kernel - Fix read event on file for select/poll API
[dragonfly.git] / sys / vfs / hammer / hammer_vnops.c
1 /*
2  * Copyright (c) 2007-2008 The DragonFly Project.  All rights reserved.
3  * 
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  * 
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  * 
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  * 
34  * $DragonFly: src/sys/vfs/hammer/hammer_vnops.c,v 1.102 2008/10/16 17:24:16 dillon Exp $
35  */
36
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
40 #include <sys/fcntl.h>
41 #include <sys/namecache.h>
42 #include <sys/vnode.h>
43 #include <sys/lockf.h>
44 #include <sys/event.h>
45 #include <sys/stat.h>
46 #include <sys/dirent.h>
47 #include <sys/file.h>
48 #include <vm/vm_extern.h>
49 #include <vfs/fifofs/fifo.h>
50
51 #include <sys/mplock2.h>
52
53 #include "hammer.h"
54
55 /*
56  * USERFS VNOPS
57  */
58 /*static int hammer_vop_vnoperate(struct vop_generic_args *);*/
59 static int hammer_vop_fsync(struct vop_fsync_args *);
60 static int hammer_vop_read(struct vop_read_args *);
61 static int hammer_vop_write(struct vop_write_args *);
62 static int hammer_vop_access(struct vop_access_args *);
63 static int hammer_vop_advlock(struct vop_advlock_args *);
64 static int hammer_vop_close(struct vop_close_args *);
65 static int hammer_vop_ncreate(struct vop_ncreate_args *);
66 static int hammer_vop_getattr(struct vop_getattr_args *);
67 static int hammer_vop_nresolve(struct vop_nresolve_args *);
68 static int hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *);
69 static int hammer_vop_nlink(struct vop_nlink_args *);
70 static int hammer_vop_nmkdir(struct vop_nmkdir_args *);
71 static int hammer_vop_nmknod(struct vop_nmknod_args *);
72 static int hammer_vop_open(struct vop_open_args *);
73 static int hammer_vop_print(struct vop_print_args *);
74 static int hammer_vop_readdir(struct vop_readdir_args *);
75 static int hammer_vop_readlink(struct vop_readlink_args *);
76 static int hammer_vop_nremove(struct vop_nremove_args *);
77 static int hammer_vop_nrename(struct vop_nrename_args *);
78 static int hammer_vop_nrmdir(struct vop_nrmdir_args *);
79 static int hammer_vop_markatime(struct vop_markatime_args *);
80 static int hammer_vop_setattr(struct vop_setattr_args *);
81 static int hammer_vop_strategy(struct vop_strategy_args *);
82 static int hammer_vop_bmap(struct vop_bmap_args *ap);
83 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
84 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
85 static int hammer_vop_ioctl(struct vop_ioctl_args *);
86 static int hammer_vop_mountctl(struct vop_mountctl_args *);
87 static int hammer_vop_kqfilter (struct vop_kqfilter_args *);
88
89 static int hammer_vop_fifoclose (struct vop_close_args *);
90 static int hammer_vop_fiforead (struct vop_read_args *);
91 static int hammer_vop_fifowrite (struct vop_write_args *);
92 static int hammer_vop_fifokqfilter (struct vop_kqfilter_args *);
93
94 struct vop_ops hammer_vnode_vops = {
95         .vop_default =          vop_defaultop,
96         .vop_fsync =            hammer_vop_fsync,
97         .vop_getpages =         vop_stdgetpages,
98         .vop_putpages =         vop_stdputpages,
99         .vop_read =             hammer_vop_read,
100         .vop_write =            hammer_vop_write,
101         .vop_access =           hammer_vop_access,
102         .vop_advlock =          hammer_vop_advlock,
103         .vop_close =            hammer_vop_close,
104         .vop_ncreate =          hammer_vop_ncreate,
105         .vop_getattr =          hammer_vop_getattr,
106         .vop_inactive =         hammer_vop_inactive,
107         .vop_reclaim =          hammer_vop_reclaim,
108         .vop_nresolve =         hammer_vop_nresolve,
109         .vop_nlookupdotdot =    hammer_vop_nlookupdotdot,
110         .vop_nlink =            hammer_vop_nlink,
111         .vop_nmkdir =           hammer_vop_nmkdir,
112         .vop_nmknod =           hammer_vop_nmknod,
113         .vop_open =             hammer_vop_open,
114         .vop_pathconf =         vop_stdpathconf,
115         .vop_print =            hammer_vop_print,
116         .vop_readdir =          hammer_vop_readdir,
117         .vop_readlink =         hammer_vop_readlink,
118         .vop_nremove =          hammer_vop_nremove,
119         .vop_nrename =          hammer_vop_nrename,
120         .vop_nrmdir =           hammer_vop_nrmdir,
121         .vop_markatime =        hammer_vop_markatime,
122         .vop_setattr =          hammer_vop_setattr,
123         .vop_bmap =             hammer_vop_bmap,
124         .vop_strategy =         hammer_vop_strategy,
125         .vop_nsymlink =         hammer_vop_nsymlink,
126         .vop_nwhiteout =        hammer_vop_nwhiteout,
127         .vop_ioctl =            hammer_vop_ioctl,
128         .vop_mountctl =         hammer_vop_mountctl,
129         .vop_kqfilter =         hammer_vop_kqfilter
130 };
131
132 struct vop_ops hammer_spec_vops = {
133         .vop_default =          vop_defaultop,
134         .vop_fsync =            hammer_vop_fsync,
135         .vop_read =             vop_stdnoread,
136         .vop_write =            vop_stdnowrite,
137         .vop_access =           hammer_vop_access,
138         .vop_close =            hammer_vop_close,
139         .vop_markatime =        hammer_vop_markatime,
140         .vop_getattr =          hammer_vop_getattr,
141         .vop_inactive =         hammer_vop_inactive,
142         .vop_reclaim =          hammer_vop_reclaim,
143         .vop_setattr =          hammer_vop_setattr
144 };
145
146 struct vop_ops hammer_fifo_vops = {
147         .vop_default =          fifo_vnoperate,
148         .vop_fsync =            hammer_vop_fsync,
149         .vop_read =             hammer_vop_fiforead,
150         .vop_write =            hammer_vop_fifowrite,
151         .vop_access =           hammer_vop_access,
152         .vop_close =            hammer_vop_fifoclose,
153         .vop_markatime =        hammer_vop_markatime,
154         .vop_getattr =          hammer_vop_getattr,
155         .vop_inactive =         hammer_vop_inactive,
156         .vop_reclaim =          hammer_vop_reclaim,
157         .vop_setattr =          hammer_vop_setattr,
158         .vop_kqfilter =         hammer_vop_fifokqfilter
159 };
160
161 static __inline
162 void
163 hammer_knote(struct vnode *vp, int flags)
164 {
165         if (flags)
166                 KNOTE(&vp->v_pollinfo.vpi_kqinfo.ki_note, flags);
167 }
168
169 #ifdef DEBUG_TRUNCATE
170 struct hammer_inode *HammerTruncIp;
171 #endif
172
173 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
174                            struct vnode *dvp, struct ucred *cred,
175                            int flags, int isdir);
176 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
177 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
178
179 #if 0
180 static
181 int
182 hammer_vop_vnoperate(struct vop_generic_args *)
183 {
184         return (VOCALL(&hammer_vnode_vops, ap));
185 }
186 #endif
187
188 /*
189  * hammer_vop_fsync { vp, waitfor }
190  *
191  * fsync() an inode to disk and wait for it to be completely committed
192  * such that the information would not be undone if a crash occured after
193  * return.
194  *
195  * NOTE: HAMMER's fsync()'s are going to remain expensive until we implement
196  *       a REDO log.  A sysctl is provided to relax HAMMER's fsync()
197  *       operation.
198  *
199  *       Ultimately the combination of a REDO log and use of fast storage
200  *       to front-end cluster caches will make fsync fast, but it aint
201  *       here yet.  And, in anycase, we need real transactional
202  *       all-or-nothing features which are not restricted to a single file.
203  */
204 static
205 int
206 hammer_vop_fsync(struct vop_fsync_args *ap)
207 {
208         hammer_inode_t ip = VTOI(ap->a_vp);
209         hammer_mount_t hmp = ip->hmp;
210         int waitfor = ap->a_waitfor;
211         int mode;
212
213         /*
214          * Fsync rule relaxation (default is either full synchronous flush
215          * or REDO semantics with synchronous flush).
216          */
217         if (ap->a_flags & VOP_FSYNC_SYSCALL) {
218                 switch(hammer_fsync_mode) {
219                 case 0:
220 mode0:
221                         /* no REDO, full synchronous flush */
222                         goto skip;
223                 case 1:
224 mode1:
225                         /* no REDO, full asynchronous flush */
226                         if (waitfor == MNT_WAIT)
227                                 waitfor = MNT_NOWAIT;
228                         goto skip;
229                 case 2:
230                         /* REDO semantics, synchronous flush */
231                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
232                                 goto mode0;
233                         mode = HAMMER_FLUSH_UNDOS_AUTO;
234                         break;
235                 case 3:
236                         /* REDO semantics, relaxed asynchronous flush */
237                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
238                                 goto mode1;
239                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
240                         if (waitfor == MNT_WAIT)
241                                 waitfor = MNT_NOWAIT;
242                         break;
243                 case 4:
244                         /* ignore the fsync() system call */
245                         return(0);
246                 default:
247                         /* we have to do something */
248                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
249                         if (waitfor == MNT_WAIT)
250                                 waitfor = MNT_NOWAIT;
251                         break;
252                 }
253
254                 /*
255                  * Fast fsync only needs to flush the UNDO/REDO fifo if
256                  * HAMMER_INODE_REDO is non-zero and the only modifications
257                  * made to the file are write or write-extends.
258                  */
259                 if ((ip->flags & HAMMER_INODE_REDO) &&
260                     (ip->flags & HAMMER_INODE_MODMASK_NOREDO) == 0
261                 ) {
262                         ++hammer_count_fsyncs;
263                         hammer_flusher_flush_undos(hmp, mode);
264                         ip->redo_count = 0;
265                         return(0);
266                 }
267
268                 /*
269                  * REDO is enabled by fsync(), the idea being we really only
270                  * want to lay down REDO records when programs are using
271                  * fsync() heavily.  The first fsync() on the file starts
272                  * the gravy train going and later fsync()s keep it hot by
273                  * resetting the redo_count.
274                  *
275                  * We weren't running REDOs before now so we have to fall
276                  * through and do a full fsync of what we have.
277                  */
278                 if (hmp->version >= HAMMER_VOL_VERSION_FOUR &&
279                     (hmp->flags & HAMMER_MOUNT_REDO_RECOVERY_RUN) == 0) {
280                         ip->flags |= HAMMER_INODE_REDO;
281                         ip->redo_count = 0;
282                 }
283         }
284 skip:
285
286         /*
287          * Do a full flush sequence.
288          */
289         ++hammer_count_fsyncs;
290         vfsync(ap->a_vp, waitfor, 1, NULL, NULL);
291         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
292         if (waitfor == MNT_WAIT) {
293                 vn_unlock(ap->a_vp);
294                 hammer_wait_inode(ip);
295                 vn_lock(ap->a_vp, LK_EXCLUSIVE | LK_RETRY);
296         }
297         return (ip->error);
298 }
299
300 /*
301  * hammer_vop_read { vp, uio, ioflag, cred }
302  *
303  * MPALMOSTSAFE
304  */
305 static
306 int
307 hammer_vop_read(struct vop_read_args *ap)
308 {
309         struct hammer_transaction trans;
310         hammer_inode_t ip;
311         off_t offset;
312         struct buf *bp;
313         struct uio *uio;
314         int error;
315         int n;
316         int seqcount;
317         int ioseqcount;
318         int blksize;
319         int got_mplock;
320         int bigread;
321
322         if (ap->a_vp->v_type != VREG)
323                 return (EINVAL);
324         ip = VTOI(ap->a_vp);
325         error = 0;
326         uio = ap->a_uio;
327
328         /*
329          * Allow the UIO's size to override the sequential heuristic.
330          */
331         blksize = hammer_blocksize(uio->uio_offset);
332         seqcount = (uio->uio_resid + (BKVASIZE - 1)) / BKVASIZE;
333         ioseqcount = (ap->a_ioflag >> 16);
334         if (seqcount < ioseqcount)
335                 seqcount = ioseqcount;
336
337         /*
338          * Temporary hack until more of HAMMER can be made MPSAFE.
339          */
340 #ifdef SMP
341         if (curthread->td_mpcount) {
342                 got_mplock = -1;
343                 hammer_start_transaction(&trans, ip->hmp);
344         } else {
345                 got_mplock = 0;
346         }
347 #else
348         hammer_start_transaction(&trans, ip->hmp);
349         got_mplock = -1;
350 #endif
351
352         /*
353          * If reading or writing a huge amount of data we have to break
354          * atomicy and allow the operation to be interrupted by a signal
355          * or it can DOS the machine.
356          */
357         bigread = (uio->uio_resid > 100 * 1024 * 1024);
358
359         /*
360          * Access the data typically in HAMMER_BUFSIZE blocks via the
361          * buffer cache, but HAMMER may use a variable block size based
362          * on the offset.
363          *
364          * XXX Temporary hack, delay the start transaction while we remain
365          *     MPSAFE.  NOTE: ino_data.size cannot change while vnode is
366          *     locked-shared.
367          */
368         while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
369                 int64_t base_offset;
370                 int64_t file_limit;
371
372                 blksize = hammer_blocksize(uio->uio_offset);
373                 offset = (int)uio->uio_offset & (blksize - 1);
374                 base_offset = uio->uio_offset - offset;
375
376                 if (bigread && (error = hammer_signal_check(ip->hmp)) != 0)
377                         break;
378
379                 /*
380                  * MPSAFE
381                  */
382                 bp = getcacheblk(ap->a_vp, base_offset);
383                 if (bp) {
384                         error = 0;
385                         goto skip;
386                 }
387
388                 /*
389                  * MPUNSAFE
390                  */
391                 if (got_mplock == 0) {
392                         got_mplock = 1;
393                         get_mplock();
394                         hammer_start_transaction(&trans, ip->hmp);
395                 }
396
397                 if (hammer_cluster_enable) {
398                         /*
399                          * Use file_limit to prevent cluster_read() from
400                          * creating buffers of the wrong block size past
401                          * the demarc.
402                          */
403                         file_limit = ip->ino_data.size;
404                         if (base_offset < HAMMER_XDEMARC &&
405                             file_limit > HAMMER_XDEMARC) {
406                                 file_limit = HAMMER_XDEMARC;
407                         }
408                         error = cluster_read(ap->a_vp,
409                                              file_limit, base_offset,
410                                              blksize, uio->uio_resid,
411                                              seqcount * BKVASIZE, &bp);
412                 } else {
413                         error = bread(ap->a_vp, base_offset, blksize, &bp);
414                 }
415                 if (error) {
416                         brelse(bp);
417                         break;
418                 }
419 skip:
420                 if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IODEBUG)) {
421                         kprintf("doff %016jx read file %016jx@%016jx\n",
422                                 (intmax_t)bp->b_bio2.bio_offset,
423                                 (intmax_t)ip->obj_id,
424                                 (intmax_t)bp->b_loffset);
425                 }
426                 bp->b_flags &= ~B_IODEBUG;
427
428                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
429                 n = blksize - offset;
430                 if (n > uio->uio_resid)
431                         n = uio->uio_resid;
432                 if (n > ip->ino_data.size - uio->uio_offset)
433                         n = (int)(ip->ino_data.size - uio->uio_offset);
434                 error = uiomove((char *)bp->b_data + offset, n, uio);
435
436                 /* data has a lower priority then meta-data */
437                 bp->b_flags |= B_AGE;
438                 bqrelse(bp);
439                 if (error)
440                         break;
441                 hammer_stats_file_read += n;
442         }
443
444         /*
445          * XXX only update the atime if we had to get the MP lock.
446          * XXX hack hack hack, fixme.
447          */
448         if (got_mplock) {
449                 if ((ip->flags & HAMMER_INODE_RO) == 0 &&
450                     (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
451                         ip->ino_data.atime = trans.time;
452                         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
453                 }
454                 hammer_done_transaction(&trans);
455                 if (got_mplock > 0)
456                         rel_mplock();
457         }
458         return (error);
459 }
460
461 /*
462  * hammer_vop_write { vp, uio, ioflag, cred }
463  */
464 static
465 int
466 hammer_vop_write(struct vop_write_args *ap)
467 {
468         struct hammer_transaction trans;
469         struct hammer_inode *ip;
470         hammer_mount_t hmp;
471         struct uio *uio;
472         int offset;
473         off_t base_offset;
474         struct buf *bp;
475         int kflags;
476         int error;
477         int n;
478         int flags;
479         int seqcount;
480         int bigwrite;
481
482         if (ap->a_vp->v_type != VREG)
483                 return (EINVAL);
484         ip = VTOI(ap->a_vp);
485         hmp = ip->hmp;
486         error = 0;
487         kflags = 0;
488         seqcount = ap->a_ioflag >> 16;
489
490         if (ip->flags & HAMMER_INODE_RO)
491                 return (EROFS);
492
493         /*
494          * Create a transaction to cover the operations we perform.
495          */
496         hammer_start_transaction(&trans, hmp);
497         uio = ap->a_uio;
498
499         /*
500          * Check append mode
501          */
502         if (ap->a_ioflag & IO_APPEND)
503                 uio->uio_offset = ip->ino_data.size;
504
505         /*
506          * Check for illegal write offsets.  Valid range is 0...2^63-1.
507          *
508          * NOTE: the base_off assignment is required to work around what
509          * I consider to be a GCC-4 optimization bug.
510          */
511         if (uio->uio_offset < 0) {
512                 hammer_done_transaction(&trans);
513                 return (EFBIG);
514         }
515         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
516         if (uio->uio_resid > 0 && base_offset <= uio->uio_offset) {
517                 hammer_done_transaction(&trans);
518                 return (EFBIG);
519         }
520
521         /*
522          * If reading or writing a huge amount of data we have to break
523          * atomicy and allow the operation to be interrupted by a signal
524          * or it can DOS the machine.
525          *
526          * Preset redo_count so we stop generating REDOs earlier if the
527          * limit is exceeded.
528          */
529         bigwrite = (uio->uio_resid > 100 * 1024 * 1024);
530         if ((ip->flags & HAMMER_INODE_REDO) &&
531             ip->redo_count < hammer_limit_redo) {
532                 ip->redo_count += uio->uio_resid;
533         }
534
535         /*
536          * Access the data typically in HAMMER_BUFSIZE blocks via the
537          * buffer cache, but HAMMER may use a variable block size based
538          * on the offset.
539          */
540         while (uio->uio_resid > 0) {
541                 int fixsize = 0;
542                 int blksize;
543                 int blkmask;
544                 int trivial;
545                 int endofblk;
546                 off_t nsize;
547
548                 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
549                         break;
550                 if (bigwrite && (error = hammer_signal_check(hmp)) != 0)
551                         break;
552
553                 blksize = hammer_blocksize(uio->uio_offset);
554
555                 /*
556                  * Do not allow HAMMER to blow out the buffer cache.  Very
557                  * large UIOs can lockout other processes due to bwillwrite()
558                  * mechanics.
559                  *
560                  * The hammer inode is not locked during these operations.
561                  * The vnode is locked which can interfere with the pageout
562                  * daemon for non-UIO_NOCOPY writes but should not interfere
563                  * with the buffer cache.  Even so, we cannot afford to
564                  * allow the pageout daemon to build up too many dirty buffer
565                  * cache buffers.
566                  *
567                  * Only call this if we aren't being recursively called from
568                  * a virtual disk device (vn), else we may deadlock.
569                  */
570                 if ((ap->a_ioflag & IO_RECURSE) == 0)
571                         bwillwrite(blksize);
572
573                 /*
574                  * Control the number of pending records associated with
575                  * this inode.  If too many have accumulated start a
576                  * flush.  Try to maintain a pipeline with the flusher.
577                  */
578                 if (ip->rsv_recs >= hammer_limit_inode_recs) {
579                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
580                 }
581                 if (ip->rsv_recs >= hammer_limit_inode_recs * 2) {
582                         while (ip->rsv_recs >= hammer_limit_inode_recs) {
583                                 tsleep(&ip->rsv_recs, 0, "hmrwww", hz);
584                         }
585                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
586                 }
587
588 #if 0
589                 /*
590                  * Do not allow HAMMER to blow out system memory by
591                  * accumulating too many records.   Records are so well
592                  * decoupled from the buffer cache that it is possible
593                  * for userland to push data out to the media via
594                  * direct-write, but build up the records queued to the
595                  * backend faster then the backend can flush them out.
596                  * HAMMER has hit its write limit but the frontend has
597                  * no pushback to slow it down.
598                  */
599                 if (hmp->rsv_recs > hammer_limit_recs / 2) {
600                         /*
601                          * Get the inode on the flush list
602                          */
603                         if (ip->rsv_recs >= 64)
604                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
605                         else if (ip->rsv_recs >= 16)
606                                 hammer_flush_inode(ip, 0);
607
608                         /*
609                          * Keep the flusher going if the system keeps
610                          * queueing records.
611                          */
612                         delta = hmp->count_newrecords -
613                                 hmp->last_newrecords;
614                         if (delta < 0 || delta > hammer_limit_recs / 2) {
615                                 hmp->last_newrecords = hmp->count_newrecords;
616                                 hammer_sync_hmp(hmp, MNT_NOWAIT);
617                         }
618
619                         /*
620                          * If we have gotten behind start slowing
621                          * down the writers.
622                          */
623                         delta = (hmp->rsv_recs - hammer_limit_recs) *
624                                 hz / hammer_limit_recs;
625                         if (delta > 0)
626                                 tsleep(&trans, 0, "hmrslo", delta);
627                 }
628 #endif
629
630                 /*
631                  * Calculate the blocksize at the current offset and figure
632                  * out how much we can actually write.
633                  */
634                 blkmask = blksize - 1;
635                 offset = (int)uio->uio_offset & blkmask;
636                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
637                 n = blksize - offset;
638                 if (n > uio->uio_resid) {
639                         n = uio->uio_resid;
640                         endofblk = 0;
641                 } else {
642                         endofblk = 1;
643                 }
644                 nsize = uio->uio_offset + n;
645                 if (nsize > ip->ino_data.size) {
646                         if (uio->uio_offset > ip->ino_data.size)
647                                 trivial = 0;
648                         else
649                                 trivial = 1;
650                         nvextendbuf(ap->a_vp,
651                                     ip->ino_data.size,
652                                     nsize,
653                                     hammer_blocksize(ip->ino_data.size),
654                                     hammer_blocksize(nsize),
655                                     hammer_blockoff(ip->ino_data.size),
656                                     hammer_blockoff(nsize),
657                                     trivial);
658                         fixsize = 1;
659                         kflags |= NOTE_EXTEND;
660                 }
661
662                 if (uio->uio_segflg == UIO_NOCOPY) {
663                         /*
664                          * Issuing a write with the same data backing the
665                          * buffer.  Instantiate the buffer to collect the
666                          * backing vm pages, then read-in any missing bits.
667                          *
668                          * This case is used by vop_stdputpages().
669                          */
670                         bp = getblk(ap->a_vp, base_offset,
671                                     blksize, GETBLK_BHEAVY, 0);
672                         if ((bp->b_flags & B_CACHE) == 0) {
673                                 bqrelse(bp);
674                                 error = bread(ap->a_vp, base_offset,
675                                               blksize, &bp);
676                         }
677                 } else if (offset == 0 && uio->uio_resid >= blksize) {
678                         /*
679                          * Even though we are entirely overwriting the buffer
680                          * we may still have to zero it out to avoid a 
681                          * mmap/write visibility issue.
682                          */
683                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
684                         if ((bp->b_flags & B_CACHE) == 0)
685                                 vfs_bio_clrbuf(bp);
686                 } else if (base_offset >= ip->ino_data.size) {
687                         /*
688                          * If the base offset of the buffer is beyond the
689                          * file EOF, we don't have to issue a read.
690                          */
691                         bp = getblk(ap->a_vp, base_offset,
692                                     blksize, GETBLK_BHEAVY, 0);
693                         vfs_bio_clrbuf(bp);
694                 } else {
695                         /*
696                          * Partial overwrite, read in any missing bits then
697                          * replace the portion being written.
698                          */
699                         error = bread(ap->a_vp, base_offset, blksize, &bp);
700                         if (error == 0)
701                                 bheavy(bp);
702                 }
703                 if (error == 0)
704                         error = uiomove(bp->b_data + offset, n, uio);
705
706                 /*
707                  * Generate REDO records if enabled and redo_count will not
708                  * exceeded the limit.
709                  *
710                  * If redo_count exceeds the limit we stop generating records
711                  * and clear HAMMER_INODE_REDO.  This will cause the next
712                  * fsync() to do a full meta-data sync instead of just an
713                  * UNDO/REDO fifo update.
714                  *
715                  * When clearing HAMMER_INODE_REDO any pre-existing REDOs
716                  * will still be tracked.  The tracks will be terminated
717                  * when the related meta-data (including possible data
718                  * modifications which are not tracked via REDO) is
719                  * flushed.
720                  */
721                 if ((ip->flags & HAMMER_INODE_REDO) && error == 0) {
722                         if (ip->redo_count < hammer_limit_redo) {
723                                 bp->b_flags |= B_VFSFLAG1;
724                                 error = hammer_generate_redo(&trans, ip,
725                                                      base_offset + offset,
726                                                      HAMMER_REDO_WRITE,
727                                                      bp->b_data + offset,
728                                                      (size_t)n);
729                         } else {
730                                 ip->flags &= ~HAMMER_INODE_REDO;
731                         }
732                 }
733
734                 /*
735                  * If we screwed up we have to undo any VM size changes we
736                  * made.
737                  */
738                 if (error) {
739                         brelse(bp);
740                         if (fixsize) {
741                                 nvtruncbuf(ap->a_vp, ip->ino_data.size,
742                                           hammer_blocksize(ip->ino_data.size),
743                                           hammer_blockoff(ip->ino_data.size));
744                         }
745                         break;
746                 }
747                 kflags |= NOTE_WRITE;
748                 hammer_stats_file_write += n;
749                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
750                 if (ip->ino_data.size < uio->uio_offset) {
751                         ip->ino_data.size = uio->uio_offset;
752                         flags = HAMMER_INODE_SDIRTY;
753                 } else {
754                         flags = 0;
755                 }
756                 ip->ino_data.mtime = trans.time;
757                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
758                 hammer_modify_inode(&trans, ip, flags);
759
760                 /*
761                  * Once we dirty the buffer any cached zone-X offset
762                  * becomes invalid.  HAMMER NOTE: no-history mode cannot 
763                  * allow overwriting over the same data sector unless
764                  * we provide UNDOs for the old data, which we don't.
765                  */
766                 bp->b_bio2.bio_offset = NOOFFSET;
767
768                 /*
769                  * Final buffer disposition.
770                  *
771                  * Because meta-data updates are deferred, HAMMER is
772                  * especially sensitive to excessive bdwrite()s because
773                  * the I/O stream is not broken up by disk reads.  So the
774                  * buffer cache simply cannot keep up.
775                  *
776                  * WARNING!  blksize is variable.  cluster_write() is
777                  *           expected to not blow up if it encounters
778                  *           buffers that do not match the passed blksize.
779                  *
780                  * NOTE!  Hammer shouldn't need to bawrite()/cluster_write().
781                  *        The ip->rsv_recs check should burst-flush the data.
782                  *        If we queue it immediately the buf could be left
783                  *        locked on the device queue for a very long time.
784                  *
785                  * NOTE!  To avoid degenerate stalls due to mismatched block
786                  *        sizes we only honor IO_DIRECT on the write which
787                  *        abuts the end of the buffer.  However, we must
788                  *        honor IO_SYNC in case someone is silly enough to
789                  *        configure a HAMMER file as swap, or when HAMMER
790                  *        is serving NFS (for commits).  Ick ick.
791                  */
792                 bp->b_flags |= B_AGE;
793                 if (ap->a_ioflag & IO_SYNC) {
794                         bwrite(bp);
795                 } else if ((ap->a_ioflag & IO_DIRECT) && endofblk) {
796                         bawrite(bp);
797                 } else {
798 #if 0
799                 if (offset + n == blksize) {
800                         if (hammer_cluster_enable == 0 ||
801                             (ap->a_vp->v_mount->mnt_flag & MNT_NOCLUSTERW)) {
802                                 bawrite(bp);
803                         } else {
804                                 cluster_write(bp, ip->ino_data.size,
805                                               blksize, seqcount);
806                         }
807                 } else {
808 #endif
809                         bdwrite(bp);
810                 }
811         }
812         hammer_done_transaction(&trans);
813         hammer_knote(ap->a_vp, kflags);
814         return (error);
815 }
816
817 /*
818  * hammer_vop_access { vp, mode, cred }
819  */
820 static
821 int
822 hammer_vop_access(struct vop_access_args *ap)
823 {
824         struct hammer_inode *ip = VTOI(ap->a_vp);
825         uid_t uid;
826         gid_t gid;
827         int error;
828
829         ++hammer_stats_file_iopsr;
830         uid = hammer_to_unix_xid(&ip->ino_data.uid);
831         gid = hammer_to_unix_xid(&ip->ino_data.gid);
832
833         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
834                                   ip->ino_data.uflags);
835         return (error);
836 }
837
838 /*
839  * hammer_vop_advlock { vp, id, op, fl, flags }
840  */
841 static
842 int
843 hammer_vop_advlock(struct vop_advlock_args *ap)
844 {
845         hammer_inode_t ip = VTOI(ap->a_vp);
846
847         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
848 }
849
850 /*
851  * hammer_vop_close { vp, fflag }
852  *
853  * We can only sync-on-close for normal closes.
854  */
855 static
856 int
857 hammer_vop_close(struct vop_close_args *ap)
858 {
859 #if 0
860         struct vnode *vp = ap->a_vp;
861         hammer_inode_t ip = VTOI(vp);
862         int waitfor;
863         if (ip->flags & (HAMMER_INODE_CLOSESYNC|HAMMER_INODE_CLOSEASYNC)) {
864                 if (vn_islocked(vp) == LK_EXCLUSIVE &&
865                     (vp->v_flag & (VINACTIVE|VRECLAIMED)) == 0) {
866                         if (ip->flags & HAMMER_INODE_CLOSESYNC)
867                                 waitfor = MNT_WAIT;
868                         else
869                                 waitfor = MNT_NOWAIT;
870                         ip->flags &= ~(HAMMER_INODE_CLOSESYNC |
871                                        HAMMER_INODE_CLOSEASYNC);
872                         VOP_FSYNC(vp, MNT_NOWAIT, waitfor);
873                 }
874         }
875 #endif
876         return (vop_stdclose(ap));
877 }
878
879 /*
880  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
881  *
882  * The operating system has already ensured that the directory entry
883  * does not exist and done all appropriate namespace locking.
884  */
885 static
886 int
887 hammer_vop_ncreate(struct vop_ncreate_args *ap)
888 {
889         struct hammer_transaction trans;
890         struct hammer_inode *dip;
891         struct hammer_inode *nip;
892         struct nchandle *nch;
893         int error;
894
895         nch = ap->a_nch;
896         dip = VTOI(ap->a_dvp);
897
898         if (dip->flags & HAMMER_INODE_RO)
899                 return (EROFS);
900         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
901                 return (error);
902
903         /*
904          * Create a transaction to cover the operations we perform.
905          */
906         hammer_start_transaction(&trans, dip->hmp);
907         ++hammer_stats_file_iopsw;
908
909         /*
910          * Create a new filesystem object of the requested type.  The
911          * returned inode will be referenced and shared-locked to prevent
912          * it from being moved to the flusher.
913          */
914         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
915                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
916                                     NULL, &nip);
917         if (error) {
918                 hkprintf("hammer_create_inode error %d\n", error);
919                 hammer_done_transaction(&trans);
920                 *ap->a_vpp = NULL;
921                 return (error);
922         }
923
924         /*
925          * Add the new filesystem object to the directory.  This will also
926          * bump the inode's link count.
927          */
928         error = hammer_ip_add_directory(&trans, dip,
929                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
930                                         nip);
931         if (error)
932                 hkprintf("hammer_ip_add_directory error %d\n", error);
933
934         /*
935          * Finish up.
936          */
937         if (error) {
938                 hammer_rel_inode(nip, 0);
939                 hammer_done_transaction(&trans);
940                 *ap->a_vpp = NULL;
941         } else {
942                 error = hammer_get_vnode(nip, ap->a_vpp);
943                 hammer_done_transaction(&trans);
944                 hammer_rel_inode(nip, 0);
945                 if (error == 0) {
946                         cache_setunresolved(ap->a_nch);
947                         cache_setvp(ap->a_nch, *ap->a_vpp);
948                 }
949                 hammer_knote(ap->a_dvp, NOTE_WRITE);
950         }
951         return (error);
952 }
953
954 /*
955  * hammer_vop_getattr { vp, vap }
956  *
957  * Retrieve an inode's attribute information.  When accessing inodes
958  * historically we fake the atime field to ensure consistent results.
959  * The atime field is stored in the B-Tree element and allowed to be
960  * updated without cycling the element.
961  *
962  * MPSAFE
963  */
964 static
965 int
966 hammer_vop_getattr(struct vop_getattr_args *ap)
967 {
968         struct hammer_inode *ip = VTOI(ap->a_vp);
969         struct vattr *vap = ap->a_vap;
970
971         /*
972          * We want the fsid to be different when accessing a filesystem
973          * with different as-of's so programs like diff don't think
974          * the files are the same.
975          *
976          * We also want the fsid to be the same when comparing snapshots,
977          * or when comparing mirrors (which might be backed by different
978          * physical devices).  HAMMER fsids are based on the PFS's
979          * shared_uuid field.
980          *
981          * XXX there is a chance of collision here.  The va_fsid reported
982          * by stat is different from the more involved fsid used in the
983          * mount structure.
984          */
985         ++hammer_stats_file_iopsr;
986         hammer_lock_sh(&ip->lock);
987         vap->va_fsid = ip->pfsm->fsid_udev ^ (u_int32_t)ip->obj_asof ^
988                        (u_int32_t)(ip->obj_asof >> 32);
989
990         vap->va_fileid = ip->ino_leaf.base.obj_id;
991         vap->va_mode = ip->ino_data.mode;
992         vap->va_nlink = ip->ino_data.nlinks;
993         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
994         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
995         vap->va_rmajor = 0;
996         vap->va_rminor = 0;
997         vap->va_size = ip->ino_data.size;
998
999         /*
1000          * Special case for @@PFS softlinks.  The actual size of the
1001          * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
1002          * or for MAX_TID is    "@@-1:%05d" == 10 bytes.
1003          */
1004         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
1005             ip->ino_data.size == 10 &&
1006             ip->obj_asof == HAMMER_MAX_TID &&
1007             ip->obj_localization == 0 &&
1008             strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
1009                     if (ip->pfsm->pfsd.mirror_flags & HAMMER_PFSD_SLAVE)
1010                             vap->va_size = 26;
1011                     else
1012                             vap->va_size = 10;
1013         }
1014
1015         /*
1016          * We must provide a consistent atime and mtime for snapshots
1017          * so people can do a 'tar cf - ... | md5' on them and get
1018          * consistent results.
1019          */
1020         if (ip->flags & HAMMER_INODE_RO) {
1021                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
1022                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
1023         } else {
1024                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
1025                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
1026         }
1027         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
1028         vap->va_flags = ip->ino_data.uflags;
1029         vap->va_gen = 1;        /* hammer inums are unique for all time */
1030         vap->va_blocksize = HAMMER_BUFSIZE;
1031         if (ip->ino_data.size >= HAMMER_XDEMARC) {
1032                 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
1033                                 ~HAMMER_XBUFMASK64;
1034         } else if (ip->ino_data.size > HAMMER_BUFSIZE / 2) {
1035                 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
1036                                 ~HAMMER_BUFMASK64;
1037         } else {
1038                 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
1039         }
1040
1041         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
1042         vap->va_filerev = 0;    /* XXX */
1043         vap->va_uid_uuid = ip->ino_data.uid;
1044         vap->va_gid_uuid = ip->ino_data.gid;
1045         vap->va_fsid_uuid = ip->hmp->fsid;
1046         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
1047                           VA_FSID_UUID_VALID;
1048
1049         switch (ip->ino_data.obj_type) {
1050         case HAMMER_OBJTYPE_CDEV:
1051         case HAMMER_OBJTYPE_BDEV:
1052                 vap->va_rmajor = ip->ino_data.rmajor;
1053                 vap->va_rminor = ip->ino_data.rminor;
1054                 break;
1055         default:
1056                 break;
1057         }
1058         hammer_unlock(&ip->lock);
1059         return(0);
1060 }
1061
1062 /*
1063  * hammer_vop_nresolve { nch, dvp, cred }
1064  *
1065  * Locate the requested directory entry.
1066  */
1067 static
1068 int
1069 hammer_vop_nresolve(struct vop_nresolve_args *ap)
1070 {
1071         struct hammer_transaction trans;
1072         struct namecache *ncp;
1073         hammer_inode_t dip;
1074         hammer_inode_t ip;
1075         hammer_tid_t asof;
1076         struct hammer_cursor cursor;
1077         struct vnode *vp;
1078         int64_t namekey;
1079         int error;
1080         int i;
1081         int nlen;
1082         int flags;
1083         int ispfs;
1084         int64_t obj_id;
1085         u_int32_t localization;
1086         u_int32_t max_iterations;
1087
1088         /*
1089          * Misc initialization, plus handle as-of name extensions.  Look for
1090          * the '@@' extension.  Note that as-of files and directories cannot
1091          * be modified.
1092          */
1093         dip = VTOI(ap->a_dvp);
1094         ncp = ap->a_nch->ncp;
1095         asof = dip->obj_asof;
1096         localization = dip->obj_localization;   /* for code consistency */
1097         nlen = ncp->nc_nlen;
1098         flags = dip->flags & HAMMER_INODE_RO;
1099         ispfs = 0;
1100
1101         hammer_simple_transaction(&trans, dip->hmp);
1102         ++hammer_stats_file_iopsr;
1103
1104         for (i = 0; i < nlen; ++i) {
1105                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
1106                         error = hammer_str_to_tid(ncp->nc_name + i + 2,
1107                                                   &ispfs, &asof, &localization);
1108                         if (error != 0) {
1109                                 i = nlen;
1110                                 break;
1111                         }
1112                         if (asof != HAMMER_MAX_TID)
1113                                 flags |= HAMMER_INODE_RO;
1114                         break;
1115                 }
1116         }
1117         nlen = i;
1118
1119         /*
1120          * If this is a PFS softlink we dive into the PFS
1121          */
1122         if (ispfs && nlen == 0) {
1123                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1124                                       asof, localization,
1125                                       flags, &error);
1126                 if (error == 0) {
1127                         error = hammer_get_vnode(ip, &vp);
1128                         hammer_rel_inode(ip, 0);
1129                 } else {
1130                         vp = NULL;
1131                 }
1132                 if (error == 0) {
1133                         vn_unlock(vp);
1134                         cache_setvp(ap->a_nch, vp);
1135                         vrele(vp);
1136                 }
1137                 goto done;
1138         }
1139
1140         /*
1141          * If there is no path component the time extension is relative to dip.
1142          * e.g. "fubar/@@<snapshot>"
1143          *
1144          * "." is handled by the kernel, but ".@@<snapshot>" is not.
1145          * e.g. "fubar/.@@<snapshot>"
1146          *
1147          * ".." is handled by the kernel.  We do not currently handle
1148          * "..@<snapshot>".
1149          */
1150         if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1151                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
1152                                       asof, dip->obj_localization,
1153                                       flags, &error);
1154                 if (error == 0) {
1155                         error = hammer_get_vnode(ip, &vp);
1156                         hammer_rel_inode(ip, 0);
1157                 } else {
1158                         vp = NULL;
1159                 }
1160                 if (error == 0) {
1161                         vn_unlock(vp);
1162                         cache_setvp(ap->a_nch, vp);
1163                         vrele(vp);
1164                 }
1165                 goto done;
1166         }
1167
1168         /*
1169          * Calculate the namekey and setup the key range for the scan.  This
1170          * works kinda like a chained hash table where the lower 32 bits
1171          * of the namekey synthesize the chain.
1172          *
1173          * The key range is inclusive of both key_beg and key_end.
1174          */
1175         namekey = hammer_directory_namekey(dip, ncp->nc_name, nlen,
1176                                            &max_iterations);
1177
1178         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1179         cursor.key_beg.localization = dip->obj_localization +
1180                                       hammer_dir_localization(dip);
1181         cursor.key_beg.obj_id = dip->obj_id;
1182         cursor.key_beg.key = namekey;
1183         cursor.key_beg.create_tid = 0;
1184         cursor.key_beg.delete_tid = 0;
1185         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1186         cursor.key_beg.obj_type = 0;
1187
1188         cursor.key_end = cursor.key_beg;
1189         cursor.key_end.key += max_iterations;
1190         cursor.asof = asof;
1191         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1192
1193         /*
1194          * Scan all matching records (the chain), locate the one matching
1195          * the requested path component.
1196          *
1197          * The hammer_ip_*() functions merge in-memory records with on-disk
1198          * records for the purposes of the search.
1199          */
1200         obj_id = 0;
1201         localization = HAMMER_DEF_LOCALIZATION;
1202
1203         if (error == 0) {
1204                 error = hammer_ip_first(&cursor);
1205                 while (error == 0) {
1206                         error = hammer_ip_resolve_data(&cursor);
1207                         if (error)
1208                                 break;
1209                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1210                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1211                                 obj_id = cursor.data->entry.obj_id;
1212                                 localization = cursor.data->entry.localization;
1213                                 break;
1214                         }
1215                         error = hammer_ip_next(&cursor);
1216                 }
1217         }
1218         hammer_done_cursor(&cursor);
1219
1220         /*
1221          * Lookup the obj_id.  This should always succeed.  If it does not
1222          * the filesystem may be damaged and we return a dummy inode.
1223          */
1224         if (error == 0) {
1225                 ip = hammer_get_inode(&trans, dip, obj_id,
1226                                       asof, localization,
1227                                       flags, &error);
1228                 if (error == ENOENT) {
1229                         kprintf("HAMMER: WARNING: Missing "
1230                                 "inode for dirent \"%s\"\n"
1231                                 "\tobj_id = %016llx, asof=%016llx, lo=%08x\n",
1232                                 ncp->nc_name,
1233                                 (long long)obj_id, (long long)asof,
1234                                 localization);
1235                         error = 0;
1236                         ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1237                                                     asof, localization,
1238                                                     flags, &error);
1239                 }
1240                 if (error == 0) {
1241                         error = hammer_get_vnode(ip, &vp);
1242                         hammer_rel_inode(ip, 0);
1243                 } else {
1244                         vp = NULL;
1245                 }
1246                 if (error == 0) {
1247                         vn_unlock(vp);
1248                         cache_setvp(ap->a_nch, vp);
1249                         vrele(vp);
1250                 }
1251         } else if (error == ENOENT) {
1252                 cache_setvp(ap->a_nch, NULL);
1253         }
1254 done:
1255         hammer_done_transaction(&trans);
1256         return (error);
1257 }
1258
1259 /*
1260  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1261  *
1262  * Locate the parent directory of a directory vnode.
1263  *
1264  * dvp is referenced but not locked.  *vpp must be returned referenced and
1265  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
1266  * at the root, instead it could indicate that the directory we were in was
1267  * removed.
1268  *
1269  * NOTE: as-of sequences are not linked into the directory structure.  If
1270  * we are at the root with a different asof then the mount point, reload
1271  * the same directory with the mount point's asof.   I'm not sure what this
1272  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
1273  * get confused, but it hasn't been tested.
1274  */
1275 static
1276 int
1277 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1278 {
1279         struct hammer_transaction trans;
1280         struct hammer_inode *dip;
1281         struct hammer_inode *ip;
1282         int64_t parent_obj_id;
1283         u_int32_t parent_obj_localization;
1284         hammer_tid_t asof;
1285         int error;
1286
1287         dip = VTOI(ap->a_dvp);
1288         asof = dip->obj_asof;
1289
1290         /*
1291          * Whos are parent?  This could be the root of a pseudo-filesystem
1292          * whos parent is in another localization domain.
1293          */
1294         parent_obj_id = dip->ino_data.parent_obj_id;
1295         if (dip->obj_id == HAMMER_OBJID_ROOT)
1296                 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
1297         else
1298                 parent_obj_localization = dip->obj_localization;
1299
1300         if (parent_obj_id == 0) {
1301                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1302                    asof != dip->hmp->asof) {
1303                         parent_obj_id = dip->obj_id;
1304                         asof = dip->hmp->asof;
1305                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1306                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
1307                                   (long long)dip->obj_asof);
1308                 } else {
1309                         *ap->a_vpp = NULL;
1310                         return ENOENT;
1311                 }
1312         }
1313
1314         hammer_simple_transaction(&trans, dip->hmp);
1315         ++hammer_stats_file_iopsr;
1316
1317         ip = hammer_get_inode(&trans, dip, parent_obj_id,
1318                               asof, parent_obj_localization,
1319                               dip->flags, &error);
1320         if (ip) {
1321                 error = hammer_get_vnode(ip, ap->a_vpp);
1322                 hammer_rel_inode(ip, 0);
1323         } else {
1324                 *ap->a_vpp = NULL;
1325         }
1326         hammer_done_transaction(&trans);
1327         return (error);
1328 }
1329
1330 /*
1331  * hammer_vop_nlink { nch, dvp, vp, cred }
1332  */
1333 static
1334 int
1335 hammer_vop_nlink(struct vop_nlink_args *ap)
1336 {
1337         struct hammer_transaction trans;
1338         struct hammer_inode *dip;
1339         struct hammer_inode *ip;
1340         struct nchandle *nch;
1341         int error;
1342
1343         if (ap->a_dvp->v_mount != ap->a_vp->v_mount)    
1344                 return(EXDEV);
1345
1346         nch = ap->a_nch;
1347         dip = VTOI(ap->a_dvp);
1348         ip = VTOI(ap->a_vp);
1349
1350         if (dip->obj_localization != ip->obj_localization)
1351                 return(EXDEV);
1352
1353         if (dip->flags & HAMMER_INODE_RO)
1354                 return (EROFS);
1355         if (ip->flags & HAMMER_INODE_RO)
1356                 return (EROFS);
1357         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1358                 return (error);
1359
1360         /*
1361          * Create a transaction to cover the operations we perform.
1362          */
1363         hammer_start_transaction(&trans, dip->hmp);
1364         ++hammer_stats_file_iopsw;
1365
1366         /*
1367          * Add the filesystem object to the directory.  Note that neither
1368          * dip nor ip are referenced or locked, but their vnodes are
1369          * referenced.  This function will bump the inode's link count.
1370          */
1371         error = hammer_ip_add_directory(&trans, dip,
1372                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1373                                         ip);
1374
1375         /*
1376          * Finish up.
1377          */
1378         if (error == 0) {
1379                 cache_setunresolved(nch);
1380                 cache_setvp(nch, ap->a_vp);
1381         }
1382         hammer_done_transaction(&trans);
1383         hammer_knote(ap->a_vp, NOTE_LINK);
1384         hammer_knote(ap->a_dvp, NOTE_WRITE);
1385         return (error);
1386 }
1387
1388 /*
1389  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1390  *
1391  * The operating system has already ensured that the directory entry
1392  * does not exist and done all appropriate namespace locking.
1393  */
1394 static
1395 int
1396 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1397 {
1398         struct hammer_transaction trans;
1399         struct hammer_inode *dip;
1400         struct hammer_inode *nip;
1401         struct nchandle *nch;
1402         int error;
1403
1404         nch = ap->a_nch;
1405         dip = VTOI(ap->a_dvp);
1406
1407         if (dip->flags & HAMMER_INODE_RO)
1408                 return (EROFS);
1409         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1410                 return (error);
1411
1412         /*
1413          * Create a transaction to cover the operations we perform.
1414          */
1415         hammer_start_transaction(&trans, dip->hmp);
1416         ++hammer_stats_file_iopsw;
1417
1418         /*
1419          * Create a new filesystem object of the requested type.  The
1420          * returned inode will be referenced but not locked.
1421          */
1422         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1423                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1424                                     NULL, &nip);
1425         if (error) {
1426                 hkprintf("hammer_mkdir error %d\n", error);
1427                 hammer_done_transaction(&trans);
1428                 *ap->a_vpp = NULL;
1429                 return (error);
1430         }
1431         /*
1432          * Add the new filesystem object to the directory.  This will also
1433          * bump the inode's link count.
1434          */
1435         error = hammer_ip_add_directory(&trans, dip,
1436                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1437                                         nip);
1438         if (error)
1439                 hkprintf("hammer_mkdir (add) error %d\n", error);
1440
1441         /*
1442          * Finish up.
1443          */
1444         if (error) {
1445                 hammer_rel_inode(nip, 0);
1446                 *ap->a_vpp = NULL;
1447         } else {
1448                 error = hammer_get_vnode(nip, ap->a_vpp);
1449                 hammer_rel_inode(nip, 0);
1450                 if (error == 0) {
1451                         cache_setunresolved(ap->a_nch);
1452                         cache_setvp(ap->a_nch, *ap->a_vpp);
1453                 }
1454         }
1455         hammer_done_transaction(&trans);
1456         if (error == 0)
1457                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1458         return (error);
1459 }
1460
1461 /*
1462  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1463  *
1464  * The operating system has already ensured that the directory entry
1465  * does not exist and done all appropriate namespace locking.
1466  */
1467 static
1468 int
1469 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1470 {
1471         struct hammer_transaction trans;
1472         struct hammer_inode *dip;
1473         struct hammer_inode *nip;
1474         struct nchandle *nch;
1475         int error;
1476
1477         nch = ap->a_nch;
1478         dip = VTOI(ap->a_dvp);
1479
1480         if (dip->flags & HAMMER_INODE_RO)
1481                 return (EROFS);
1482         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1483                 return (error);
1484
1485         /*
1486          * Create a transaction to cover the operations we perform.
1487          */
1488         hammer_start_transaction(&trans, dip->hmp);
1489         ++hammer_stats_file_iopsw;
1490
1491         /*
1492          * Create a new filesystem object of the requested type.  The
1493          * returned inode will be referenced but not locked.
1494          *
1495          * If mknod specifies a directory a pseudo-fs is created.
1496          */
1497         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1498                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1499                                     NULL, &nip);
1500         if (error) {
1501                 hammer_done_transaction(&trans);
1502                 *ap->a_vpp = NULL;
1503                 return (error);
1504         }
1505
1506         /*
1507          * Add the new filesystem object to the directory.  This will also
1508          * bump the inode's link count.
1509          */
1510         error = hammer_ip_add_directory(&trans, dip,
1511                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1512                                         nip);
1513
1514         /*
1515          * Finish up.
1516          */
1517         if (error) {
1518                 hammer_rel_inode(nip, 0);
1519                 *ap->a_vpp = NULL;
1520         } else {
1521                 error = hammer_get_vnode(nip, ap->a_vpp);
1522                 hammer_rel_inode(nip, 0);
1523                 if (error == 0) {
1524                         cache_setunresolved(ap->a_nch);
1525                         cache_setvp(ap->a_nch, *ap->a_vpp);
1526                 }
1527         }
1528         hammer_done_transaction(&trans);
1529         if (error == 0)
1530                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1531         return (error);
1532 }
1533
1534 /*
1535  * hammer_vop_open { vp, mode, cred, fp }
1536  */
1537 static
1538 int
1539 hammer_vop_open(struct vop_open_args *ap)
1540 {
1541         hammer_inode_t ip;
1542
1543         ++hammer_stats_file_iopsr;
1544         ip = VTOI(ap->a_vp);
1545
1546         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1547                 return (EROFS);
1548         return(vop_stdopen(ap));
1549 }
1550
1551 /*
1552  * hammer_vop_print { vp }
1553  */
1554 static
1555 int
1556 hammer_vop_print(struct vop_print_args *ap)
1557 {
1558         return EOPNOTSUPP;
1559 }
1560
1561 /*
1562  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1563  */
1564 static
1565 int
1566 hammer_vop_readdir(struct vop_readdir_args *ap)
1567 {
1568         struct hammer_transaction trans;
1569         struct hammer_cursor cursor;
1570         struct hammer_inode *ip;
1571         struct uio *uio;
1572         hammer_base_elm_t base;
1573         int error;
1574         int cookie_index;
1575         int ncookies;
1576         off_t *cookies;
1577         off_t saveoff;
1578         int r;
1579         int dtype;
1580
1581         ++hammer_stats_file_iopsr;
1582         ip = VTOI(ap->a_vp);
1583         uio = ap->a_uio;
1584         saveoff = uio->uio_offset;
1585
1586         if (ap->a_ncookies) {
1587                 ncookies = uio->uio_resid / 16 + 1;
1588                 if (ncookies > 1024)
1589                         ncookies = 1024;
1590                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1591                 cookie_index = 0;
1592         } else {
1593                 ncookies = -1;
1594                 cookies = NULL;
1595                 cookie_index = 0;
1596         }
1597
1598         hammer_simple_transaction(&trans, ip->hmp);
1599
1600         /*
1601          * Handle artificial entries
1602          *
1603          * It should be noted that the minimum value for a directory
1604          * hash key on-media is 0x0000000100000000, so we can use anything
1605          * less then that to represent our 'special' key space.
1606          */
1607         error = 0;
1608         if (saveoff == 0) {
1609                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1610                 if (r)
1611                         goto done;
1612                 if (cookies)
1613                         cookies[cookie_index] = saveoff;
1614                 ++saveoff;
1615                 ++cookie_index;
1616                 if (cookie_index == ncookies)
1617                         goto done;
1618         }
1619         if (saveoff == 1) {
1620                 if (ip->ino_data.parent_obj_id) {
1621                         r = vop_write_dirent(&error, uio,
1622                                              ip->ino_data.parent_obj_id,
1623                                              DT_DIR, 2, "..");
1624                 } else {
1625                         r = vop_write_dirent(&error, uio,
1626                                              ip->obj_id, DT_DIR, 2, "..");
1627                 }
1628                 if (r)
1629                         goto done;
1630                 if (cookies)
1631                         cookies[cookie_index] = saveoff;
1632                 ++saveoff;
1633                 ++cookie_index;
1634                 if (cookie_index == ncookies)
1635                         goto done;
1636         }
1637
1638         /*
1639          * Key range (begin and end inclusive) to scan.  Directory keys
1640          * directly translate to a 64 bit 'seek' position.
1641          */
1642         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1643         cursor.key_beg.localization = ip->obj_localization +
1644                                       hammer_dir_localization(ip);
1645         cursor.key_beg.obj_id = ip->obj_id;
1646         cursor.key_beg.create_tid = 0;
1647         cursor.key_beg.delete_tid = 0;
1648         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1649         cursor.key_beg.obj_type = 0;
1650         cursor.key_beg.key = saveoff;
1651
1652         cursor.key_end = cursor.key_beg;
1653         cursor.key_end.key = HAMMER_MAX_KEY;
1654         cursor.asof = ip->obj_asof;
1655         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1656
1657         error = hammer_ip_first(&cursor);
1658
1659         while (error == 0) {
1660                 error = hammer_ip_resolve_data(&cursor);
1661                 if (error)
1662                         break;
1663                 base = &cursor.leaf->base;
1664                 saveoff = base->key;
1665                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1666
1667                 if (base->obj_id != ip->obj_id)
1668                         panic("readdir: bad record at %p", cursor.node);
1669
1670                 /*
1671                  * Convert pseudo-filesystems into softlinks
1672                  */
1673                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1674                 r = vop_write_dirent(
1675                              &error, uio, cursor.data->entry.obj_id,
1676                              dtype,
1677                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1678                              (void *)cursor.data->entry.name);
1679                 if (r)
1680                         break;
1681                 ++saveoff;
1682                 if (cookies)
1683                         cookies[cookie_index] = base->key;
1684                 ++cookie_index;
1685                 if (cookie_index == ncookies)
1686                         break;
1687                 error = hammer_ip_next(&cursor);
1688         }
1689         hammer_done_cursor(&cursor);
1690
1691 done:
1692         hammer_done_transaction(&trans);
1693
1694         if (ap->a_eofflag)
1695                 *ap->a_eofflag = (error == ENOENT);
1696         uio->uio_offset = saveoff;
1697         if (error && cookie_index == 0) {
1698                 if (error == ENOENT)
1699                         error = 0;
1700                 if (cookies) {
1701                         kfree(cookies, M_TEMP);
1702                         *ap->a_ncookies = 0;
1703                         *ap->a_cookies = NULL;
1704                 }
1705         } else {
1706                 if (error == ENOENT)
1707                         error = 0;
1708                 if (cookies) {
1709                         *ap->a_ncookies = cookie_index;
1710                         *ap->a_cookies = cookies;
1711                 }
1712         }
1713         return(error);
1714 }
1715
1716 /*
1717  * hammer_vop_readlink { vp, uio, cred }
1718  */
1719 static
1720 int
1721 hammer_vop_readlink(struct vop_readlink_args *ap)
1722 {
1723         struct hammer_transaction trans;
1724         struct hammer_cursor cursor;
1725         struct hammer_inode *ip;
1726         char buf[32];
1727         u_int32_t localization;
1728         hammer_pseudofs_inmem_t pfsm;
1729         int error;
1730
1731         ip = VTOI(ap->a_vp);
1732
1733         /*
1734          * Shortcut if the symlink data was stuffed into ino_data.
1735          *
1736          * Also expand special "@@PFS%05d" softlinks (expansion only
1737          * occurs for non-historical (current) accesses made from the
1738          * primary filesystem).
1739          */
1740         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1741                 char *ptr;
1742                 int bytes;
1743
1744                 ptr = ip->ino_data.ext.symlink;
1745                 bytes = (int)ip->ino_data.size;
1746                 if (bytes == 10 &&
1747                     ip->obj_asof == HAMMER_MAX_TID &&
1748                     ip->obj_localization == 0 &&
1749                     strncmp(ptr, "@@PFS", 5) == 0) {
1750                         hammer_simple_transaction(&trans, ip->hmp);
1751                         bcopy(ptr + 5, buf, 5);
1752                         buf[5] = 0;
1753                         localization = strtoul(buf, NULL, 10) << 16;
1754                         pfsm = hammer_load_pseudofs(&trans, localization,
1755                                                     &error);
1756                         if (error == 0) {
1757                                 if (pfsm->pfsd.mirror_flags &
1758                                     HAMMER_PFSD_SLAVE) {
1759                                         /* vap->va_size == 26 */
1760                                         ksnprintf(buf, sizeof(buf),
1761                                                   "@@0x%016llx:%05d",
1762                                                   (long long)pfsm->pfsd.sync_end_tid,
1763                                                   localization >> 16);
1764                                 } else {
1765                                         /* vap->va_size == 10 */
1766                                         ksnprintf(buf, sizeof(buf),
1767                                                   "@@-1:%05d",
1768                                                   localization >> 16);
1769 #if 0
1770                                         ksnprintf(buf, sizeof(buf),
1771                                                   "@@0x%016llx:%05d",
1772                                                   (long long)HAMMER_MAX_TID,
1773                                                   localization >> 16);
1774 #endif
1775                                 }
1776                                 ptr = buf;
1777                                 bytes = strlen(buf);
1778                         }
1779                         if (pfsm)
1780                                 hammer_rel_pseudofs(trans.hmp, pfsm);
1781                         hammer_done_transaction(&trans);
1782                 }
1783                 error = uiomove(ptr, bytes, ap->a_uio);
1784                 return(error);
1785         }
1786
1787         /*
1788          * Long version
1789          */
1790         hammer_simple_transaction(&trans, ip->hmp);
1791         ++hammer_stats_file_iopsr;
1792         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1793
1794         /*
1795          * Key range (begin and end inclusive) to scan.  Directory keys
1796          * directly translate to a 64 bit 'seek' position.
1797          */
1798         cursor.key_beg.localization = ip->obj_localization +
1799                                       HAMMER_LOCALIZE_MISC;
1800         cursor.key_beg.obj_id = ip->obj_id;
1801         cursor.key_beg.create_tid = 0;
1802         cursor.key_beg.delete_tid = 0;
1803         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1804         cursor.key_beg.obj_type = 0;
1805         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1806         cursor.asof = ip->obj_asof;
1807         cursor.flags |= HAMMER_CURSOR_ASOF;
1808
1809         error = hammer_ip_lookup(&cursor);
1810         if (error == 0) {
1811                 error = hammer_ip_resolve_data(&cursor);
1812                 if (error == 0) {
1813                         KKASSERT(cursor.leaf->data_len >=
1814                                  HAMMER_SYMLINK_NAME_OFF);
1815                         error = uiomove(cursor.data->symlink.name,
1816                                         cursor.leaf->data_len -
1817                                                 HAMMER_SYMLINK_NAME_OFF,
1818                                         ap->a_uio);
1819                 }
1820         }
1821         hammer_done_cursor(&cursor);
1822         hammer_done_transaction(&trans);
1823         return(error);
1824 }
1825
1826 /*
1827  * hammer_vop_nremove { nch, dvp, cred }
1828  */
1829 static
1830 int
1831 hammer_vop_nremove(struct vop_nremove_args *ap)
1832 {
1833         struct hammer_transaction trans;
1834         struct hammer_inode *dip;
1835         int error;
1836
1837         dip = VTOI(ap->a_dvp);
1838
1839         if (hammer_nohistory(dip) == 0 &&
1840             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1841                 return (error);
1842         }
1843
1844         hammer_start_transaction(&trans, dip->hmp);
1845         ++hammer_stats_file_iopsw;
1846         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1847         hammer_done_transaction(&trans);
1848         if (error == 0)
1849                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1850         return (error);
1851 }
1852
1853 /*
1854  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1855  */
1856 static
1857 int
1858 hammer_vop_nrename(struct vop_nrename_args *ap)
1859 {
1860         struct hammer_transaction trans;
1861         struct namecache *fncp;
1862         struct namecache *tncp;
1863         struct hammer_inode *fdip;
1864         struct hammer_inode *tdip;
1865         struct hammer_inode *ip;
1866         struct hammer_cursor cursor;
1867         int64_t namekey;
1868         u_int32_t max_iterations;
1869         int nlen, error;
1870
1871         if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount) 
1872                 return(EXDEV);
1873         if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1874                 return(EXDEV);
1875
1876         fdip = VTOI(ap->a_fdvp);
1877         tdip = VTOI(ap->a_tdvp);
1878         fncp = ap->a_fnch->ncp;
1879         tncp = ap->a_tnch->ncp;
1880         ip = VTOI(fncp->nc_vp);
1881         KKASSERT(ip != NULL);
1882
1883         if (fdip->obj_localization != tdip->obj_localization)
1884                 return(EXDEV);
1885         if (fdip->obj_localization != ip->obj_localization)
1886                 return(EXDEV);
1887
1888         if (fdip->flags & HAMMER_INODE_RO)
1889                 return (EROFS);
1890         if (tdip->flags & HAMMER_INODE_RO)
1891                 return (EROFS);
1892         if (ip->flags & HAMMER_INODE_RO)
1893                 return (EROFS);
1894         if ((error = hammer_checkspace(fdip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1895                 return (error);
1896
1897         hammer_start_transaction(&trans, fdip->hmp);
1898         ++hammer_stats_file_iopsw;
1899
1900         /*
1901          * Remove tncp from the target directory and then link ip as
1902          * tncp. XXX pass trans to dounlink
1903          *
1904          * Force the inode sync-time to match the transaction so it is
1905          * in-sync with the creation of the target directory entry.
1906          */
1907         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
1908                                 ap->a_cred, 0, -1);
1909         if (error == 0 || error == ENOENT) {
1910                 error = hammer_ip_add_directory(&trans, tdip,
1911                                                 tncp->nc_name, tncp->nc_nlen,
1912                                                 ip);
1913                 if (error == 0) {
1914                         ip->ino_data.parent_obj_id = tdip->obj_id;
1915                         ip->ino_data.ctime = trans.time;
1916                         hammer_modify_inode(&trans, ip, HAMMER_INODE_DDIRTY);
1917                 }
1918         }
1919         if (error)
1920                 goto failed; /* XXX */
1921
1922         /*
1923          * Locate the record in the originating directory and remove it.
1924          *
1925          * Calculate the namekey and setup the key range for the scan.  This
1926          * works kinda like a chained hash table where the lower 32 bits
1927          * of the namekey synthesize the chain.
1928          *
1929          * The key range is inclusive of both key_beg and key_end.
1930          */
1931         namekey = hammer_directory_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
1932                                            &max_iterations);
1933 retry:
1934         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
1935         cursor.key_beg.localization = fdip->obj_localization +
1936                                       hammer_dir_localization(fdip);
1937         cursor.key_beg.obj_id = fdip->obj_id;
1938         cursor.key_beg.key = namekey;
1939         cursor.key_beg.create_tid = 0;
1940         cursor.key_beg.delete_tid = 0;
1941         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1942         cursor.key_beg.obj_type = 0;
1943
1944         cursor.key_end = cursor.key_beg;
1945         cursor.key_end.key += max_iterations;
1946         cursor.asof = fdip->obj_asof;
1947         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1948
1949         /*
1950          * Scan all matching records (the chain), locate the one matching
1951          * the requested path component.
1952          *
1953          * The hammer_ip_*() functions merge in-memory records with on-disk
1954          * records for the purposes of the search.
1955          */
1956         error = hammer_ip_first(&cursor);
1957         while (error == 0) {
1958                 if (hammer_ip_resolve_data(&cursor) != 0)
1959                         break;
1960                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
1961                 KKASSERT(nlen > 0);
1962                 if (fncp->nc_nlen == nlen &&
1963                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1964                         break;
1965                 }
1966                 error = hammer_ip_next(&cursor);
1967         }
1968
1969         /*
1970          * If all is ok we have to get the inode so we can adjust nlinks.
1971          *
1972          * WARNING: hammer_ip_del_directory() may have to terminate the
1973          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
1974          * twice.
1975          */
1976         if (error == 0)
1977                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
1978
1979         /*
1980          * XXX A deadlock here will break rename's atomicy for the purposes
1981          * of crash recovery.
1982          */
1983         if (error == EDEADLK) {
1984                 hammer_done_cursor(&cursor);
1985                 goto retry;
1986         }
1987
1988         /*
1989          * Cleanup and tell the kernel that the rename succeeded.
1990          */
1991         hammer_done_cursor(&cursor);
1992         if (error == 0) {
1993                 cache_rename(ap->a_fnch, ap->a_tnch);
1994                 hammer_knote(ap->a_fdvp, NOTE_WRITE);
1995                 hammer_knote(ap->a_tdvp, NOTE_WRITE);
1996                 if (ip->vp)
1997                         hammer_knote(ip->vp, NOTE_RENAME);
1998         }
1999
2000 failed:
2001         hammer_done_transaction(&trans);
2002         return (error);
2003 }
2004
2005 /*
2006  * hammer_vop_nrmdir { nch, dvp, cred }
2007  */
2008 static
2009 int
2010 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
2011 {
2012         struct hammer_transaction trans;
2013         struct hammer_inode *dip;
2014         int error;
2015
2016         dip = VTOI(ap->a_dvp);
2017
2018         if (hammer_nohistory(dip) == 0 &&
2019             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2020                 return (error);
2021         }
2022
2023         hammer_start_transaction(&trans, dip->hmp);
2024         ++hammer_stats_file_iopsw;
2025         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
2026         hammer_done_transaction(&trans);
2027         if (error == 0)
2028                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
2029         return (error);
2030 }
2031
2032 /*
2033  * hammer_vop_markatime { vp, cred }
2034  */
2035 static
2036 int
2037 hammer_vop_markatime(struct vop_markatime_args *ap)
2038 {
2039         struct hammer_transaction trans;
2040         struct hammer_inode *ip;
2041
2042         ip = VTOI(ap->a_vp);
2043         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2044                 return (EROFS);
2045         if (ip->flags & HAMMER_INODE_RO)
2046                 return (EROFS);
2047         if (ip->hmp->mp->mnt_flag & MNT_NOATIME)
2048                 return (0);
2049         hammer_start_transaction(&trans, ip->hmp);
2050         ++hammer_stats_file_iopsw;
2051
2052         ip->ino_data.atime = trans.time;
2053         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
2054         hammer_done_transaction(&trans);
2055         hammer_knote(ap->a_vp, NOTE_ATTRIB);
2056         return (0);
2057 }
2058
2059 /*
2060  * hammer_vop_setattr { vp, vap, cred }
2061  */
2062 static
2063 int
2064 hammer_vop_setattr(struct vop_setattr_args *ap)
2065 {
2066         struct hammer_transaction trans;
2067         struct vattr *vap;
2068         struct hammer_inode *ip;
2069         int modflags;
2070         int error;
2071         int truncating;
2072         int blksize;
2073         int kflags;
2074 #if 0
2075         int64_t aligned_size;
2076 #endif
2077         u_int32_t flags;
2078
2079         vap = ap->a_vap;
2080         ip = ap->a_vp->v_data;
2081         modflags = 0;
2082         kflags = 0;
2083
2084         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2085                 return(EROFS);
2086         if (ip->flags & HAMMER_INODE_RO)
2087                 return (EROFS);
2088         if (hammer_nohistory(ip) == 0 &&
2089             (error = hammer_checkspace(ip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2090                 return (error);
2091         }
2092
2093         hammer_start_transaction(&trans, ip->hmp);
2094         ++hammer_stats_file_iopsw;
2095         error = 0;
2096
2097         if (vap->va_flags != VNOVAL) {
2098                 flags = ip->ino_data.uflags;
2099                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
2100                                          hammer_to_unix_xid(&ip->ino_data.uid),
2101                                          ap->a_cred);
2102                 if (error == 0) {
2103                         if (ip->ino_data.uflags != flags) {
2104                                 ip->ino_data.uflags = flags;
2105                                 ip->ino_data.ctime = trans.time;
2106                                 modflags |= HAMMER_INODE_DDIRTY;
2107                                 kflags |= NOTE_ATTRIB;
2108                         }
2109                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2110                                 error = 0;
2111                                 goto done;
2112                         }
2113                 }
2114                 goto done;
2115         }
2116         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2117                 error = EPERM;
2118                 goto done;
2119         }
2120         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2121                 mode_t cur_mode = ip->ino_data.mode;
2122                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2123                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2124                 uuid_t uuid_uid;
2125                 uuid_t uuid_gid;
2126
2127                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2128                                          ap->a_cred,
2129                                          &cur_uid, &cur_gid, &cur_mode);
2130                 if (error == 0) {
2131                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
2132                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
2133                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
2134                                  sizeof(uuid_uid)) ||
2135                             bcmp(&uuid_gid, &ip->ino_data.gid,
2136                                  sizeof(uuid_gid)) ||
2137                             ip->ino_data.mode != cur_mode
2138                         ) {
2139                                 ip->ino_data.uid = uuid_uid;
2140                                 ip->ino_data.gid = uuid_gid;
2141                                 ip->ino_data.mode = cur_mode;
2142                                 ip->ino_data.ctime = trans.time;
2143                                 modflags |= HAMMER_INODE_DDIRTY;
2144                         }
2145                         kflags |= NOTE_ATTRIB;
2146                 }
2147         }
2148         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2149                 switch(ap->a_vp->v_type) {
2150                 case VREG:
2151                         if (vap->va_size == ip->ino_data.size)
2152                                 break;
2153
2154                         /*
2155                          * Log the operation if in fast-fsync mode or if
2156                          * there are unterminated redo write records present.
2157                          *
2158                          * The second check is needed so the recovery code
2159                          * properly truncates write redos even if nominal
2160                          * REDO operations is turned off due to excessive
2161                          * writes, because the related records might be
2162                          * destroyed and never lay down a TERM_WRITE.
2163                          */
2164                         if ((ip->flags & HAMMER_INODE_REDO) ||
2165                             (ip->flags & HAMMER_INODE_RDIRTY)) {
2166                                 error = hammer_generate_redo(&trans, ip,
2167                                                              vap->va_size,
2168                                                              HAMMER_REDO_TRUNC,
2169                                                              NULL, 0);
2170                         }
2171                         blksize = hammer_blocksize(vap->va_size);
2172
2173                         /*
2174                          * XXX break atomicy, we can deadlock the backend
2175                          * if we do not release the lock.  Probably not a
2176                          * big deal here.
2177                          */
2178                         if (vap->va_size < ip->ino_data.size) {
2179                                 nvtruncbuf(ap->a_vp, vap->va_size,
2180                                            blksize,
2181                                            hammer_blockoff(vap->va_size));
2182                                 truncating = 1;
2183                                 kflags |= NOTE_WRITE;
2184                         } else {
2185                                 nvextendbuf(ap->a_vp,
2186                                             ip->ino_data.size,
2187                                             vap->va_size,
2188                                             hammer_blocksize(ip->ino_data.size),
2189                                             hammer_blocksize(vap->va_size),
2190                                             hammer_blockoff(ip->ino_data.size),
2191                                             hammer_blockoff(vap->va_size),
2192                                             0);
2193                                 truncating = 0;
2194                                 kflags |= NOTE_WRITE | NOTE_EXTEND;
2195                         }
2196                         ip->ino_data.size = vap->va_size;
2197                         ip->ino_data.mtime = trans.time;
2198                         /* XXX safe to use SDIRTY instead of DDIRTY here? */
2199                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2200
2201                         /*
2202                          * On-media truncation is cached in the inode until
2203                          * the inode is synchronized.  We must immediately
2204                          * handle any frontend records.
2205                          */
2206                         if (truncating) {
2207                                 hammer_ip_frontend_trunc(ip, vap->va_size);
2208 #ifdef DEBUG_TRUNCATE
2209                                 if (HammerTruncIp == NULL)
2210                                         HammerTruncIp = ip;
2211 #endif
2212                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2213                                         ip->flags |= HAMMER_INODE_TRUNCATED;
2214                                         ip->trunc_off = vap->va_size;
2215 #ifdef DEBUG_TRUNCATE
2216                                         if (ip == HammerTruncIp)
2217                                         kprintf("truncate1 %016llx\n",
2218                                                 (long long)ip->trunc_off);
2219 #endif
2220                                 } else if (ip->trunc_off > vap->va_size) {
2221                                         ip->trunc_off = vap->va_size;
2222 #ifdef DEBUG_TRUNCATE
2223                                         if (ip == HammerTruncIp)
2224                                         kprintf("truncate2 %016llx\n",
2225                                                 (long long)ip->trunc_off);
2226 #endif
2227                                 } else {
2228 #ifdef DEBUG_TRUNCATE
2229                                         if (ip == HammerTruncIp)
2230                                         kprintf("truncate3 %016llx (ignored)\n",
2231                                                 (long long)vap->va_size);
2232 #endif
2233                                 }
2234                         }
2235
2236 #if 0
2237                         /*
2238                          * When truncating, nvtruncbuf() may have cleaned out
2239                          * a portion of the last block on-disk in the buffer
2240                          * cache.  We must clean out any frontend records
2241                          * for blocks beyond the new last block.
2242                          */
2243                         aligned_size = (vap->va_size + (blksize - 1)) &
2244                                        ~(int64_t)(blksize - 1);
2245                         if (truncating && vap->va_size < aligned_size) {
2246                                 aligned_size -= blksize;
2247                                 hammer_ip_frontend_trunc(ip, aligned_size);
2248                         }
2249 #endif
2250                         break;
2251                 case VDATABASE:
2252                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2253                                 ip->flags |= HAMMER_INODE_TRUNCATED;
2254                                 ip->trunc_off = vap->va_size;
2255                         } else if (ip->trunc_off > vap->va_size) {
2256                                 ip->trunc_off = vap->va_size;
2257                         }
2258                         hammer_ip_frontend_trunc(ip, vap->va_size);
2259                         ip->ino_data.size = vap->va_size;
2260                         ip->ino_data.mtime = trans.time;
2261                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2262                         kflags |= NOTE_ATTRIB;
2263                         break;
2264                 default:
2265                         error = EINVAL;
2266                         goto done;
2267                 }
2268                 break;
2269         }
2270         if (vap->va_atime.tv_sec != VNOVAL) {
2271                 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2272                 modflags |= HAMMER_INODE_ATIME;
2273                 kflags |= NOTE_ATTRIB;
2274         }
2275         if (vap->va_mtime.tv_sec != VNOVAL) {
2276                 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2277                 modflags |= HAMMER_INODE_MTIME;
2278                 kflags |= NOTE_ATTRIB;
2279         }
2280         if (vap->va_mode != (mode_t)VNOVAL) {
2281                 mode_t   cur_mode = ip->ino_data.mode;
2282                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2283                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2284
2285                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2286                                          cur_uid, cur_gid, &cur_mode);
2287                 if (error == 0 && ip->ino_data.mode != cur_mode) {
2288                         ip->ino_data.mode = cur_mode;
2289                         ip->ino_data.ctime = trans.time;
2290                         modflags |= HAMMER_INODE_DDIRTY;
2291                         kflags |= NOTE_ATTRIB;
2292                 }
2293         }
2294 done:
2295         if (error == 0)
2296                 hammer_modify_inode(&trans, ip, modflags);
2297         hammer_done_transaction(&trans);
2298         hammer_knote(ap->a_vp, kflags);
2299         return (error);
2300 }
2301
2302 /*
2303  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2304  */
2305 static
2306 int
2307 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2308 {
2309         struct hammer_transaction trans;
2310         struct hammer_inode *dip;
2311         struct hammer_inode *nip;
2312         struct nchandle *nch;
2313         hammer_record_t record;
2314         int error;
2315         int bytes;
2316
2317         ap->a_vap->va_type = VLNK;
2318
2319         nch = ap->a_nch;
2320         dip = VTOI(ap->a_dvp);
2321
2322         if (dip->flags & HAMMER_INODE_RO)
2323                 return (EROFS);
2324         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
2325                 return (error);
2326
2327         /*
2328          * Create a transaction to cover the operations we perform.
2329          */
2330         hammer_start_transaction(&trans, dip->hmp);
2331         ++hammer_stats_file_iopsw;
2332
2333         /*
2334          * Create a new filesystem object of the requested type.  The
2335          * returned inode will be referenced but not locked.
2336          */
2337
2338         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2339                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2340                                     NULL, &nip);
2341         if (error) {
2342                 hammer_done_transaction(&trans);
2343                 *ap->a_vpp = NULL;
2344                 return (error);
2345         }
2346
2347         /*
2348          * Add a record representing the symlink.  symlink stores the link
2349          * as pure data, not a string, and is no \0 terminated.
2350          */
2351         if (error == 0) {
2352                 bytes = strlen(ap->a_target);
2353
2354                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2355                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2356                 } else {
2357                         record = hammer_alloc_mem_record(nip, bytes);
2358                         record->type = HAMMER_MEM_RECORD_GENERAL;
2359
2360                         record->leaf.base.localization = nip->obj_localization +
2361                                                          HAMMER_LOCALIZE_MISC;
2362                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2363                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2364                         record->leaf.data_len = bytes;
2365                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2366                         bcopy(ap->a_target, record->data->symlink.name, bytes);
2367                         error = hammer_ip_add_record(&trans, record);
2368                 }
2369
2370                 /*
2371                  * Set the file size to the length of the link.
2372                  */
2373                 if (error == 0) {
2374                         nip->ino_data.size = bytes;
2375                         hammer_modify_inode(&trans, nip, HAMMER_INODE_DDIRTY);
2376                 }
2377         }
2378         if (error == 0)
2379                 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
2380                                                 nch->ncp->nc_nlen, nip);
2381
2382         /*
2383          * Finish up.
2384          */
2385         if (error) {
2386                 hammer_rel_inode(nip, 0);
2387                 *ap->a_vpp = NULL;
2388         } else {
2389                 error = hammer_get_vnode(nip, ap->a_vpp);
2390                 hammer_rel_inode(nip, 0);
2391                 if (error == 0) {
2392                         cache_setunresolved(ap->a_nch);
2393                         cache_setvp(ap->a_nch, *ap->a_vpp);
2394                         hammer_knote(ap->a_dvp, NOTE_WRITE);
2395                 }
2396         }
2397         hammer_done_transaction(&trans);
2398         return (error);
2399 }
2400
2401 /*
2402  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2403  */
2404 static
2405 int
2406 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2407 {
2408         struct hammer_transaction trans;
2409         struct hammer_inode *dip;
2410         int error;
2411
2412         dip = VTOI(ap->a_dvp);
2413
2414         if (hammer_nohistory(dip) == 0 &&
2415             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2416                 return (error);
2417         }
2418
2419         hammer_start_transaction(&trans, dip->hmp);
2420         ++hammer_stats_file_iopsw;
2421         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2422                                 ap->a_cred, ap->a_flags, -1);
2423         hammer_done_transaction(&trans);
2424
2425         return (error);
2426 }
2427
2428 /*
2429  * hammer_vop_ioctl { vp, command, data, fflag, cred }
2430  */
2431 static
2432 int
2433 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2434 {
2435         struct hammer_inode *ip = ap->a_vp->v_data;
2436
2437         ++hammer_stats_file_iopsr;
2438         return(hammer_ioctl(ip, ap->a_command, ap->a_data,
2439                             ap->a_fflag, ap->a_cred));
2440 }
2441
2442 static
2443 int
2444 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2445 {
2446         static const struct mountctl_opt extraopt[] = {
2447                 { HMNT_NOHISTORY,       "nohistory" },
2448                 { HMNT_MASTERID,        "master" },
2449                 { 0, NULL}
2450
2451         };
2452         struct hammer_mount *hmp;
2453         struct mount *mp;
2454         int usedbytes;
2455         int error;
2456
2457         error = 0;
2458         usedbytes = 0;
2459         mp = ap->a_head.a_ops->head.vv_mount;
2460         KKASSERT(mp->mnt_data != NULL);
2461         hmp = (struct hammer_mount *)mp->mnt_data;
2462
2463         switch(ap->a_op) {
2464
2465         case MOUNTCTL_SET_EXPORT:
2466                 if (ap->a_ctllen != sizeof(struct export_args))
2467                         error = EINVAL;
2468                 else
2469                         error = hammer_vfs_export(mp, ap->a_op,
2470                                       (const struct export_args *)ap->a_ctl);
2471                 break;
2472         case MOUNTCTL_MOUNTFLAGS:
2473         {
2474                 /*
2475                  * Call standard mountctl VOP function
2476                  * so we get user mount flags.
2477                  */
2478                 error = vop_stdmountctl(ap);
2479                 if (error)
2480                         break;
2481
2482                 usedbytes = *ap->a_res;
2483
2484                 if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2485                         usedbytes += vfs_flagstostr(hmp->hflags, extraopt, ap->a_buf,
2486                                                     ap->a_buflen - usedbytes,
2487                                                     &error);
2488                 }
2489
2490                 *ap->a_res += usedbytes;
2491                 break;
2492         }
2493         default:
2494                 error = vop_stdmountctl(ap);
2495                 break;
2496         }
2497         return(error);
2498 }
2499
2500 /*
2501  * hammer_vop_strategy { vp, bio }
2502  *
2503  * Strategy call, used for regular file read & write only.  Note that the
2504  * bp may represent a cluster.
2505  *
2506  * To simplify operation and allow better optimizations in the future,
2507  * this code does not make any assumptions with regards to buffer alignment
2508  * or size.
2509  */
2510 static
2511 int
2512 hammer_vop_strategy(struct vop_strategy_args *ap)
2513 {
2514         struct buf *bp;
2515         int error;
2516
2517         bp = ap->a_bio->bio_buf;
2518
2519         switch(bp->b_cmd) {
2520         case BUF_CMD_READ:
2521                 error = hammer_vop_strategy_read(ap);
2522                 break;
2523         case BUF_CMD_WRITE:
2524                 error = hammer_vop_strategy_write(ap);
2525                 break;
2526         default:
2527                 bp->b_error = error = EINVAL;
2528                 bp->b_flags |= B_ERROR;
2529                 biodone(ap->a_bio);
2530                 break;
2531         }
2532         return (error);
2533 }
2534
2535 /*
2536  * Read from a regular file.  Iterate the related records and fill in the
2537  * BIO/BUF.  Gaps are zero-filled.
2538  *
2539  * The support code in hammer_object.c should be used to deal with mixed
2540  * in-memory and on-disk records.
2541  *
2542  * NOTE: Can be called from the cluster code with an oversized buf.
2543  *
2544  * XXX atime update
2545  */
2546 static
2547 int
2548 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2549 {
2550         struct hammer_transaction trans;
2551         struct hammer_inode *ip;
2552         struct hammer_inode *dip;
2553         struct hammer_cursor cursor;
2554         hammer_base_elm_t base;
2555         hammer_off_t disk_offset;
2556         struct bio *bio;
2557         struct bio *nbio;
2558         struct buf *bp;
2559         int64_t rec_offset;
2560         int64_t ran_end;
2561         int64_t tmp64;
2562         int error;
2563         int boff;
2564         int roff;
2565         int n;
2566
2567         bio = ap->a_bio;
2568         bp = bio->bio_buf;
2569         ip = ap->a_vp->v_data;
2570
2571         /*
2572          * The zone-2 disk offset may have been set by the cluster code via
2573          * a BMAP operation, or else should be NOOFFSET.
2574          *
2575          * Checking the high bits for a match against zone-2 should suffice.
2576          */
2577         nbio = push_bio(bio);
2578         if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2579             HAMMER_ZONE_LARGE_DATA) {
2580                 error = hammer_io_direct_read(ip->hmp, nbio, NULL);
2581                 return (error);
2582         }
2583
2584         /*
2585          * Well, that sucked.  Do it the hard way.  If all the stars are
2586          * aligned we may still be able to issue a direct-read.
2587          */
2588         hammer_simple_transaction(&trans, ip->hmp);
2589         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2590
2591         /*
2592          * Key range (begin and end inclusive) to scan.  Note that the key's
2593          * stored in the actual records represent BASE+LEN, not BASE.  The
2594          * first record containing bio_offset will have a key > bio_offset.
2595          */
2596         cursor.key_beg.localization = ip->obj_localization +
2597                                       HAMMER_LOCALIZE_MISC;
2598         cursor.key_beg.obj_id = ip->obj_id;
2599         cursor.key_beg.create_tid = 0;
2600         cursor.key_beg.delete_tid = 0;
2601         cursor.key_beg.obj_type = 0;
2602         cursor.key_beg.key = bio->bio_offset + 1;
2603         cursor.asof = ip->obj_asof;
2604         cursor.flags |= HAMMER_CURSOR_ASOF;
2605
2606         cursor.key_end = cursor.key_beg;
2607         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2608 #if 0
2609         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2610                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2611                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2612                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2613         } else
2614 #endif
2615         {
2616                 ran_end = bio->bio_offset + bp->b_bufsize;
2617                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2618                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2619                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2620                 if (tmp64 < ran_end)
2621                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2622                 else
2623                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2624         }
2625         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2626
2627         error = hammer_ip_first(&cursor);
2628         boff = 0;
2629
2630         while (error == 0) {
2631                 /*
2632                  * Get the base file offset of the record.  The key for
2633                  * data records is (base + bytes) rather then (base).
2634                  */
2635                 base = &cursor.leaf->base;
2636                 rec_offset = base->key - cursor.leaf->data_len;
2637
2638                 /*
2639                  * Calculate the gap, if any, and zero-fill it.
2640                  *
2641                  * n is the offset of the start of the record verses our
2642                  * current seek offset in the bio.
2643                  */
2644                 n = (int)(rec_offset - (bio->bio_offset + boff));
2645                 if (n > 0) {
2646                         if (n > bp->b_bufsize - boff)
2647                                 n = bp->b_bufsize - boff;
2648                         bzero((char *)bp->b_data + boff, n);
2649                         boff += n;
2650                         n = 0;
2651                 }
2652
2653                 /*
2654                  * Calculate the data offset in the record and the number
2655                  * of bytes we can copy.
2656                  *
2657                  * There are two degenerate cases.  First, boff may already
2658                  * be at bp->b_bufsize.  Secondly, the data offset within
2659                  * the record may exceed the record's size.
2660                  */
2661                 roff = -n;
2662                 rec_offset += roff;
2663                 n = cursor.leaf->data_len - roff;
2664                 if (n <= 0) {
2665                         kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2666                         n = 0;
2667                 } else if (n > bp->b_bufsize - boff) {
2668                         n = bp->b_bufsize - boff;
2669                 }
2670
2671                 /*
2672                  * Deal with cached truncations.  This cool bit of code
2673                  * allows truncate()/ftruncate() to avoid having to sync
2674                  * the file.
2675                  *
2676                  * If the frontend is truncated then all backend records are
2677                  * subject to the frontend's truncation.
2678                  *
2679                  * If the backend is truncated then backend records on-disk
2680                  * (but not in-memory) are subject to the backend's
2681                  * truncation.  In-memory records owned by the backend
2682                  * represent data written after the truncation point on the
2683                  * backend and must not be truncated.
2684                  *
2685                  * Truncate operations deal with frontend buffer cache
2686                  * buffers and frontend-owned in-memory records synchronously.
2687                  */
2688                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2689                         if (hammer_cursor_ondisk(&cursor)/* ||
2690                             cursor.iprec->flush_state == HAMMER_FST_FLUSH*/) {
2691                                 if (ip->trunc_off <= rec_offset)
2692                                         n = 0;
2693                                 else if (ip->trunc_off < rec_offset + n)
2694                                         n = (int)(ip->trunc_off - rec_offset);
2695                         }
2696                 }
2697                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2698                         if (hammer_cursor_ondisk(&cursor)) {
2699                                 if (ip->sync_trunc_off <= rec_offset)
2700                                         n = 0;
2701                                 else if (ip->sync_trunc_off < rec_offset + n)
2702                                         n = (int)(ip->sync_trunc_off - rec_offset);
2703                         }
2704                 }
2705
2706                 /*
2707                  * Try to issue a direct read into our bio if possible,
2708                  * otherwise resolve the element data into a hammer_buffer
2709                  * and copy.
2710                  *
2711                  * The buffer on-disk should be zerod past any real
2712                  * truncation point, but may not be for any synthesized
2713                  * truncation point from above.
2714                  */
2715                 disk_offset = cursor.leaf->data_offset + roff;
2716                 if (boff == 0 && n == bp->b_bufsize &&
2717                     hammer_cursor_ondisk(&cursor) &&
2718                     (disk_offset & HAMMER_BUFMASK) == 0) {
2719                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2720                                  HAMMER_ZONE_LARGE_DATA);
2721                         nbio->bio_offset = disk_offset;
2722                         error = hammer_io_direct_read(trans.hmp, nbio,
2723                                                       cursor.leaf);
2724                         goto done;
2725                 } else if (n) {
2726                         error = hammer_ip_resolve_data(&cursor);
2727                         if (error == 0) {
2728                                 bcopy((char *)cursor.data + roff,
2729                                       (char *)bp->b_data + boff, n);
2730                         }
2731                 }
2732                 if (error)
2733                         break;
2734
2735                 /*
2736                  * Iterate until we have filled the request.
2737                  */
2738                 boff += n;
2739                 if (boff == bp->b_bufsize)
2740                         break;
2741                 error = hammer_ip_next(&cursor);
2742         }
2743
2744         /*
2745          * There may have been a gap after the last record
2746          */
2747         if (error == ENOENT)
2748                 error = 0;
2749         if (error == 0 && boff != bp->b_bufsize) {
2750                 KKASSERT(boff < bp->b_bufsize);
2751                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2752                 /* boff = bp->b_bufsize; */
2753         }
2754         bp->b_resid = 0;
2755         bp->b_error = error;
2756         if (error)
2757                 bp->b_flags |= B_ERROR;
2758         biodone(ap->a_bio);
2759
2760 done:
2761         /*
2762          * Cache the b-tree node for the last data read in cache[1].
2763          *
2764          * If we hit the file EOF then also cache the node in the
2765          * governing director's cache[3], it will be used to initialize
2766          * the inode's cache[1] for any inodes looked up via the directory.
2767          *
2768          * This doesn't reduce disk accesses since the B-Tree chain is
2769          * likely cached, but it does reduce cpu overhead when looking
2770          * up file offsets for cpdup/tar/cpio style iterations.
2771          */
2772         if (cursor.node)
2773                 hammer_cache_node(&ip->cache[1], cursor.node);
2774         if (ran_end >= ip->ino_data.size) {
2775                 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
2776                                         ip->obj_asof, ip->obj_localization);
2777                 if (dip) {
2778                         hammer_cache_node(&dip->cache[3], cursor.node);
2779                         hammer_rel_inode(dip, 0);
2780                 }
2781         }
2782         hammer_done_cursor(&cursor);
2783         hammer_done_transaction(&trans);
2784         return(error);
2785 }
2786
2787 /*
2788  * BMAP operation - used to support cluster_read() only.
2789  *
2790  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2791  *
2792  * This routine may return EOPNOTSUPP if the opration is not supported for
2793  * the specified offset.  The contents of the pointer arguments do not
2794  * need to be initialized in that case. 
2795  *
2796  * If a disk address is available and properly aligned return 0 with 
2797  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2798  * to the run-length relative to that offset.  Callers may assume that
2799  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2800  * large, so return EOPNOTSUPP if it is not sufficiently large.
2801  */
2802 static
2803 int
2804 hammer_vop_bmap(struct vop_bmap_args *ap)
2805 {
2806         struct hammer_transaction trans;
2807         struct hammer_inode *ip;
2808         struct hammer_cursor cursor;
2809         hammer_base_elm_t base;
2810         int64_t rec_offset;
2811         int64_t ran_end;
2812         int64_t tmp64;
2813         int64_t base_offset;
2814         int64_t base_disk_offset;
2815         int64_t last_offset;
2816         hammer_off_t last_disk_offset;
2817         hammer_off_t disk_offset;
2818         int     rec_len;
2819         int     error;
2820         int     blksize;
2821
2822         ++hammer_stats_file_iopsr;
2823         ip = ap->a_vp->v_data;
2824
2825         /*
2826          * We can only BMAP regular files.  We can't BMAP database files,
2827          * directories, etc.
2828          */
2829         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
2830                 return(EOPNOTSUPP);
2831
2832         /*
2833          * bmap is typically called with runp/runb both NULL when used
2834          * for writing.  We do not support BMAP for writing atm.
2835          */
2836         if (ap->a_cmd != BUF_CMD_READ)
2837                 return(EOPNOTSUPP);
2838
2839         /*
2840          * Scan the B-Tree to acquire blockmap addresses, then translate
2841          * to raw addresses.
2842          */
2843         hammer_simple_transaction(&trans, ip->hmp);
2844 #if 0
2845         kprintf("bmap_beg %016llx ip->cache %p\n",
2846                 (long long)ap->a_loffset, ip->cache[1]);
2847 #endif
2848         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2849
2850         /*
2851          * Key range (begin and end inclusive) to scan.  Note that the key's
2852          * stored in the actual records represent BASE+LEN, not BASE.  The
2853          * first record containing bio_offset will have a key > bio_offset.
2854          */
2855         cursor.key_beg.localization = ip->obj_localization +
2856                                       HAMMER_LOCALIZE_MISC;
2857         cursor.key_beg.obj_id = ip->obj_id;
2858         cursor.key_beg.create_tid = 0;
2859         cursor.key_beg.delete_tid = 0;
2860         cursor.key_beg.obj_type = 0;
2861         if (ap->a_runb)
2862                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
2863         else
2864                 cursor.key_beg.key = ap->a_loffset + 1;
2865         if (cursor.key_beg.key < 0)
2866                 cursor.key_beg.key = 0;
2867         cursor.asof = ip->obj_asof;
2868         cursor.flags |= HAMMER_CURSOR_ASOF;
2869
2870         cursor.key_end = cursor.key_beg;
2871         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2872
2873         ran_end = ap->a_loffset + MAXPHYS;
2874         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2875         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2876         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2877         if (tmp64 < ran_end)
2878                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2879         else
2880                 cursor.key_end.key = ran_end + MAXPHYS + 1;
2881
2882         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2883
2884         error = hammer_ip_first(&cursor);
2885         base_offset = last_offset = 0;
2886         base_disk_offset = last_disk_offset = 0;
2887
2888         while (error == 0) {
2889                 /*
2890                  * Get the base file offset of the record.  The key for
2891                  * data records is (base + bytes) rather then (base).
2892                  *
2893                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
2894                  * The extra bytes should be zero on-disk and the BMAP op
2895                  * should still be ok.
2896                  */
2897                 base = &cursor.leaf->base;
2898                 rec_offset = base->key - cursor.leaf->data_len;
2899                 rec_len    = cursor.leaf->data_len;
2900
2901                 /*
2902                  * Incorporate any cached truncation.
2903                  *
2904                  * NOTE: Modifications to rec_len based on synthesized
2905                  * truncation points remove the guarantee that any extended
2906                  * data on disk is zero (since the truncations may not have
2907                  * taken place on-media yet).
2908                  */
2909                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2910                         if (hammer_cursor_ondisk(&cursor) ||
2911                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2912                                 if (ip->trunc_off <= rec_offset)
2913                                         rec_len = 0;
2914                                 else if (ip->trunc_off < rec_offset + rec_len)
2915                                         rec_len = (int)(ip->trunc_off - rec_offset);
2916                         }
2917                 }
2918                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2919                         if (hammer_cursor_ondisk(&cursor)) {
2920                                 if (ip->sync_trunc_off <= rec_offset)
2921                                         rec_len = 0;
2922                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
2923                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
2924                         }
2925                 }
2926
2927                 /*
2928                  * Accumulate information.  If we have hit a discontiguous
2929                  * block reset base_offset unless we are already beyond the
2930                  * requested offset.  If we are, that's it, we stop.
2931                  */
2932                 if (error)
2933                         break;
2934                 if (hammer_cursor_ondisk(&cursor)) {
2935                         disk_offset = cursor.leaf->data_offset;
2936                         if (rec_offset != last_offset ||
2937                             disk_offset != last_disk_offset) {
2938                                 if (rec_offset > ap->a_loffset)
2939                                         break;
2940                                 base_offset = rec_offset;
2941                                 base_disk_offset = disk_offset;
2942                         }
2943                         last_offset = rec_offset + rec_len;
2944                         last_disk_offset = disk_offset + rec_len;
2945                 }
2946                 error = hammer_ip_next(&cursor);
2947         }
2948
2949 #if 0
2950         kprintf("BMAP %016llx:  %016llx - %016llx\n",
2951                 (long long)ap->a_loffset,
2952                 (long long)base_offset,
2953                 (long long)last_offset);
2954         kprintf("BMAP %16s:  %016llx - %016llx\n", "",
2955                 (long long)base_disk_offset,
2956                 (long long)last_disk_offset);
2957 #endif
2958
2959         if (cursor.node) {
2960                 hammer_cache_node(&ip->cache[1], cursor.node);
2961 #if 0
2962                 kprintf("bmap_end2 %016llx ip->cache %p\n",
2963                         (long long)ap->a_loffset, ip->cache[1]);
2964 #endif
2965         }
2966         hammer_done_cursor(&cursor);
2967         hammer_done_transaction(&trans);
2968
2969         /*
2970          * If we couldn't find any records or the records we did find were
2971          * all behind the requested offset, return failure.  A forward
2972          * truncation can leave a hole w/ no on-disk records.
2973          */
2974         if (last_offset == 0 || last_offset < ap->a_loffset)
2975                 return (EOPNOTSUPP);
2976
2977         /*
2978          * Figure out the block size at the requested offset and adjust
2979          * our limits so the cluster_read() does not create inappropriately
2980          * sized buffer cache buffers.
2981          */
2982         blksize = hammer_blocksize(ap->a_loffset);
2983         if (hammer_blocksize(base_offset) != blksize) {
2984                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
2985         }
2986         if (last_offset != ap->a_loffset &&
2987             hammer_blocksize(last_offset - 1) != blksize) {
2988                 last_offset = hammer_blockdemarc(ap->a_loffset,
2989                                                  last_offset - 1);
2990         }
2991
2992         /*
2993          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
2994          * from occuring.
2995          */
2996         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
2997
2998         if ((disk_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_LARGE_DATA) {
2999                 /*
3000                  * Only large-data zones can be direct-IOd
3001                  */
3002                 error = EOPNOTSUPP;
3003         } else if ((disk_offset & HAMMER_BUFMASK) ||
3004                    (last_offset - ap->a_loffset) < blksize) {
3005                 /*
3006                  * doffsetp is not aligned or the forward run size does
3007                  * not cover a whole buffer, disallow the direct I/O.
3008                  */
3009                 error = EOPNOTSUPP;
3010         } else {
3011                 /*
3012                  * We're good.
3013                  */
3014                 *ap->a_doffsetp = disk_offset;
3015                 if (ap->a_runb) {
3016                         *ap->a_runb = ap->a_loffset - base_offset;
3017                         KKASSERT(*ap->a_runb >= 0);
3018                 }
3019                 if (ap->a_runp) {
3020                         *ap->a_runp = last_offset - ap->a_loffset;
3021                         KKASSERT(*ap->a_runp >= 0);
3022                 }
3023                 error = 0;
3024         }
3025         return(error);
3026 }
3027
3028 /*
3029  * Write to a regular file.   Because this is a strategy call the OS is
3030  * trying to actually get data onto the media.
3031  */
3032 static
3033 int
3034 hammer_vop_strategy_write(struct vop_strategy_args *ap)
3035 {
3036         hammer_record_t record;
3037         hammer_mount_t hmp;
3038         hammer_inode_t ip;
3039         struct bio *bio;
3040         struct buf *bp;
3041         int blksize;
3042         int bytes;
3043         int error;
3044
3045         bio = ap->a_bio;
3046         bp = bio->bio_buf;
3047         ip = ap->a_vp->v_data;
3048         hmp = ip->hmp;
3049
3050         blksize = hammer_blocksize(bio->bio_offset);
3051         KKASSERT(bp->b_bufsize == blksize);
3052
3053         if (ip->flags & HAMMER_INODE_RO) {
3054                 bp->b_error = EROFS;
3055                 bp->b_flags |= B_ERROR;
3056                 biodone(ap->a_bio);
3057                 return(EROFS);
3058         }
3059
3060         /*
3061          * Interlock with inode destruction (no in-kernel or directory
3062          * topology visibility).  If we queue new IO while trying to
3063          * destroy the inode we can deadlock the vtrunc call in
3064          * hammer_inode_unloadable_check().
3065          *
3066          * Besides, there's no point flushing a bp associated with an
3067          * inode that is being destroyed on-media and has no kernel
3068          * references.
3069          */
3070         if ((ip->flags | ip->sync_flags) &
3071             (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
3072                 bp->b_resid = 0;
3073                 biodone(ap->a_bio);
3074                 return(0);
3075         }
3076
3077         /*
3078          * Reserve space and issue a direct-write from the front-end. 
3079          * NOTE: The direct_io code will hammer_bread/bcopy smaller
3080          * allocations.
3081          *
3082          * An in-memory record will be installed to reference the storage
3083          * until the flusher can get to it.
3084          *
3085          * Since we own the high level bio the front-end will not try to
3086          * do a direct-read until the write completes.
3087          *
3088          * NOTE: The only time we do not reserve a full-sized buffers
3089          * worth of data is if the file is small.  We do not try to
3090          * allocate a fragment (from the small-data zone) at the end of
3091          * an otherwise large file as this can lead to wildly separated
3092          * data.
3093          */
3094         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
3095         KKASSERT(bio->bio_offset < ip->ino_data.size);
3096         if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
3097                 bytes = bp->b_bufsize;
3098         else
3099                 bytes = ((int)ip->ino_data.size + 15) & ~15;
3100
3101         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
3102                                     bytes, &error);
3103
3104         /*
3105          * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3106          * in hammer_vop_write().  We must flag the record so the proper
3107          * REDO_TERM_WRITE entry is generated during the flush.
3108          */
3109         if (record) {
3110                 if (bp->b_flags & B_VFSFLAG1) {
3111                         record->flags |= HAMMER_RECF_REDO;
3112                         bp->b_flags &= ~B_VFSFLAG1;
3113                 }
3114                 hammer_io_direct_write(hmp, bio, record);
3115                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
3116                         hammer_flush_inode(ip, 0);
3117         } else {
3118                 bp->b_bio2.bio_offset = NOOFFSET;
3119                 bp->b_error = error;
3120                 bp->b_flags |= B_ERROR;
3121                 biodone(ap->a_bio);
3122         }
3123         return(error);
3124 }
3125
3126 /*
3127  * dounlink - disconnect a directory entry
3128  *
3129  * XXX whiteout support not really in yet
3130  */
3131 static int
3132 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
3133                 struct vnode *dvp, struct ucred *cred, 
3134                 int flags, int isdir)
3135 {
3136         struct namecache *ncp;
3137         hammer_inode_t dip;
3138         hammer_inode_t ip;
3139         struct hammer_cursor cursor;
3140         int64_t namekey;
3141         u_int32_t max_iterations;
3142         int nlen, error;
3143
3144         /*
3145          * Calculate the namekey and setup the key range for the scan.  This
3146          * works kinda like a chained hash table where the lower 32 bits
3147          * of the namekey synthesize the chain.
3148          *
3149          * The key range is inclusive of both key_beg and key_end.
3150          */
3151         dip = VTOI(dvp);
3152         ncp = nch->ncp;
3153
3154         if (dip->flags & HAMMER_INODE_RO)
3155                 return (EROFS);
3156
3157         namekey = hammer_directory_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3158                                            &max_iterations);
3159 retry:
3160         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3161         cursor.key_beg.localization = dip->obj_localization +
3162                                       hammer_dir_localization(dip);
3163         cursor.key_beg.obj_id = dip->obj_id;
3164         cursor.key_beg.key = namekey;
3165         cursor.key_beg.create_tid = 0;
3166         cursor.key_beg.delete_tid = 0;
3167         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3168         cursor.key_beg.obj_type = 0;
3169
3170         cursor.key_end = cursor.key_beg;
3171         cursor.key_end.key += max_iterations;
3172         cursor.asof = dip->obj_asof;
3173         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3174
3175         /*
3176          * Scan all matching records (the chain), locate the one matching
3177          * the requested path component.  info->last_error contains the
3178          * error code on search termination and could be 0, ENOENT, or
3179          * something else.
3180          *
3181          * The hammer_ip_*() functions merge in-memory records with on-disk
3182          * records for the purposes of the search.
3183          */
3184         error = hammer_ip_first(&cursor);
3185
3186         while (error == 0) {
3187                 error = hammer_ip_resolve_data(&cursor);
3188                 if (error)
3189                         break;
3190                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3191                 KKASSERT(nlen > 0);
3192                 if (ncp->nc_nlen == nlen &&
3193                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3194                         break;
3195                 }
3196                 error = hammer_ip_next(&cursor);
3197         }
3198
3199         /*
3200          * If all is ok we have to get the inode so we can adjust nlinks.
3201          * To avoid a deadlock with the flusher we must release the inode
3202          * lock on the directory when acquiring the inode for the entry.
3203          *
3204          * If the target is a directory, it must be empty.
3205          */
3206         if (error == 0) {
3207                 hammer_unlock(&cursor.ip->lock);
3208                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3209                                       dip->hmp->asof,
3210                                       cursor.data->entry.localization,
3211                                       0, &error);
3212                 hammer_lock_sh(&cursor.ip->lock);
3213                 if (error == ENOENT) {
3214                         kprintf("HAMMER: WARNING: Removing "
3215                                 "dirent w/missing inode \"%s\"\n"
3216                                 "\tobj_id = %016llx\n",
3217                                 ncp->nc_name,
3218                                 (long long)cursor.data->entry.obj_id);
3219                         error = 0;
3220                 }
3221
3222                 /*
3223                  * If isdir >= 0 we validate that the entry is or is not a
3224                  * directory.  If isdir < 0 we don't care.
3225                  */
3226                 if (error == 0 && isdir >= 0 && ip) {
3227                         if (isdir &&
3228                             ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3229                                 error = ENOTDIR;
3230                         } else if (isdir == 0 &&
3231                             ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3232                                 error = EISDIR;
3233                         }
3234                 }
3235
3236                 /*
3237                  * If we are trying to remove a directory the directory must
3238                  * be empty.
3239                  *
3240                  * The check directory code can loop and deadlock/retry.  Our
3241                  * own cursor's node locks must be released to avoid a 3-way
3242                  * deadlock with the flusher if the check directory code
3243                  * blocks.
3244                  *
3245                  * If any changes whatsoever have been made to the cursor
3246                  * set EDEADLK and retry.
3247                  *
3248                  * WARNING: See warnings in hammer_unlock_cursor()
3249                  *          function.
3250                  */
3251                 if (error == 0 && ip && ip->ino_data.obj_type ==
3252                                         HAMMER_OBJTYPE_DIRECTORY) {
3253                         hammer_unlock_cursor(&cursor);
3254                         error = hammer_ip_check_directory_empty(trans, ip);
3255                         hammer_lock_cursor(&cursor);
3256                         if (cursor.flags & HAMMER_CURSOR_RETEST) {
3257                                 kprintf("HAMMER: Warning: avoided deadlock "
3258                                         "on rmdir '%s'\n",
3259                                         ncp->nc_name);
3260                                 error = EDEADLK;
3261                         }
3262                 }
3263
3264                 /*
3265                  * Delete the directory entry.
3266                  *
3267                  * WARNING: hammer_ip_del_directory() may have to terminate
3268                  * the cursor to avoid a deadlock.  It is ok to call
3269                  * hammer_done_cursor() twice.
3270                  */
3271                 if (error == 0) {
3272                         error = hammer_ip_del_directory(trans, &cursor,
3273                                                         dip, ip);
3274                 }
3275                 hammer_done_cursor(&cursor);
3276                 if (error == 0) {
3277                         cache_setunresolved(nch);
3278                         cache_setvp(nch, NULL);
3279
3280                         /*
3281                          * XXX locking.  Note: ip->vp might get ripped out
3282                          * when we setunresolved() the nch since we had
3283                          * no other reference to it.  In that case ip->vp
3284                          * will be NULL.
3285                          */
3286                         if (ip && ip->vp) {
3287                                 hammer_knote(ip->vp, NOTE_DELETE);
3288                                 cache_inval_vp(ip->vp, CINV_DESTROY);
3289                         }
3290                 }
3291                 if (ip)
3292                         hammer_rel_inode(ip, 0);
3293         } else {
3294                 hammer_done_cursor(&cursor);
3295         }
3296         if (error == EDEADLK)
3297                 goto retry;
3298
3299         return (error);
3300 }
3301
3302 /************************************************************************
3303  *                          FIFO AND SPECFS OPS                         *
3304  ************************************************************************
3305  *
3306  */
3307
3308 static int
3309 hammer_vop_fifoclose (struct vop_close_args *ap)
3310 {
3311         /* XXX update itimes */
3312         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3313 }
3314
3315 static int
3316 hammer_vop_fiforead (struct vop_read_args *ap)
3317 {
3318         int error;
3319
3320         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3321         /* XXX update access time */
3322         return (error);
3323 }
3324
3325 static int
3326 hammer_vop_fifowrite (struct vop_write_args *ap)
3327 {
3328         int error;
3329
3330         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3331         /* XXX update access time */
3332         return (error);
3333 }
3334
3335 static
3336 int
3337 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3338 {
3339         int error;
3340
3341         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3342         if (error)
3343                 error = hammer_vop_kqfilter(ap);
3344         return(error);
3345 }
3346
3347 /************************************************************************
3348  *                          KQFILTER OPS                                *
3349  ************************************************************************
3350  *
3351  */
3352 static void filt_hammerdetach(struct knote *kn);
3353 static int filt_hammerread(struct knote *kn, long hint);
3354 static int filt_hammerwrite(struct knote *kn, long hint);
3355 static int filt_hammervnode(struct knote *kn, long hint);
3356
3357 static struct filterops hammerread_filtops =
3358         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammerread };
3359 static struct filterops hammerwrite_filtops =
3360         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammerwrite };
3361 static struct filterops hammervnode_filtops =
3362         { FILTEROP_ISFD, NULL, filt_hammerdetach, filt_hammervnode };
3363
3364 static
3365 int
3366 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3367 {
3368         struct vnode *vp = ap->a_vp;
3369         struct knote *kn = ap->a_kn;
3370
3371         switch (kn->kn_filter) {
3372         case EVFILT_READ:
3373                 kn->kn_fop = &hammerread_filtops;
3374                 break;
3375         case EVFILT_WRITE:
3376                 kn->kn_fop = &hammerwrite_filtops;
3377                 break;
3378         case EVFILT_VNODE:
3379                 kn->kn_fop = &hammervnode_filtops;
3380                 break;
3381         default:
3382                 return (EOPNOTSUPP);
3383         }
3384
3385         kn->kn_hook = (caddr_t)vp;
3386
3387         knote_insert(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3388
3389         return(0);
3390 }
3391
3392 static void
3393 filt_hammerdetach(struct knote *kn)
3394 {
3395         struct vnode *vp = (void *)kn->kn_hook;
3396
3397         knote_remove(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3398 }
3399
3400 static int
3401 filt_hammerread(struct knote *kn, long hint)
3402 {
3403         struct vnode *vp = (void *)kn->kn_hook;
3404         hammer_inode_t ip = VTOI(vp);
3405         off_t off;
3406
3407         if (hint == NOTE_REVOKE) {
3408                 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3409                 return(1);
3410         }
3411         off = ip->ino_data.size - kn->kn_fp->f_offset;
3412         kn->kn_data = (off < INTPTR_MAX) ? off : INTPTR_MAX;
3413         if (kn->kn_sfflags & NOTE_OLDAPI)
3414                 return(1);
3415         return (kn->kn_data != 0);
3416 }
3417
3418 static int
3419 filt_hammerwrite(struct knote *kn, long hint)
3420 {
3421         if (hint == NOTE_REVOKE)
3422                 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3423         kn->kn_data = 0;
3424         return (1);
3425 }
3426
3427 static int
3428 filt_hammervnode(struct knote *kn, long hint)
3429 {
3430         if (kn->kn_sfflags & hint)
3431                 kn->kn_fflags |= hint;
3432         if (hint == NOTE_REVOKE) {
3433                 kn->kn_flags |= EV_EOF;
3434                 return (1);
3435         }
3436         return (kn->kn_fflags != 0);
3437 }
3438