kernel - Cluster fixes + Enable clustering for HAMMER1
[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
436                 bp->b_flags |= B_CLUSTEROK;
437                 n = blksize - offset;
438                 if (n > uio->uio_resid)
439                         n = uio->uio_resid;
440                 if (n > ip->ino_data.size - uio->uio_offset)
441                         n = (int)(ip->ino_data.size - uio->uio_offset);
442                 if (got_fstoken)
443                         lwkt_reltoken(&hmp->fs_token);
444
445                 /*
446                  * Set B_AGE, data has a lower priority than meta-data.
447                  *
448                  * Use a hold/unlock/drop sequence to run the uiomove
449                  * with the buffer unlocked, avoiding deadlocks against
450                  * read()s on mmap()'d spaces.
451                  */
452                 bp->b_flags |= B_AGE;
453                 bqhold(bp);
454                 bqrelse(bp);
455                 error = uiomove((char *)bp->b_data + offset, n, uio);
456                 bqdrop(bp);
457
458                 if (got_fstoken)
459                         lwkt_gettoken(&hmp->fs_token);
460
461                 if (error)
462                         break;
463                 hammer_stats_file_read += n;
464         }
465
466         /*
467          * Try to update the atime with just the inode lock for maximum
468          * concurrency.  If we can't shortcut it we have to get the full
469          * blown transaction.
470          */
471         if (got_fstoken == 0 && hammer_update_atime_quick(ip) < 0) {
472                 lwkt_gettoken(&hmp->fs_token);
473                 got_fstoken = 1;
474                 hammer_start_transaction(&trans, ip->hmp);
475         }
476
477         if (got_fstoken) {
478                 if ((ip->flags & HAMMER_INODE_RO) == 0 &&
479                     (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
480                         ip->ino_data.atime = trans.time;
481                         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
482                 }
483                 hammer_done_transaction(&trans);
484                 lwkt_reltoken(&hmp->fs_token);
485         }
486         return (error);
487 }
488
489 /*
490  * hammer_vop_write { vp, uio, ioflag, cred }
491  */
492 static
493 int
494 hammer_vop_write(struct vop_write_args *ap)
495 {
496         struct hammer_transaction trans;
497         struct hammer_inode *ip;
498         hammer_mount_t hmp;
499         thread_t td;
500         struct uio *uio;
501         int offset;
502         off_t base_offset;
503         int64_t cluster_eof;
504         struct buf *bp;
505         int kflags;
506         int error;
507         int n;
508         int flags;
509         int seqcount;
510         int bigwrite;
511
512         if (ap->a_vp->v_type != VREG)
513                 return (EINVAL);
514         ip = VTOI(ap->a_vp);
515         hmp = ip->hmp;
516         error = 0;
517         kflags = 0;
518         seqcount = ap->a_ioflag >> 16;
519
520         if (ip->flags & HAMMER_INODE_RO)
521                 return (EROFS);
522
523         /*
524          * Create a transaction to cover the operations we perform.
525          */
526         lwkt_gettoken(&hmp->fs_token);
527         hammer_start_transaction(&trans, hmp);
528         uio = ap->a_uio;
529
530         /*
531          * Check append mode
532          */
533         if (ap->a_ioflag & IO_APPEND)
534                 uio->uio_offset = ip->ino_data.size;
535
536         /*
537          * Check for illegal write offsets.  Valid range is 0...2^63-1.
538          *
539          * NOTE: the base_off assignment is required to work around what
540          * I consider to be a GCC-4 optimization bug.
541          */
542         if (uio->uio_offset < 0) {
543                 hammer_done_transaction(&trans);
544                 lwkt_reltoken(&hmp->fs_token);
545                 return (EFBIG);
546         }
547         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
548         if (uio->uio_resid > 0 && base_offset <= uio->uio_offset) {
549                 hammer_done_transaction(&trans);
550                 lwkt_reltoken(&hmp->fs_token);
551                 return (EFBIG);
552         }
553
554         if (uio->uio_resid > 0 && (td = uio->uio_td) != NULL && td->td_proc &&
555             base_offset > td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
556                 hammer_done_transaction(&trans);
557                 lwkt_reltoken(&hmp->fs_token);
558                 lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ);
559                 return (EFBIG);
560         }
561
562         /*
563          * If reading or writing a huge amount of data we have to break
564          * atomicy and allow the operation to be interrupted by a signal
565          * or it can DOS the machine.
566          *
567          * Preset redo_count so we stop generating REDOs earlier if the
568          * limit is exceeded.
569          */
570         bigwrite = (uio->uio_resid > 100 * 1024 * 1024);
571         if ((ip->flags & HAMMER_INODE_REDO) &&
572             ip->redo_count < hammer_limit_redo) {
573                 ip->redo_count += uio->uio_resid;
574         }
575
576         /*
577          * Access the data typically in HAMMER_BUFSIZE blocks via the
578          * buffer cache, but HAMMER may use a variable block size based
579          * on the offset.
580          */
581         while (uio->uio_resid > 0) {
582                 int fixsize = 0;
583                 int blksize;
584                 int blkmask;
585                 int trivial;
586                 int endofblk;
587                 off_t nsize;
588
589                 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
590                         break;
591                 if (bigwrite && (error = hammer_signal_check(hmp)) != 0)
592                         break;
593
594                 blksize = hammer_blocksize(uio->uio_offset);
595
596                 /*
597                  * Do not allow HAMMER to blow out the buffer cache.  Very
598                  * large UIOs can lockout other processes due to bwillwrite()
599                  * mechanics.
600                  *
601                  * The hammer inode is not locked during these operations.
602                  * The vnode is locked which can interfere with the pageout
603                  * daemon for non-UIO_NOCOPY writes but should not interfere
604                  * with the buffer cache.  Even so, we cannot afford to
605                  * allow the pageout daemon to build up too many dirty buffer
606                  * cache buffers.
607                  *
608                  * Only call this if we aren't being recursively called from
609                  * a virtual disk device (vn), else we may deadlock.
610                  */
611                 if ((ap->a_ioflag & IO_RECURSE) == 0)
612                         bwillwrite(blksize);
613
614                 /*
615                  * Control the number of pending records associated with
616                  * this inode.  If too many have accumulated start a
617                  * flush.  Try to maintain a pipeline with the flusher.
618                  *
619                  * NOTE: It is possible for other sources to grow the
620                  *       records but not necessarily issue another flush,
621                  *       so use a timeout and ensure that a re-flush occurs.
622                  */
623                 if (ip->rsv_recs >= hammer_limit_inode_recs) {
624                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
625                         while (ip->rsv_recs >= hammer_limit_inode_recs * 2) {
626                                 ip->flags |= HAMMER_INODE_RECSW;
627                                 tsleep(&ip->rsv_recs, 0, "hmrwww", hz);
628                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
629                         }
630                 }
631
632 #if 0
633                 /*
634                  * Do not allow HAMMER to blow out system memory by
635                  * accumulating too many records.   Records are so well
636                  * decoupled from the buffer cache that it is possible
637                  * for userland to push data out to the media via
638                  * direct-write, but build up the records queued to the
639                  * backend faster then the backend can flush them out.
640                  * HAMMER has hit its write limit but the frontend has
641                  * no pushback to slow it down.
642                  */
643                 if (hmp->rsv_recs > hammer_limit_recs / 2) {
644                         /*
645                          * Get the inode on the flush list
646                          */
647                         if (ip->rsv_recs >= 64)
648                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
649                         else if (ip->rsv_recs >= 16)
650                                 hammer_flush_inode(ip, 0);
651
652                         /*
653                          * Keep the flusher going if the system keeps
654                          * queueing records.
655                          */
656                         delta = hmp->count_newrecords -
657                                 hmp->last_newrecords;
658                         if (delta < 0 || delta > hammer_limit_recs / 2) {
659                                 hmp->last_newrecords = hmp->count_newrecords;
660                                 hammer_sync_hmp(hmp, MNT_NOWAIT);
661                         }
662
663                         /*
664                          * If we have gotten behind start slowing
665                          * down the writers.
666                          */
667                         delta = (hmp->rsv_recs - hammer_limit_recs) *
668                                 hz / hammer_limit_recs;
669                         if (delta > 0)
670                                 tsleep(&trans, 0, "hmrslo", delta);
671                 }
672 #endif
673
674                 /*
675                  * Calculate the blocksize at the current offset and figure
676                  * out how much we can actually write.
677                  */
678                 blkmask = blksize - 1;
679                 offset = (int)uio->uio_offset & blkmask;
680                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
681                 n = blksize - offset;
682                 if (n > uio->uio_resid) {
683                         n = uio->uio_resid;
684                         endofblk = 0;
685                 } else {
686                         endofblk = 1;
687                 }
688                 nsize = uio->uio_offset + n;
689                 if (nsize > ip->ino_data.size) {
690                         if (uio->uio_offset > ip->ino_data.size)
691                                 trivial = 0;
692                         else
693                                 trivial = 1;
694                         nvextendbuf(ap->a_vp,
695                                     ip->ino_data.size,
696                                     nsize,
697                                     hammer_blocksize(ip->ino_data.size),
698                                     hammer_blocksize(nsize),
699                                     hammer_blockoff(ip->ino_data.size),
700                                     hammer_blockoff(nsize),
701                                     trivial);
702                         fixsize = 1;
703                         kflags |= NOTE_EXTEND;
704                 }
705
706                 if (uio->uio_segflg == UIO_NOCOPY) {
707                         /*
708                          * Issuing a write with the same data backing the
709                          * buffer.  Instantiate the buffer to collect the
710                          * backing vm pages, then read-in any missing bits.
711                          *
712                          * This case is used by vop_stdputpages().
713                          */
714                         bp = getblk(ap->a_vp, base_offset,
715                                     blksize, GETBLK_BHEAVY, 0);
716                         if ((bp->b_flags & B_CACHE) == 0) {
717                                 bqrelse(bp);
718                                 error = bread(ap->a_vp, base_offset,
719                                               blksize, &bp);
720                         }
721                 } else if (offset == 0 && uio->uio_resid >= blksize) {
722                         /*
723                          * Even though we are entirely overwriting the buffer
724                          * we may still have to zero it out to avoid a 
725                          * mmap/write visibility issue.
726                          */
727                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
728                         if ((bp->b_flags & B_CACHE) == 0)
729                                 vfs_bio_clrbuf(bp);
730                 } else if (base_offset >= ip->ino_data.size) {
731                         /*
732                          * If the base offset of the buffer is beyond the
733                          * file EOF, we don't have to issue a read.
734                          */
735                         bp = getblk(ap->a_vp, base_offset,
736                                     blksize, GETBLK_BHEAVY, 0);
737                         vfs_bio_clrbuf(bp);
738                 } else {
739                         /*
740                          * Partial overwrite, read in any missing bits then
741                          * replace the portion being written.
742                          */
743                         error = bread(ap->a_vp, base_offset, blksize, &bp);
744                         if (error == 0)
745                                 bheavy(bp);
746                 }
747                 if (error == 0) {
748                         lwkt_reltoken(&hmp->fs_token);
749                         error = uiomove(bp->b_data + offset, n, uio);
750                         lwkt_gettoken(&hmp->fs_token);
751                 }
752
753                 /*
754                  * Generate REDO records if enabled and redo_count will not
755                  * exceeded the limit.
756                  *
757                  * If redo_count exceeds the limit we stop generating records
758                  * and clear HAMMER_INODE_REDO.  This will cause the next
759                  * fsync() to do a full meta-data sync instead of just an
760                  * UNDO/REDO fifo update.
761                  *
762                  * When clearing HAMMER_INODE_REDO any pre-existing REDOs
763                  * will still be tracked.  The tracks will be terminated
764                  * when the related meta-data (including possible data
765                  * modifications which are not tracked via REDO) is
766                  * flushed.
767                  */
768                 if ((ip->flags & HAMMER_INODE_REDO) && error == 0) {
769                         if (ip->redo_count < hammer_limit_redo) {
770                                 bp->b_flags |= B_VFSFLAG1;
771                                 error = hammer_generate_redo(&trans, ip,
772                                                      base_offset + offset,
773                                                      HAMMER_REDO_WRITE,
774                                                      bp->b_data + offset,
775                                                      (size_t)n);
776                         } else {
777                                 ip->flags &= ~HAMMER_INODE_REDO;
778                         }
779                 }
780
781                 /*
782                  * If we screwed up we have to undo any VM size changes we
783                  * made.
784                  */
785                 if (error) {
786                         brelse(bp);
787                         if (fixsize) {
788                                 nvtruncbuf(ap->a_vp, ip->ino_data.size,
789                                           hammer_blocksize(ip->ino_data.size),
790                                           hammer_blockoff(ip->ino_data.size));
791                         }
792                         break;
793                 }
794                 kflags |= NOTE_WRITE;
795                 hammer_stats_file_write += n;
796                 bp->b_flags |= B_CLUSTEROK;
797                 if (ip->ino_data.size < uio->uio_offset) {
798                         ip->ino_data.size = uio->uio_offset;
799                         flags = HAMMER_INODE_SDIRTY;
800                 } else {
801                         flags = 0;
802                 }
803                 ip->ino_data.mtime = trans.time;
804                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
805                 hammer_modify_inode(&trans, ip, flags);
806
807                 /*
808                  * Once we dirty the buffer any cached zone-X offset
809                  * becomes invalid.  HAMMER NOTE: no-history mode cannot 
810                  * allow overwriting over the same data sector unless
811                  * we provide UNDOs for the old data, which we don't.
812                  */
813                 bp->b_bio2.bio_offset = NOOFFSET;
814
815                 /*
816                  * Final buffer disposition.
817                  *
818                  * Because meta-data updates are deferred, HAMMER is
819                  * especially sensitive to excessive bdwrite()s because
820                  * the I/O stream is not broken up by disk reads.  So the
821                  * buffer cache simply cannot keep up.
822                  *
823                  * WARNING!  blksize is variable.  cluster_write() is
824                  *           expected to not blow up if it encounters
825                  *           buffers that do not match the passed blksize.
826                  *
827                  * NOTE!  Hammer shouldn't need to bawrite()/cluster_write().
828                  *        The ip->rsv_recs check should burst-flush the data.
829                  *        If we queue it immediately the buf could be left
830                  *        locked on the device queue for a very long time.
831                  *
832                  *        However, failing to flush a dirty buffer out when
833                  *        issued from the pageout daemon can result in a low
834                  *        memory deadlock against bio_page_alloc(), so we
835                  *        have to bawrite() on IO_ASYNC as well.
836                  *
837                  * NOTE!  To avoid degenerate stalls due to mismatched block
838                  *        sizes we only honor IO_DIRECT on the write which
839                  *        abuts the end of the buffer.  However, we must
840                  *        honor IO_SYNC in case someone is silly enough to
841                  *        configure a HAMMER file as swap, or when HAMMER
842                  *        is serving NFS (for commits).  Ick ick.
843                  */
844                 bp->b_flags |= B_AGE | B_CLUSTEROK;
845                 if (ap->a_ioflag & IO_SYNC) {
846                         bwrite(bp);
847                 } else if ((ap->a_ioflag & IO_DIRECT) && endofblk) {
848                         bawrite(bp);
849                 } else if (ap->a_ioflag & IO_ASYNC) {
850                         bawrite(bp);
851                 } else if (hammer_cluster_enable &&
852                            !(ap->a_vp->v_mount->mnt_flag & MNT_NOCLUSTERW)) {
853                         if (base_offset < HAMMER_XDEMARC)
854                                 cluster_eof = hammer_blockdemarc(base_offset,
855                                                          ip->ino_data.size);
856                         else
857                                 cluster_eof = ip->ino_data.size;
858                         cluster_write(bp, cluster_eof, blksize, seqcount);
859                 } else {
860                         bdwrite(bp);
861                 }
862         }
863         hammer_done_transaction(&trans);
864         hammer_knote(ap->a_vp, kflags);
865         lwkt_reltoken(&hmp->fs_token);
866         return (error);
867 }
868
869 /*
870  * hammer_vop_access { vp, mode, cred }
871  *
872  * MPSAFE - does not require fs_token
873  */
874 static
875 int
876 hammer_vop_access(struct vop_access_args *ap)
877 {
878         struct hammer_inode *ip = VTOI(ap->a_vp);
879         uid_t uid;
880         gid_t gid;
881         int error;
882
883         ++hammer_stats_file_iopsr;
884         uid = hammer_to_unix_xid(&ip->ino_data.uid);
885         gid = hammer_to_unix_xid(&ip->ino_data.gid);
886
887         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
888                                   ip->ino_data.uflags);
889         return (error);
890 }
891
892 /*
893  * hammer_vop_advlock { vp, id, op, fl, flags }
894  *
895  * MPSAFE - does not require fs_token
896  */
897 static
898 int
899 hammer_vop_advlock(struct vop_advlock_args *ap)
900 {
901         hammer_inode_t ip = VTOI(ap->a_vp);
902
903         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
904 }
905
906 /*
907  * hammer_vop_close { vp, fflag }
908  *
909  * We can only sync-on-close for normal closes.  XXX disabled for now.
910  */
911 static
912 int
913 hammer_vop_close(struct vop_close_args *ap)
914 {
915 #if 0
916         struct vnode *vp = ap->a_vp;
917         hammer_inode_t ip = VTOI(vp);
918         int waitfor;
919         if (ip->flags & (HAMMER_INODE_CLOSESYNC|HAMMER_INODE_CLOSEASYNC)) {
920                 if (vn_islocked(vp) == LK_EXCLUSIVE &&
921                     (vp->v_flag & (VINACTIVE|VRECLAIMED)) == 0) {
922                         if (ip->flags & HAMMER_INODE_CLOSESYNC)
923                                 waitfor = MNT_WAIT;
924                         else
925                                 waitfor = MNT_NOWAIT;
926                         ip->flags &= ~(HAMMER_INODE_CLOSESYNC |
927                                        HAMMER_INODE_CLOSEASYNC);
928                         VOP_FSYNC(vp, MNT_NOWAIT, waitfor);
929                 }
930         }
931 #endif
932         return (vop_stdclose(ap));
933 }
934
935 /*
936  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
937  *
938  * The operating system has already ensured that the directory entry
939  * does not exist and done all appropriate namespace locking.
940  */
941 static
942 int
943 hammer_vop_ncreate(struct vop_ncreate_args *ap)
944 {
945         struct hammer_transaction trans;
946         struct hammer_inode *dip;
947         struct hammer_inode *nip;
948         struct nchandle *nch;
949         hammer_mount_t hmp;
950         int error;
951
952         nch = ap->a_nch;
953         dip = VTOI(ap->a_dvp);
954         hmp = dip->hmp;
955
956         if (dip->flags & HAMMER_INODE_RO)
957                 return (EROFS);
958         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
959                 return (error);
960
961         /*
962          * Create a transaction to cover the operations we perform.
963          */
964         lwkt_gettoken(&hmp->fs_token);
965         hammer_start_transaction(&trans, hmp);
966         ++hammer_stats_file_iopsw;
967
968         /*
969          * Create a new filesystem object of the requested type.  The
970          * returned inode will be referenced and shared-locked to prevent
971          * it from being moved to the flusher.
972          */
973         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
974                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
975                                     NULL, &nip);
976         if (error) {
977                 hkprintf("hammer_create_inode error %d\n", error);
978                 hammer_done_transaction(&trans);
979                 *ap->a_vpp = NULL;
980                 lwkt_reltoken(&hmp->fs_token);
981                 return (error);
982         }
983
984         /*
985          * Add the new filesystem object to the directory.  This will also
986          * bump the inode's link count.
987          */
988         error = hammer_ip_add_directory(&trans, dip,
989                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
990                                         nip);
991         if (error)
992                 hkprintf("hammer_ip_add_directory error %d\n", error);
993
994         /*
995          * Finish up.
996          */
997         if (error) {
998                 hammer_rel_inode(nip, 0);
999                 hammer_done_transaction(&trans);
1000                 *ap->a_vpp = NULL;
1001         } else {
1002                 error = hammer_get_vnode(nip, ap->a_vpp);
1003                 hammer_done_transaction(&trans);
1004                 hammer_rel_inode(nip, 0);
1005                 if (error == 0) {
1006                         cache_setunresolved(ap->a_nch);
1007                         cache_setvp(ap->a_nch, *ap->a_vpp);
1008                 }
1009                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1010         }
1011         lwkt_reltoken(&hmp->fs_token);
1012         return (error);
1013 }
1014
1015 /*
1016  * hammer_vop_getattr { vp, vap }
1017  *
1018  * Retrieve an inode's attribute information.  When accessing inodes
1019  * historically we fake the atime field to ensure consistent results.
1020  * The atime field is stored in the B-Tree element and allowed to be
1021  * updated without cycling the element.
1022  *
1023  * MPSAFE - does not require fs_token
1024  */
1025 static
1026 int
1027 hammer_vop_getattr(struct vop_getattr_args *ap)
1028 {
1029         struct hammer_inode *ip = VTOI(ap->a_vp);
1030         struct vattr *vap = ap->a_vap;
1031
1032         /*
1033          * We want the fsid to be different when accessing a filesystem
1034          * with different as-of's so programs like diff don't think
1035          * the files are the same.
1036          *
1037          * We also want the fsid to be the same when comparing snapshots,
1038          * or when comparing mirrors (which might be backed by different
1039          * physical devices).  HAMMER fsids are based on the PFS's
1040          * shared_uuid field.
1041          *
1042          * XXX there is a chance of collision here.  The va_fsid reported
1043          * by stat is different from the more involved fsid used in the
1044          * mount structure.
1045          */
1046         ++hammer_stats_file_iopsr;
1047         hammer_lock_sh(&ip->lock);
1048         vap->va_fsid = ip->pfsm->fsid_udev ^ (u_int32_t)ip->obj_asof ^
1049                        (u_int32_t)(ip->obj_asof >> 32);
1050
1051         vap->va_fileid = ip->ino_leaf.base.obj_id;
1052         vap->va_mode = ip->ino_data.mode;
1053         vap->va_nlink = ip->ino_data.nlinks;
1054         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1055         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1056         vap->va_rmajor = 0;
1057         vap->va_rminor = 0;
1058         vap->va_size = ip->ino_data.size;
1059
1060         /*
1061          * Special case for @@PFS softlinks.  The actual size of the
1062          * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
1063          * or for MAX_TID is    "@@-1:%05d" == 10 bytes.
1064          */
1065         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
1066             ip->ino_data.size == 10 &&
1067             ip->obj_asof == HAMMER_MAX_TID &&
1068             ip->obj_localization == 0 &&
1069             strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
1070                     if (ip->pfsm->pfsd.mirror_flags & HAMMER_PFSD_SLAVE)
1071                             vap->va_size = 26;
1072                     else
1073                             vap->va_size = 10;
1074         }
1075
1076         /*
1077          * We must provide a consistent atime and mtime for snapshots
1078          * so people can do a 'tar cf - ... | md5' on them and get
1079          * consistent results.
1080          */
1081         if (ip->flags & HAMMER_INODE_RO) {
1082                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
1083                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
1084         } else {
1085                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
1086                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
1087         }
1088         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
1089         vap->va_flags = ip->ino_data.uflags;
1090         vap->va_gen = 1;        /* hammer inums are unique for all time */
1091         vap->va_blocksize = HAMMER_BUFSIZE;
1092         if (ip->ino_data.size >= HAMMER_XDEMARC) {
1093                 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
1094                                 ~HAMMER_XBUFMASK64;
1095         } else if (ip->ino_data.size > HAMMER_BUFSIZE / 2) {
1096                 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
1097                                 ~HAMMER_BUFMASK64;
1098         } else {
1099                 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
1100         }
1101
1102         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
1103         vap->va_filerev = 0;    /* XXX */
1104         vap->va_uid_uuid = ip->ino_data.uid;
1105         vap->va_gid_uuid = ip->ino_data.gid;
1106         vap->va_fsid_uuid = ip->hmp->fsid;
1107         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
1108                           VA_FSID_UUID_VALID;
1109
1110         switch (ip->ino_data.obj_type) {
1111         case HAMMER_OBJTYPE_CDEV:
1112         case HAMMER_OBJTYPE_BDEV:
1113                 vap->va_rmajor = ip->ino_data.rmajor;
1114                 vap->va_rminor = ip->ino_data.rminor;
1115                 break;
1116         default:
1117                 break;
1118         }
1119         hammer_unlock(&ip->lock);
1120         return(0);
1121 }
1122
1123 /*
1124  * hammer_vop_nresolve { nch, dvp, cred }
1125  *
1126  * Locate the requested directory entry.
1127  */
1128 static
1129 int
1130 hammer_vop_nresolve(struct vop_nresolve_args *ap)
1131 {
1132         struct hammer_transaction trans;
1133         struct namecache *ncp;
1134         hammer_mount_t hmp;
1135         hammer_inode_t dip;
1136         hammer_inode_t ip;
1137         hammer_tid_t asof;
1138         struct hammer_cursor cursor;
1139         struct vnode *vp;
1140         int64_t namekey;
1141         int error;
1142         int i;
1143         int nlen;
1144         int flags;
1145         int ispfs;
1146         int64_t obj_id;
1147         u_int32_t localization;
1148         u_int32_t max_iterations;
1149
1150         /*
1151          * Misc initialization, plus handle as-of name extensions.  Look for
1152          * the '@@' extension.  Note that as-of files and directories cannot
1153          * be modified.
1154          */
1155         dip = VTOI(ap->a_dvp);
1156         ncp = ap->a_nch->ncp;
1157         asof = dip->obj_asof;
1158         localization = dip->obj_localization;   /* for code consistency */
1159         nlen = ncp->nc_nlen;
1160         flags = dip->flags & HAMMER_INODE_RO;
1161         ispfs = 0;
1162         hmp = dip->hmp;
1163
1164         lwkt_gettoken(&hmp->fs_token);
1165         hammer_simple_transaction(&trans, hmp);
1166         ++hammer_stats_file_iopsr;
1167
1168         for (i = 0; i < nlen; ++i) {
1169                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
1170                         error = hammer_str_to_tid(ncp->nc_name + i + 2,
1171                                                   &ispfs, &asof, &localization);
1172                         if (error != 0) {
1173                                 i = nlen;
1174                                 break;
1175                         }
1176                         if (asof != HAMMER_MAX_TID)
1177                                 flags |= HAMMER_INODE_RO;
1178                         break;
1179                 }
1180         }
1181         nlen = i;
1182
1183         /*
1184          * If this is a PFS softlink we dive into the PFS
1185          */
1186         if (ispfs && nlen == 0) {
1187                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1188                                       asof, localization,
1189                                       flags, &error);
1190                 if (error == 0) {
1191                         error = hammer_get_vnode(ip, &vp);
1192                         hammer_rel_inode(ip, 0);
1193                 } else {
1194                         vp = NULL;
1195                 }
1196                 if (error == 0) {
1197                         vn_unlock(vp);
1198                         cache_setvp(ap->a_nch, vp);
1199                         vrele(vp);
1200                 }
1201                 goto done;
1202         }
1203
1204         /*
1205          * If there is no path component the time extension is relative to dip.
1206          * e.g. "fubar/@@<snapshot>"
1207          *
1208          * "." is handled by the kernel, but ".@@<snapshot>" is not.
1209          * e.g. "fubar/.@@<snapshot>"
1210          *
1211          * ".." is handled by the kernel.  We do not currently handle
1212          * "..@<snapshot>".
1213          */
1214         if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1215                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
1216                                       asof, dip->obj_localization,
1217                                       flags, &error);
1218                 if (error == 0) {
1219                         error = hammer_get_vnode(ip, &vp);
1220                         hammer_rel_inode(ip, 0);
1221                 } else {
1222                         vp = NULL;
1223                 }
1224                 if (error == 0) {
1225                         vn_unlock(vp);
1226                         cache_setvp(ap->a_nch, vp);
1227                         vrele(vp);
1228                 }
1229                 goto done;
1230         }
1231
1232         /*
1233          * Calculate the namekey and setup the key range for the scan.  This
1234          * works kinda like a chained hash table where the lower 32 bits
1235          * of the namekey synthesize the chain.
1236          *
1237          * The key range is inclusive of both key_beg and key_end.
1238          */
1239         namekey = hammer_directory_namekey(dip, ncp->nc_name, nlen,
1240                                            &max_iterations);
1241
1242         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1243         cursor.key_beg.localization = dip->obj_localization +
1244                                       hammer_dir_localization(dip);
1245         cursor.key_beg.obj_id = dip->obj_id;
1246         cursor.key_beg.key = namekey;
1247         cursor.key_beg.create_tid = 0;
1248         cursor.key_beg.delete_tid = 0;
1249         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1250         cursor.key_beg.obj_type = 0;
1251
1252         cursor.key_end = cursor.key_beg;
1253         cursor.key_end.key += max_iterations;
1254         cursor.asof = asof;
1255         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1256
1257         /*
1258          * Scan all matching records (the chain), locate the one matching
1259          * the requested path component.
1260          *
1261          * The hammer_ip_*() functions merge in-memory records with on-disk
1262          * records for the purposes of the search.
1263          */
1264         obj_id = 0;
1265         localization = HAMMER_DEF_LOCALIZATION;
1266
1267         if (error == 0) {
1268                 error = hammer_ip_first(&cursor);
1269                 while (error == 0) {
1270                         error = hammer_ip_resolve_data(&cursor);
1271                         if (error)
1272                                 break;
1273                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1274                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1275                                 obj_id = cursor.data->entry.obj_id;
1276                                 localization = cursor.data->entry.localization;
1277                                 break;
1278                         }
1279                         error = hammer_ip_next(&cursor);
1280                 }
1281         }
1282         hammer_done_cursor(&cursor);
1283
1284         /*
1285          * Lookup the obj_id.  This should always succeed.  If it does not
1286          * the filesystem may be damaged and we return a dummy inode.
1287          */
1288         if (error == 0) {
1289                 ip = hammer_get_inode(&trans, dip, obj_id,
1290                                       asof, localization,
1291                                       flags, &error);
1292                 if (error == ENOENT) {
1293                         kprintf("HAMMER: WARNING: Missing "
1294                                 "inode for dirent \"%s\"\n"
1295                                 "\tobj_id = %016llx, asof=%016llx, lo=%08x\n",
1296                                 ncp->nc_name,
1297                                 (long long)obj_id, (long long)asof,
1298                                 localization);
1299                         error = 0;
1300                         ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1301                                                     asof, localization,
1302                                                     flags, &error);
1303                 }
1304                 if (error == 0) {
1305                         error = hammer_get_vnode(ip, &vp);
1306                         hammer_rel_inode(ip, 0);
1307                 } else {
1308                         vp = NULL;
1309                 }
1310                 if (error == 0) {
1311                         vn_unlock(vp);
1312                         cache_setvp(ap->a_nch, vp);
1313                         vrele(vp);
1314                 }
1315         } else if (error == ENOENT) {
1316                 cache_setvp(ap->a_nch, NULL);
1317         }
1318 done:
1319         hammer_done_transaction(&trans);
1320         lwkt_reltoken(&hmp->fs_token);
1321         return (error);
1322 }
1323
1324 /*
1325  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1326  *
1327  * Locate the parent directory of a directory vnode.
1328  *
1329  * dvp is referenced but not locked.  *vpp must be returned referenced and
1330  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
1331  * at the root, instead it could indicate that the directory we were in was
1332  * removed.
1333  *
1334  * NOTE: as-of sequences are not linked into the directory structure.  If
1335  * we are at the root with a different asof then the mount point, reload
1336  * the same directory with the mount point's asof.   I'm not sure what this
1337  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
1338  * get confused, but it hasn't been tested.
1339  */
1340 static
1341 int
1342 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1343 {
1344         struct hammer_transaction trans;
1345         struct hammer_inode *dip;
1346         struct hammer_inode *ip;
1347         hammer_mount_t hmp;
1348         int64_t parent_obj_id;
1349         u_int32_t parent_obj_localization;
1350         hammer_tid_t asof;
1351         int error;
1352
1353         dip = VTOI(ap->a_dvp);
1354         asof = dip->obj_asof;
1355         hmp = dip->hmp;
1356
1357         /*
1358          * Whos are parent?  This could be the root of a pseudo-filesystem
1359          * whos parent is in another localization domain.
1360          */
1361         lwkt_gettoken(&hmp->fs_token);
1362         parent_obj_id = dip->ino_data.parent_obj_id;
1363         if (dip->obj_id == HAMMER_OBJID_ROOT)
1364                 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
1365         else
1366                 parent_obj_localization = dip->obj_localization;
1367
1368         if (parent_obj_id == 0) {
1369                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1370                    asof != hmp->asof) {
1371                         parent_obj_id = dip->obj_id;
1372                         asof = hmp->asof;
1373                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1374                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
1375                                   (long long)dip->obj_asof);
1376                 } else {
1377                         *ap->a_vpp = NULL;
1378                         lwkt_reltoken(&hmp->fs_token);
1379                         return ENOENT;
1380                 }
1381         }
1382
1383         hammer_simple_transaction(&trans, hmp);
1384         ++hammer_stats_file_iopsr;
1385
1386         ip = hammer_get_inode(&trans, dip, parent_obj_id,
1387                               asof, parent_obj_localization,
1388                               dip->flags, &error);
1389         if (ip) {
1390                 error = hammer_get_vnode(ip, ap->a_vpp);
1391                 hammer_rel_inode(ip, 0);
1392         } else {
1393                 *ap->a_vpp = NULL;
1394         }
1395         hammer_done_transaction(&trans);
1396         lwkt_reltoken(&hmp->fs_token);
1397         return (error);
1398 }
1399
1400 /*
1401  * hammer_vop_nlink { nch, dvp, vp, cred }
1402  */
1403 static
1404 int
1405 hammer_vop_nlink(struct vop_nlink_args *ap)
1406 {
1407         struct hammer_transaction trans;
1408         struct hammer_inode *dip;
1409         struct hammer_inode *ip;
1410         struct nchandle *nch;
1411         hammer_mount_t hmp;
1412         int error;
1413
1414         if (ap->a_dvp->v_mount != ap->a_vp->v_mount)    
1415                 return(EXDEV);
1416
1417         nch = ap->a_nch;
1418         dip = VTOI(ap->a_dvp);
1419         ip = VTOI(ap->a_vp);
1420         hmp = dip->hmp;
1421
1422         if (dip->obj_localization != ip->obj_localization)
1423                 return(EXDEV);
1424
1425         if (dip->flags & HAMMER_INODE_RO)
1426                 return (EROFS);
1427         if (ip->flags & HAMMER_INODE_RO)
1428                 return (EROFS);
1429         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1430                 return (error);
1431
1432         /*
1433          * Create a transaction to cover the operations we perform.
1434          */
1435         lwkt_gettoken(&hmp->fs_token);
1436         hammer_start_transaction(&trans, hmp);
1437         ++hammer_stats_file_iopsw;
1438
1439         /*
1440          * Add the filesystem object to the directory.  Note that neither
1441          * dip nor ip are referenced or locked, but their vnodes are
1442          * referenced.  This function will bump the inode's link count.
1443          */
1444         error = hammer_ip_add_directory(&trans, dip,
1445                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1446                                         ip);
1447
1448         /*
1449          * Finish up.
1450          */
1451         if (error == 0) {
1452                 cache_setunresolved(nch);
1453                 cache_setvp(nch, ap->a_vp);
1454         }
1455         hammer_done_transaction(&trans);
1456         hammer_knote(ap->a_vp, NOTE_LINK);
1457         hammer_knote(ap->a_dvp, NOTE_WRITE);
1458         lwkt_reltoken(&hmp->fs_token);
1459         return (error);
1460 }
1461
1462 /*
1463  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1464  *
1465  * The operating system has already ensured that the directory entry
1466  * does not exist and done all appropriate namespace locking.
1467  */
1468 static
1469 int
1470 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1471 {
1472         struct hammer_transaction trans;
1473         struct hammer_inode *dip;
1474         struct hammer_inode *nip;
1475         struct nchandle *nch;
1476         hammer_mount_t hmp;
1477         int error;
1478
1479         nch = ap->a_nch;
1480         dip = VTOI(ap->a_dvp);
1481         hmp = dip->hmp;
1482
1483         if (dip->flags & HAMMER_INODE_RO)
1484                 return (EROFS);
1485         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1486                 return (error);
1487
1488         /*
1489          * Create a transaction to cover the operations we perform.
1490          */
1491         lwkt_gettoken(&hmp->fs_token);
1492         hammer_start_transaction(&trans, hmp);
1493         ++hammer_stats_file_iopsw;
1494
1495         /*
1496          * Create a new filesystem object of the requested type.  The
1497          * returned inode will be referenced but not locked.
1498          */
1499         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1500                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1501                                     NULL, &nip);
1502         if (error) {
1503                 hkprintf("hammer_mkdir error %d\n", error);
1504                 hammer_done_transaction(&trans);
1505                 *ap->a_vpp = NULL;
1506                 lwkt_reltoken(&hmp->fs_token);
1507                 return (error);
1508         }
1509         /*
1510          * Add the new filesystem object to the directory.  This will also
1511          * bump the inode's link count.
1512          */
1513         error = hammer_ip_add_directory(&trans, dip,
1514                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1515                                         nip);
1516         if (error)
1517                 hkprintf("hammer_mkdir (add) error %d\n", error);
1518
1519         /*
1520          * Finish up.
1521          */
1522         if (error) {
1523                 hammer_rel_inode(nip, 0);
1524                 *ap->a_vpp = NULL;
1525         } else {
1526                 error = hammer_get_vnode(nip, ap->a_vpp);
1527                 hammer_rel_inode(nip, 0);
1528                 if (error == 0) {
1529                         cache_setunresolved(ap->a_nch);
1530                         cache_setvp(ap->a_nch, *ap->a_vpp);
1531                 }
1532         }
1533         hammer_done_transaction(&trans);
1534         if (error == 0)
1535                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1536         lwkt_reltoken(&hmp->fs_token);
1537         return (error);
1538 }
1539
1540 /*
1541  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1542  *
1543  * The operating system has already ensured that the directory entry
1544  * does not exist and done all appropriate namespace locking.
1545  */
1546 static
1547 int
1548 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1549 {
1550         struct hammer_transaction trans;
1551         struct hammer_inode *dip;
1552         struct hammer_inode *nip;
1553         struct nchandle *nch;
1554         hammer_mount_t hmp;
1555         int error;
1556
1557         nch = ap->a_nch;
1558         dip = VTOI(ap->a_dvp);
1559         hmp = dip->hmp;
1560
1561         if (dip->flags & HAMMER_INODE_RO)
1562                 return (EROFS);
1563         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1564                 return (error);
1565
1566         /*
1567          * Create a transaction to cover the operations we perform.
1568          */
1569         lwkt_gettoken(&hmp->fs_token);
1570         hammer_start_transaction(&trans, hmp);
1571         ++hammer_stats_file_iopsw;
1572
1573         /*
1574          * Create a new filesystem object of the requested type.  The
1575          * returned inode will be referenced but not locked.
1576          *
1577          * If mknod specifies a directory a pseudo-fs is created.
1578          */
1579         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1580                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1581                                     NULL, &nip);
1582         if (error) {
1583                 hammer_done_transaction(&trans);
1584                 *ap->a_vpp = NULL;
1585                 lwkt_reltoken(&hmp->fs_token);
1586                 return (error);
1587         }
1588
1589         /*
1590          * Add the new filesystem object to the directory.  This will also
1591          * bump the inode's link count.
1592          */
1593         error = hammer_ip_add_directory(&trans, dip,
1594                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1595                                         nip);
1596
1597         /*
1598          * Finish up.
1599          */
1600         if (error) {
1601                 hammer_rel_inode(nip, 0);
1602                 *ap->a_vpp = NULL;
1603         } else {
1604                 error = hammer_get_vnode(nip, ap->a_vpp);
1605                 hammer_rel_inode(nip, 0);
1606                 if (error == 0) {
1607                         cache_setunresolved(ap->a_nch);
1608                         cache_setvp(ap->a_nch, *ap->a_vpp);
1609                 }
1610         }
1611         hammer_done_transaction(&trans);
1612         if (error == 0)
1613                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1614         lwkt_reltoken(&hmp->fs_token);
1615         return (error);
1616 }
1617
1618 /*
1619  * hammer_vop_open { vp, mode, cred, fp }
1620  *
1621  * MPSAFE (does not require fs_token)
1622  */
1623 static
1624 int
1625 hammer_vop_open(struct vop_open_args *ap)
1626 {
1627         hammer_inode_t ip;
1628
1629         ++hammer_stats_file_iopsr;
1630         ip = VTOI(ap->a_vp);
1631
1632         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1633                 return (EROFS);
1634         return(vop_stdopen(ap));
1635 }
1636
1637 /*
1638  * hammer_vop_print { vp }
1639  */
1640 static
1641 int
1642 hammer_vop_print(struct vop_print_args *ap)
1643 {
1644         return EOPNOTSUPP;
1645 }
1646
1647 /*
1648  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1649  */
1650 static
1651 int
1652 hammer_vop_readdir(struct vop_readdir_args *ap)
1653 {
1654         struct hammer_transaction trans;
1655         struct hammer_cursor cursor;
1656         struct hammer_inode *ip;
1657         hammer_mount_t hmp;
1658         struct uio *uio;
1659         hammer_base_elm_t base;
1660         int error;
1661         int cookie_index;
1662         int ncookies;
1663         off_t *cookies;
1664         off_t saveoff;
1665         int r;
1666         int dtype;
1667
1668         ++hammer_stats_file_iopsr;
1669         ip = VTOI(ap->a_vp);
1670         uio = ap->a_uio;
1671         saveoff = uio->uio_offset;
1672         hmp = ip->hmp;
1673
1674         if (ap->a_ncookies) {
1675                 ncookies = uio->uio_resid / 16 + 1;
1676                 if (ncookies > 1024)
1677                         ncookies = 1024;
1678                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1679                 cookie_index = 0;
1680         } else {
1681                 ncookies = -1;
1682                 cookies = NULL;
1683                 cookie_index = 0;
1684         }
1685
1686         lwkt_gettoken(&hmp->fs_token);
1687         hammer_simple_transaction(&trans, hmp);
1688
1689         /*
1690          * Handle artificial entries
1691          *
1692          * It should be noted that the minimum value for a directory
1693          * hash key on-media is 0x0000000100000000, so we can use anything
1694          * less then that to represent our 'special' key space.
1695          */
1696         error = 0;
1697         if (saveoff == 0) {
1698                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1699                 if (r)
1700                         goto done;
1701                 if (cookies)
1702                         cookies[cookie_index] = saveoff;
1703                 ++saveoff;
1704                 ++cookie_index;
1705                 if (cookie_index == ncookies)
1706                         goto done;
1707         }
1708         if (saveoff == 1) {
1709                 if (ip->ino_data.parent_obj_id) {
1710                         r = vop_write_dirent(&error, uio,
1711                                              ip->ino_data.parent_obj_id,
1712                                              DT_DIR, 2, "..");
1713                 } else {
1714                         r = vop_write_dirent(&error, uio,
1715                                              ip->obj_id, DT_DIR, 2, "..");
1716                 }
1717                 if (r)
1718                         goto done;
1719                 if (cookies)
1720                         cookies[cookie_index] = saveoff;
1721                 ++saveoff;
1722                 ++cookie_index;
1723                 if (cookie_index == ncookies)
1724                         goto done;
1725         }
1726
1727         /*
1728          * Key range (begin and end inclusive) to scan.  Directory keys
1729          * directly translate to a 64 bit 'seek' position.
1730          */
1731         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1732         cursor.key_beg.localization = ip->obj_localization +
1733                                       hammer_dir_localization(ip);
1734         cursor.key_beg.obj_id = ip->obj_id;
1735         cursor.key_beg.create_tid = 0;
1736         cursor.key_beg.delete_tid = 0;
1737         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1738         cursor.key_beg.obj_type = 0;
1739         cursor.key_beg.key = saveoff;
1740
1741         cursor.key_end = cursor.key_beg;
1742         cursor.key_end.key = HAMMER_MAX_KEY;
1743         cursor.asof = ip->obj_asof;
1744         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1745
1746         error = hammer_ip_first(&cursor);
1747
1748         while (error == 0) {
1749                 error = hammer_ip_resolve_data(&cursor);
1750                 if (error)
1751                         break;
1752                 base = &cursor.leaf->base;
1753                 saveoff = base->key;
1754                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1755
1756                 if (base->obj_id != ip->obj_id)
1757                         panic("readdir: bad record at %p", cursor.node);
1758
1759                 /*
1760                  * Convert pseudo-filesystems into softlinks
1761                  */
1762                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1763                 r = vop_write_dirent(
1764                              &error, uio, cursor.data->entry.obj_id,
1765                              dtype,
1766                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1767                              (void *)cursor.data->entry.name);
1768                 if (r)
1769                         break;
1770                 ++saveoff;
1771                 if (cookies)
1772                         cookies[cookie_index] = base->key;
1773                 ++cookie_index;
1774                 if (cookie_index == ncookies)
1775                         break;
1776                 error = hammer_ip_next(&cursor);
1777         }
1778         hammer_done_cursor(&cursor);
1779
1780 done:
1781         hammer_done_transaction(&trans);
1782
1783         if (ap->a_eofflag)
1784                 *ap->a_eofflag = (error == ENOENT);
1785         uio->uio_offset = saveoff;
1786         if (error && cookie_index == 0) {
1787                 if (error == ENOENT)
1788                         error = 0;
1789                 if (cookies) {
1790                         kfree(cookies, M_TEMP);
1791                         *ap->a_ncookies = 0;
1792                         *ap->a_cookies = NULL;
1793                 }
1794         } else {
1795                 if (error == ENOENT)
1796                         error = 0;
1797                 if (cookies) {
1798                         *ap->a_ncookies = cookie_index;
1799                         *ap->a_cookies = cookies;
1800                 }
1801         }
1802         lwkt_reltoken(&hmp->fs_token);
1803         return(error);
1804 }
1805
1806 /*
1807  * hammer_vop_readlink { vp, uio, cred }
1808  */
1809 static
1810 int
1811 hammer_vop_readlink(struct vop_readlink_args *ap)
1812 {
1813         struct hammer_transaction trans;
1814         struct hammer_cursor cursor;
1815         struct hammer_inode *ip;
1816         hammer_mount_t hmp;
1817         char buf[32];
1818         u_int32_t localization;
1819         hammer_pseudofs_inmem_t pfsm;
1820         int error;
1821
1822         ip = VTOI(ap->a_vp);
1823         hmp = ip->hmp;
1824
1825         lwkt_gettoken(&hmp->fs_token);
1826
1827         /*
1828          * Shortcut if the symlink data was stuffed into ino_data.
1829          *
1830          * Also expand special "@@PFS%05d" softlinks (expansion only
1831          * occurs for non-historical (current) accesses made from the
1832          * primary filesystem).
1833          */
1834         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1835                 char *ptr;
1836                 int bytes;
1837
1838                 ptr = ip->ino_data.ext.symlink;
1839                 bytes = (int)ip->ino_data.size;
1840                 if (bytes == 10 &&
1841                     ip->obj_asof == HAMMER_MAX_TID &&
1842                     ip->obj_localization == 0 &&
1843                     strncmp(ptr, "@@PFS", 5) == 0) {
1844                         hammer_simple_transaction(&trans, hmp);
1845                         bcopy(ptr + 5, buf, 5);
1846                         buf[5] = 0;
1847                         localization = strtoul(buf, NULL, 10) << 16;
1848                         pfsm = hammer_load_pseudofs(&trans, localization,
1849                                                     &error);
1850                         if (error == 0) {
1851                                 if (pfsm->pfsd.mirror_flags &
1852                                     HAMMER_PFSD_SLAVE) {
1853                                         /* vap->va_size == 26 */
1854                                         ksnprintf(buf, sizeof(buf),
1855                                                   "@@0x%016llx:%05d",
1856                                                   (long long)pfsm->pfsd.sync_end_tid,
1857                                                   localization >> 16);
1858                                 } else {
1859                                         /* vap->va_size == 10 */
1860                                         ksnprintf(buf, sizeof(buf),
1861                                                   "@@-1:%05d",
1862                                                   localization >> 16);
1863 #if 0
1864                                         ksnprintf(buf, sizeof(buf),
1865                                                   "@@0x%016llx:%05d",
1866                                                   (long long)HAMMER_MAX_TID,
1867                                                   localization >> 16);
1868 #endif
1869                                 }
1870                                 ptr = buf;
1871                                 bytes = strlen(buf);
1872                         }
1873                         if (pfsm)
1874                                 hammer_rel_pseudofs(hmp, pfsm);
1875                         hammer_done_transaction(&trans);
1876                 }
1877                 error = uiomove(ptr, bytes, ap->a_uio);
1878                 lwkt_reltoken(&hmp->fs_token);
1879                 return(error);
1880         }
1881
1882         /*
1883          * Long version
1884          */
1885         hammer_simple_transaction(&trans, hmp);
1886         ++hammer_stats_file_iopsr;
1887         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1888
1889         /*
1890          * Key range (begin and end inclusive) to scan.  Directory keys
1891          * directly translate to a 64 bit 'seek' position.
1892          */
1893         cursor.key_beg.localization = ip->obj_localization +
1894                                       HAMMER_LOCALIZE_MISC;
1895         cursor.key_beg.obj_id = ip->obj_id;
1896         cursor.key_beg.create_tid = 0;
1897         cursor.key_beg.delete_tid = 0;
1898         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1899         cursor.key_beg.obj_type = 0;
1900         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1901         cursor.asof = ip->obj_asof;
1902         cursor.flags |= HAMMER_CURSOR_ASOF;
1903
1904         error = hammer_ip_lookup(&cursor);
1905         if (error == 0) {
1906                 error = hammer_ip_resolve_data(&cursor);
1907                 if (error == 0) {
1908                         KKASSERT(cursor.leaf->data_len >=
1909                                  HAMMER_SYMLINK_NAME_OFF);
1910                         error = uiomove(cursor.data->symlink.name,
1911                                         cursor.leaf->data_len -
1912                                                 HAMMER_SYMLINK_NAME_OFF,
1913                                         ap->a_uio);
1914                 }
1915         }
1916         hammer_done_cursor(&cursor);
1917         hammer_done_transaction(&trans);
1918         lwkt_reltoken(&hmp->fs_token);
1919         return(error);
1920 }
1921
1922 /*
1923  * hammer_vop_nremove { nch, dvp, cred }
1924  */
1925 static
1926 int
1927 hammer_vop_nremove(struct vop_nremove_args *ap)
1928 {
1929         struct hammer_transaction trans;
1930         struct hammer_inode *dip;
1931         hammer_mount_t hmp;
1932         int error;
1933
1934         dip = VTOI(ap->a_dvp);
1935         hmp = dip->hmp;
1936
1937         if (hammer_nohistory(dip) == 0 &&
1938             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1939                 return (error);
1940         }
1941
1942         lwkt_gettoken(&hmp->fs_token);
1943         hammer_start_transaction(&trans, hmp);
1944         ++hammer_stats_file_iopsw;
1945         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1946         hammer_done_transaction(&trans);
1947         if (error == 0)
1948                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1949         lwkt_reltoken(&hmp->fs_token);
1950         return (error);
1951 }
1952
1953 /*
1954  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1955  */
1956 static
1957 int
1958 hammer_vop_nrename(struct vop_nrename_args *ap)
1959 {
1960         struct hammer_transaction trans;
1961         struct namecache *fncp;
1962         struct namecache *tncp;
1963         struct hammer_inode *fdip;
1964         struct hammer_inode *tdip;
1965         struct hammer_inode *ip;
1966         hammer_mount_t hmp;
1967         struct hammer_cursor cursor;
1968         int64_t namekey;
1969         u_int32_t max_iterations;
1970         int nlen, error;
1971
1972         if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount) 
1973                 return(EXDEV);
1974         if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1975                 return(EXDEV);
1976
1977         fdip = VTOI(ap->a_fdvp);
1978         tdip = VTOI(ap->a_tdvp);
1979         fncp = ap->a_fnch->ncp;
1980         tncp = ap->a_tnch->ncp;
1981         ip = VTOI(fncp->nc_vp);
1982         KKASSERT(ip != NULL);
1983
1984         hmp = ip->hmp;
1985
1986         if (fdip->obj_localization != tdip->obj_localization)
1987                 return(EXDEV);
1988         if (fdip->obj_localization != ip->obj_localization)
1989                 return(EXDEV);
1990
1991         if (fdip->flags & HAMMER_INODE_RO)
1992                 return (EROFS);
1993         if (tdip->flags & HAMMER_INODE_RO)
1994                 return (EROFS);
1995         if (ip->flags & HAMMER_INODE_RO)
1996                 return (EROFS);
1997         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1998                 return (error);
1999
2000         lwkt_gettoken(&hmp->fs_token);
2001         hammer_start_transaction(&trans, hmp);
2002         ++hammer_stats_file_iopsw;
2003
2004         /*
2005          * Remove tncp from the target directory and then link ip as
2006          * tncp. XXX pass trans to dounlink
2007          *
2008          * Force the inode sync-time to match the transaction so it is
2009          * in-sync with the creation of the target directory entry.
2010          */
2011         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
2012                                 ap->a_cred, 0, -1);
2013         if (error == 0 || error == ENOENT) {
2014                 error = hammer_ip_add_directory(&trans, tdip,
2015                                                 tncp->nc_name, tncp->nc_nlen,
2016                                                 ip);
2017                 if (error == 0) {
2018                         ip->ino_data.parent_obj_id = tdip->obj_id;
2019                         ip->ino_data.ctime = trans.time;
2020                         hammer_modify_inode(&trans, ip, HAMMER_INODE_DDIRTY);
2021                 }
2022         }
2023         if (error)
2024                 goto failed; /* XXX */
2025
2026         /*
2027          * Locate the record in the originating directory and remove it.
2028          *
2029          * Calculate the namekey and setup the key range for the scan.  This
2030          * works kinda like a chained hash table where the lower 32 bits
2031          * of the namekey synthesize the chain.
2032          *
2033          * The key range is inclusive of both key_beg and key_end.
2034          */
2035         namekey = hammer_directory_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
2036                                            &max_iterations);
2037 retry:
2038         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
2039         cursor.key_beg.localization = fdip->obj_localization +
2040                                       hammer_dir_localization(fdip);
2041         cursor.key_beg.obj_id = fdip->obj_id;
2042         cursor.key_beg.key = namekey;
2043         cursor.key_beg.create_tid = 0;
2044         cursor.key_beg.delete_tid = 0;
2045         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2046         cursor.key_beg.obj_type = 0;
2047
2048         cursor.key_end = cursor.key_beg;
2049         cursor.key_end.key += max_iterations;
2050         cursor.asof = fdip->obj_asof;
2051         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2052
2053         /*
2054          * Scan all matching records (the chain), locate the one matching
2055          * the requested path component.
2056          *
2057          * The hammer_ip_*() functions merge in-memory records with on-disk
2058          * records for the purposes of the search.
2059          */
2060         error = hammer_ip_first(&cursor);
2061         while (error == 0) {
2062                 if (hammer_ip_resolve_data(&cursor) != 0)
2063                         break;
2064                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2065                 KKASSERT(nlen > 0);
2066                 if (fncp->nc_nlen == nlen &&
2067                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2068                         break;
2069                 }
2070                 error = hammer_ip_next(&cursor);
2071         }
2072
2073         /*
2074          * If all is ok we have to get the inode so we can adjust nlinks.
2075          *
2076          * WARNING: hammer_ip_del_directory() may have to terminate the
2077          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
2078          * twice.
2079          */
2080         if (error == 0)
2081                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
2082
2083         /*
2084          * XXX A deadlock here will break rename's atomicy for the purposes
2085          * of crash recovery.
2086          */
2087         if (error == EDEADLK) {
2088                 hammer_done_cursor(&cursor);
2089                 goto retry;
2090         }
2091
2092         /*
2093          * Cleanup and tell the kernel that the rename succeeded.
2094          *
2095          * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2096          *       without formally acquiring the vp since the vp might
2097          *       have zero refs on it, or in the middle of a reclaim,
2098          *       etc.
2099          */
2100         hammer_done_cursor(&cursor);
2101         if (error == 0) {
2102                 cache_rename(ap->a_fnch, ap->a_tnch);
2103                 hammer_knote(ap->a_fdvp, NOTE_WRITE);
2104                 hammer_knote(ap->a_tdvp, NOTE_WRITE);
2105                 while (ip->vp) {
2106                         struct vnode *vp;
2107
2108                         error = hammer_get_vnode(ip, &vp);
2109                         if (error == 0 && vp) {
2110                                 vn_unlock(vp);
2111                                 hammer_knote(ip->vp, NOTE_RENAME);
2112                                 vrele(vp);
2113                                 break;
2114                         }
2115                         kprintf("Debug: HAMMER ip/vp race2 avoided\n");
2116                 }
2117         }
2118
2119 failed:
2120         hammer_done_transaction(&trans);
2121         lwkt_reltoken(&hmp->fs_token);
2122         return (error);
2123 }
2124
2125 /*
2126  * hammer_vop_nrmdir { nch, dvp, cred }
2127  */
2128 static
2129 int
2130 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
2131 {
2132         struct hammer_transaction trans;
2133         struct hammer_inode *dip;
2134         hammer_mount_t hmp;
2135         int error;
2136
2137         dip = VTOI(ap->a_dvp);
2138         hmp = dip->hmp;
2139
2140         if (hammer_nohistory(dip) == 0 &&
2141             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2142                 return (error);
2143         }
2144
2145         lwkt_gettoken(&hmp->fs_token);
2146         hammer_start_transaction(&trans, hmp);
2147         ++hammer_stats_file_iopsw;
2148         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
2149         hammer_done_transaction(&trans);
2150         if (error == 0)
2151                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
2152         lwkt_reltoken(&hmp->fs_token);
2153         return (error);
2154 }
2155
2156 /*
2157  * hammer_vop_markatime { vp, cred }
2158  */
2159 static
2160 int
2161 hammer_vop_markatime(struct vop_markatime_args *ap)
2162 {
2163         struct hammer_transaction trans;
2164         struct hammer_inode *ip;
2165         hammer_mount_t hmp;
2166
2167         ip = VTOI(ap->a_vp);
2168         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2169                 return (EROFS);
2170         if (ip->flags & HAMMER_INODE_RO)
2171                 return (EROFS);
2172         hmp = ip->hmp;
2173         if (hmp->mp->mnt_flag & MNT_NOATIME)
2174                 return (0);
2175         lwkt_gettoken(&hmp->fs_token);
2176         hammer_start_transaction(&trans, hmp);
2177         ++hammer_stats_file_iopsw;
2178
2179         ip->ino_data.atime = trans.time;
2180         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
2181         hammer_done_transaction(&trans);
2182         hammer_knote(ap->a_vp, NOTE_ATTRIB);
2183         lwkt_reltoken(&hmp->fs_token);
2184         return (0);
2185 }
2186
2187 /*
2188  * hammer_vop_setattr { vp, vap, cred }
2189  */
2190 static
2191 int
2192 hammer_vop_setattr(struct vop_setattr_args *ap)
2193 {
2194         struct hammer_transaction trans;
2195         struct hammer_inode *ip;
2196         struct vattr *vap;
2197         hammer_mount_t hmp;
2198         int modflags;
2199         int error;
2200         int truncating;
2201         int blksize;
2202         int kflags;
2203 #if 0
2204         int64_t aligned_size;
2205 #endif
2206         u_int32_t flags;
2207
2208         vap = ap->a_vap;
2209         ip = ap->a_vp->v_data;
2210         modflags = 0;
2211         kflags = 0;
2212         hmp = ip->hmp;
2213
2214         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2215                 return(EROFS);
2216         if (ip->flags & HAMMER_INODE_RO)
2217                 return (EROFS);
2218         if (hammer_nohistory(ip) == 0 &&
2219             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2220                 return (error);
2221         }
2222
2223         lwkt_gettoken(&hmp->fs_token);
2224         hammer_start_transaction(&trans, hmp);
2225         ++hammer_stats_file_iopsw;
2226         error = 0;
2227
2228         if (vap->va_flags != VNOVAL) {
2229                 flags = ip->ino_data.uflags;
2230                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
2231                                          hammer_to_unix_xid(&ip->ino_data.uid),
2232                                          ap->a_cred);
2233                 if (error == 0) {
2234                         if (ip->ino_data.uflags != flags) {
2235                                 ip->ino_data.uflags = flags;
2236                                 ip->ino_data.ctime = trans.time;
2237                                 modflags |= HAMMER_INODE_DDIRTY;
2238                                 kflags |= NOTE_ATTRIB;
2239                         }
2240                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2241                                 error = 0;
2242                                 goto done;
2243                         }
2244                 }
2245                 goto done;
2246         }
2247         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2248                 error = EPERM;
2249                 goto done;
2250         }
2251         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2252                 mode_t cur_mode = ip->ino_data.mode;
2253                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2254                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2255                 uuid_t uuid_uid;
2256                 uuid_t uuid_gid;
2257
2258                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2259                                          ap->a_cred,
2260                                          &cur_uid, &cur_gid, &cur_mode);
2261                 if (error == 0) {
2262                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
2263                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
2264                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
2265                                  sizeof(uuid_uid)) ||
2266                             bcmp(&uuid_gid, &ip->ino_data.gid,
2267                                  sizeof(uuid_gid)) ||
2268                             ip->ino_data.mode != cur_mode
2269                         ) {
2270                                 ip->ino_data.uid = uuid_uid;
2271                                 ip->ino_data.gid = uuid_gid;
2272                                 ip->ino_data.mode = cur_mode;
2273                                 ip->ino_data.ctime = trans.time;
2274                                 modflags |= HAMMER_INODE_DDIRTY;
2275                         }
2276                         kflags |= NOTE_ATTRIB;
2277                 }
2278         }
2279         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2280                 switch(ap->a_vp->v_type) {
2281                 case VREG:
2282                         if (vap->va_size == ip->ino_data.size)
2283                                 break;
2284
2285                         /*
2286                          * Log the operation if in fast-fsync mode or if
2287                          * there are unterminated redo write records present.
2288                          *
2289                          * The second check is needed so the recovery code
2290                          * properly truncates write redos even if nominal
2291                          * REDO operations is turned off due to excessive
2292                          * writes, because the related records might be
2293                          * destroyed and never lay down a TERM_WRITE.
2294                          */
2295                         if ((ip->flags & HAMMER_INODE_REDO) ||
2296                             (ip->flags & HAMMER_INODE_RDIRTY)) {
2297                                 error = hammer_generate_redo(&trans, ip,
2298                                                              vap->va_size,
2299                                                              HAMMER_REDO_TRUNC,
2300                                                              NULL, 0);
2301                         }
2302                         blksize = hammer_blocksize(vap->va_size);
2303
2304                         /*
2305                          * XXX break atomicy, we can deadlock the backend
2306                          * if we do not release the lock.  Probably not a
2307                          * big deal here.
2308                          */
2309                         if (vap->va_size < ip->ino_data.size) {
2310                                 nvtruncbuf(ap->a_vp, vap->va_size,
2311                                            blksize,
2312                                            hammer_blockoff(vap->va_size));
2313                                 truncating = 1;
2314                                 kflags |= NOTE_WRITE;
2315                         } else {
2316                                 nvextendbuf(ap->a_vp,
2317                                             ip->ino_data.size,
2318                                             vap->va_size,
2319                                             hammer_blocksize(ip->ino_data.size),
2320                                             hammer_blocksize(vap->va_size),
2321                                             hammer_blockoff(ip->ino_data.size),
2322                                             hammer_blockoff(vap->va_size),
2323                                             0);
2324                                 truncating = 0;
2325                                 kflags |= NOTE_WRITE | NOTE_EXTEND;
2326                         }
2327                         ip->ino_data.size = vap->va_size;
2328                         ip->ino_data.mtime = trans.time;
2329                         /* XXX safe to use SDIRTY instead of DDIRTY here? */
2330                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2331
2332                         /*
2333                          * On-media truncation is cached in the inode until
2334                          * the inode is synchronized.  We must immediately
2335                          * handle any frontend records.
2336                          */
2337                         if (truncating) {
2338                                 hammer_ip_frontend_trunc(ip, vap->va_size);
2339 #ifdef DEBUG_TRUNCATE
2340                                 if (HammerTruncIp == NULL)
2341                                         HammerTruncIp = ip;
2342 #endif
2343                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2344                                         ip->flags |= HAMMER_INODE_TRUNCATED;
2345                                         ip->trunc_off = vap->va_size;
2346 #ifdef DEBUG_TRUNCATE
2347                                         if (ip == HammerTruncIp)
2348                                         kprintf("truncate1 %016llx\n",
2349                                                 (long long)ip->trunc_off);
2350 #endif
2351                                 } else if (ip->trunc_off > vap->va_size) {
2352                                         ip->trunc_off = vap->va_size;
2353 #ifdef DEBUG_TRUNCATE
2354                                         if (ip == HammerTruncIp)
2355                                         kprintf("truncate2 %016llx\n",
2356                                                 (long long)ip->trunc_off);
2357 #endif
2358                                 } else {
2359 #ifdef DEBUG_TRUNCATE
2360                                         if (ip == HammerTruncIp)
2361                                         kprintf("truncate3 %016llx (ignored)\n",
2362                                                 (long long)vap->va_size);
2363 #endif
2364                                 }
2365                         }
2366
2367 #if 0
2368                         /*
2369                          * When truncating, nvtruncbuf() may have cleaned out
2370                          * a portion of the last block on-disk in the buffer
2371                          * cache.  We must clean out any frontend records
2372                          * for blocks beyond the new last block.
2373                          */
2374                         aligned_size = (vap->va_size + (blksize - 1)) &
2375                                        ~(int64_t)(blksize - 1);
2376                         if (truncating && vap->va_size < aligned_size) {
2377                                 aligned_size -= blksize;
2378                                 hammer_ip_frontend_trunc(ip, aligned_size);
2379                         }
2380 #endif
2381                         break;
2382                 case VDATABASE:
2383                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2384                                 ip->flags |= HAMMER_INODE_TRUNCATED;
2385                                 ip->trunc_off = vap->va_size;
2386                         } else if (ip->trunc_off > vap->va_size) {
2387                                 ip->trunc_off = vap->va_size;
2388                         }
2389                         hammer_ip_frontend_trunc(ip, vap->va_size);
2390                         ip->ino_data.size = vap->va_size;
2391                         ip->ino_data.mtime = trans.time;
2392                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2393                         kflags |= NOTE_ATTRIB;
2394                         break;
2395                 default:
2396                         error = EINVAL;
2397                         goto done;
2398                 }
2399                 break;
2400         }
2401         if (vap->va_atime.tv_sec != VNOVAL) {
2402                 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2403                 modflags |= HAMMER_INODE_ATIME;
2404                 kflags |= NOTE_ATTRIB;
2405         }
2406         if (vap->va_mtime.tv_sec != VNOVAL) {
2407                 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2408                 modflags |= HAMMER_INODE_MTIME;
2409                 kflags |= NOTE_ATTRIB;
2410         }
2411         if (vap->va_mode != (mode_t)VNOVAL) {
2412                 mode_t   cur_mode = ip->ino_data.mode;
2413                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2414                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2415
2416                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2417                                          cur_uid, cur_gid, &cur_mode);
2418                 if (error == 0 && ip->ino_data.mode != cur_mode) {
2419                         ip->ino_data.mode = cur_mode;
2420                         ip->ino_data.ctime = trans.time;
2421                         modflags |= HAMMER_INODE_DDIRTY;
2422                         kflags |= NOTE_ATTRIB;
2423                 }
2424         }
2425 done:
2426         if (error == 0)
2427                 hammer_modify_inode(&trans, ip, modflags);
2428         hammer_done_transaction(&trans);
2429         hammer_knote(ap->a_vp, kflags);
2430         lwkt_reltoken(&hmp->fs_token);
2431         return (error);
2432 }
2433
2434 /*
2435  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2436  */
2437 static
2438 int
2439 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2440 {
2441         struct hammer_transaction trans;
2442         struct hammer_inode *dip;
2443         struct hammer_inode *nip;
2444         hammer_record_t record;
2445         struct nchandle *nch;
2446         hammer_mount_t hmp;
2447         int error;
2448         int bytes;
2449
2450         ap->a_vap->va_type = VLNK;
2451
2452         nch = ap->a_nch;
2453         dip = VTOI(ap->a_dvp);
2454         hmp = dip->hmp;
2455
2456         if (dip->flags & HAMMER_INODE_RO)
2457                 return (EROFS);
2458         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
2459                 return (error);
2460
2461         /*
2462          * Create a transaction to cover the operations we perform.
2463          */
2464         lwkt_gettoken(&hmp->fs_token);
2465         hammer_start_transaction(&trans, hmp);
2466         ++hammer_stats_file_iopsw;
2467
2468         /*
2469          * Create a new filesystem object of the requested type.  The
2470          * returned inode will be referenced but not locked.
2471          */
2472
2473         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2474                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2475                                     NULL, &nip);
2476         if (error) {
2477                 hammer_done_transaction(&trans);
2478                 *ap->a_vpp = NULL;
2479                 lwkt_reltoken(&hmp->fs_token);
2480                 return (error);
2481         }
2482
2483         /*
2484          * Add a record representing the symlink.  symlink stores the link
2485          * as pure data, not a string, and is no \0 terminated.
2486          */
2487         if (error == 0) {
2488                 bytes = strlen(ap->a_target);
2489
2490                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2491                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2492                 } else {
2493                         record = hammer_alloc_mem_record(nip, bytes);
2494                         record->type = HAMMER_MEM_RECORD_GENERAL;
2495
2496                         record->leaf.base.localization = nip->obj_localization +
2497                                                          HAMMER_LOCALIZE_MISC;
2498                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2499                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2500                         record->leaf.data_len = bytes;
2501                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2502                         bcopy(ap->a_target, record->data->symlink.name, bytes);
2503                         error = hammer_ip_add_record(&trans, record);
2504                 }
2505
2506                 /*
2507                  * Set the file size to the length of the link.
2508                  */
2509                 if (error == 0) {
2510                         nip->ino_data.size = bytes;
2511                         hammer_modify_inode(&trans, nip, HAMMER_INODE_DDIRTY);
2512                 }
2513         }
2514         if (error == 0)
2515                 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
2516                                                 nch->ncp->nc_nlen, nip);
2517
2518         /*
2519          * Finish up.
2520          */
2521         if (error) {
2522                 hammer_rel_inode(nip, 0);
2523                 *ap->a_vpp = NULL;
2524         } else {
2525                 error = hammer_get_vnode(nip, ap->a_vpp);
2526                 hammer_rel_inode(nip, 0);
2527                 if (error == 0) {
2528                         cache_setunresolved(ap->a_nch);
2529                         cache_setvp(ap->a_nch, *ap->a_vpp);
2530                         hammer_knote(ap->a_dvp, NOTE_WRITE);
2531                 }
2532         }
2533         hammer_done_transaction(&trans);
2534         lwkt_reltoken(&hmp->fs_token);
2535         return (error);
2536 }
2537
2538 /*
2539  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2540  */
2541 static
2542 int
2543 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2544 {
2545         struct hammer_transaction trans;
2546         struct hammer_inode *dip;
2547         hammer_mount_t hmp;
2548         int error;
2549
2550         dip = VTOI(ap->a_dvp);
2551         hmp = dip->hmp;
2552
2553         if (hammer_nohistory(dip) == 0 &&
2554             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2555                 return (error);
2556         }
2557
2558         lwkt_gettoken(&hmp->fs_token);
2559         hammer_start_transaction(&trans, hmp);
2560         ++hammer_stats_file_iopsw;
2561         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2562                                 ap->a_cred, ap->a_flags, -1);
2563         hammer_done_transaction(&trans);
2564         lwkt_reltoken(&hmp->fs_token);
2565
2566         return (error);
2567 }
2568
2569 /*
2570  * hammer_vop_ioctl { vp, command, data, fflag, cred }
2571  */
2572 static
2573 int
2574 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2575 {
2576         struct hammer_inode *ip = ap->a_vp->v_data;
2577         hammer_mount_t hmp = ip->hmp;
2578         int error;
2579
2580         ++hammer_stats_file_iopsr;
2581         lwkt_gettoken(&hmp->fs_token);
2582         error = hammer_ioctl(ip, ap->a_command, ap->a_data,
2583                              ap->a_fflag, ap->a_cred);
2584         lwkt_reltoken(&hmp->fs_token);
2585         return (error);
2586 }
2587
2588 static
2589 int
2590 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2591 {
2592         static const struct mountctl_opt extraopt[] = {
2593                 { HMNT_NOHISTORY,       "nohistory" },
2594                 { HMNT_MASTERID,        "master" },
2595                 { 0, NULL}
2596
2597         };
2598         struct hammer_mount *hmp;
2599         struct mount *mp;
2600         int usedbytes;
2601         int error;
2602
2603         error = 0;
2604         usedbytes = 0;
2605         mp = ap->a_head.a_ops->head.vv_mount;
2606         KKASSERT(mp->mnt_data != NULL);
2607         hmp = (struct hammer_mount *)mp->mnt_data;
2608
2609         lwkt_gettoken(&hmp->fs_token);
2610
2611         switch(ap->a_op) {
2612         case MOUNTCTL_SET_EXPORT:
2613                 if (ap->a_ctllen != sizeof(struct export_args))
2614                         error = EINVAL;
2615                 else
2616                         error = hammer_vfs_export(mp, ap->a_op,
2617                                       (const struct export_args *)ap->a_ctl);
2618                 break;
2619         case MOUNTCTL_MOUNTFLAGS:
2620         {
2621                 /*
2622                  * Call standard mountctl VOP function
2623                  * so we get user mount flags.
2624                  */
2625                 error = vop_stdmountctl(ap);
2626                 if (error)
2627                         break;
2628
2629                 usedbytes = *ap->a_res;
2630
2631                 if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2632                         usedbytes += vfs_flagstostr(hmp->hflags, extraopt,
2633                                                     ap->a_buf,
2634                                                     ap->a_buflen - usedbytes,
2635                                                     &error);
2636                 }
2637
2638                 *ap->a_res += usedbytes;
2639                 break;
2640         }
2641         default:
2642                 error = vop_stdmountctl(ap);
2643                 break;
2644         }
2645         lwkt_reltoken(&hmp->fs_token);
2646         return(error);
2647 }
2648
2649 /*
2650  * hammer_vop_strategy { vp, bio }
2651  *
2652  * Strategy call, used for regular file read & write only.  Note that the
2653  * bp may represent a cluster.
2654  *
2655  * To simplify operation and allow better optimizations in the future,
2656  * this code does not make any assumptions with regards to buffer alignment
2657  * or size.
2658  */
2659 static
2660 int
2661 hammer_vop_strategy(struct vop_strategy_args *ap)
2662 {
2663         struct buf *bp;
2664         int error;
2665
2666         bp = ap->a_bio->bio_buf;
2667
2668         switch(bp->b_cmd) {
2669         case BUF_CMD_READ:
2670                 error = hammer_vop_strategy_read(ap);
2671                 break;
2672         case BUF_CMD_WRITE:
2673                 error = hammer_vop_strategy_write(ap);
2674                 break;
2675         default:
2676                 bp->b_error = error = EINVAL;
2677                 bp->b_flags |= B_ERROR;
2678                 biodone(ap->a_bio);
2679                 break;
2680         }
2681
2682         /* hammer_dump_dedup_cache(((hammer_inode_t)ap->a_vp->v_data)->hmp); */
2683
2684         return (error);
2685 }
2686
2687 /*
2688  * Read from a regular file.  Iterate the related records and fill in the
2689  * BIO/BUF.  Gaps are zero-filled.
2690  *
2691  * The support code in hammer_object.c should be used to deal with mixed
2692  * in-memory and on-disk records.
2693  *
2694  * NOTE: Can be called from the cluster code with an oversized buf.
2695  *
2696  * XXX atime update
2697  */
2698 static
2699 int
2700 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2701 {
2702         struct hammer_transaction trans;
2703         struct hammer_inode *ip;
2704         struct hammer_inode *dip;
2705         hammer_mount_t hmp;
2706         struct hammer_cursor cursor;
2707         hammer_base_elm_t base;
2708         hammer_off_t disk_offset;
2709         struct bio *bio;
2710         struct bio *nbio;
2711         struct buf *bp;
2712         int64_t rec_offset;
2713         int64_t ran_end;
2714         int64_t tmp64;
2715         int error;
2716         int boff;
2717         int roff;
2718         int n;
2719         int isdedupable;
2720
2721         bio = ap->a_bio;
2722         bp = bio->bio_buf;
2723         ip = ap->a_vp->v_data;
2724         hmp = ip->hmp;
2725
2726         /*
2727          * The zone-2 disk offset may have been set by the cluster code via
2728          * a BMAP operation, or else should be NOOFFSET.
2729          *
2730          * Checking the high bits for a match against zone-2 should suffice.
2731          *
2732          * In cases where a lot of data duplication is present it may be
2733          * more beneficial to drop through and doubule-buffer through the
2734          * device.
2735          */
2736         nbio = push_bio(bio);
2737         if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2738             HAMMER_ZONE_LARGE_DATA) {
2739                 if (hammer_double_buffer == 0) {
2740                         lwkt_gettoken(&hmp->fs_token);
2741                         error = hammer_io_direct_read(hmp, nbio, NULL);
2742                         lwkt_reltoken(&hmp->fs_token);
2743                         return (error);
2744                 }
2745
2746                 /*
2747                  * Try to shortcut requests for double_buffer mode too.
2748                  * Since this mode runs through the device buffer cache
2749                  * only compatible buffer sizes (meaning those generated
2750                  * by normal filesystem buffers) are legal.
2751                  */
2752                 if (hammer_live_dedup == 0 && (bp->b_flags & B_PAGING) == 0) {
2753                         error = hammer_io_indirect_read(hmp, nbio, NULL);
2754                         return (error);
2755                 }
2756         }
2757
2758         /*
2759          * Well, that sucked.  Do it the hard way.  If all the stars are
2760          * aligned we may still be able to issue a direct-read.
2761          */
2762         lwkt_gettoken(&hmp->fs_token);
2763         hammer_simple_transaction(&trans, hmp);
2764         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2765
2766         /*
2767          * Key range (begin and end inclusive) to scan.  Note that the key's
2768          * stored in the actual records represent BASE+LEN, not BASE.  The
2769          * first record containing bio_offset will have a key > bio_offset.
2770          */
2771         cursor.key_beg.localization = ip->obj_localization +
2772                                       HAMMER_LOCALIZE_MISC;
2773         cursor.key_beg.obj_id = ip->obj_id;
2774         cursor.key_beg.create_tid = 0;
2775         cursor.key_beg.delete_tid = 0;
2776         cursor.key_beg.obj_type = 0;
2777         cursor.key_beg.key = bio->bio_offset + 1;
2778         cursor.asof = ip->obj_asof;
2779         cursor.flags |= HAMMER_CURSOR_ASOF;
2780
2781         cursor.key_end = cursor.key_beg;
2782         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2783 #if 0
2784         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2785                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2786                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2787                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2788         } else
2789 #endif
2790         {
2791                 ran_end = bio->bio_offset + bp->b_bufsize;
2792                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2793                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2794                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2795                 if (tmp64 < ran_end)
2796                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2797                 else
2798                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2799         }
2800         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2801
2802         /*
2803          * Set NOSWAPCACHE for cursor data extraction if double buffering
2804          * is disabled or (if the file is not marked cacheable via chflags
2805          * and vm.swapcache_use_chflags is enabled).
2806          */
2807         if (hammer_double_buffer == 0 ||
2808             ((ap->a_vp->v_flag & VSWAPCACHE) == 0 &&
2809              vm_swapcache_use_chflags)) {
2810                 cursor.flags |= HAMMER_CURSOR_NOSWAPCACHE;
2811         }
2812
2813         error = hammer_ip_first(&cursor);
2814         boff = 0;
2815
2816         while (error == 0) {
2817                 /*
2818                  * Get the base file offset of the record.  The key for
2819                  * data records is (base + bytes) rather then (base).
2820                  */
2821                 base = &cursor.leaf->base;
2822                 rec_offset = base->key - cursor.leaf->data_len;
2823
2824                 /*
2825                  * Calculate the gap, if any, and zero-fill it.
2826                  *
2827                  * n is the offset of the start of the record verses our
2828                  * current seek offset in the bio.
2829                  */
2830                 n = (int)(rec_offset - (bio->bio_offset + boff));
2831                 if (n > 0) {
2832                         if (n > bp->b_bufsize - boff)
2833                                 n = bp->b_bufsize - boff;
2834                         bzero((char *)bp->b_data + boff, n);
2835                         boff += n;
2836                         n = 0;
2837                 }
2838
2839                 /*
2840                  * Calculate the data offset in the record and the number
2841                  * of bytes we can copy.
2842                  *
2843                  * There are two degenerate cases.  First, boff may already
2844                  * be at bp->b_bufsize.  Secondly, the data offset within
2845                  * the record may exceed the record's size.
2846                  */
2847                 roff = -n;
2848                 rec_offset += roff;
2849                 n = cursor.leaf->data_len - roff;
2850                 if (n <= 0) {
2851                         kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2852                         n = 0;
2853                 } else if (n > bp->b_bufsize - boff) {
2854                         n = bp->b_bufsize - boff;
2855                 }
2856
2857                 /*
2858                  * Deal with cached truncations.  This cool bit of code
2859                  * allows truncate()/ftruncate() to avoid having to sync
2860                  * the file.
2861                  *
2862                  * If the frontend is truncated then all backend records are
2863                  * subject to the frontend's truncation.
2864                  *
2865                  * If the backend is truncated then backend records on-disk
2866                  * (but not in-memory) are subject to the backend's
2867                  * truncation.  In-memory records owned by the backend
2868                  * represent data written after the truncation point on the
2869                  * backend and must not be truncated.
2870                  *
2871                  * Truncate operations deal with frontend buffer cache
2872                  * buffers and frontend-owned in-memory records synchronously.
2873                  */
2874                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2875                         if (hammer_cursor_ondisk(&cursor)/* ||
2876                             cursor.iprec->flush_state == HAMMER_FST_FLUSH*/) {
2877                                 if (ip->trunc_off <= rec_offset)
2878                                         n = 0;
2879                                 else if (ip->trunc_off < rec_offset + n)
2880                                         n = (int)(ip->trunc_off - rec_offset);
2881                         }
2882                 }
2883                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2884                         if (hammer_cursor_ondisk(&cursor)) {
2885                                 if (ip->sync_trunc_off <= rec_offset)
2886                                         n = 0;
2887                                 else if (ip->sync_trunc_off < rec_offset + n)
2888                                         n = (int)(ip->sync_trunc_off - rec_offset);
2889                         }
2890                 }
2891
2892                 /*
2893                  * Try to issue a direct read into our bio if possible,
2894                  * otherwise resolve the element data into a hammer_buffer
2895                  * and copy.
2896                  *
2897                  * The buffer on-disk should be zerod past any real
2898                  * truncation point, but may not be for any synthesized
2899                  * truncation point from above.
2900                  *
2901                  * NOTE: disk_offset is only valid if the cursor data is
2902                  *       on-disk.
2903                  */
2904                 disk_offset = cursor.leaf->data_offset + roff;
2905                 isdedupable = (boff == 0 && n == bp->b_bufsize &&
2906                                hammer_cursor_ondisk(&cursor) &&
2907                                ((int)disk_offset & HAMMER_BUFMASK) == 0);
2908
2909                 if (isdedupable && hammer_double_buffer == 0) {
2910                         /*
2911                          * Direct read case
2912                          */
2913                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2914                                  HAMMER_ZONE_LARGE_DATA);
2915                         nbio->bio_offset = disk_offset;
2916                         error = hammer_io_direct_read(hmp, nbio, cursor.leaf);
2917                         if (hammer_live_dedup && error == 0)
2918                                 hammer_dedup_cache_add(ip, cursor.leaf);
2919                         goto done;
2920                 } else if (isdedupable) {
2921                         /*
2922                          * Async I/O case for reading from backing store
2923                          * and copying the data to the filesystem buffer.
2924                          * live-dedup has to verify the data anyway if it
2925                          * gets a hit later so we can just add the entry
2926                          * now.
2927                          */
2928                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2929                                  HAMMER_ZONE_LARGE_DATA);
2930                         nbio->bio_offset = disk_offset;
2931                         if (hammer_live_dedup)
2932                                 hammer_dedup_cache_add(ip, cursor.leaf);
2933                         error = hammer_io_indirect_read(hmp, nbio, cursor.leaf);
2934                         goto done;
2935                 } else if (n) {
2936                         error = hammer_ip_resolve_data(&cursor);
2937                         if (error == 0) {
2938                                 if (hammer_live_dedup && isdedupable)
2939                                         hammer_dedup_cache_add(ip, cursor.leaf);
2940                                 bcopy((char *)cursor.data + roff,
2941                                       (char *)bp->b_data + boff, n);
2942                         }
2943                 }
2944                 if (error)
2945                         break;
2946
2947                 /*
2948                  * We have to be sure that the only elements added to the
2949                  * dedup cache are those which are already on-media.
2950                  */
2951                 if (hammer_live_dedup && hammer_cursor_ondisk(&cursor))
2952                         hammer_dedup_cache_add(ip, cursor.leaf);
2953
2954                 /*
2955                  * Iterate until we have filled the request.
2956                  */
2957                 boff += n;
2958                 if (boff == bp->b_bufsize)
2959                         break;
2960                 error = hammer_ip_next(&cursor);
2961         }
2962
2963         /*
2964          * There may have been a gap after the last record
2965          */
2966         if (error == ENOENT)
2967                 error = 0;
2968         if (error == 0 && boff != bp->b_bufsize) {
2969                 KKASSERT(boff < bp->b_bufsize);
2970                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2971                 /* boff = bp->b_bufsize; */
2972         }
2973
2974         /*
2975          * Disallow swapcache operation on the vnode buffer if double
2976          * buffering is enabled, the swapcache will get the data via
2977          * the block device buffer.
2978          */
2979         if (hammer_double_buffer)
2980                 bp->b_flags |= B_NOTMETA;
2981
2982         /*
2983          * Cleanup
2984          */
2985         bp->b_resid = 0;
2986         bp->b_error = error;
2987         if (error)
2988                 bp->b_flags |= B_ERROR;
2989         biodone(ap->a_bio);
2990
2991 done:
2992         /*
2993          * Cache the b-tree node for the last data read in cache[1].
2994          *
2995          * If we hit the file EOF then also cache the node in the
2996          * governing director's cache[3], it will be used to initialize
2997          * the inode's cache[1] for any inodes looked up via the directory.
2998          *
2999          * This doesn't reduce disk accesses since the B-Tree chain is
3000          * likely cached, but it does reduce cpu overhead when looking
3001          * up file offsets for cpdup/tar/cpio style iterations.
3002          */
3003         if (cursor.node)
3004                 hammer_cache_node(&ip->cache[1], cursor.node);
3005         if (ran_end >= ip->ino_data.size) {
3006                 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
3007                                         ip->obj_asof, ip->obj_localization);
3008                 if (dip) {
3009                         hammer_cache_node(&dip->cache[3], cursor.node);
3010                         hammer_rel_inode(dip, 0);
3011                 }
3012         }
3013         hammer_done_cursor(&cursor);
3014         hammer_done_transaction(&trans);
3015         lwkt_reltoken(&hmp->fs_token);
3016         return(error);
3017 }
3018
3019 /*
3020  * BMAP operation - used to support cluster_read() only.
3021  *
3022  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
3023  *
3024  * This routine may return EOPNOTSUPP if the opration is not supported for
3025  * the specified offset.  The contents of the pointer arguments do not
3026  * need to be initialized in that case. 
3027  *
3028  * If a disk address is available and properly aligned return 0 with 
3029  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
3030  * to the run-length relative to that offset.  Callers may assume that
3031  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
3032  * large, so return EOPNOTSUPP if it is not sufficiently large.
3033  */
3034 static
3035 int
3036 hammer_vop_bmap(struct vop_bmap_args *ap)
3037 {
3038         struct hammer_transaction trans;
3039         struct hammer_inode *ip;
3040         hammer_mount_t hmp;
3041         struct hammer_cursor cursor;
3042         hammer_base_elm_t base;
3043         int64_t rec_offset;
3044         int64_t ran_end;
3045         int64_t tmp64;
3046         int64_t base_offset;
3047         int64_t base_disk_offset;
3048         int64_t last_offset;
3049         hammer_off_t last_disk_offset;
3050         hammer_off_t disk_offset;
3051         int     rec_len;
3052         int     error;
3053         int     blksize;
3054
3055         ++hammer_stats_file_iopsr;
3056         ip = ap->a_vp->v_data;
3057         hmp = ip->hmp;
3058
3059         /*
3060          * We can only BMAP regular files.  We can't BMAP database files,
3061          * directories, etc.
3062          */
3063         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
3064                 return(EOPNOTSUPP);
3065
3066         /*
3067          * bmap is typically called with runp/runb both NULL when used
3068          * for writing.  We do not support BMAP for writing atm.
3069          */
3070         if (ap->a_cmd != BUF_CMD_READ)
3071                 return(EOPNOTSUPP);
3072
3073         /*
3074          * Scan the B-Tree to acquire blockmap addresses, then translate
3075          * to raw addresses.
3076          */
3077         lwkt_gettoken(&hmp->fs_token);
3078         hammer_simple_transaction(&trans, hmp);
3079 #if 0
3080         kprintf("bmap_beg %016llx ip->cache %p\n",
3081                 (long long)ap->a_loffset, ip->cache[1]);
3082 #endif
3083         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
3084
3085         /*
3086          * Key range (begin and end inclusive) to scan.  Note that the key's
3087          * stored in the actual records represent BASE+LEN, not BASE.  The
3088          * first record containing bio_offset will have a key > bio_offset.
3089          */
3090         cursor.key_beg.localization = ip->obj_localization +
3091                                       HAMMER_LOCALIZE_MISC;
3092         cursor.key_beg.obj_id = ip->obj_id;
3093         cursor.key_beg.create_tid = 0;
3094         cursor.key_beg.delete_tid = 0;
3095         cursor.key_beg.obj_type = 0;
3096         if (ap->a_runb)
3097                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
3098         else
3099                 cursor.key_beg.key = ap->a_loffset + 1;
3100         if (cursor.key_beg.key < 0)
3101                 cursor.key_beg.key = 0;
3102         cursor.asof = ip->obj_asof;
3103         cursor.flags |= HAMMER_CURSOR_ASOF;
3104
3105         cursor.key_end = cursor.key_beg;
3106         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
3107
3108         ran_end = ap->a_loffset + MAXPHYS;
3109         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
3110         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
3111         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
3112         if (tmp64 < ran_end)
3113                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
3114         else
3115                 cursor.key_end.key = ran_end + MAXPHYS + 1;
3116
3117         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
3118
3119         error = hammer_ip_first(&cursor);
3120         base_offset = last_offset = 0;
3121         base_disk_offset = last_disk_offset = 0;
3122
3123         while (error == 0) {
3124                 /*
3125                  * Get the base file offset of the record.  The key for
3126                  * data records is (base + bytes) rather then (base).
3127                  *
3128                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
3129                  * The extra bytes should be zero on-disk and the BMAP op
3130                  * should still be ok.
3131                  */
3132                 base = &cursor.leaf->base;
3133                 rec_offset = base->key - cursor.leaf->data_len;
3134                 rec_len    = cursor.leaf->data_len;
3135
3136                 /*
3137                  * Incorporate any cached truncation.
3138                  *
3139                  * NOTE: Modifications to rec_len based on synthesized
3140                  * truncation points remove the guarantee that any extended
3141                  * data on disk is zero (since the truncations may not have
3142                  * taken place on-media yet).
3143                  */
3144                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
3145                         if (hammer_cursor_ondisk(&cursor) ||
3146                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
3147                                 if (ip->trunc_off <= rec_offset)
3148                                         rec_len = 0;
3149                                 else if (ip->trunc_off < rec_offset + rec_len)
3150                                         rec_len = (int)(ip->trunc_off - rec_offset);
3151                         }
3152                 }
3153                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
3154                         if (hammer_cursor_ondisk(&cursor)) {
3155                                 if (ip->sync_trunc_off <= rec_offset)
3156                                         rec_len = 0;
3157                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
3158                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
3159                         }
3160                 }
3161
3162                 /*
3163                  * Accumulate information.  If we have hit a discontiguous
3164                  * block reset base_offset unless we are already beyond the
3165                  * requested offset.  If we are, that's it, we stop.
3166                  */
3167                 if (error)
3168                         break;
3169                 if (hammer_cursor_ondisk(&cursor)) {
3170                         disk_offset = cursor.leaf->data_offset;
3171                         if (rec_offset != last_offset ||
3172                             disk_offset != last_disk_offset) {
3173                                 if (rec_offset > ap->a_loffset)
3174                                         break;
3175                                 base_offset = rec_offset;
3176                                 base_disk_offset = disk_offset;
3177                         }
3178                         last_offset = rec_offset + rec_len;
3179                         last_disk_offset = disk_offset + rec_len;
3180
3181                         if (hammer_live_dedup)
3182                                 hammer_dedup_cache_add(ip, cursor.leaf);
3183                 }
3184                 
3185                 error = hammer_ip_next(&cursor);
3186         }
3187
3188 #if 0
3189         kprintf("BMAP %016llx:  %016llx - %016llx\n",
3190                 (long long)ap->a_loffset,
3191                 (long long)base_offset,
3192                 (long long)last_offset);
3193         kprintf("BMAP %16s:  %016llx - %016llx\n", "",
3194                 (long long)base_disk_offset,
3195                 (long long)last_disk_offset);
3196 #endif
3197
3198         if (cursor.node) {
3199                 hammer_cache_node(&ip->cache[1], cursor.node);
3200 #if 0
3201                 kprintf("bmap_end2 %016llx ip->cache %p\n",
3202                         (long long)ap->a_loffset, ip->cache[1]);
3203 #endif
3204         }
3205         hammer_done_cursor(&cursor);
3206         hammer_done_transaction(&trans);
3207         lwkt_reltoken(&hmp->fs_token);
3208
3209         /*
3210          * If we couldn't find any records or the records we did find were
3211          * all behind the requested offset, return failure.  A forward
3212          * truncation can leave a hole w/ no on-disk records.
3213          */
3214         if (last_offset == 0 || last_offset < ap->a_loffset)
3215                 return (EOPNOTSUPP);
3216
3217         /*
3218          * Figure out the block size at the requested offset and adjust
3219          * our limits so the cluster_read() does not create inappropriately
3220          * sized buffer cache buffers.
3221          */
3222         blksize = hammer_blocksize(ap->a_loffset);
3223         if (hammer_blocksize(base_offset) != blksize) {
3224                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
3225         }
3226         if (last_offset != ap->a_loffset &&
3227             hammer_blocksize(last_offset - 1) != blksize) {
3228                 last_offset = hammer_blockdemarc(ap->a_loffset,
3229                                                  last_offset - 1);
3230         }
3231
3232         /*
3233          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3234          * from occuring.
3235          */
3236         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
3237
3238         if ((disk_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_LARGE_DATA) {
3239                 /*
3240                  * Only large-data zones can be direct-IOd
3241                  */
3242                 error = EOPNOTSUPP;
3243         } else if ((disk_offset & HAMMER_BUFMASK) ||
3244                    (last_offset - ap->a_loffset) < blksize) {
3245                 /*
3246                  * doffsetp is not aligned or the forward run size does
3247                  * not cover a whole buffer, disallow the direct I/O.
3248                  */
3249                 error = EOPNOTSUPP;
3250         } else {
3251                 /*
3252                  * We're good.
3253                  */
3254                 *ap->a_doffsetp = disk_offset;
3255                 if (ap->a_runb) {
3256                         *ap->a_runb = ap->a_loffset - base_offset;
3257                         KKASSERT(*ap->a_runb >= 0);
3258                 }
3259                 if (ap->a_runp) {
3260                         *ap->a_runp = last_offset - ap->a_loffset;
3261                         KKASSERT(*ap->a_runp >= 0);
3262                 }
3263                 error = 0;
3264         }
3265         return(error);
3266 }
3267
3268 /*
3269  * Write to a regular file.   Because this is a strategy call the OS is
3270  * trying to actually get data onto the media.
3271  */
3272 static
3273 int
3274 hammer_vop_strategy_write(struct vop_strategy_args *ap)
3275 {
3276         hammer_record_t record;
3277         hammer_mount_t hmp;
3278         hammer_inode_t ip;
3279         struct bio *bio;
3280         struct buf *bp;
3281         int blksize;
3282         int bytes;
3283         int error;
3284
3285         bio = ap->a_bio;
3286         bp = bio->bio_buf;
3287         ip = ap->a_vp->v_data;
3288         hmp = ip->hmp;
3289
3290         blksize = hammer_blocksize(bio->bio_offset);
3291         KKASSERT(bp->b_bufsize == blksize);
3292
3293         if (ip->flags & HAMMER_INODE_RO) {
3294                 bp->b_error = EROFS;
3295                 bp->b_flags |= B_ERROR;
3296                 biodone(ap->a_bio);
3297                 return(EROFS);
3298         }
3299
3300         lwkt_gettoken(&hmp->fs_token);
3301
3302         /*
3303          * Disallow swapcache operation on the vnode buffer if double
3304          * buffering is enabled, the swapcache will get the data via
3305          * the block device buffer.
3306          */
3307         if (hammer_double_buffer)
3308                 bp->b_flags |= B_NOTMETA;
3309
3310         /*
3311          * Interlock with inode destruction (no in-kernel or directory
3312          * topology visibility).  If we queue new IO while trying to
3313          * destroy the inode we can deadlock the vtrunc call in
3314          * hammer_inode_unloadable_check().
3315          *
3316          * Besides, there's no point flushing a bp associated with an
3317          * inode that is being destroyed on-media and has no kernel
3318          * references.
3319          */
3320         if ((ip->flags | ip->sync_flags) &
3321             (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
3322                 bp->b_resid = 0;
3323                 biodone(ap->a_bio);
3324                 lwkt_reltoken(&hmp->fs_token);
3325                 return(0);
3326         }
3327
3328         /*
3329          * Reserve space and issue a direct-write from the front-end. 
3330          * NOTE: The direct_io code will hammer_bread/bcopy smaller
3331          * allocations.
3332          *
3333          * An in-memory record will be installed to reference the storage
3334          * until the flusher can get to it.
3335          *
3336          * Since we own the high level bio the front-end will not try to
3337          * do a direct-read until the write completes.
3338          *
3339          * NOTE: The only time we do not reserve a full-sized buffers
3340          * worth of data is if the file is small.  We do not try to
3341          * allocate a fragment (from the small-data zone) at the end of
3342          * an otherwise large file as this can lead to wildly separated
3343          * data.
3344          */
3345         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
3346         KKASSERT(bio->bio_offset < ip->ino_data.size);
3347         if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
3348                 bytes = bp->b_bufsize;
3349         else
3350                 bytes = ((int)ip->ino_data.size + 15) & ~15;
3351
3352         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
3353                                     bytes, &error);
3354
3355         /*
3356          * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3357          * in hammer_vop_write().  We must flag the record so the proper
3358          * REDO_TERM_WRITE entry is generated during the flush.
3359          */
3360         if (record) {
3361                 if (bp->b_flags & B_VFSFLAG1) {
3362                         record->flags |= HAMMER_RECF_REDO;
3363                         bp->b_flags &= ~B_VFSFLAG1;
3364                 }
3365                 if (record->flags & HAMMER_RECF_DEDUPED) {
3366                         bp->b_resid = 0;
3367                         hammer_ip_replace_bulk(hmp, record);
3368                         biodone(ap->a_bio);
3369                 } else {
3370                         hammer_io_direct_write(hmp, bio, record);
3371                 }
3372                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
3373                         hammer_flush_inode(ip, 0);
3374         } else {
3375                 bp->b_bio2.bio_offset = NOOFFSET;
3376                 bp->b_error = error;
3377                 bp->b_flags |= B_ERROR;
3378                 biodone(ap->a_bio);
3379         }
3380         lwkt_reltoken(&hmp->fs_token);
3381         return(error);
3382 }
3383
3384 /*
3385  * dounlink - disconnect a directory entry
3386  *
3387  * XXX whiteout support not really in yet
3388  */
3389 static int
3390 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
3391                 struct vnode *dvp, struct ucred *cred, 
3392                 int flags, int isdir)
3393 {
3394         struct namecache *ncp;
3395         hammer_inode_t dip;
3396         hammer_inode_t ip;
3397         hammer_mount_t hmp;
3398         struct hammer_cursor cursor;
3399         int64_t namekey;
3400         u_int32_t max_iterations;
3401         int nlen, error;
3402
3403         /*
3404          * Calculate the namekey and setup the key range for the scan.  This
3405          * works kinda like a chained hash table where the lower 32 bits
3406          * of the namekey synthesize the chain.
3407          *
3408          * The key range is inclusive of both key_beg and key_end.
3409          */
3410         dip = VTOI(dvp);
3411         ncp = nch->ncp;
3412         hmp = dip->hmp;
3413
3414         if (dip->flags & HAMMER_INODE_RO)
3415                 return (EROFS);
3416
3417         namekey = hammer_directory_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3418                                            &max_iterations);
3419 retry:
3420         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3421         cursor.key_beg.localization = dip->obj_localization +
3422                                       hammer_dir_localization(dip);
3423         cursor.key_beg.obj_id = dip->obj_id;
3424         cursor.key_beg.key = namekey;
3425         cursor.key_beg.create_tid = 0;
3426         cursor.key_beg.delete_tid = 0;
3427         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3428         cursor.key_beg.obj_type = 0;
3429
3430         cursor.key_end = cursor.key_beg;
3431         cursor.key_end.key += max_iterations;
3432         cursor.asof = dip->obj_asof;
3433         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3434
3435         /*
3436          * Scan all matching records (the chain), locate the one matching
3437          * the requested path component.  info->last_error contains the
3438          * error code on search termination and could be 0, ENOENT, or
3439          * something else.
3440          *
3441          * The hammer_ip_*() functions merge in-memory records with on-disk
3442          * records for the purposes of the search.
3443          */
3444         error = hammer_ip_first(&cursor);
3445
3446         while (error == 0) {
3447                 error = hammer_ip_resolve_data(&cursor);
3448                 if (error)
3449                         break;
3450                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3451                 KKASSERT(nlen > 0);
3452                 if (ncp->nc_nlen == nlen &&
3453                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3454                         break;
3455                 }
3456                 error = hammer_ip_next(&cursor);
3457         }
3458
3459         /*
3460          * If all is ok we have to get the inode so we can adjust nlinks.
3461          * To avoid a deadlock with the flusher we must release the inode
3462          * lock on the directory when acquiring the inode for the entry.
3463          *
3464          * If the target is a directory, it must be empty.
3465          */
3466         if (error == 0) {
3467                 hammer_unlock(&cursor.ip->lock);
3468                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3469                                       hmp->asof,
3470                                       cursor.data->entry.localization,
3471                                       0, &error);
3472                 hammer_lock_sh(&cursor.ip->lock);
3473                 if (error == ENOENT) {
3474                         kprintf("HAMMER: WARNING: Removing "
3475                                 "dirent w/missing inode \"%s\"\n"
3476                                 "\tobj_id = %016llx\n",
3477                                 ncp->nc_name,
3478                                 (long long)cursor.data->entry.obj_id);
3479                         error = 0;
3480                 }
3481
3482                 /*
3483                  * If isdir >= 0 we validate that the entry is or is not a
3484                  * directory.  If isdir < 0 we don't care.
3485                  */
3486                 if (error == 0 && isdir >= 0 && ip) {
3487                         if (isdir &&
3488                             ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3489                                 error = ENOTDIR;
3490                         } else if (isdir == 0 &&
3491                             ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3492                                 error = EISDIR;
3493                         }
3494                 }
3495
3496                 /*
3497                  * If we are trying to remove a directory the directory must
3498                  * be empty.
3499                  *
3500                  * The check directory code can loop and deadlock/retry.  Our
3501                  * own cursor's node locks must be released to avoid a 3-way
3502                  * deadlock with the flusher if the check directory code
3503                  * blocks.
3504                  *
3505                  * If any changes whatsoever have been made to the cursor
3506                  * set EDEADLK and retry.
3507                  *
3508                  * WARNING: See warnings in hammer_unlock_cursor()
3509                  *          function.
3510                  */
3511                 if (error == 0 && ip && ip->ino_data.obj_type ==
3512                                         HAMMER_OBJTYPE_DIRECTORY) {
3513                         hammer_unlock_cursor(&cursor);
3514                         error = hammer_ip_check_directory_empty(trans, ip);
3515                         hammer_lock_cursor(&cursor);
3516                         if (cursor.flags & HAMMER_CURSOR_RETEST) {
3517                                 kprintf("HAMMER: Warning: avoided deadlock "
3518                                         "on rmdir '%s'\n",
3519                                         ncp->nc_name);
3520                                 error = EDEADLK;
3521                         }
3522                 }
3523
3524                 /*
3525                  * Delete the directory entry.
3526                  *
3527                  * WARNING: hammer_ip_del_directory() may have to terminate
3528                  * the cursor to avoid a deadlock.  It is ok to call
3529                  * hammer_done_cursor() twice.
3530                  */
3531                 if (error == 0) {
3532                         error = hammer_ip_del_directory(trans, &cursor,
3533                                                         dip, ip);
3534                 }
3535                 hammer_done_cursor(&cursor);
3536                 if (error == 0) {
3537                         cache_setunresolved(nch);
3538                         cache_setvp(nch, NULL);
3539
3540                         /*
3541                          * NOTE: ip->vp, if non-NULL, cannot be directly
3542                          *       referenced without formally acquiring the
3543                          *       vp since the vp might have zero refs on it,
3544                          *       or in the middle of a reclaim, etc.
3545                          *
3546                          * NOTE: The cache_setunresolved() can rip the vp
3547                          *       out from under us since the vp may not have
3548                          *       any refs, in which case ip->vp will be NULL
3549                          *       from the outset.
3550                          */
3551                         while (ip && ip->vp) {
3552                                 struct vnode *vp;
3553
3554                                 error = hammer_get_vnode(ip, &vp);
3555                                 if (error == 0 && vp) {
3556                                         vn_unlock(vp);
3557                                         hammer_knote(ip->vp, NOTE_DELETE);
3558                                         cache_inval_vp(ip->vp, CINV_DESTROY);
3559                                         vrele(vp);
3560                                         break;
3561                                 }
3562                                 kprintf("Debug: HAMMER ip/vp race1 avoided\n");
3563                         }
3564                 }
3565                 if (ip)
3566                         hammer_rel_inode(ip, 0);
3567         } else {
3568                 hammer_done_cursor(&cursor);
3569         }
3570         if (error == EDEADLK)
3571                 goto retry;
3572
3573         return (error);
3574 }
3575
3576 /************************************************************************
3577  *                          FIFO AND SPECFS OPS                         *
3578  ************************************************************************
3579  *
3580  */
3581 static int
3582 hammer_vop_fifoclose (struct vop_close_args *ap)
3583 {
3584         /* XXX update itimes */
3585         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3586 }
3587
3588 static int
3589 hammer_vop_fiforead (struct vop_read_args *ap)
3590 {
3591         int error;
3592
3593         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3594         /* XXX update access time */
3595         return (error);
3596 }
3597
3598 static int
3599 hammer_vop_fifowrite (struct vop_write_args *ap)
3600 {
3601         int error;
3602
3603         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3604         /* XXX update access time */
3605         return (error);
3606 }
3607
3608 static
3609 int
3610 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3611 {
3612         int error;
3613
3614         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3615         if (error)
3616                 error = hammer_vop_kqfilter(ap);
3617         return(error);
3618 }
3619
3620 /************************************************************************
3621  *                          KQFILTER OPS                                *
3622  ************************************************************************
3623  *
3624  */
3625 static void filt_hammerdetach(struct knote *kn);
3626 static int filt_hammerread(struct knote *kn, long hint);
3627 static int filt_hammerwrite(struct knote *kn, long hint);
3628 static int filt_hammervnode(struct knote *kn, long hint);
3629
3630 static struct filterops hammerread_filtops =
3631         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammerread };
3632 static struct filterops hammerwrite_filtops =
3633         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammerwrite };
3634 static struct filterops hammervnode_filtops =
3635         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammervnode };
3636
3637 static
3638 int
3639 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3640 {
3641         struct vnode *vp = ap->a_vp;
3642         struct knote *kn = ap->a_kn;
3643
3644         switch (kn->kn_filter) {
3645         case EVFILT_READ:
3646                 kn->kn_fop = &hammerread_filtops;
3647                 break;
3648         case EVFILT_WRITE:
3649                 kn->kn_fop = &hammerwrite_filtops;
3650                 break;
3651         case EVFILT_VNODE:
3652                 kn->kn_fop = &hammervnode_filtops;
3653                 break;
3654         default:
3655                 return (EOPNOTSUPP);
3656         }
3657
3658         kn->kn_hook = (caddr_t)vp;
3659
3660         knote_insert(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3661
3662         return(0);
3663 }
3664
3665 static void
3666 filt_hammerdetach(struct knote *kn)
3667 {
3668         struct vnode *vp = (void *)kn->kn_hook;
3669
3670         knote_remove(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3671 }
3672
3673 static int
3674 filt_hammerread(struct knote *kn, long hint)
3675 {
3676         struct vnode *vp = (void *)kn->kn_hook;
3677         hammer_inode_t ip = VTOI(vp);
3678         hammer_mount_t hmp = ip->hmp;
3679         off_t off;
3680
3681         if (hint == NOTE_REVOKE) {
3682                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3683                 return(1);
3684         }
3685         lwkt_gettoken(&hmp->fs_token);  /* XXX use per-ip-token */
3686         off = ip->ino_data.size - kn->kn_fp->f_offset;
3687         kn->kn_data = (off < INTPTR_MAX) ? off : INTPTR_MAX;
3688         lwkt_reltoken(&hmp->fs_token);
3689         if (kn->kn_sfflags & NOTE_OLDAPI)
3690                 return(1);
3691         return (kn->kn_data != 0);
3692 }
3693
3694 static int
3695 filt_hammerwrite(struct knote *kn, long hint)
3696 {
3697         if (hint == NOTE_REVOKE)
3698                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3699         kn->kn_data = 0;
3700         return (1);
3701 }
3702
3703 static int
3704 filt_hammervnode(struct knote *kn, long hint)
3705 {
3706         if (kn->kn_sfflags & hint)
3707                 kn->kn_fflags |= hint;
3708         if (hint == NOTE_REVOKE) {
3709                 kn->kn_flags |= (EV_EOF | EV_NODATA);
3710                 return (1);
3711         }
3712         return (kn->kn_fflags != 0);
3713 }
3714