ip->error);
308 }
309
310 /*
311  * hammer_vop_read { vp, uio, ioflag, cred }
312  *
313  * MPSAFE (for the cache safe does not require fs_token)
314  */
315 static
316 int
317 hammer_vop_read(struct vop_read_args *ap)
318 {
319         struct hammer_transaction trans;
320         hammer_inode_t ip;
321         hammer_mount_t hmp;
322         off_t offset;
323         struct buf *bp;
324         struct uio *uio;
325         int error;
326         int n;
327         int seqcount;
328         int ioseqcount;
329         int blksize;
330         int bigread;
331         int got_fstoken;
332
333         if (ap->a_vp->v_type != VREG)
334                 return (EINVAL);
335         ip = VTOI(ap->a_vp);
336         hmp = ip->hmp;
337         error = 0;
338         uio = ap->a_uio;
339
340         /*
341          * Allow the UIO's size to override the sequential heuristic.
342          */
343         blksize = hammer_blocksize(uio->uio_offset);
344         seqcount = (uio->uio_resid + (BKVASIZE - 1)) / BKVASIZE;
345         ioseqcount = (ap->a_ioflag >> 16);
346         if (seqcount < ioseqcount)
347                 seqcount = ioseqcount;
348
349         /*
350          * If reading or writing a huge amount of data we have to break
351          * atomicy and allow the operation to be interrupted by a signal
352          * or it can DOS the machine.
353          */
354         bigread = (uio->uio_resid > 100 * 1024 * 1024);
355         got_fstoken = 0;
356
357         /*
358          * Access the data typically in HAMMER_BUFSIZE blocks via the
359          * buffer cache, but HAMMER may use a variable block size based
360          * on the offset.
361          *
362          * XXX Temporary hack, delay the start transaction while we remain
363          *     MPSAFE.  NOTE: ino_data.size cannot change while vnode is
364          *     locked-shared.
365          */
366         while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
367                 int64_t base_offset;
368                 int64_t file_limit;
369
370                 blksize = hammer_blocksize(uio->uio_offset);
371                 offset = (int)uio->uio_offset & (blksize - 1);
372                 base_offset = uio->uio_offset - offset;
373
374                 if (bigread && (error = hammer_signal_check(ip->hmp)) != 0)
375                         break;
376
377                 /*
378                  * MPSAFE
379                  */
380                 bp = getblk(ap->a_vp, base_offset, blksize, 0, 0);
381                 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == B_CACHE) {
382                         bp->b_flags &= ~B_AGE;
383                         error = 0;
384                         goto skip;
385                 }
386                 if (ap->a_ioflag & IO_NRDELAY) {
387                         bqrelse(bp);
388                         return (EWOULDBLOCK);
389                 }
390
391                 /*
392                  * MPUNSAFE
393                  */
394                 if (got_fstoken == 0) {
395                         lwkt_gettoken(&hmp->fs_token);
396                         got_fstoken = 1;
397                         hammer_start_transaction(&trans, ip->hmp);
398                 }
399
400                 /*
401                  * NOTE: A valid bp has already been acquired, but was not
402                  *       B_CACHE.
403                  */
404                 if (hammer_cluster_enable) {
405                         /*
406                          * Use file_limit to prevent cluster_read() from
407                          * creating buffers of the wrong block size past
408                          * the demarc.
409                          */
410                         file_limit = ip->ino_data.size;
411                         if (base_offset < HAMMER_XDEMARC &&
412                             file_limit > HAMMER_XDEMARC) {
413                                 file_limit = HAMMER_XDEMARC;
414                         }
415                         error = cluster_readx(ap->a_vp,
416                                              file_limit, base_offset,
417                                              blksize, uio->uio_resid,
418                                              seqcount * BKVASIZE, &bp);
419                 } else {
420                         error = breadnx(ap->a_vp, base_offset, blksize,
421                                         NULL, NULL, 0, &bp);
422                 }
423                 if (error) {
424                         brelse(bp);
425                         break;
426                 }
427 skip:
428                 if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IODEBUG)) {
429                         kprintf("doff %016jx read file %016jx@%016jx\n",
430                                 (intmax_t)bp->b_bio2.bio_offset,
431                                 (intmax_t)ip->obj_id,
432                                 (intmax_t)bp->b_loffset);
433                 }
434                 bp->b_flags &= ~B_IODEBUG;
435                 if (blksize == HAMMER_XBUFSIZE)
436                         bp->b_flags |= B_CLUSTEROK;
437
438                 n = blksize - offset;
439                 if (n > uio->uio_resid)
440                         n = uio->uio_resid;
441                 if (n > ip->ino_data.size - uio->uio_offset)
442                         n = (int)(ip->ino_data.size - uio->uio_offset);
443                 if (got_fstoken)
444                         lwkt_reltoken(&hmp->fs_token);
445
446                 /*
447                  * Set B_AGE, data has a lower priority than meta-data.
448                  *
449                  * Use a hold/unlock/drop sequence to run the uiomove
450                  * with the buffer unlocked, avoiding deadlocks against
451                  * read()s on mmap()'d spaces.
452                  */
453                 bp->b_flags |= B_AGE;
454                 bqhold(bp);
455                 bqrelse(bp);
456                 error = uiomove((char *)bp->b_data + offset, n, uio);
457                 bqdrop(bp);
458
459                 if (got_fstoken)
460                         lwkt_gettoken(&hmp->fs_token);
461
462                 if (error)
463                         break;
464                 hammer_stats_file_read += n;
465         }
466
467         /*
468          * Try to update the atime with just the inode lock for maximum
469          * concurrency.  If we can't shortcut it we have to get the full
470          * blown transaction.
471          */
472         if (got_fstoken == 0 && hammer_update_atime_quick(ip) < 0) {
473                 lwkt_gettoken(&hmp->fs_token);
474                 got_fstoken = 1;
475                 hammer_start_transaction(&trans, ip->hmp);
476         }
477
478         if (got_fstoken) {
479                 if ((ip->flags & HAMMER_INODE_RO) == 0 &&
480                     (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
481                         ip->ino_data.atime = trans.time;
482                         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
483                 }
484                 hammer_done_transaction(&trans);
485                 lwkt_reltoken(&hmp->fs_token);
486         }
487         return (error);
488 }
489
490 /*
491  * hammer_vop_write { vp, uio, ioflag, cred }
492  */
493 static
494 int
495 hammer_vop_write(struct vop_write_args *ap)
496 {
497         struct hammer_transaction trans;
498         struct hammer_inode *ip;
499         hammer_mount_t hmp;
500         thread_t td;
501         struct uio *uio;
502         int offset;
503         off_t base_offset;
504         int64_t cluster_eof;
505         struct buf *bp;
506         int kflags;
507         int error;
508         int n;
509         int flags;
510         int seqcount;
511         int bigwrite;
512
513         if (ap->a_vp->v_type != VREG)
514                 return (EINVAL);
515         ip = VTOI(ap->a_vp);
516         hmp = ip->hmp;
517         error = 0;
518         kflags = 0;
519         seqcount = ap->a_ioflag >> 16;
520
521         if (ip->flags & HAMMER_INODE_RO)
522                 return (EROFS);
523
524         /*
525          * Create a transaction to cover the operations we perform.
526          */
527         lwkt_gettoken(&hmp->fs_token);
528         hammer_start_transaction(&trans, hmp);
529         uio = ap->a_uio;
530
531         /*
532          * Check append mode
533          */
534         if (ap->a_ioflag & IO_APPEND)
535                 uio->uio_offset = ip->ino_data.size;
536
537         /*
538          * Check for illegal write offsets.  Valid range is 0...2^63-1.
539          *
540          * NOTE: the base_off assignment is required to work around what
541          * I consider to be a GCC-4 optimization bug.
542          */
543         if (uio->uio_offset < 0) {
544                 hammer_done_transaction(&trans);
545                 lwkt_reltoken(&hmp->fs_token);
546                 return (EFBIG);
547         }
548         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
549         if (uio->uio_resid > 0 && base_offset <= uio->uio_offset) {
550                 hammer_done_transaction(&trans);
551                 lwkt_reltoken(&hmp->fs_token);
552                 return (EFBIG);
553         }
554
555         if (uio->uio_resid > 0 && (td = uio->uio_td) != NULL && td->td_proc &&
556             base_offset > td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
557                 hammer_done_transaction(&trans);
558                 lwkt_reltoken(&hmp->fs_token);
559                 lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ);
560                 return (EFBIG);
561         }
562
563         /*
564          * If reading or writing a huge amount of data we have to break
565          * atomicy and allow the operation to be interrupted by a signal
566          * or it can DOS the machine.
567          *
568          * Preset redo_count so we stop generating REDOs earlier if the
569          * limit is exceeded.
570          */
571         bigwrite = (uio->uio_resid > 100 * 1024 * 1024);
572         if ((ip->flags & HAMMER_INODE_REDO) &&
573             ip->redo_count < hammer_limit_redo) {
574                 ip->redo_count += uio->uio_resid;
575         }
576
577         /*
578          * Access the data typically in HAMMER_BUFSIZE blocks via the
579          * buffer cache, but HAMMER may use a variable block size based
580          * on the offset.
581          */
582         while (uio->uio_resid > 0) {
583                 int fixsize = 0;
584                 int blksize;
585                 int blkmask;
586                 int trivial;
587                 int endofblk;
588                 off_t nsize;
589
590                 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
591                         break;
592                 if (bigwrite && (error = hammer_signal_check(hmp)) != 0)
593                         break;
594
595                 blksize = hammer_blocksize(uio->uio_offset);
596
597                 /*
598                  * Do not allow HAMMER to blow out the buffer cache.  Very
599                  * large UIOs can lockout other processes due to bwillwrite()
600                  * mechanics.
601                  *
602                  * The hammer inode is not locked during these operations.
603                  * The vnode is locked which can interfere with the pageout
604                  * daemon for non-UIO_NOCOPY writes but should not interfere
605                  * with the buffer cache.  Even so, we cannot afford to
606                  * allow the pageout daemon to build up too many dirty buffer
607                  * cache buffers.
608                  *
609                  * Only call this if we aren't being recursively called from
610                  * a virtual disk device (vn), else we may deadlock.
611                  */
612                 if ((ap->a_ioflag & IO_RECURSE) == 0)
613                         bwillwrite(blksize);
614
615                 /*
616                  * Control the number of pending records associated with
617                  * this inode.  If too many have accumulated start a
618                  * flush.  Try to maintain a pipeline with the flusher.
619                  *
620                  * NOTE: It is possible for other sources to grow the
621                  *       records but not necessarily issue another flush,
622                  *       so use a timeout and ensure that a re-flush occurs.
623                  */
624                 if (ip->rsv_recs >= hammer_limit_inode_recs) {
625                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
626                         while (ip->rsv_recs >= hammer_limit_inode_recs * 2) {
627                                 ip->flags |= HAMMER_INODE_RECSW;
628                                 tsleep(&ip->rsv_recs, 0, "hmrwww", hz);
629                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
630                         }
631                 }
632
633 #if 0
634                 /*
635                  * Do not allow HAMMER to blow out system memory by
636                  * accumulating too many records.   Records are so well
637                  * decoupled from the buffer cache that it is possible
638                  * for userland to push data out to the media via
639                  * direct-write, but build up the records queued to the
640                  * backend faster then the backend can flush them out.
641                  * HAMMER has hit its write limit but the frontend has
642                  * no pushback to slow it down.
643                  */
644                 if (hmp->rsv_recs > hammer_limit_recs / 2) {
645                         /*
646                          * Get the inode on the flush list
647                          */
648                         if (ip->rsv_recs >= 64)
649                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
650                         else if (ip->rsv_recs >= 16)
651                                 hammer_flush_inode(ip, 0);
652
653                         /*
654                          * Keep the flusher going if the system keeps
655                          * queueing records.
656                          */
657                         delta = hmp->count_newrecords -
658                                 hmp->last_newrecords;
659                         if (delta < 0 || delta > hammer_limit_recs / 2) {
660                                 hmp->last_newrecords = hmp->count_newrecords;
661                                 hammer_sync_hmp(hmp, MNT_NOWAIT);
662                         }
663
664                         /*
665                          * If we have gotten behind start slowing
666                          * down the writers.
667                          */
668                         delta = (hmp->rsv_recs - hammer_limit_recs) *
669                                 hz / hammer_limit_recs;
670                         if (delta > 0)
671                                 tsleep(&trans, 0, "hmrslo", delta);
672                 }
673 #endif
674
675                 /*
676                  * Calculate the blocksize at the current offset and figure
677                  * out how much we can actually write.
678                  */
679                 blkmask = blksize - 1;
680                 offset = (int)uio->uio_offset & blkmask;
681                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
682                 n = blksize - offset;
683                 if (n > uio->uio_resid) {
684                         n = uio->uio_resid;
685                         endofblk = 0;
686                 } else {
687                         endofblk = 1;
688                 }
689                 nsize = uio->uio_offset + n;
690                 if (nsize > ip->ino_data.size) {
691                         if (uio->uio_offset > ip->ino_data.size)
692                                 trivial = 0;
693                         else
694                                 trivial = 1;
695                         nvextendbuf(ap->a_vp,
696                                     ip->ino_data.size,
697                                     nsize,
698                                     hammer_blocksize(ip->ino_data.size),
699                                     hammer_blocksize(nsize),
700                                     hammer_blockoff(ip->ino_data.size),
701                                     hammer_blockoff(nsize),
702                                     trivial);
703                         fixsize = 1;
704                         kflags |= NOTE_EXTEND;
705                 }
706
707                 if (uio->uio_segflg == UIO_NOCOPY) {
708                         /*
709                          * Issuing a write with the same data backing the
710                          * buffer.  Instantiate the buffer to collect the
711                          * backing vm pages, then read-in any missing bits.
712                          *
713                          * This case is used by vop_stdputpages().
714                          */
715                         bp = getblk(ap->a_vp, base_offset,
716                                     blksize, GETBLK_BHEAVY, 0);
717                         if ((bp->b_flags & B_CACHE) == 0) {
718                                 bqrelse(bp);
719                                 error = bread(ap->a_vp, base_offset,
720                                               blksize, &bp);
721                         }
722                 } else if (offset == 0 && uio->uio_resid >= blksize) {
723                         /*
724                          * Even though we are entirely overwriting the buffer
725                          * we may still have to zero it out to avoid a 
726                          * mmap/write visibility issue.
727                          */
728                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
729                         if ((bp->b_flags & B_CACHE) == 0)
730                                 vfs_bio_clrbuf(bp);
731                 } else if (base_offset >= ip->ino_data.size) {
732                         /*
733                          * If the base offset of the buffer is beyond the
734                          * file EOF, we don't have to issue a read.
735                          */
736                         bp = getblk(ap->a_vp, base_offset,
737                                     blksize, GETBLK_BHEAVY, 0);
738                         vfs_bio_clrbuf(bp);
739                 } else {
740                         /*
741                          * Partial overwrite, read in any missing bits then
742                          * replace the portion being written.
743                          */
744                         error = bread(ap->a_vp, base_offset, blksize, &bp);
745                         if (error == 0)
746                                 bheavy(bp);
747                 }
748                 if (error == 0) {
749                         lwkt_reltoken(&hmp->fs_token);
750                         error = uiomove(bp->b_data + offset, n, uio);
751                         lwkt_gettoken(&hmp->fs_token);
752                 }
753
754                 /*
755                  * Generate REDO records if enabled and redo_count will not
756                  * exceeded the limit.
757                  *
758                  * If redo_count exceeds the limit we stop generating records
759                  * and clear HAMMER_INODE_REDO.  This will cause the next
760                  * fsync() to do a full meta-data sync instead of just an
761                  * UNDO/REDO fifo update.
762                  *
763                  * When clearing HAMMER_INODE_REDO any pre-existing REDOs
764                  * will still be tracked.  The tracks will be terminated
765                  * when the related meta-data (including possible data
766                  * modifications which are not tracked via REDO) is
767                  * flushed.
768                  */
769                 if ((ip->flags & HAMMER_INODE_REDO) && error == 0) {
770                         if (ip->redo_count < hammer_limit_redo) {
771                                 bp->b_flags |= B_VFSFLAG1;
772                                 error = hammer_generate_redo(&trans, ip,
773                                                      base_offset + offset,
774                                                      HAMMER_REDO_WRITE,
775                                                      bp->b_data + offset,
776                                                      (size_t)n);
777                         } else {
778                                 ip->flags &= ~HAMMER_INODE_REDO;
779                         }
780                 }
781
782                 /*
783                  * If we screwed up we have to undo any VM size changes we
784                  * made.
785                  */
786                 if (error) {
787                         brelse(bp);
788                         if (fixsize) {
789                                 nvtruncbuf(ap->a_vp, ip->ino_data.size,
790                                           hammer_blocksize(ip->ino_data.size),
791                                           hammer_blockoff(ip->ino_data.size),
792                                           0);
793                         }
794                         break;
795                 }
796                 kflags |= NOTE_WRITE;
797                 hammer_stats_file_write += n;
798                 if (blksize == HAMMER_XBUFSIZE)
799                         bp->b_flags |= B_CLUSTEROK;
800                 if (ip->ino_data.size < uio->uio_offset) {
801                         ip->ino_data.size = uio->uio_offset;
802                         flags = HAMMER_INODE_SDIRTY;
803                 } else {
804                         flags = 0;
805                 }
806                 ip->ino_data.mtime = trans.time;
807                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
808                 hammer_modify_inode(&trans, ip, flags);
809
810                 /*
811                  * Once we dirty the buffer any cached zone-X offset
812                  * becomes invalid.  HAMMER NOTE: no-history mode cannot 
813                  * allow overwriting over the same data sector unless
814                  * we provide UNDOs for the old data, which we don't.
815                  */
816                 bp->b_bio2.bio_offset = NOOFFSET;
817
818                 /*
819                  * Final buffer disposition.
820                  *
821                  * Because meta-data updates are deferred, HAMMER is
822                  * especially sensitive to excessive bdwrite()s because
823                  * the I/O stream is not broken up by disk reads.  So the
824                  * buffer cache simply cannot keep up.
825                  *
826                  * WARNING!  blksize is variable.  cluster_write() is
827                  *           expected to not blow up if it encounters
828                  *           buffers that do not match the passed blksize.
829                  *
830                  * NOTE!  Hammer shouldn't need to bawrite()/cluster_write().
831                  *        The ip->rsv_recs check should burst-flush the data.
832                  *        If we queue it immediately the buf could be left
833                  *        locked on the device queue for a very long time.
834                  *
835                  *        However, failing to flush a dirty buffer out when
836                  *        issued from the pageout daemon can result in a low
837                  *        memory deadlock against bio_page_alloc(), so we
838                  *        have to bawrite() on IO_ASYNC as well.
839                  *
840                  * NOTE!  To avoid degenerate stalls due to mismatched block
841                  *        sizes we only honor IO_DIRECT on the write which
842                  *        abuts the end of the buffer.  However, we must
843                  *        honor IO_SYNC in case someone is silly enough to
844                  *        configure a HAMMER file as swap, or when HAMMER
845                  *        is serving NFS (for commits).  Ick ick.
846                  */
847                 bp->b_flags |= B_AGE;
848                 if (blksize == HAMMER_XBUFSIZE)
849                         bp->b_flags |= B_CLUSTEROK;
850
851                 if (ap->a_ioflag & IO_SYNC) {
852                         bwrite(bp);
853                 } else if ((ap->a_ioflag & IO_DIRECT) && endofblk) {
854                         bawrite(bp);
855                 } else if (ap->a_ioflag & IO_ASYNC) {
856                         bawrite(bp);
857                 } else if (hammer_cluster_enable &&
858                            !(ap->a_vp->v_mount->mnt_flag & MNT_NOCLUSTERW)) {
859                         if (base_offset < HAMMER_XDEMARC)
860                                 cluster_eof = hammer_blockdemarc(base_offset,
861                                                          ip->ino_data.size);
862                         else
863                                 cluster_eof = ip->ino_data.size;
864                         cluster_write(bp, cluster_eof, blksize, seqcount);
865                 } else {
866                         bdwrite(bp);
867                 }
868         }
869         hammer_done_transaction(&trans);
870         hammer_knote(ap->a_vp, kflags);
871         lwkt_reltoken(&hmp->fs_token);
872         return (error);
873 }
874
875 /*
876  * hammer_vop_access { vp, mode, cred }
877  *
878  * MPSAFE - does not require fs_token
879  */
880 static
881 int
882 hammer_vop_access(struct vop_access_args *ap)
883 {
884         struct hammer_inode *ip = VTOI(ap->a_vp);
885         uid_t uid;
886         gid_t gid;
887         int error;
888
889         ++hammer_stats_file_iopsr;
890         uid = hammer_to_unix_xid(&ip->ino_data.uid);
891         gid = hammer_to_unix_xid(&ip->ino_data.gid);
892
893         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
894                                   ip->ino_data.uflags);
895         return (error);
896 }
897
898 /*
899  * hammer_vop_advlock { vp, id, op, fl, flags }
900  *
901  * MPSAFE - does not require fs_token
902  */
903 static
904 int
905 hammer_vop_advlock(struct vop_advlock_args *ap)
906 {
907         hammer_inode_t ip = VTOI(ap->a_vp);
908
909         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
910 }
911
912 /*
913  * hammer_vop_close { vp, fflag }
914  *
915  * We can only sync-on-close for normal closes.  XXX disabled for now.
916  */
917 static
918 int
919 hammer_vop_close(struct vop_close_args *ap)
920 {
921 #if 0
922         struct vnode *vp = ap->a_vp;
923         hammer_inode_t ip = VTOI(vp);
924         int waitfor;
925         if (ip->flags & (HAMMER_INODE_CLOSESYNC|HAMMER_INODE_CLOSEASYNC)) {
926                 if (vn_islocked(vp) == LK_EXCLUSIVE &&
927                     (vp->v_flag & (VINACTIVE|VRECLAIMED)) == 0) {
928                         if (ip->flags & HAMMER_INODE_CLOSESYNC)
929                                 waitfor = MNT_WAIT;
930                         else
931                                 waitfor = MNT_NOWAIT;
932                         ip->flags &= ~(HAMMER_INODE_CLOSESYNC |
933                                        HAMMER_INODE_CLOSEASYNC);
934                         VOP_FSYNC(vp, MNT_NOWAIT, waitfor);
935                 }
936         }
937 #endif
938         return (vop_stdclose(ap));
939 }
940
941 /*
942  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
943  *
944  * The operating system has already ensured that the directory entry
945  * does not exist and done all appropriate namespace locking.
946  */
947 static
948 int
949 hammer_vop_ncreate(struct vop_ncreate_args *ap)
950 {
951         struct hammer_transaction trans;
952         struct hammer_inode *dip;
953         struct hammer_inode *nip;
954         struct nchandle *nch;
955         hammer_mount_t hmp;
956         int error;
957
958         nch = ap->a_nch;
959         dip = VTOI(ap->a_dvp);
960         hmp = dip->hmp;
961
962         if (dip->flags & HAMMER_INODE_RO)
963                 return (EROFS);
964         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
965                 return (error);
966
967         /*
968          * Create a transaction to cover the operations we perform.
969          */
970         lwkt_gettoken(&hmp->fs_token);
971         hammer_start_transaction(&trans, hmp);
972         ++hammer_stats_file_iopsw;
973
974         /*
975          * Create a new filesystem object of the requested type.  The
976          * returned inode will be referenced and shared-locked to prevent
977          * it from being moved to the flusher.
978          */
979         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
980                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
981                                     NULL, &nip);
982         if (error) {
983                 hkprintf("hammer_create_inode error %d\n", error);
984                 hammer_done_transaction(&trans);
985                 *ap->a_vpp = NULL;
986                 lwkt_reltoken(&hmp->fs_token);
987                 return (error);
988         }
989
990         /*
991          * Add the new filesystem object to the directory.  This will also
992          * bump the inode's link count.
993          */
994         error = hammer_ip_add_directory(&trans, dip,
995                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
996                                         nip);
997         if (error)
998                 hkprintf("hammer_ip_add_directory error %d\n", error);
999
1000         /*
1001          * Finish up.
1002          */
1003         if (error) {
1004                 hammer_rel_inode(nip, 0);
1005                 hammer_done_transaction(&trans);
1006                 *ap->a_vpp = NULL;
1007         } else {
1008                 error = hammer_get_vnode(nip, ap->a_vpp);
1009                 hammer_done_transaction(&trans);
1010                 hammer_rel_inode(nip, 0);
1011                 if (error == 0) {
1012                         cache_setunresolved(ap->a_nch);
1013                         cache_setvp(ap->a_nch, *ap->a_vpp);
1014                 }
1015                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1016         }
1017         lwkt_reltoken(&hmp->fs_token);
1018         return (error);
1019 }
1020
1021 /*
1022  * hammer_vop_getattr { vp, vap }
1023  *
1024  * Retrieve an inode's attribute information.  When accessing inodes
1025  * historically we fake the atime field to ensure consistent results.
1026  * The atime field is stored in the B-Tree element and allowed to be
1027  * updated without cycling the element.
1028  *
1029  * MPSAFE - does not require fs_token
1030  */
1031 static
1032 int
1033 hammer_vop_getattr(struct vop_getattr_args *ap)
1034 {
1035         struct hammer_inode *ip = VTOI(ap->a_vp);
1036         struct vattr *vap = ap->a_vap;
1037
1038         /*
1039          * We want the fsid to be different when accessing a filesystem
1040          * with different as-of's so programs like diff don't think
1041          * the files are the same.
1042          *
1043          * We also want the fsid to be the same when comparing snapshots,
1044          * or when comparing mirrors (which might be backed by different
1045          * physical devices).  HAMMER fsids are based on the PFS's
1046          * shared_uuid field.
1047          *
1048          * XXX there is a chance of collision here.  The va_fsid reported
1049          * by stat is different from the more involved fsid used in the
1050          * mount structure.
1051          */
1052         ++hammer_stats_file_iopsr;
1053         hammer_lock_sh(&ip->lock);
1054         vap->va_fsid = ip->pfsm->fsid_udev ^ (u_int32_t)ip->obj_asof ^
1055                        (u_int32_t)(ip->obj_asof >> 32);
1056
1057         vap->va_fileid = ip->ino_leaf.base.obj_id;
1058         vap->va_mode = ip->ino_data.mode;
1059         vap->va_nlink = ip->ino_data.nlinks;
1060         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1061         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1062         vap->va_rmajor = 0;
1063         vap->va_rminor = 0;
1064         vap->va_size = ip->ino_data.size;
1065
1066         /*
1067          * Special case for @@PFS softlinks.  The actual size of the
1068          * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
1069          * or for MAX_TID is    "@@-1:%05d" == 10 bytes.
1070          */
1071         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
1072             ip->ino_data.size == 10 &&
1073             ip->obj_asof == HAMMER_MAX_TID &&
1074             ip->obj_localization == 0 &&
1075             strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
1076                     if (ip->pfsm->pfsd.mirror_flags & HAMMER_PFSD_SLAVE)
1077                             vap->va_size = 26;
1078                     else
1079                             vap->va_size = 10;
1080         }
1081
1082         /*
1083          * We must provide a consistent atime and mtime for snapshots
1084          * so people can do a 'tar cf - ... | md5' on them and get
1085          * consistent results.
1086          */
1087         if (ip->flags & HAMMER_INODE_RO) {
1088                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
1089                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
1090         } else {
1091                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
1092                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
1093         }
1094         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
1095         vap->va_flags = ip->ino_data.uflags;
1096         vap->va_gen = 1;        /* hammer inums are unique for all time */
1097         vap->va_blocksize = HAMMER_BUFSIZE;
1098         if (ip->ino_data.size >= HAMMER_XDEMARC) {
1099                 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
1100                                 ~HAMMER_XBUFMASK64;
1101         } else if (ip->ino_data.size > HAMMER_BUFSIZE / 2) {
1102                 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
1103                                 ~HAMMER_BUFMASK64;
1104         } else {
1105                 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
1106         }
1107
1108         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
1109         vap->va_filerev = 0;    /* XXX */
1110         vap->va_uid_uuid = ip->ino_data.uid;
1111         vap->va_gid_uuid = ip->ino_data.gid;
1112         vap->va_fsid_uuid = ip->hmp->fsid;
1113         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
1114                           VA_FSID_UUID_VALID;
1115
1116         switch (ip->ino_data.obj_type) {
1117         case HAMMER_OBJTYPE_CDEV:
1118         case HAMMER_OBJTYPE_BDEV:
1119                 vap->va_rmajor = ip->ino_data.rmajor;
1120                 vap->va_rminor = ip->ino_data.rminor;
1121                 break;
1122         default:
1123                 break;
1124         }
1125         hammer_unlock(&ip->lock);
1126         return(0);
1127 }
1128
1129 /*
1130  * hammer_vop_nresolve { nch, dvp, cred }
1131  *
1132  * Locate the requested directory entry.
1133  */
1134 static
1135 int
1136 hammer_vop_nresolve(struct vop_nresolve_args *ap)
1137 {
1138         struct hammer_transaction trans;
1139         struct namecache *ncp;
1140         hammer_mount_t hmp;
1141         hammer_inode_t dip;
1142         hammer_inode_t ip;
1143         hammer_tid_t asof;
1144         struct hammer_cursor cursor;
1145         struct vnode *vp;
1146         int64_t namekey;
1147         int error;
1148         int i;
1149         int nlen;
1150         int flags;
1151         int ispfs;
1152         int64_t obj_id;
1153         u_int32_t localization;
1154         u_int32_t max_iterations;
1155
1156         /*
1157          * Misc initialization, plus handle as-of name extensions.  Look for
1158          * the '@@' extension.  Note that as-of files and directories cannot
1159          * be modified.
1160          */
1161         dip = VTOI(ap->a_dvp);
1162         ncp = ap->a_nch->ncp;
1163         asof = dip->obj_asof;
1164         localization = dip->obj_localization;   /* for code consistency */
1165         nlen = ncp->nc_nlen;
1166         flags = dip->flags & HAMMER_INODE_RO;
1167         ispfs = 0;
1168         hmp = dip->hmp;
1169
1170         lwkt_gettoken(&hmp->fs_token);
1171         hammer_simple_transaction(&trans, hmp);
1172         ++hammer_stats_file_iopsr;
1173
1174         for (i = 0; i < nlen; ++i) {
1175                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
1176                         error = hammer_str_to_tid(ncp->nc_name + i + 2,
1177                                                   &ispfs, &asof, &localization);
1178                         if (error != 0) {
1179                                 i = nlen;
1180                                 break;
1181                         }
1182                         if (asof != HAMMER_MAX_TID)
1183                                 flags |= HAMMER_INODE_RO;
1184                         break;
1185                 }
1186         }
1187         nlen = i;
1188
1189         /*
1190          * If this is a PFS softlink we dive into the PFS
1191          */
1192         if (ispfs && nlen == 0) {
1193                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1194                                       asof, localization,
1195                                       flags, &error);
1196                 if (error == 0) {
1197                         error = hammer_get_vnode(ip, &vp);
1198                         hammer_rel_inode(ip, 0);
1199                 } else {
1200                         vp = NULL;
1201                 }
1202                 if (error == 0) {
1203                         vn_unlock(vp);
1204                         cache_setvp(ap->a_nch, vp);
1205                         vrele(vp);
1206                 }
1207                 goto done;
1208         }
1209
1210         /*
1211          * If there is no path component the time extension is relative to dip.
1212          * e.g. "fubar/@@<snapshot>"
1213          *
1214          * "." is handled by the kernel, but ".@@<snapshot>" is not.
1215          * e.g. "fubar/.@@<snapshot>"
1216          *
1217          * ".." is handled by the kernel.  We do not currently handle
1218          * "..@<snapshot>".
1219          */
1220         if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1221                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
1222                                       asof, dip->obj_localization,
1223                                       flags, &error);
1224                 if (error == 0) {
1225                         error = hammer_get_vnode(ip, &vp);
1226                         hammer_rel_inode(ip, 0);
1227                 } else {
1228                         vp = NULL;
1229                 }
1230                 if (error == 0) {
1231                         vn_unlock(vp);
1232                         cache_setvp(ap->a_nch, vp);
1233                         vrele(vp);
1234                 }
1235                 goto done;
1236         }
1237
1238         /*
1239          * Calculate the namekey and setup the key range for the scan.  This
1240          * works kinda like a chained hash table where the lower 32 bits
1241          * of the namekey synthesize the chain.
1242          *
1243          * The key range is inclusive of both key_beg and key_end.
1244          */
1245         namekey = hammer_directory_namekey(dip, ncp->nc_name, nlen,
1246                                            &max_iterations);
1247
1248         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1249         cursor.key_beg.localization = dip->obj_localization +
1250                                       hammer_dir_localization(dip);
1251         cursor.key_beg.obj_id = dip->obj_id;
1252         cursor.key_beg.key = namekey;
1253         cursor.key_beg.create_tid = 0;
1254         cursor.key_beg.delete_tid = 0;
1255         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1256         cursor.key_beg.obj_type = 0;
1257
1258         cursor.key_end = cursor.key_beg;
1259         cursor.key_end.key += max_iterations;
1260         cursor.asof = asof;
1261         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1262
1263         /*
1264          * Scan all matching records (the chain), locate the one matching
1265          * the requested path component.
1266          *
1267          * The hammer_ip_*() functions merge in-memory records with on-disk
1268          * records for the purposes of the search.
1269          */
1270         obj_id = 0;
1271         localization = HAMMER_DEF_LOCALIZATION;
1272
1273         if (error == 0) {
1274                 error = hammer_ip_first(&cursor);
1275                 while (error == 0) {
1276                         error = hammer_ip_resolve_data(&cursor);
1277                         if (error)
1278                                 break;
1279                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1280                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1281                                 obj_id = cursor.data->entry.obj_id;
1282                                 localization = cursor.data->entry.localization;
1283                                 break;
1284                         }
1285                         error = hammer_ip_next(&cursor);
1286                 }
1287         }
1288         hammer_done_cursor(&cursor);
1289
1290         /*
1291          * Lookup the obj_id.  This should always succeed.  If it does not
1292          * the filesystem may be damaged and we return a dummy inode.
1293          */
1294         if (error == 0) {
1295                 ip = hammer_get_inode(&trans, dip, obj_id,
1296                                       asof, localization,
1297                                       flags, &error);
1298                 if (error == ENOENT) {
1299                         kprintf("HAMMER: WARNING: Missing "
1300                                 "inode for dirent \"%s\"\n"
1301                                 "\tobj_id = %016llx, asof=%016llx, lo=%08x\n",
1302                                 ncp->nc_name,
1303                                 (long long)obj_id, (long long)asof,
1304                                 localization);
1305                         error = 0;
1306                         ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1307                                                     asof, localization,
1308                                                     flags, &error);
1309                 }
1310                 if (error == 0) {
1311                         error = hammer_get_vnode(ip, &vp);
1312                         hammer_rel_inode(ip, 0);
1313                 } else {
1314                         vp = NULL;
1315                 }
1316                 if (error == 0) {
1317                         vn_unlock(vp);
1318                         cache_setvp(ap->a_nch, vp);
1319                         vrele(vp);
1320                 }
1321         } else if (error == ENOENT) {
1322                 cache_setvp(ap->a_nch, NULL);
1323         }
1324 done:
1325         hammer_done_transaction(&trans);
1326         lwkt_reltoken(&hmp->fs_token);
1327         return (error);
1328 }
1329
1330 /*
1331  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1332  *
1333  * Locate the parent directory of a directory vnode.
1334  *
1335  * dvp is referenced but not locked.  *vpp must be returned referenced and
1336  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
1337  * at the root, instead it could indicate that the directory we were in was
1338  * removed.
1339  *
1340  * NOTE: as-of sequences are not linked into the directory structure.  If
1341  * we are at the root with a different asof then the mount point, reload
1342  * the same directory with the mount point's asof.   I'm not sure what this
1343  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
1344  * get confused, but it hasn't been tested.
1345  */
1346 static
1347 int
1348 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1349 {
1350         struct hammer_transaction trans;
1351         struct hammer_inode *dip;
1352         struct hammer_inode *ip;
1353         hammer_mount_t hmp;
1354         int64_t parent_obj_id;
1355         u_int32_t parent_obj_localization;
1356         hammer_tid_t asof;
1357         int error;
1358
1359         dip = VTOI(ap->a_dvp);
1360         asof = dip->obj_asof;
1361         hmp = dip->hmp;
1362
1363         /*
1364          * Whos are parent?  This could be the root of a pseudo-filesystem
1365          * whos parent is in another localization domain.
1366          */
1367         lwkt_gettoken(&hmp->fs_token);
1368         parent_obj_id = dip->ino_data.parent_obj_id;
1369         if (dip->obj_id == HAMMER_OBJID_ROOT)
1370                 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
1371         else
1372                 parent_obj_localization = dip->obj_localization;
1373
1374         if (parent_obj_id == 0) {
1375                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1376                    asof != hmp->asof) {
1377                         parent_obj_id = dip->obj_id;
1378                         asof = hmp->asof;
1379                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1380                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
1381                                   (long long)dip->obj_asof);
1382                 } else {
1383                         *ap->a_vpp = NULL;
1384                         lwkt_reltoken(&hmp->fs_token);
1385                         return ENOENT;
1386                 }
1387         }
1388
1389         hammer_simple_transaction(&trans, hmp);
1390         ++hammer_stats_file_iopsr;
1391
1392         ip = hammer_get_inode(&trans, dip, parent_obj_id,
1393                               asof, parent_obj_localization,
1394                               dip->flags, &error);
1395         if (ip) {
1396                 error = hammer_get_vnode(ip, ap->a_vpp);
1397                 hammer_rel_inode(ip, 0);
1398         } else {
1399                 *ap->a_vpp = NULL;
1400         }
1401         hammer_done_transaction(&trans);
1402         lwkt_reltoken(&hmp->fs_token);
1403         return (error);
1404 }
1405
1406 /*
1407  * hammer_vop_nlink { nch, dvp, vp, cred }
1408  */
1409 static
1410 int
1411 hammer_vop_nlink(struct vop_nlink_args *ap)
1412 {
1413         struct hammer_transaction trans;
1414         struct hammer_inode *dip;
1415         struct hammer_inode *ip;
1416         struct nchandle *nch;
1417         hammer_mount_t hmp;
1418         int error;
1419
1420         if (ap->a_dvp->v_mount != ap->a_vp->v_mount)    
1421                 return(EXDEV);
1422
1423         nch = ap->a_nch;
1424         dip = VTOI(ap->a_dvp);
1425         ip = VTOI(ap->a_vp);
1426         hmp = dip->hmp;
1427
1428         if (dip->obj_localization != ip->obj_localization)
1429                 return(EXDEV);
1430
1431         if (dip->flags & HAMMER_INODE_RO)
1432                 return (EROFS);
1433         if (ip->flags & HAMMER_INODE_RO)
1434                 return (EROFS);
1435         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1436                 return (error);
1437
1438         /*
1439          * Create a transaction to cover the operations we perform.
1440          */
1441         lwkt_gettoken(&hmp->fs_token);
1442         hammer_start_transaction(&trans, hmp);
1443         ++hammer_stats_file_iopsw;
1444
1445         /*
1446          * Add the filesystem object to the directory.  Note that neither
1447          * dip nor ip are referenced or locked, but their vnodes are
1448          * referenced.  This function will bump the inode's link count.
1449          */
1450         error = hammer_ip_add_directory(&trans, dip,
1451                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1452                                         ip);
1453
1454         /*
1455          * Finish up.
1456          */
1457         if (error == 0) {
1458                 cache_setunresolved(nch);
1459                 cache_setvp(nch, ap->a_vp);
1460         }
1461         hammer_done_transaction(&trans);
1462         hammer_knote(ap->a_vp, NOTE_LINK);
1463         hammer_knote(ap->a_dvp, NOTE_WRITE);
1464         lwkt_reltoken(&hmp->fs_token);
1465         return (error);
1466 }
1467
1468 /*
1469  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1470  *
1471  * The operating system has already ensured that the directory entry
1472  * does not exist and done all appropriate namespace locking.
1473  */
1474 static
1475 int
1476 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1477 {
1478         struct hammer_transaction trans;
1479         struct hammer_inode *dip;
1480         struct hammer_inode *nip;
1481         struct nchandle *nch;
1482         hammer_mount_t hmp;
1483         int error;
1484
1485         nch = ap->a_nch;
1486         dip = VTOI(ap->a_dvp);
1487         hmp = dip->hmp;
1488
1489         if (dip->flags & HAMMER_INODE_RO)
1490                 return (EROFS);
1491         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1492                 return (error);
1493
1494         /*
1495          * Create a transaction to cover the operations we perform.
1496          */
1497         lwkt_gettoken(&hmp->fs_token);
1498         hammer_start_transaction(&trans, hmp);
1499         ++hammer_stats_file_iopsw;
1500
1501         /*
1502          * Create a new filesystem object of the requested type.  The
1503          * returned inode will be referenced but not locked.
1504          */
1505         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1506                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1507                                     NULL, &nip);
1508         if (error) {
1509                 hkprintf("hammer_mkdir error %d\n", error);
1510                 hammer_done_transaction(&trans);
1511                 *ap->a_vpp = NULL;
1512                 lwkt_reltoken(&hmp->fs_token);
1513                 return (error);
1514         }
1515         /*
1516          * Add the new filesystem object to the directory.  This will also
1517          * bump the inode's link count.
1518          */
1519         error = hammer_ip_add_directory(&trans, dip,
1520                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1521                                         nip);
1522         if (error)
1523                 hkprintf("hammer_mkdir (add) error %d\n", error);
1524
1525         /*
1526          * Finish up.
1527          */
1528         if (error) {
1529                 hammer_rel_inode(nip, 0);
1530                 *ap->a_vpp = NULL;
1531         } else {
1532                 error = hammer_get_vnode(nip, ap->a_vpp);
1533                 hammer_rel_inode(nip, 0);
1534                 if (error == 0) {
1535                         cache_setunresolved(ap->a_nch);
1536                         cache_setvp(ap->a_nch, *ap->a_vpp);
1537                 }
1538         }
1539         hammer_done_transaction(&trans);
1540         if (error == 0)
1541                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1542         lwkt_reltoken(&hmp->fs_token);
1543         return (error);
1544 }
1545
1546 /*
1547  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1548  *
1549  * The operating system has already ensured that the directory entry
1550  * does not exist and done all appropriate namespace locking.
1551  */
1552 static
1553 int
1554 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1555 {
1556         struct hammer_transaction trans;
1557         struct hammer_inode *dip;
1558         struct hammer_inode *nip;
1559         struct nchandle *nch;
1560         hammer_mount_t hmp;
1561         int error;
1562
1563         nch = ap->a_nch;
1564         dip = VTOI(ap->a_dvp);
1565         hmp = dip->hmp;
1566
1567         if (dip->flags & HAMMER_INODE_RO)
1568                 return (EROFS);
1569         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1570                 return (error);
1571
1572         /*
1573          * Create a transaction to cover the operations we perform.
1574          */
1575         lwkt_gettoken(&hmp->fs_token);
1576         hammer_start_transaction(&trans, hmp);
1577         ++hammer_stats_file_iopsw;
1578
1579         /*
1580          * Create a new filesystem object of the requested type.  The
1581          * returned inode will be referenced but not locked.
1582          *
1583          * If mknod specifies a directory a pseudo-fs is created.
1584          */
1585         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1586                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1587                                     NULL, &nip);
1588         if (error) {
1589                 hammer_done_transaction(&trans);
1590                 *ap->a_vpp = NULL;
1591                 lwkt_reltoken(&hmp->fs_token);
1592                 return (error);
1593         }
1594
1595         /*
1596          * Add the new filesystem object to the directory.  This will also
1597          * bump the inode's link count.
1598          */
1599         error = hammer_ip_add_directory(&trans, dip,
1600                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1601                                         nip);
1602
1603         /*
1604          * Finish up.
1605          */
1606         if (error) {
1607                 hammer_rel_inode(nip, 0);
1608                 *ap->a_vpp = NULL;
1609         } else {
1610                 error = hammer_get_vnode(nip, ap->a_vpp);
1611                 hammer_rel_inode(nip, 0);
1612                 if (error == 0) {
1613                         cache_setunresolved(ap->a_nch);
1614                         cache_setvp(ap->a_nch, *ap->a_vpp);
1615                 }
1616         }
1617         hammer_done_transaction(&trans);
1618         if (error == 0)
1619                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1620         lwkt_reltoken(&hmp->fs_token);
1621         return (error);
1622 }
1623
1624 /*
1625  * hammer_vop_open { vp, mode, cred, fp }
1626  *
1627  * MPSAFE (does not require fs_token)
1628  */
1629 static
1630 int
1631 hammer_vop_open(struct vop_open_args *ap)
1632 {
1633         hammer_inode_t ip;
1634
1635         ++hammer_stats_file_iopsr;
1636         ip = VTOI(ap->a_vp);
1637
1638         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1639                 return (EROFS);
1640         return(vop_stdopen(ap));
1641 }
1642
1643 /*
1644  * hammer_vop_print { vp }
1645  */
1646 static
1647 int
1648 hammer_vop_print(struct vop_print_args *ap)
1649 {
1650         return EOPNOTSUPP;
1651 }
1652
1653 /*
1654  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1655  */
1656 static
1657 int
1658 hammer_vop_readdir(struct vop_readdir_args *ap)
1659 {
1660         struct hammer_transaction trans;
1661         struct hammer_cursor cursor;
1662         struct hammer_inode *ip;
1663         hammer_mount_t hmp;
1664         struct uio *uio;
1665         hammer_base_elm_t base;
1666         int error;
1667         int cookie_index;
1668         int ncookies;
1669         off_t *cookies;
1670         off_t saveoff;
1671         int r;
1672         int dtype;
1673
1674         ++hammer_stats_file_iopsr;
1675         ip = VTOI(ap->a_vp);
1676         uio = ap->a_uio;
1677         saveoff = uio->uio_offset;
1678         hmp = ip->hmp;
1679
1680         if (ap->a_ncookies) {
1681                 ncookies = uio->uio_resid / 16 + 1;
1682                 if (ncookies > 1024)
1683                         ncookies = 1024;
1684                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1685                 cookie_index = 0;
1686         } else {
1687                 ncookies = -1;
1688                 cookies = NULL;
1689                 cookie_index = 0;
1690         }
1691
1692         lwkt_gettoken(&hmp->fs_token);
1693         hammer_simple_transaction(&trans, hmp);
1694
1695         /*
1696          * Handle artificial entries
1697          *
1698          * It should be noted that the minimum value for a directory
1699          * hash key on-media is 0x0000000100000000, so we can use anything
1700          * less then that to represent our 'special' key space.
1701          */
1702         error = 0;
1703         if (saveoff == 0) {
1704                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1705                 if (r)
1706                         goto done;
1707                 if (cookies)
1708                         cookies[cookie_index] = saveoff;
1709                 ++saveoff;
1710                 ++cookie_index;
1711                 if (cookie_index == ncookies)
1712                         goto done;
1713         }
1714         if (saveoff == 1) {
1715                 if (ip->ino_data.parent_obj_id) {
1716                         r = vop_write_dirent(&error, uio,
1717                                              ip->ino_data.parent_obj_id,
1718                                              DT_DIR, 2, "..");
1719                 } else {
1720                         r = vop_write_dirent(&error, uio,
1721                                              ip->obj_id, DT_DIR, 2, "..");
1722                 }
1723                 if (r)
1724                         goto done;
1725                 if (cookies)
1726                         cookies[cookie_index] = saveoff;
1727                 ++saveoff;
1728                 ++cookie_index;
1729                 if (cookie_index == ncookies)
1730                         goto done;
1731         }
1732
1733         /*
1734          * Key range (begin and end inclusive) to scan.  Directory keys
1735          * directly translate to a 64 bit 'seek' position.
1736          */
1737         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1738         cursor.key_beg.localization = ip->obj_localization +
1739                                       hammer_dir_localization(ip);
1740         cursor.key_beg.obj_id = ip->obj_id;
1741         cursor.key_beg.create_tid = 0;
1742         cursor.key_beg.delete_tid = 0;
1743         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1744         cursor.key_beg.obj_type = 0;
1745         cursor.key_beg.key = saveoff;
1746
1747         cursor.key_end = cursor.key_beg;
1748         cursor.key_end.key = HAMMER_MAX_KEY;
1749         cursor.asof = ip->obj_asof;
1750         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1751
1752         error = hammer_ip_first(&cursor);
1753
1754         while (error == 0) {
1755                 error = hammer_ip_resolve_data(&cursor);
1756                 if (error)
1757                         break;
1758                 base = &cursor.leaf->base;
1759                 saveoff = base->key;
1760                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1761
1762                 if (base->obj_id != ip->obj_id)
1763                         panic("readdir: bad record at %p", cursor.node);
1764
1765                 /*
1766                  * Convert pseudo-filesystems into softlinks
1767                  */
1768                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1769                 r = vop_write_dirent(
1770                              &error, uio, cursor.data->entry.obj_id,
1771                              dtype,
1772                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1773                              (void *)cursor.data->entry.name);
1774                 if (r)
1775                         break;
1776                 ++saveoff;
1777                 if (cookies)
1778                         cookies[cookie_index] = base->key;
1779                 ++cookie_index;
1780                 if (cookie_index == ncookies)
1781                         break;
1782                 error = hammer_ip_next(&cursor);
1783         }
1784         hammer_done_cursor(&cursor);
1785
1786 done:
1787         hammer_done_transaction(&trans);
1788
1789         if (ap->a_eofflag)
1790                 *ap->a_eofflag = (error == ENOENT);
1791         uio->uio_offset = saveoff;
1792         if (error && cookie_index == 0) {
1793                 if (error == ENOENT)
1794                         error = 0;
1795                 if (cookies) {
1796                         kfree(cookies, M_TEMP);
1797                         *ap->a_ncookies = 0;
1798                         *ap->a_cookies = NULL;
1799                 }
1800         } else {
1801                 if (error == ENOENT)
1802                         error = 0;
1803                 if (cookies) {
1804                         *ap->a_ncookies = cookie_index;
1805                         *ap->a_cookies = cookies;
1806                 }
1807         }
1808         lwkt_reltoken(&hmp->fs_token);
1809         return(error);
1810 }
1811
1812 /*
1813  * hammer_vop_readlink { vp, uio, cred }
1814  */
1815 static
1816 int
1817 hammer_vop_readlink(struct vop_readlink_args *ap)
1818 {
1819         struct hammer_transaction trans;
1820         struct hammer_cursor cursor;
1821         struct hammer_inode *ip;
1822         hammer_mount_t hmp;
1823         char buf[32];
1824         u_int32_t localization;
1825         hammer_pseudofs_inmem_t pfsm;
1826         int error;
1827
1828         ip = VTOI(ap->a_vp);
1829         hmp = ip->hmp;
1830
1831         lwkt_gettoken(&hmp->fs_token);
1832
1833         /*
1834          * Shortcut if the symlink data was stuffed into ino_data.
1835          *
1836          * Also expand special "@@PFS%05d" softlinks (expansion only
1837          * occurs for non-historical (current) accesses made from the
1838          * primary filesystem).
1839          */
1840         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1841                 char *ptr;
1842                 int bytes;
1843
1844                 ptr = ip->ino_data.ext.symlink;
1845                 bytes = (int)ip->ino_data.size;
1846                 if (bytes == 10 &&
1847                     ip->obj_asof == HAMMER_MAX_TID &&
1848                     ip->obj_localization == 0 &&
1849                     strncmp(ptr, "@@PFS", 5) == 0) {
1850                         hammer_simple_transaction(&trans, hmp);
1851                         bcopy(ptr + 5, buf, 5);
1852                         buf[5] = 0;
1853                         localization = strtoul(buf, NULL, 10) << 16;
1854                         pfsm = hammer_load_pseudofs(&trans, localization,
1855                                                     &error);
1856                         if (error == 0) {
1857                                 if (pfsm->pfsd.mirror_flags &
1858                                     HAMMER_PFSD_SLAVE) {
1859                                         /* vap->va_size == 26 */
1860                                         ksnprintf(buf, sizeof(buf),
1861                                                   "@@0x%016llx:%05d",
1862                                                   (long long)pfsm->pfsd.sync_end_tid,
1863                                                   localization >> 16);
1864                                 } else {
1865                                         /* vap->va_size == 10 */
1866                                         ksnprintf(buf, sizeof(buf),
1867                                                   "@@-1:%05d",
1868                                                   localization >> 16);
1869 #if 0
1870                                         ksnprintf(buf, sizeof(buf),
1871                                                   "@@0x%016llx:%05d",
1872                                                   (long long)HAMMER_MAX_TID,
1873                                                   localization >> 16);
1874 #endif
1875                                 }
1876                                 ptr = buf;
1877                                 bytes = strlen(buf);
1878                         }
1879                         if (pfsm)
1880                                 hammer_rel_pseudofs(hmp, pfsm);
1881                         hammer_done_transaction(&trans);
1882                 }
1883                 error = uiomove(ptr, bytes, ap->a_uio);
1884                 lwkt_reltoken(&hmp->fs_token);
1885                 return(error);
1886         }
1887
1888         /*
1889          * Long version
1890          */
1891         hammer_simple_transaction(&trans, hmp);
1892         ++hammer_stats_file_iopsr;
1893         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1894
1895         /*
1896          * Key range (begin and end inclusive) to scan.  Directory keys
1897          * directly translate to a 64 bit 'seek' position.
1898          */
1899         cursor.key_beg.localization = ip->obj_localization +
1900                                       HAMMER_LOCALIZE_MISC;
1901         cursor.key_beg.obj_id = ip->obj_id;
1902         cursor.key_beg.create_tid = 0;
1903         cursor.key_beg.delete_tid = 0;
1904         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1905         cursor.key_beg.obj_type = 0;
1906         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1907         cursor.asof = ip->obj_asof;
1908         cursor.flags |= HAMMER_CURSOR_ASOF;
1909
1910         error = hammer_ip_lookup(&cursor);
1911         if (error == 0) {
1912                 error = hammer_ip_resolve_data(&cursor);
1913                 if (error == 0) {
1914                         KKASSERT(cursor.leaf->data_len >=
1915                                  HAMMER_SYMLINK_NAME_OFF);
1916                         error = uiomove(cursor.data->symlink.name,
1917                                         cursor.leaf->data_len -
1918                                                 HAMMER_SYMLINK_NAME_OFF,
1919                                         ap->a_uio);
1920                 }
1921         }
1922         hammer_done_cursor(&cursor);
1923         hammer_done_transaction(&trans);
1924         lwkt_reltoken(&hmp->fs_token);
1925         return(error);
1926 }
1927
1928 /*
1929  * hammer_vop_nremove { nch, dvp, cred }
1930  */
1931 static
1932 int
1933 hammer_vop_nremove(struct vop_nremove_args *ap)
1934 {
1935         struct hammer_transaction trans;
1936         struct hammer_inode *dip;
1937         hammer_mount_t hmp;
1938         int error;
1939
1940         dip = VTOI(ap->a_dvp);
1941         hmp = dip->hmp;
1942
1943         if (hammer_nohistory(dip) == 0 &&
1944             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1945                 return (error);
1946         }
1947
1948         lwkt_gettoken(&hmp->fs_token);
1949         hammer_start_transaction(&trans, hmp);
1950         ++hammer_stats_file_iopsw;
1951         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1952         hammer_done_transaction(&trans);
1953         if (error == 0)
1954                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1955         lwkt_reltoken(&hmp->fs_token);
1956         return (error);
1957 }
1958
1959 /*
1960  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1961  */
1962 static
1963 int
1964 hammer_vop_nrename(struct vop_nrename_args *ap)
1965 {
1966         struct hammer_transaction trans;
1967         struct namecache *fncp;
1968         struct namecache *tncp;
1969         struct hammer_inode *fdip;
1970         struct hammer_inode *tdip;
1971         struct hammer_inode *ip;
1972         hammer_mount_t hmp;
1973         struct hammer_cursor cursor;
1974         int64_t namekey;
1975         u_int32_t max_iterations;
1976         int nlen, error;
1977
1978         if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount) 
1979                 return(EXDEV);
1980         if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1981                 return(EXDEV);
1982
1983         fdip = VTOI(ap->a_fdvp);
1984         tdip = VTOI(ap->a_tdvp);
1985         fncp = ap->a_fnch->ncp;
1986         tncp = ap->a_tnch->ncp;
1987         ip = VTOI(fncp->nc_vp);
1988         KKASSERT(ip != NULL);
1989
1990         hmp = ip->hmp;
1991
1992         if (fdip->obj_localization != tdip->obj_localization)
1993                 return(EXDEV);
1994         if (fdip->obj_localization != ip->obj_localization)
1995                 return(EXDEV);
1996
1997         if (fdip->flags & HAMMER_INODE_RO)
1998                 return (EROFS);
1999         if (tdip->flags & HAMMER_INODE_RO)
2000                 return (EROFS);
2001         if (ip->flags & HAMMER_INODE_RO)
2002                 return (EROFS);
2003         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
2004                 return (error);
2005
2006         lwkt_gettoken(&hmp->fs_token);
2007         hammer_start_transaction(&trans, hmp);
2008         ++hammer_stats_file_iopsw;
2009
2010         /*
2011          * Remove tncp from the target directory and then link ip as
2012          * tncp. XXX pass trans to dounlink
2013          *
2014          * Force the inode sync-time to match the transaction so it is
2015          * in-sync with the creation of the target directory entry.
2016          */
2017         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
2018                                 ap->a_cred, 0, -1);
2019         if (error == 0 || error == ENOENT) {
2020                 error = hammer_ip_add_directory(&trans, tdip,
2021                                                 tncp->nc_name, tncp->nc_nlen,
2022                                                 ip);
2023                 if (error == 0) {
2024                         ip->ino_data.parent_obj_id = tdip->obj_id;
2025                         ip->ino_data.ctime = trans.time;
2026                         hammer_modify_inode(&trans, ip, HAMMER_INODE_DDIRTY);
2027                 }
2028         }
2029         if (error)
2030                 goto failed; /* XXX */
2031
2032         /*
2033          * Locate the record in the originating directory and remove it.
2034          *
2035          * Calculate the namekey and setup the key range for the scan.  This
2036          * works kinda like a chained hash table where the lower 32 bits
2037          * of the namekey synthesize the chain.
2038          *
2039          * The key range is inclusive of both key_beg and key_end.
2040          */
2041         namekey = hammer_directory_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
2042                                            &max_iterations);
2043 retry:
2044         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
2045         cursor.key_beg.localization = fdip->obj_localization +
2046                                       hammer_dir_localization(fdip);
2047         cursor.key_beg.obj_id = fdip->obj_id;
2048         cursor.key_beg.key = namekey;
2049         cursor.key_beg.create_tid = 0;
2050         cursor.key_beg.delete_tid = 0;
2051         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2052         cursor.key_beg.obj_type = 0;
2053
2054         cursor.key_end = cursor.key_beg;
2055         cursor.key_end.key += max_iterations;
2056         cursor.asof = fdip->obj_asof;
2057         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2058
2059         /*
2060          * Scan all matching records (the chain), locate the one matching
2061          * the requested path component.
2062          *
2063          * The hammer_ip_*() functions merge in-memory records with on-disk
2064          * records for the purposes of the search.
2065          */
2066         error = hammer_ip_first(&cursor);
2067         while (error == 0) {
2068                 if (hammer_ip_resolve_data(&cursor) != 0)
2069                         break;
2070                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2071                 KKASSERT(nlen > 0);
2072                 if (fncp->nc_nlen == nlen &&
2073                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2074                         break;
2075                 }
2076                 error = hammer_ip_next(&cursor);
2077         }
2078
2079         /*
2080          * If all is ok we have to get the inode so we can adjust nlinks.
2081          *
2082          * WARNING: hammer_ip_del_directory() may have to terminate the
2083          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
2084          * twice.
2085          */
2086         if (error == 0)
2087                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
2088
2089         /*
2090          * XXX A deadlock here will break rename's atomicy for the purposes
2091          * of crash recovery.
2092          */
2093         if (error == EDEADLK) {
2094                 hammer_done_cursor(&cursor);
2095                 goto retry;
2096         }
2097
2098         /*
2099          * Cleanup and tell the kernel that the rename succeeded.
2100          *
2101          * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2102          *       without formally acquiring the vp since the vp might
2103          *       have zero refs on it, or in the middle of a reclaim,
2104          *       etc.
2105          */
2106         hammer_done_cursor(&cursor);
2107         if (error == 0) {
2108                 cache_rename(ap->a_fnch, ap->a_tnch);
2109                 hammer_knote(ap->a_fdvp, NOTE_WRITE);
2110                 hammer_knote(ap->a_tdvp, NOTE_WRITE);
2111                 while (ip->vp) {
2112                         struct vnode *vp;
2113
2114                         error = hammer_get_vnode(ip, &vp);
2115                         if (error == 0 && vp) {
2116                                 vn_unlock(vp);
2117                                 hammer_knote(ip->vp, NOTE_RENAME);
2118                                 vrele(vp);
2119                                 break;
2120                         }
2121                         kprintf("Debug: HAMMER ip/vp race2 avoided\n");
2122                 }
2123         }
2124
2125 failed:
2126         hammer_done_transaction(&trans);
2127         lwkt_reltoken(&hmp->fs_token);
2128         return (error);
2129 }
2130
2131 /*
2132  * hammer_vop_nrmdir { nch, dvp, cred }
2133  */
2134 static
2135 int
2136 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
2137 {
2138         struct hammer_transaction trans;
2139         struct hammer_inode *dip;
2140         hammer_mount_t hmp;
2141         int error;
2142
2143         dip = VTOI(ap->a_dvp);
2144         hmp = dip->hmp;
2145
2146         if (hammer_nohistory(dip) == 0 &&
2147             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2148                 return (error);
2149         }
2150
2151         lwkt_gettoken(&hmp->fs_token);
2152         hammer_start_transaction(&trans, hmp);
2153         ++hammer_stats_file_iopsw;
2154         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
2155         hammer_done_transaction(&trans);
2156         if (error == 0)
2157                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
2158         lwkt_reltoken(&hmp->fs_token);
2159         return (error);
2160 }
2161
2162 /*
2163  * hammer_vop_markatime { vp, cred }
2164  */
2165 static
2166 int
2167 hammer_vop_markatime(struct vop_markatime_args *ap)
2168 {
2169         struct hammer_transaction trans;
2170         struct hammer_inode *ip;
2171         hammer_mount_t hmp;
2172
2173         ip = VTOI(ap->a_vp);
2174         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2175                 return (EROFS);
2176         if (ip->flags & HAMMER_INODE_RO)
2177                 return (EROFS);
2178         hmp = ip->hmp;
2179         if (hmp->mp->mnt_flag & MNT_NOATIME)
2180                 return (0);
2181         lwkt_gettoken(&hmp->fs_token);
2182         hammer_start_transaction(&trans, hmp);
2183         ++hammer_stats_file_iopsw;
2184
2185         ip->ino_data.atime = trans.time;
2186         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
2187         hammer_done_transaction(&trans);
2188         hammer_knote(ap->a_vp, NOTE_ATTRIB);
2189         lwkt_reltoken(&hmp->fs_token);
2190         return (0);
2191 }
2192
2193 /*
2194  * hammer_vop_setattr { vp, vap, cred }
2195  */
2196 static
2197 int
2198 hammer_vop_setattr(struct vop_setattr_args *ap)
2199 {
2200         struct hammer_transaction trans;
2201         struct hammer_inode *ip;
2202         struct vattr *vap;
2203         hammer_mount_t hmp;
2204         int modflags;
2205         int error;
2206         int truncating;
2207         int blksize;
2208         int kflags;
2209 #if 0
2210         int64_t aligned_size;
2211 #endif
2212         u_int32_t flags;
2213
2214         vap = ap->a_vap;
2215         ip = ap->a_vp->v_data;
2216         modflags = 0;
2217         kflags = 0;
2218         hmp = ip->hmp;
2219
2220         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2221                 return(EROFS);
2222         if (ip->flags & HAMMER_INODE_RO)
2223                 return (EROFS);
2224         if (hammer_nohistory(ip) == 0 &&
2225             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2226                 return (error);
2227         }
2228
2229         lwkt_gettoken(&hmp->fs_token);
2230         hammer_start_transaction(&trans, hmp);
2231         ++hammer_stats_file_iopsw;
2232         error = 0;
2233
2234         if (vap->va_flags != VNOVAL) {
2235                 flags = ip->ino_data.uflags;
2236                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
2237                                          hammer_to_unix_xid(&ip->ino_data.uid),
2238                                          ap->a_cred);
2239                 if (error == 0) {
2240                         if (ip->ino_data.uflags != flags) {
2241                                 ip->ino_data.uflags = flags;
2242                                 ip->ino_data.ctime = trans.time;
2243                                 modflags |= HAMMER_INODE_DDIRTY;
2244                                 kflags |= NOTE_ATTRIB;
2245                         }
2246                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2247                                 error = 0;
2248                                 goto done;
2249                         }
2250                 }
2251                 goto done;
2252         }
2253         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2254                 error = EPERM;
2255                 goto done;
2256         }
2257         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2258                 mode_t cur_mode = ip->ino_data.mode;
2259                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2260                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2261                 uuid_t uuid_uid;
2262                 uuid_t uuid_gid;
2263
2264                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2265                                          ap->a_cred,
2266                                          &cur_uid, &cur_gid, &cur_mode);
2267                 if (error == 0) {
2268                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
2269                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
2270                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
2271                                  sizeof(uuid_uid)) ||
2272                             bcmp(&uuid_gid, &ip->ino_data.gid,
2273                                  sizeof(uuid_gid)) ||
2274                             ip->ino_data.mode != cur_mode
2275                         ) {
2276                                 ip->ino_data.uid = uuid_uid;
2277                                 ip->ino_data.gid = uuid_gid;
2278                                 ip->ino_data.mode = cur_mode;
2279                                 ip->ino_data.ctime = trans.time;
2280                                 modflags |= HAMMER_INODE_DDIRTY;
2281                         }
2282                         kflags |= NOTE_ATTRIB;
2283                 }
2284         }
2285         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2286                 switch(ap->a_vp->v_type) {
2287                 case VREG:
2288                         if (vap->va_size == ip->ino_data.size)
2289                                 break;
2290
2291                         /*
2292                          * Log the operation if in fast-fsync mode or if
2293                          * there are unterminated redo write records present.
2294                          *
2295                          * The second check is needed so the recovery code
2296                          * properly truncates write redos even if nominal
2297                          * REDO operations is turned off due to excessive
2298                          * writes, because the related records might be
2299                          * destroyed and never lay down a TERM_WRITE.
2300                          */
2301                         if ((ip->flags & HAMMER_INODE_REDO) ||
2302                             (ip->flags & HAMMER_INODE_RDIRTY)) {
2303                                 error = hammer_generate_redo(&trans, ip,
2304                                                              vap->va_size,
2305                                                              HAMMER_REDO_TRUNC,
2306                                                              NULL, 0);
2307                         }
2308                         blksize = hammer_blocksize(vap->va_size);
2309
2310                         /*
2311                          * XXX break atomicy, we can deadlock the backend
2312                          * if we do not release the lock.  Probably not a
2313                          * big deal here.
2314                          */
2315                         if (vap->va_size < ip->ino_data.size) {
2316                                 nvtruncbuf(ap->a_vp, vap->va_size,
2317                                            blksize,
2318                                            hammer_blockoff(vap->va_size),
2319                                            0);
2320                                 truncating = 1;
2321                                 kflags |= NOTE_WRITE;
2322                         } else {
2323                                 nvextendbuf(ap->a_vp,
2324                                             ip->ino_data.size,
2325                                             vap->va_size,
2326                                             hammer_blocksize(ip->ino_data.size),
2327                                             hammer_blocksize(vap->va_size),
2328                                             hammer_blockoff(ip->ino_data.size),
2329                                             hammer_blockoff(vap->va_size),
2330                                             0);
2331                                 truncating = 0;
2332                                 kflags |= NOTE_WRITE | NOTE_EXTEND;
2333                         }
2334                         ip->ino_data.size = vap->va_size;
2335                         ip->ino_data.mtime = trans.time;
2336                         /* XXX safe to use SDIRTY instead of DDIRTY here? */
2337                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2338
2339                         /*
2340                          * On-media truncation is cached in the inode until
2341                          * the inode is synchronized.  We must immediately
2342                          * handle any frontend records.
2343                          */
2344                         if (truncating) {
2345                                 hammer_ip_frontend_trunc(ip, vap->va_size);
2346 #ifdef DEBUG_TRUNCATE
2347                                 if (HammerTruncIp == NULL)
2348                                         HammerTruncIp = ip;
2349 #endif
2350                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2351                                         ip->flags |= HAMMER_INODE_TRUNCATED;
2352                                         ip->trunc_off = vap->va_size;
2353 #ifdef DEBUG_TRUNCATE
2354                                         if (ip == HammerTruncIp)
2355                                         kprintf("truncate1 %016llx\n",
2356                                                 (long long)ip->trunc_off);
2357 #endif
2358                                 } else if (ip->trunc_off > vap->va_size) {
2359                                         ip->trunc_off = vap->va_size;
2360 #ifdef DEBUG_TRUNCATE
2361                                         if (ip == HammerTruncIp)
2362                                         kprintf("truncate2 %016llx\n",
2363                                                 (long long)ip->trunc_off);
2364 #endif
2365                                 } else {
2366 #ifdef DEBUG_TRUNCATE
2367                                         if (ip == HammerTruncIp)
2368                                         kprintf("truncate3 %016llx (ignored)\n",
2369                                                 (long long)vap->va_size);
2370 #endif
2371                                 }
2372                         }
2373
2374 #if 0
2375                         /*
2376                          * When truncating, nvtruncbuf() may have cleaned out
2377                          * a portion of the last block on-disk in the buffer
2378                          * cache.  We must clean out any frontend records
2379                          * for blocks beyond the new last block.
2380                          */
2381                         aligned_size = (vap->va_size + (blksize - 1)) &
2382                                        ~(int64_t)(blksize - 1);
2383                         if (truncating && vap->va_size < aligned_size) {
2384                                 aligned_size -= blksize;
2385                                 hammer_ip_frontend_trunc(ip, aligned_size);
2386                         }
2387 #endif
2388                         break;
2389                 case VDATABASE:
2390                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2391                                 ip->flags |= HAMMER_INODE_TRUNCATED;
2392                                 ip->trunc_off = vap->va_size;
2393                         } else if (ip->trunc_off > vap->va_size) {
2394                                 ip->trunc_off = vap->va_size;
2395                         }
2396                         hammer_ip_frontend_trunc(ip, vap->va_size);
2397                         ip->ino_data.size = vap->va_size;
2398                         ip->ino_data.mtime = trans.time;
2399                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2400                         kflags |= NOTE_ATTRIB;
2401                         break;
2402                 default:
2403                         error = EINVAL;
2404                         goto done;
2405                 }
2406                 break;
2407         }
2408         if (vap->va_atime.tv_sec != VNOVAL) {
2409                 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2410                 modflags |= HAMMER_INODE_ATIME;
2411                 kflags |= NOTE_ATTRIB;
2412         }
2413         if (vap->va_mtime.tv_sec != VNOVAL) {
2414                 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2415                 modflags |= HAMMER_INODE_MTIME;
2416                 kflags |= NOTE_ATTRIB;
2417         }
2418         if (vap->va_mode != (mode_t)VNOVAL) {
2419                 mode_t   cur_mode = ip->ino_data.mode;
2420                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2421                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2422
2423                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2424                                          cur_uid, cur_gid, &cur_mode);
2425                 if (error == 0 && ip->ino_data.mode != cur_mode) {
2426                         ip->ino_data.mode = cur_mode;
2427                         ip->ino_data.ctime = trans.time;
2428                         modflags |= HAMMER_INODE_DDIRTY;
2429                         kflags |= NOTE_ATTRIB;
2430                 }
2431         }
2432 done:
2433         if (error == 0)
2434                 hammer_modify_inode(&trans, ip, modflags);
2435         hammer_done_transaction(&trans);
2436         hammer_knote(ap->a_vp, kflags);
2437         lwkt_reltoken(&hmp->fs_token);
2438         return (error);
2439 }
2440
2441 /*
2442  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2443  */
2444 static
2445 int
2446 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2447 {
2448         struct hammer_transaction trans;
2449         struct hammer_inode *dip;
2450         struct hammer_inode *nip;
2451         hammer_record_t record;
2452         struct nchandle *nch;
2453         hammer_mount_t hmp;
2454         int error;
2455         int bytes;
2456
2457         ap->a_vap->va_type = VLNK;
2458
2459         nch = ap->a_nch;
2460         dip = VTOI(ap->a_dvp);
2461         hmp = dip->hmp;
2462
2463         if (dip->flags & HAMMER_INODE_RO)
2464                 return (EROFS);
2465         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
2466                 return (error);
2467
2468         /*
2469          * Create a transaction to cover the operations we perform.
2470          */
2471         lwkt_gettoken(&hmp->fs_token);
2472         hammer_start_transaction(&trans, hmp);
2473         ++hammer_stats_file_iopsw;
2474
2475         /*
2476          * Create a new filesystem object of the requested type.  The
2477          * returned inode will be referenced but not locked.
2478          */
2479
2480         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2481                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2482                                     NULL, &nip);
2483         if (error) {
2484                 hammer_done_transaction(&trans);
2485                 *ap->a_vpp = NULL;
2486                 lwkt_reltoken(&hmp->fs_token);
2487                 return (error);
2488         }
2489
2490         /*
2491          * Add a record representing the symlink.  symlink stores the link
2492          * as pure data, not a string, and is no \0 terminated.
2493          */
2494         if (error == 0) {
2495                 bytes = strlen(ap->a_target);
2496
2497                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2498                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2499                 } else {
2500                         record = hammer_alloc_mem_record(nip, bytes);
2501                         record->type = HAMMER_MEM_RECORD_GENERAL;
2502
2503                         record->leaf.base.localization = nip->obj_localization +
2504                                                          HAMMER_LOCALIZE_MISC;
2505                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2506                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2507                         record->leaf.data_len = bytes;
2508                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2509                         bcopy(ap->a_target, record->data->symlink.name, bytes);
2510                         error = hammer_ip_add_record(&trans, record);
2511                 }
2512
2513                 /*
2514                  * Set the file size to the length of the link.
2515                  */
2516                 if (error == 0) {
2517                         nip->ino_data.size = bytes;
2518                         hammer_modify_inode(&trans, nip, HAMMER_INODE_DDIRTY);
2519                 }
2520         }
2521         if (error == 0)
2522                 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
2523                                                 nch->ncp->nc_nlen, nip);
2524
2525         /*
2526          * Finish up.
2527          */
2528         if (error) {
2529                 hammer_rel_inode(nip, 0);
2530                 *ap->a_vpp = NULL;
2531         } else {
2532                 error = hammer_get_vnode(nip, ap->a_vpp);
2533                 hammer_rel_inode(nip, 0);
2534                 if (error == 0) {
2535                         cache_setunresolved(ap->a_nch);
2536                         cache_setvp(ap->a_nch, *ap->a_vpp);
2537                         hammer_knote(ap->a_dvp, NOTE_WRITE);
2538                 }
2539         }
2540         hammer_done_transaction(&trans);
2541         lwkt_reltoken(&hmp->fs_token);
2542         return (error);
2543 }
2544
2545 /*
2546  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2547  */
2548 static
2549 int
2550 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2551 {
2552         struct hammer_transaction trans;
2553         struct hammer_inode *dip;
2554         hammer_mount_t hmp;
2555         int error;
2556
2557         dip = VTOI(ap->a_dvp);
2558         hmp = dip->hmp;
2559
2560         if (hammer_nohistory(dip) == 0 &&
2561             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2562                 return (error);
2563         }
2564
2565         lwkt_gettoken(&hmp->fs_token);
2566         hammer_start_transaction(&trans, hmp);
2567         ++hammer_stats_file_iopsw;
2568         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2569                                 ap->a_cred, ap->a_flags, -1);
2570         hammer_done_transaction(&trans);
2571         lwkt_reltoken(&hmp->fs_token);
2572
2573         return (error);
2574 }
2575
2576 /*
2577  * hammer_vop_ioctl { vp, command, data, fflag, cred }
2578  */
2579 static
2580 int
2581 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2582 {
2583         struct hammer_inode *ip = ap->a_vp->v_data;
2584         hammer_mount_t hmp = ip->hmp;
2585         int error;
2586
2587         ++hammer_stats_file_iopsr;
2588         lwkt_gettoken(&hmp->fs_token);
2589         error = hammer_ioctl(ip, ap->a_command, ap->a_data,
2590                              ap->a_fflag, ap->a_cred);
2591         lwkt_reltoken(&hmp->fs_token);
2592         return (error);
2593 }
2594
2595 static
2596 int
2597 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2598 {
2599         static const struct mountctl_opt extraopt[] = {
2600                 { HMNT_NOHISTORY,       "nohistory" },
2601                 { HMNT_MASTERID,        "master" },
2602                 { 0, NULL}
2603
2604         };
2605         struct hammer_mount *hmp;
2606         struct mount *mp;
2607         int usedbytes;
2608         int error;
2609
2610         error = 0;
2611         usedbytes = 0;
2612         mp = ap->a_head.a_ops->head.vv_mount;
2613         KKASSERT(mp->mnt_data != NULL);
2614         hmp = (struct hammer_mount *)mp->mnt_data;
2615
2616         lwkt_gettoken(&hmp->fs_token);
2617
2618         switch(ap->a_op) {
2619         case MOUNTCTL_SET_EXPORT:
2620                 if (ap->a_ctllen != sizeof(struct export_args))
2621                         error = EINVAL;
2622                 else
2623                         error = hammer_vfs_export(mp, ap->a_op,
2624                                       (const struct export_args *)ap->a_ctl);
2625                 break;
2626         case MOUNTCTL_MOUNTFLAGS:
2627         {
2628                 /*
2629                  * Call standard mountctl VOP function
2630                  * so we get user mount flags.
2631                  */
2632                 error = vop_stdmountctl(ap);
2633                 if (error)
2634                         break;
2635
2636                 usedbytes = *ap->a_res;
2637
2638                 if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2639                         usedbytes += vfs_flagstostr(hmp->hflags, extraopt,
2640                                                     ap->a_buf,
2641                                                     ap->a_buflen - usedbytes,
2642                                                     &error);
2643                 }
2644
2645                 *ap->a_res += usedbytes;
2646                 break;
2647         }
2648         default:
2649                 error = vop_stdmountctl(ap);
2650                 break;
2651         }
2652         lwkt_reltoken(&hmp->fs_token);
2653         return(error);
2654 }
2655
2656 /*
2657  * hammer_vop_strategy { vp, bio }
2658  *
2659  * Strategy call, used for regular file read & write only.  Note that the
2660  * bp may represent a cluster.
2661  *
2662  * To simplify operation and allow better optimizations in the future,
2663  * this code does not make any assumptions with regards to buffer alignment
2664  * or size.
2665  */
2666 static
2667 int
2668 hammer_vop_strategy(struct vop_strategy_args *ap)
2669 {
2670         struct buf *bp;
2671         int error;
2672
2673         bp = ap->a_bio->bio_buf;
2674
2675         switch(bp->b_cmd) {
2676         case BUF_CMD_READ:
2677                 error = hammer_vop_strategy_read(ap);
2678                 break;
2679         case BUF_CMD_WRITE:
2680                 error = hammer_vop_strategy_write(ap);
2681                 break;
2682         default:
2683                 bp->b_error = error = EINVAL;
2684                 bp->b_flags |= B_ERROR;
2685                 biodone(ap->a_bio);
2686                 break;
2687         }
2688
2689         /* hammer_dump_dedup_cache(((hammer_inode_t)ap->a_vp->v_data)->hmp); */
2690
2691         return (error);
2692 }
2693
2694 /*
2695  * Read from a regular file.  Iterate the related records and fill in the
2696  * BIO/BUF.  Gaps are zero-filled.
2697  *
2698  * The support code in hammer_object.c should be used to deal with mixed
2699  * in-memory and on-disk records.
2700  *
2701  * NOTE: Can be called from the cluster code with an oversized buf.
2702  *
2703  * XXX atime update
2704  */
2705 static
2706 int
2707 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2708 {
2709         struct hammer_transaction trans;
2710         struct hammer_inode *ip;
2711         struct hammer_inode *dip;
2712         hammer_mount_t hmp;
2713         struct hammer_cursor cursor;
2714         hammer_base_elm_t base;
2715         hammer_off_t disk_offset;
2716         struct bio *bio;
2717         struct bio *nbio;
2718         struct buf *bp;
2719         int64_t rec_offset;
2720         int64_t ran_end;
2721         int64_t tmp64;
2722         int error;
2723         int boff;
2724         int roff;
2725         int n;
2726         int isdedupable;
2727
2728         bio = ap->a_bio;
2729         bp = bio->bio_buf;
2730         ip = ap->a_vp->v_data;
2731         hmp = ip->hmp;
2732
2733         /*
2734          * The zone-2 disk offset may have been set by the cluster code via
2735          * a BMAP operation, or else should be NOOFFSET.
2736          *
2737          * Checking the high bits for a match against zone-2 should suffice.
2738          *
2739          * In cases where a lot of data duplication is present it may be
2740          * more beneficial to drop through and doubule-buffer through the
2741          * device.
2742          */
2743         nbio = push_bio(bio);
2744         if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2745             HAMMER_ZONE_LARGE_DATA) {
2746                 if (hammer_double_buffer == 0) {
2747                         lwkt_gettoken(&hmp->fs_token);
2748                         error = hammer_io_direct_read(hmp, nbio, NULL);
2749                         lwkt_reltoken(&hmp->fs_token);
2750                         return (error);
2751                 }
2752
2753                 /*
2754                  * Try to shortcut requests for double_buffer mode too.
2755                  * Since this mode runs through the device buffer cache
2756                  * only compatible buffer sizes (meaning those generated
2757                  * by normal filesystem buffers) are legal.
2758                  */
2759                 if (hammer_live_dedup == 0 && (bp->b_flags & B_PAGING) == 0) {
2760                         error = hammer_io_indirect_read(hmp, nbio, NULL);
2761                         return (error);
2762                 }
2763         }
2764
2765         /*
2766          * Well, that sucked.  Do it the hard way.  If all the stars are
2767          * aligned we may still be able to issue a direct-read.
2768          */
2769         lwkt_gettoken(&hmp->fs_token);
2770         hammer_simple_transaction(&trans, hmp);
2771         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2772
2773         /*
2774          * Key range (begin and end inclusive) to scan.  Note that the key's
2775          * stored in the actual records represent BASE+LEN, not BASE.  The
2776          * first record containing bio_offset will have a key > bio_offset.
2777          */
2778         cursor.key_beg.localization = ip->obj_localization +
2779                                       HAMMER_LOCALIZE_MISC;
2780         cursor.key_beg.obj_id = ip->obj_id;
2781         cursor.key_beg.create_tid = 0;
2782         cursor.key_beg.delete_tid = 0;
2783         cursor.key_beg.obj_type = 0;
2784         cursor.key_beg.key = bio->bio_offset + 1;
2785         cursor.asof = ip->obj_asof;
2786         cursor.flags |= HAMMER_CURSOR_ASOF;
2787
2788         cursor.key_end = cursor.key_beg;
2789         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2790 #if 0
2791         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2792                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2793                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2794                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2795         } else
2796 #endif
2797         {
2798                 ran_end = bio->bio_offset + bp->b_bufsize;
2799                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2800                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2801                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2802                 if (tmp64 < ran_end)
2803                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2804                 else
2805                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2806         }
2807         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2808
2809         /*
2810          * Set NOSWAPCACHE for cursor data extraction if double buffering
2811          * is disabled or (if the file is not marked cacheable via chflags
2812          * and vm.swapcache_use_chflags is enabled).
2813          */
2814         if (hammer_double_buffer == 0 ||
2815             ((ap->a_vp->v_flag & VSWAPCACHE) == 0 &&
2816              vm_swapcache_use_chflags)) {
2817                 cursor.flags |= HAMMER_CURSOR_NOSWAPCACHE;
2818         }
2819
2820         error = hammer_ip_first(&cursor);
2821         boff = 0;
2822
2823         while (error == 0) {
2824                 /*
2825                  * Get the base file offset of the record.  The key for
2826                  * data records is (base + bytes) rather then (base).
2827                  */
2828                 base = &cursor.leaf->base;
2829                 rec_offset = base->key - cursor.leaf->data_len;
2830
2831                 /*
2832                  * Calculate the gap, if any, and zero-fill it.
2833                  *
2834                  * n is the offset of the start of the record verses our
2835                  * current seek offset in the bio.
2836                  */
2837                 n = (int)(rec_offset - (bio->bio_offset + boff));
2838                 if (n > 0) {
2839                         if (n > bp->b_bufsize - boff)
2840                                 n = bp->b_bufsize - boff;
2841                         bzero((char *)bp->b_data + boff, n);
2842                         boff += n;
2843                         n = 0;
2844                 }
2845
2846                 /*
2847                  * Calculate the data offset in the record and the number
2848                  * of bytes we can copy.
2849                  *
2850                  * There are two degenerate cases.  First, boff may already
2851                  * be at bp->b_bufsize.  Secondly, the data offset within
2852                  * the record may exceed the record's size.
2853                  */
2854                 roff = -n;
2855                 rec_offset += roff;
2856                 n = cursor.leaf->data_len - roff;
2857                 if (n <= 0) {
2858                         kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2859                         n = 0;
2860                 } else if (n > bp->b_bufsize - boff) {
2861                         n = bp->b_bufsize - boff;
2862                 }
2863
2864                 /*
2865                  * Deal with cached truncations.  This cool bit of code
2866                  * allows truncate()/ftruncate() to avoid having to sync
2867                  * the file.
2868                  *
2869                  * If the frontend is truncated then all backend records are
2870                  * subject to the frontend's truncation.
2871                  *
2872                  * If the backend is truncated then backend records on-disk
2873                  * (but not in-memory) are subject to the backend's
2874                  * truncation.  In-memory records owned by the backend
2875                  * represent data written after the truncation point on the
2876                  * backend and must not be truncated.
2877                  *
2878                  * Truncate operations deal with frontend buffer cache
2879                  * buffers and frontend-owned in-memory records synchronously.
2880                  */
2881                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2882                         if (hammer_cursor_ondisk(&cursor)/* ||
2883                             cursor.iprec->flush_state == HAMMER_FST_FLUSH*/) {
2884                                 if (ip->trunc_off <= rec_offset)
2885                                         n = 0;
2886                                 else if (ip->trunc_off < rec_offset + n)
2887                                         n = (int)(ip->trunc_off - rec_offset);
2888                         }
2889                 }
2890                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2891                         if (hammer_cursor_ondisk(&cursor)) {
2892                                 if (ip->sync_trunc_off <= rec_offset)
2893                                         n = 0;
2894                                 else if (ip->sync_trunc_off < rec_offset + n)
2895                                         n = (int)(ip->sync_trunc_off - rec_offset);
2896                         }
2897                 }
2898
2899                 /*
2900                  * Try to issue a direct read into our bio if possible,
2901                  * otherwise resolve the element data into a hammer_buffer
2902                  * and copy.
2903                  *
2904                  * The buffer on-disk should be zerod past any real
2905                  * truncation point, but may not be for any synthesized
2906                  * truncation point from above.
2907                  *
2908                  * NOTE: disk_offset is only valid if the cursor data is
2909                  *       on-disk.
2910                  */
2911                 disk_offset = cursor.leaf->data_offset + roff;
2912                 isdedupable = (boff == 0 && n == bp->b_bufsize &&
2913                                hammer_cursor_ondisk(&cursor) &&
2914                                ((int)disk_offset & HAMMER_BUFMASK) == 0);
2915
2916                 if (isdedupable && hammer_double_buffer == 0) {
2917                         /*
2918                          * Direct read case
2919                          */
2920                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2921                                  HAMMER_ZONE_LARGE_DATA);
2922                         nbio->bio_offset = disk_offset;
2923                         error = hammer_io_direct_read(hmp, nbio, cursor.leaf);
2924                         if (hammer_live_dedup && error == 0)
2925                                 hammer_dedup_cache_add(ip, cursor.leaf);
2926                         goto done;
2927                 } else if (isdedupable) {
2928                         /*
2929                          * Async I/O case for reading from backing store
2930                          * and copying the data to the filesystem buffer.
2931                          * live-dedup has to verify the data anyway if it
2932                          * gets a hit later so we can just add the entry
2933                          * now.
2934                          */
2935                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2936                                  HAMMER_ZONE_LARGE_DATA);
2937                         nbio->bio_offset = disk_offset;
2938                         if (hammer_live_dedup)
2939                                 hammer_dedup_cache_add(ip, cursor.leaf);
2940                         error = hammer_io_indirect_read(hmp, nbio, cursor.leaf);
2941                         goto done;
2942                 } else if (n) {
2943                         error = hammer_ip_resolve_data(&cursor);
2944                         if (error == 0) {
2945                                 if (hammer_live_dedup && isdedupable)
2946                                         hammer_dedup_cache_add(ip, cursor.leaf);
2947                                 bcopy((char *)cursor.data + roff,
2948                                       (char *)bp->b_data + boff, n);
2949                         }
2950                 }
2951                 if (error)
2952                         break;
2953
2954                 /*
2955                  * We have to be sure that the only elements added to the
2956                  * dedup cache are those which are already on-media.
2957                  */
2958                 if (hammer_live_dedup && hammer_cursor_ondisk(&cursor))
2959                         hammer_dedup_cache_add(ip, cursor.leaf);
2960
2961                 /*
2962                  * Iterate until we have filled the request.
2963                  */
2964                 boff += n;
2965                 if (boff == bp->b_bufsize)
2966                         break;
2967                 error = hammer_ip_next(&cursor);
2968         }
2969
2970         /*
2971          * There may have been a gap after the last record
2972          */
2973         if (error == ENOENT)
2974                 error = 0;
2975         if (error == 0 && boff != bp->b_bufsize) {
2976                 KKASSERT(boff < bp->b_bufsize);
2977                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2978                 /* boff = bp->b_bufsize; */
2979         }
2980
2981         /*
2982          * Disallow swapcache operation on the vnode buffer if double
2983          * buffering is enabled, the swapcache will get the data via
2984          * the block device buffer.
2985          */
2986         if (hammer_double_buffer)
2987                 bp->b_flags |= B_NOTMETA;
2988
2989         /*
2990          * Cleanup
2991          */
2992         bp->b_resid = 0;
2993         bp->b_error = error;
2994         if (error)
2995                 bp->b_flags |= B_ERROR;
2996         biodone(ap->a_bio);
2997
2998 done:
2999         /*
3000          * Cache the b-tree node for the last data read in cache[1].
3001          *
3002          * If we hit the file EOF then also cache the node in the
3003          * governing director's cache[3], it will be used to initialize
3004          * the inode's cache[1] for any inodes looked up via the directory.
3005          *
3006          * This doesn't reduce disk accesses since the B-Tree chain is
3007          * likely cached, but it does reduce cpu overhead when looking
3008          * up file offsets for cpdup/tar/cpio style iterations.
3009          */
3010         if (cursor.node)
3011                 hammer_cache_node(&ip->cache[1], cursor.node);
3012         if (ran_end >= ip->ino_data.size) {
3013                 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
3014                                         ip->obj_asof, ip->obj_localization);
3015                 if (dip) {
3016                         hammer_cache_node(&dip->cache[3], cursor.node);
3017                         hammer_rel_inode(dip, 0);
3018                 }
3019         }
3020         hammer_done_cursor(&cursor);
3021         hammer_done_transaction(&trans);
3022         lwkt_reltoken(&hmp->fs_token);
3023         return(error);
3024 }
3025
3026 /*
3027  * BMAP operation - used to support cluster_read() only.
3028  *
3029  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
3030  *
3031  * This routine may return EOPNOTSUPP if the opration is not supported for
3032  * the specified offset.  The contents of the pointer arguments do not
3033  * need to be initialized in that case. 
3034  *
3035  * If a disk address is available and properly aligned return 0 with 
3036  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
3037  * to the run-length relative to that offset.  Callers may assume that
3038  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
3039  * large, so return EOPNOTSUPP if it is not sufficiently large.
3040  */
3041 static
3042 int
3043 hammer_vop_bmap(struct vop_bmap_args *ap)
3044 {
3045         struct hammer_transaction trans;
3046         struct hammer_inode *ip;
3047         hammer_mount_t hmp;
3048         struct hammer_cursor cursor;
3049         hammer_base_elm_t base;
3050         int64_t rec_offset;
3051         int64_t ran_end;
3052         int64_t tmp64;
3053         int64_t base_offset;
3054         int64_t base_disk_offset;
3055         int64_t last_offset;
3056         hammer_off_t last_disk_offset;
3057         hammer_off_t disk_offset;
3058         int     rec_len;
3059         int     error;
3060         int     blksize;
3061
3062         ++hammer_stats_file_iopsr;
3063         ip = ap->a_vp->v_data;
3064         hmp = ip->hmp;
3065
3066         /*
3067          * We can only BMAP regular files.  We can't BMAP database files,
3068          * directories, etc.
3069          */
3070         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
3071                 return(EOPNOTSUPP);
3072
3073         /*
3074          * bmap is typically called with runp/runb both NULL when used
3075          * for writing.  We do not support BMAP for writing atm.
3076          */
3077         if (ap->a_cmd != BUF_CMD_READ)
3078                 return(EOPNOTSUPP);
3079
3080         /*
3081          * Scan the B-Tree to acquire blockmap addresses, then translate
3082          * to raw addresses.
3083          */
3084         lwkt_gettoken(&hmp->fs_token);
3085         hammer_simple_transaction(&trans, hmp);
3086 #if 0
3087         kprintf("bmap_beg %016llx ip->cache %p\n",
3088                 (long long)ap->a_loffset, ip->cache[1]);
3089 #endif
3090         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
3091
3092         /*
3093          * Key range (begin and end inclusive) to scan.  Note that the key's
3094          * stored in the actual records represent BASE+LEN, not BASE.  The
3095          * first record containing bio_offset will have a key > bio_offset.
3096          */
3097         cursor.key_beg.localization = ip->obj_localization +
3098                                       HAMMER_LOCALIZE_MISC;
3099         cursor.key_beg.obj_id = ip->obj_id;
3100         cursor.key_beg.create_tid = 0;
3101         cursor.key_beg.delete_tid = 0;
3102         cursor.key_beg.obj_type = 0;
3103         if (ap->a_runb)
3104                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
3105         else
3106                 cursor.key_beg.key = ap->a_loffset + 1;
3107         if (cursor.key_beg.key < 0)
3108                 cursor.key_beg.key = 0;
3109         cursor.asof = ip->obj_asof;
3110         cursor.flags |= HAMMER_CURSOR_ASOF;
3111
3112         cursor.key_end = cursor.key_beg;
3113         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
3114
3115         ran_end = ap->a_loffset + MAXPHYS;
3116         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
3117         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
3118         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
3119         if (tmp64 < ran_end)
3120                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
3121         else
3122                 cursor.key_end.key = ran_end + MAXPHYS + 1;
3123
3124         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
3125
3126         error = hammer_ip_first(&cursor);
3127         base_offset = last_offset = 0;
3128         base_disk_offset = last_disk_offset = 0;
3129
3130         while (error == 0) {
3131                 /*
3132                  * Get the base file offset of the record.  The key for
3133                  * data records is (base + bytes) rather then (base).
3134                  *
3135                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
3136                  * The extra bytes should be zero on-disk and the BMAP op
3137                  * should still be ok.
3138                  */
3139                 base = &cursor.leaf->base;
3140                 rec_offset = base->key - cursor.leaf->data_len;
3141                 rec_len    = cursor.leaf->data_len;
3142
3143                 /*
3144                  * Incorporate any cached truncation.
3145                  *
3146                  * NOTE: Modifications to rec_len based on synthesized
3147                  * truncation points remove the guarantee that any extended
3148                  * data on disk is zero (since the truncations may not have
3149                  * taken place on-media yet).
3150                  */
3151                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
3152                         if (hammer_cursor_ondisk(&cursor) ||
3153                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
3154                                 if (ip->trunc_off <= rec_offset)
3155                                         rec_len = 0;
3156                                 else if (ip->trunc_off < rec_offset + rec_len)
3157                                         rec_len = (int)(ip->trunc_off - rec_offset);
3158                         }
3159                 }
3160                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
3161                         if (hammer_cursor_ondisk(&cursor)) {
3162                                 if (ip->sync_trunc_off <= rec_offset)
3163                                         rec_len = 0;
3164                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
3165                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
3166                         }
3167                 }
3168
3169                 /*
3170                  * Accumulate information.  If we have hit a discontiguous
3171                  * block reset base_offset unless we are already beyond the
3172                  * requested offset.  If we are, that's it, we stop.
3173                  */
3174                 if (error)
3175                         break;
3176                 if (hammer_cursor_ondisk(&cursor)) {
3177                         disk_offset = cursor.leaf->data_offset;
3178                         if (rec_offset != last_offset ||
3179                             disk_offset != last_disk_offset) {
3180                                 if (rec_offset > ap->a_loffset)
3181                                         break;
3182                                 base_offset = rec_offset;
3183                                 base_disk_offset = disk_offset;
3184                         }
3185                         last_offset = rec_offset + rec_len;
3186                         last_disk_offset = disk_offset + rec_len;
3187
3188                         if (hammer_live_dedup)
3189                                 hammer_dedup_cache_add(ip, cursor.leaf);
3190                 }
3191                 
3192                 error = hammer_ip_next(&cursor);
3193         }
3194
3195 #if 0
3196         kprintf("BMAP %016llx:  %016llx - %016llx\n",
3197                 (long long)ap->a_loffset,
3198                 (long long)base_offset,
3199                 (long long)last_offset);
3200         kprintf("BMAP %16s:  %016llx - %016llx\n", "",
3201                 (long long)base_disk_offset,
3202                 (long long)last_disk_offset);
3203 #endif
3204
3205         if (cursor.node) {
3206                 hammer_cache_node(&ip->cache[1], cursor.node);
3207 #if 0
3208                 kprintf("bmap_end2 %016llx ip->cache %p\n",
3209                         (long long)ap->a_loffset, ip->cache[1]);
3210 #endif
3211         }
3212         hammer_done_cursor(&cursor);
3213         hammer_done_transaction(&trans);
3214         lwkt_reltoken(&hmp->fs_token);
3215
3216         /*
3217          * If we couldn't find any records or the records we did find were
3218          * all behind the requested offset, return failure.  A forward
3219          * truncation can leave a hole w/ no on-disk records.
3220          */
3221         if (last_offset == 0 || last_offset < ap->a_loffset)
3222                 return (EOPNOTSUPP);
3223
3224         /*
3225          * Figure out the block size at the requested offset and adjust
3226          * our limits so the cluster_read() does not create inappropriately
3227          * sized buffer cache buffers.
3228          */
3229         blksize = hammer_blocksize(ap->a_loffset);
3230         if (hammer_blocksize(base_offset) != blksize) {
3231                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
3232         }
3233         if (last_offset != ap->a_loffset &&
3234             hammer_blocksize(last_offset - 1) != blksize) {
3235                 last_offset = hammer_blockdemarc(ap->a_loffset,
3236                                                  last_offset - 1);
3237         }
3238
3239         /*
3240          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3241          * from occuring.
3242          */
3243         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
3244
3245         if ((disk_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_LARGE_DATA) {
3246                 /*
3247                  * Only large-data zones can be direct-IOd
3248                  */
3249                 error = EOPNOTSUPP;
3250         } else if ((disk_offset & HAMMER_BUFMASK) ||
3251                    (last_offset - ap->a_loffset) < blksize) {
3252                 /*
3253                  * doffsetp is not aligned or the forward run size does
3254                  * not cover a whole buffer, disallow the direct I/O.
3255                  */
3256                 error = EOPNOTSUPP;
3257         } else {
3258                 /*
3259                  * We're good.
3260                  */
3261                 *ap->a_doffsetp = disk_offset;
3262                 if (ap->a_runb) {
3263                         *ap->a_runb = ap->a_loffset - base_offset;
3264                         KKASSERT(*ap->a_runb >= 0);
3265                 }
3266                 if (ap->a_runp) {
3267                         *ap->a_runp = last_offset - ap->a_loffset;
3268                         KKASSERT(*ap->a_runp >= 0);
3269                 }
3270                 error = 0;
3271         }
3272         return(error);
3273 }
3274
3275 /*
3276  * Write to a regular file.   Because this is a strategy call the OS is
3277  * trying to actually get data onto the media.
3278  */
3279 static
3280 int
3281 hammer_vop_strategy_write(struct vop_strategy_args *ap)
3282 {
3283         hammer_record_t record;
3284         hammer_mount_t hmp;
3285         hammer_inode_t ip;
3286         struct bio *bio;
3287         struct buf *bp;
3288         int blksize;
3289         int bytes;
3290         int error;
3291
3292         bio = ap->a_bio;
3293         bp = bio->bio_buf;
3294         ip = ap->a_vp->v_data;
3295         hmp = ip->hmp;
3296
3297         blksize = hammer_blocksize(bio->bio_offset);
3298         KKASSERT(bp->b_bufsize == blksize);
3299
3300         if (ip->flags & HAMMER_INODE_RO) {
3301                 bp->b_error = EROFS;
3302                 bp->b_flags |= B_ERROR;
3303                 biodone(ap->a_bio);
3304                 return(EROFS);
3305         }
3306
3307         lwkt_gettoken(&hmp->fs_token);
3308
3309         /*
3310          * Disallow swapcache operation on the vnode buffer if double
3311          * buffering is enabled, the swapcache will get the data via
3312          * the block device buffer.
3313          */
3314         if (hammer_double_buffer)
3315                 bp->b_flags |= B_NOTMETA;
3316
3317         /*
3318          * Interlock with inode destruction (no in-kernel or directory
3319          * topology visibility).  If we queue new IO while trying to
3320          * destroy the inode we can deadlock the vtrunc call in
3321          * hammer_inode_unloadable_check().
3322          *
3323          * Besides, there's no point flushing a bp associated with an
3324          * inode that is being destroyed on-media and has no kernel
3325          * references.
3326          */
3327         if ((ip->flags | ip->sync_flags) &
3328             (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
3329                 bp->b_resid = 0;
3330                 biodone(ap->a_bio);
3331                 lwkt_reltoken(&hmp->fs_token);
3332                 return(0);
3333         }
3334
3335         /*
3336          * Reserve space and issue a direct-write from the front-end. 
3337          * NOTE: The direct_io code will hammer_bread/bcopy smaller
3338          * allocations.
3339          *
3340          * An in-memory record will be installed to reference the storage
3341          * until the flusher can get to it.
3342          *
3343          * Since we own the high level bio the front-end will not try to
3344          * do a direct-read until the write completes.
3345          *
3346          * NOTE: The only time we do not reserve a full-sized buffers
3347          * worth of data is if the file is small.  We do not try to
3348          * allocate a fragment (from the small-data zone) at the end of
3349          * an otherwise large file as this can lead to wildly separated
3350          * data.
3351          */
3352         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
3353         KKASSERT(bio->bio_offset < ip->ino_data.size);
3354         if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
3355                 bytes = bp->b_bufsize;
3356         else
3357                 bytes = ((int)ip->ino_data.size + 15) & ~15;
3358
3359         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
3360                                     bytes, &error);
3361
3362         /*
3363          * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3364          * in hammer_vop_write().  We must flag the record so the proper
3365          * REDO_TERM_WRITE entry is generated during the flush.
3366          */
3367         if (record) {
3368                 if (bp->b_flags & B_VFSFLAG1) {
3369                         record->flags |= HAMMER_RECF_REDO;
3370                         bp->b_flags &= ~B_VFSFLAG1;
3371                 }
3372                 if (record->flags & HAMMER_RECF_DEDUPED) {
3373                         bp->b_resid = 0;
3374                         hammer_ip_replace_bulk(hmp, record);
3375                         biodone(ap->a_bio);
3376                 } else {
3377                         hammer_io_direct_write(hmp, bio, record);
3378                 }
3379                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
3380                         hammer_flush_inode(ip, 0);
3381         } else {
3382                 bp->b_bio2.bio_offset = NOOFFSET;
3383                 bp->b_error = error;
3384                 bp->b_flags |= B_ERROR;
3385                 biodone(ap->a_bio);
3386         }
3387         lwkt_reltoken(&hmp->fs_token);
3388         return(error);
3389 }
3390
3391 /*
3392  * dounlink - disconnect a directory entry
3393  *
3394  * XXX whiteout support not really in yet
3395  */
3396 static int
3397 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
3398                 struct vnode *dvp, struct ucred *cred, 
3399                 int flags, int isdir)
3400 {
3401         struct namecache *ncp;
3402         hammer_inode_t dip;
3403         hammer_inode_t ip;
3404         hammer_mount_t hmp;
3405         struct hammer_cursor cursor;
3406         int64_t namekey;
3407         u_int32_t max_iterations;
3408         int nlen, error;
3409
3410         /*
3411          * Calculate the namekey and setup the key range for the scan.  This
3412          * works kinda like a chained hash table where the lower 32 bits
3413          * of the namekey synthesize the chain.
3414          *
3415          * The key range is inclusive of both key_beg and key_end.
3416          */
3417         dip = VTOI(dvp);
3418         ncp = nch->ncp;
3419         hmp = dip->hmp;
3420
3421         if (dip->flags & HAMMER_INODE_RO)
3422                 return (EROFS);
3423
3424         namekey = hammer_directory_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3425                                            &max_iterations);
3426 retry:
3427         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3428         cursor.key_beg.localization = dip->obj_localization +
3429                                       hammer_dir_localization(dip);
3430         cursor.key_beg.obj_id = dip->obj_id;
3431         cursor.key_beg.key = namekey;
3432         cursor.key_beg.create_tid = 0;
3433         cursor.key_beg.delete_tid = 0;
3434         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3435         cursor.key_beg.obj_type = 0;
3436
3437         cursor.key_end = cursor.key_beg;
3438         cursor.key_end.key += max_iterations;
3439         cursor.asof = dip->obj_asof;
3440         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3441
3442         /*
3443          * Scan all matching records (the chain), locate the one matching
3444          * the requested path component.  info->last_error contains the
3445          * error code on search termination and could be 0, ENOENT, or
3446          * something else.
3447          *
3448          * The hammer_ip_*() functions merge in-memory records with on-disk
3449          * records for the purposes of the search.
3450          */
3451         error = hammer_ip_first(&cursor);
3452
3453         while (error == 0) {
3454                 error = hammer_ip_resolve_data(&cursor);
3455                 if (error)
3456                         break;
3457                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3458                 KKASSERT(nlen > 0);
3459                 if (ncp->nc_nlen == nlen &&
3460                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3461                         break;
3462                 }
3463                 error = hammer_ip_next(&cursor);
3464         }
3465
3466         /*
3467          * If all is ok we have to get the inode so we can adjust nlinks.
3468          * To avoid a deadlock with the flusher we must release the inode
3469          * lock on the directory when acquiring the inode for the entry.
3470          *
3471          * If the target is a directory, it must be empty.
3472          */
3473         if (error == 0) {
3474                 hammer_unlock(&cursor.ip->lock);
3475                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3476                                       hmp->asof,
3477                                       cursor.data->entry.localization,
3478                                       0, &error);
3479                 hammer_lock_sh(&cursor.ip->lock);
3480                 if (error == ENOENT) {
3481                         kprintf("HAMMER: WARNING: Removing "
3482                                 "dirent w/missing inode \"%s\"\n"
3483                                 "\tobj_id = %016llx\n",
3484                                 ncp->nc_name,
3485                                 (long long)cursor.data->entry.obj_id);
3486                         error = 0;
3487                 }
3488
3489                 /*
3490                  * If isdir >= 0 we validate that the entry is or is not a
3491                  * directory.  If isdir < 0 we don't care.
3492                  */
3493                 if (error == 0 && isdir >= 0 && ip) {
3494                         if (isdir &&
3495                             ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3496                                 error = ENOTDIR;
3497                         } else if (isdir == 0 &&
3498                             ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3499                                 error = EISDIR;
3500                         }
3501                 }
3502
3503                 /*
3504                  * If we are trying to remove a directory the directory must
3505                  * be empty.
3506                  *
3507                  * The check directory code can loop and deadlock/retry.  Our
3508                  * own cursor's node locks must be released to avoid a 3-way
3509                  * deadlock with the flusher if the check directory code
3510                  * blocks.
3511                  *
3512                  * If any changes whatsoever have been made to the cursor
3513                  * set EDEADLK and retry.
3514                  *
3515                  * WARNING: See warnings in hammer_unlock_cursor()
3516                  *          function.
3517                  */
3518                 if (error == 0 && ip && ip->ino_data.obj_type ==
3519                                         HAMMER_OBJTYPE_DIRECTORY) {
3520                         hammer_unlock_cursor(&cursor);
3521                         error = hammer_ip_check_directory_empty(trans, ip);
3522                         hammer_lock_cursor(&cursor);
3523                         if (cursor.flags & HAMMER_CURSOR_RETEST) {
3524                                 kprintf("HAMMER: Warning: avoided deadlock "
3525                                         "on rmdir '%s'\n",
3526                                         ncp->nc_name);
3527                                 error = EDEADLK;
3528                         }
3529                 }
3530
3531                 /*
3532                  * Delete the directory entry.
3533                  *
3534                  * WARNING: hammer_ip_del_directory() may have to terminate
3535                  * the cursor to avoid a deadlock.  It is ok to call
3536                  * hammer_done_cursor() twice.
3537                  */
3538                 if (error == 0) {
3539                         error = hammer_ip_del_directory(trans, &cursor,
3540                                                         dip, ip);
3541                 }
3542                 hammer_done_cursor(&cursor);
3543                 if (error == 0) {
3544                         cache_setunresolved(nch);
3545                         cache_setvp(nch, NULL);
3546
3547                         /*
3548                          * NOTE: ip->vp, if non-NULL, cannot be directly
3549                          *       referenced without formally acquiring the
3550                          *       vp since the vp might have zero refs on it,
3551                          *       or in the middle of a reclaim, etc.
3552                          *
3553                          * NOTE: The cache_setunresolved() can rip the vp
3554                          *       out from under us since the vp may not have
3555                          *       any refs, in which case ip->vp will be NULL
3556                          *       from the outset.
3557                          */
3558                         while (ip && ip->vp) {
3559                                 struct vnode *vp;
3560
3561                                 error = hammer_get_vnode(ip, &vp);
3562                                 if (error == 0 && vp) {
3563                                         vn_unlock(vp);
3564                                         hammer_knote(ip->vp, NOTE_DELETE);
3565                                         cache_inval_vp(ip->vp, CINV_DESTROY);
3566                                         vrele(vp);
3567                                         break;
3568                                 }
3569                                 kprintf("Debug: HAMMER ip/vp race1 avoided\n");
3570                         }
3571                 }
3572                 if (ip)
3573                         hammer_rel_inode(ip, 0);
3574         } else {
3575                 hammer_done_cursor(&cursor);
3576         }
3577         if (error == EDEADLK)
3578                 goto retry;
3579
3580         return (error);
3581 }
3582
3583 /************************************************************************
3584  *                          FIFO AND SPECFS OPS                         *
3585  ************************************************************************
3586  *
3587  */
3588 static int
3589 hammer_vop_fifoclose (struct vop_close_args *ap)
3590 {
3591         /* XXX update itimes */
3592         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3593 }
3594
3595 static int
3596 hammer_vop_fiforead (struct vop_read_args *ap)
3597 {
3598         int error;
3599
3600         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3601         /* XXX update access time */
3602         return (error);
3603 }
3604
3605 static int
3606 hammer_vop_fifowrite (struct vop_write_args *ap)
3607 {
3608         int error;
3609
3610         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3611         /* XXX update access time */
3612         return (error);
3613 }
3614
3615 static
3616 int
3617 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3618 {
3619         int error;
3620
3621         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3622         if (error)
3623                 error = hammer_vop_kqfilter(ap);
3624         return(error);
3625 }
3626
3627 /************************************************************************
3628  *                          KQFILTER OPS                                *
3629  ************************************************************************
3630  *
3631  */
3632 static void filt_hammerdetach(struct knote *kn);
3633 static int filt_hammerread(struct knote *kn, long hint);
3634 static int filt_hammerwrite(struct knote *kn, long hint);
3635 static int filt_hammervnode(struct knote *kn, long hint);
3636
3637 static struct filterops hammerread_filtops =
3638         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammerread };
3639 static struct filterops hammerwrite_filtops =
3640         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammerwrite };
3641 static struct filterops hammervnode_filtops =
3642         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammervnode };
3643
3644 static
3645 int
3646 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3647 {
3648         struct vnode *vp = ap->a_vp;
3649         struct knote *kn = ap->a_kn;
3650
3651         switch (kn->kn_filter) {
3652         case EVFILT_READ:
3653                 kn->kn_fop = &hammerread_filtops;
3654                 break;
3655         case EVFILT_WRITE:
3656                 kn->kn_fop = &hammerwrite_filtops;
3657                 break;
3658         case EVFILT_VNODE:
3659                 kn->kn_fop = &hammervnode_filtops;
3660                 break;
3661         default:
3662                 return (EOPNOTSUPP);
3663         }
3664
3665         kn->kn_hook = (caddr_t)vp;
3666
3667         knote_insert(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3668
3669         return(0);
3670 }
3671
3672 static void
3673 filt_hammerdetach(struct knote *kn)
3674 {
3675         struct vnode *vp = (void *)kn->kn_hook;
3676
3677         knote_remove(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3678 }
3679
3680 static int
3681 filt_hammerread(struct knote *kn, long hint)
3682 {
3683         struct vnode *vp = (void *)kn->kn_hook;
3684         hammer_inode_t ip = VTOI(vp);
3685         hammer_mount_t hmp = ip->hmp;
3686         off_t off;
3687
3688         if (hint == NOTE_REVOKE) {
3689                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3690                 return(1);
3691         }
3692         lwkt_gettoken(&hmp->fs_token);  /* XXX use per-ip-token */
3693         off = ip->ino_data.size - kn->kn_fp->f_offset;
3694         kn->kn_data = (off < INTPTR_MAX) ? off : INTPTR_MAX;
3695         lwkt_reltoken(&hmp->fs_token);
3696         if (kn->kn_sfflags & NOTE_OLDAPI)
3697                 return(1);
3698         return (kn->kn_data != 0);
3699 }
3700
3701 static int
3702 filt_hammerwrite(struct knote *kn, long hint)
3703 {
3704         if (hint == NOTE_REVOKE)
3705                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3706         kn->kn_data = 0;
3707         return (1);
3708 }
3709
3710 static int
3711 filt_hammervnode(struct knote *kn, long hint)
3712 {
3713         if (kn->kn_sfflags & hint)
3714                 kn->kn_fflags |= hint;
3715         if (hint == NOTE_REVOKE) {
3716                 kn->kn_flags |= (EV_EOF | EV_NODATA);
3717                 return (1);
3718         }
3719         return (kn->kn_fflags != 0);
3720 }
3721