sys/vfs/hammer: Remove header includes from hammer.h
[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
35 #include <sys/mountctl.h>
36 #include <sys/namecache.h>
37 #include <sys/buf2.h>
38 #include <vfs/fifofs/fifo.h>
39
40 #include "hammer.h"
41
42 /*
43  * USERFS VNOPS
44  */
45 static int hammer_vop_fsync(struct vop_fsync_args *);
46 static int hammer_vop_read(struct vop_read_args *);
47 static int hammer_vop_write(struct vop_write_args *);
48 static int hammer_vop_access(struct vop_access_args *);
49 static int hammer_vop_advlock(struct vop_advlock_args *);
50 static int hammer_vop_close(struct vop_close_args *);
51 static int hammer_vop_ncreate(struct vop_ncreate_args *);
52 static int hammer_vop_getattr(struct vop_getattr_args *);
53 static int hammer_vop_nresolve(struct vop_nresolve_args *);
54 static int hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *);
55 static int hammer_vop_nlink(struct vop_nlink_args *);
56 static int hammer_vop_nmkdir(struct vop_nmkdir_args *);
57 static int hammer_vop_nmknod(struct vop_nmknod_args *);
58 static int hammer_vop_open(struct vop_open_args *);
59 static int hammer_vop_print(struct vop_print_args *);
60 static int hammer_vop_readdir(struct vop_readdir_args *);
61 static int hammer_vop_readlink(struct vop_readlink_args *);
62 static int hammer_vop_nremove(struct vop_nremove_args *);
63 static int hammer_vop_nrename(struct vop_nrename_args *);
64 static int hammer_vop_nrmdir(struct vop_nrmdir_args *);
65 static int hammer_vop_markatime(struct vop_markatime_args *);
66 static int hammer_vop_setattr(struct vop_setattr_args *);
67 static int hammer_vop_strategy(struct vop_strategy_args *);
68 static int hammer_vop_bmap(struct vop_bmap_args *ap);
69 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
70 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
71 static int hammer_vop_ioctl(struct vop_ioctl_args *);
72 static int hammer_vop_mountctl(struct vop_mountctl_args *);
73 static int hammer_vop_kqfilter (struct vop_kqfilter_args *);
74
75 static int hammer_vop_fifoclose (struct vop_close_args *);
76 static int hammer_vop_fiforead (struct vop_read_args *);
77 static int hammer_vop_fifowrite (struct vop_write_args *);
78 static int hammer_vop_fifokqfilter (struct vop_kqfilter_args *);
79
80 struct vop_ops hammer_vnode_vops = {
81         .vop_default =          vop_defaultop,
82         .vop_fsync =            hammer_vop_fsync,
83         .vop_getpages =         vop_stdgetpages,
84         .vop_putpages =         vop_stdputpages,
85         .vop_read =             hammer_vop_read,
86         .vop_write =            hammer_vop_write,
87         .vop_access =           hammer_vop_access,
88         .vop_advlock =          hammer_vop_advlock,
89         .vop_close =            hammer_vop_close,
90         .vop_ncreate =          hammer_vop_ncreate,
91         .vop_getattr =          hammer_vop_getattr,
92         .vop_inactive =         hammer_vop_inactive,
93         .vop_reclaim =          hammer_vop_reclaim,
94         .vop_nresolve =         hammer_vop_nresolve,
95         .vop_nlookupdotdot =    hammer_vop_nlookupdotdot,
96         .vop_nlink =            hammer_vop_nlink,
97         .vop_nmkdir =           hammer_vop_nmkdir,
98         .vop_nmknod =           hammer_vop_nmknod,
99         .vop_open =             hammer_vop_open,
100         .vop_pathconf =         vop_stdpathconf,
101         .vop_print =            hammer_vop_print,
102         .vop_readdir =          hammer_vop_readdir,
103         .vop_readlink =         hammer_vop_readlink,
104         .vop_nremove =          hammer_vop_nremove,
105         .vop_nrename =          hammer_vop_nrename,
106         .vop_nrmdir =           hammer_vop_nrmdir,
107         .vop_markatime =        hammer_vop_markatime,
108         .vop_setattr =          hammer_vop_setattr,
109         .vop_bmap =             hammer_vop_bmap,
110         .vop_strategy =         hammer_vop_strategy,
111         .vop_nsymlink =         hammer_vop_nsymlink,
112         .vop_nwhiteout =        hammer_vop_nwhiteout,
113         .vop_ioctl =            hammer_vop_ioctl,
114         .vop_mountctl =         hammer_vop_mountctl,
115         .vop_kqfilter =         hammer_vop_kqfilter
116 };
117
118 struct vop_ops hammer_spec_vops = {
119         .vop_default =          vop_defaultop,
120         .vop_fsync =            hammer_vop_fsync,
121         .vop_read =             vop_stdnoread,
122         .vop_write =            vop_stdnowrite,
123         .vop_access =           hammer_vop_access,
124         .vop_close =            hammer_vop_close,
125         .vop_markatime =        hammer_vop_markatime,
126         .vop_getattr =          hammer_vop_getattr,
127         .vop_inactive =         hammer_vop_inactive,
128         .vop_reclaim =          hammer_vop_reclaim,
129         .vop_setattr =          hammer_vop_setattr
130 };
131
132 struct vop_ops hammer_fifo_vops = {
133         .vop_default =          fifo_vnoperate,
134         .vop_fsync =            hammer_vop_fsync,
135         .vop_read =             hammer_vop_fiforead,
136         .vop_write =            hammer_vop_fifowrite,
137         .vop_access =           hammer_vop_access,
138         .vop_close =            hammer_vop_fifoclose,
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         .vop_kqfilter =         hammer_vop_fifokqfilter
145 };
146
147 static __inline
148 void
149 hammer_knote(struct vnode *vp, int flags)
150 {
151         if (flags)
152                 KNOTE(&vp->v_pollinfo.vpi_kqinfo.ki_note, flags);
153 }
154
155 #ifdef DEBUG_TRUNCATE
156 struct hammer_inode *HammerTruncIp;
157 #endif
158
159 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
160                            struct vnode *dvp, struct ucred *cred,
161                            int flags, int isdir);
162 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
163 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
164
165 /*
166  * hammer_vop_fsync { vp, waitfor }
167  *
168  * fsync() an inode to disk and wait for it to be completely committed
169  * such that the information would not be undone if a crash occured after
170  * return.
171  *
172  * NOTE: HAMMER's fsync()'s are going to remain expensive until we implement
173  *       a REDO log.  A sysctl is provided to relax HAMMER's fsync()
174  *       operation.
175  *
176  *       Ultimately the combination of a REDO log and use of fast storage
177  *       to front-end cluster caches will make fsync fast, but it aint
178  *       here yet.  And, in anycase, we need real transactional
179  *       all-or-nothing features which are not restricted to a single file.
180  */
181 static
182 int
183 hammer_vop_fsync(struct vop_fsync_args *ap)
184 {
185         hammer_inode_t ip = VTOI(ap->a_vp);
186         hammer_mount_t hmp = ip->hmp;
187         int waitfor = ap->a_waitfor;
188         int mode;
189
190         lwkt_gettoken(&hmp->fs_token);
191
192         /*
193          * Fsync rule relaxation (default is either full synchronous flush
194          * or REDO semantics with synchronous flush).
195          */
196         if (ap->a_flags & VOP_FSYNC_SYSCALL) {
197                 switch(hammer_fsync_mode) {
198                 case 0:
199 mode0:
200                         /* no REDO, full synchronous flush */
201                         goto skip;
202                 case 1:
203 mode1:
204                         /* no REDO, full asynchronous flush */
205                         if (waitfor == MNT_WAIT)
206                                 waitfor = MNT_NOWAIT;
207                         goto skip;
208                 case 2:
209                         /* REDO semantics, synchronous flush */
210                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
211                                 goto mode0;
212                         mode = HAMMER_FLUSH_UNDOS_AUTO;
213                         break;
214                 case 3:
215                         /* REDO semantics, relaxed asynchronous flush */
216                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
217                                 goto mode1;
218                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
219                         if (waitfor == MNT_WAIT)
220                                 waitfor = MNT_NOWAIT;
221                         break;
222                 case 4:
223                         /* ignore the fsync() system call */
224                         lwkt_reltoken(&hmp->fs_token);
225                         return(0);
226                 default:
227                         /* we have to do something */
228                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
229                         if (waitfor == MNT_WAIT)
230                                 waitfor = MNT_NOWAIT;
231                         break;
232                 }
233
234                 /*
235                  * Fast fsync only needs to flush the UNDO/REDO fifo if
236                  * HAMMER_INODE_REDO is non-zero and the only modifications
237                  * made to the file are write or write-extends.
238                  */
239                 if ((ip->flags & HAMMER_INODE_REDO) &&
240                     (ip->flags & HAMMER_INODE_MODMASK_NOREDO) == 0) {
241                         ++hammer_count_fsyncs;
242                         hammer_flusher_flush_undos(hmp, mode);
243                         ip->redo_count = 0;
244                         if (ip->vp && (ip->flags & HAMMER_INODE_MODMASK) == 0)
245                                 vclrisdirty(ip->vp);
246                         lwkt_reltoken(&hmp->fs_token);
247                         return(0);
248                 }
249
250                 /*
251                  * REDO is enabled by fsync(), the idea being we really only
252                  * want to lay down REDO records when programs are using
253                  * fsync() heavily.  The first fsync() on the file starts
254                  * the gravy train going and later fsync()s keep it hot by
255                  * resetting the redo_count.
256                  *
257                  * We weren't running REDOs before now so we have to fall
258                  * through and do a full fsync of what we have.
259                  */
260                 if (hmp->version >= HAMMER_VOL_VERSION_FOUR &&
261                     (hmp->flags & HAMMER_MOUNT_REDO_RECOVERY_RUN) == 0) {
262                         ip->flags |= HAMMER_INODE_REDO;
263                         ip->redo_count = 0;
264                 }
265         }
266 skip:
267
268         /*
269          * Do a full flush sequence.
270          *
271          * Attempt to release the vnode while waiting for the inode to
272          * finish flushing.  This can really mess up inactive->reclaim
273          * sequences so only do it if the vnode is active.
274          *
275          * WARNING! The VX lock functions must be used.  vn_lock() will
276          *          fail when this is part of a VOP_RECLAIM sequence.
277          */
278         ++hammer_count_fsyncs;
279         vfsync(ap->a_vp, waitfor, 1, NULL, NULL);
280         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
281         if (waitfor == MNT_WAIT) {
282                 int dorelock;
283
284                 if ((ap->a_vp->v_flag & VRECLAIMED) == 0) {
285                         vx_unlock(ap->a_vp);
286                         dorelock = 1;
287                 } else {
288                         dorelock = 0;
289                 }
290                 hammer_wait_inode(ip);
291                 if (dorelock)
292                         vx_lock(ap->a_vp);
293         }
294         if (ip->vp && (ip->flags & HAMMER_INODE_MODMASK) == 0)
295                 vclrisdirty(ip->vp);
296         lwkt_reltoken(&hmp->fs_token);
297         return (ip->error);
298 }
299
300 /*
301  * hammer_vop_read { vp, uio, ioflag, cred }
302  *
303  * MPSAFE (for the cache safe does not require fs_token)
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         hammer_mount_t hmp;
312         off_t offset;
313         struct buf *bp;
314         struct uio *uio;
315         int error;
316         int n;
317         int seqcount;
318         int ioseqcount;
319         int blksize;
320         int bigread;
321         int got_trans;
322         size_t resid;
323
324         if (ap->a_vp->v_type != VREG)
325                 return (EINVAL);
326         ip = VTOI(ap->a_vp);
327         hmp = ip->hmp;
328         error = 0;
329         got_trans = 0;
330         uio = ap->a_uio;
331
332         /*
333          * Attempt to shortcut directly to the VM object using lwbufs.
334          * This is much faster than instantiating buffer cache buffers.
335          */
336         resid = uio->uio_resid;
337         error = vop_helper_read_shortcut(ap);
338         hammer_stats_file_read += resid - uio->uio_resid;
339         if (error)
340                 return (error);
341         if (uio->uio_resid == 0)
342                 goto finished;
343
344         /*
345          * Allow the UIO's size to override the sequential heuristic.
346          */
347         blksize = hammer_blocksize(uio->uio_offset);
348         seqcount = (uio->uio_resid + (BKVASIZE - 1)) / BKVASIZE;
349         ioseqcount = (ap->a_ioflag >> 16);
350         if (seqcount < ioseqcount)
351                 seqcount = ioseqcount;
352
353         /*
354          * If reading or writing a huge amount of data we have to break
355          * atomicy and allow the operation to be interrupted by a signal
356          * or it can DOS the machine.
357          */
358         bigread = (uio->uio_resid > 100 * 1024 * 1024);
359
360         /*
361          * Access the data typically in HAMMER_BUFSIZE blocks via the
362          * buffer cache, but HAMMER may use a variable block size based
363          * on the offset.
364          *
365          * XXX Temporary hack, delay the start transaction while we remain
366          *     MPSAFE.  NOTE: ino_data.size cannot change while vnode is
367          *     locked-shared.
368          */
369         while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
370                 int64_t base_offset;
371                 int64_t file_limit;
372
373                 blksize = hammer_blocksize(uio->uio_offset);
374                 offset = (int)uio->uio_offset & (blksize - 1);
375                 base_offset = uio->uio_offset - offset;
376
377                 if (bigread && (error = hammer_signal_check(ip->hmp)) != 0)
378                         break;
379
380                 /*
381                  * MPSAFE
382                  */
383                 bp = getblk(ap->a_vp, base_offset, blksize, 0, 0);
384                 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == B_CACHE) {
385                         bp->b_flags &= ~B_AGE;
386                         error = 0;
387                         goto skip;
388                 }
389                 if (ap->a_ioflag & IO_NRDELAY) {
390                         bqrelse(bp);
391                         return (EWOULDBLOCK);
392                 }
393
394                 /*
395                  * MPUNSAFE
396                  */
397                 if (got_trans == 0) {
398                         hammer_start_transaction(&trans, ip->hmp);
399                         got_trans = 1;
400                 }
401
402                 /*
403                  * NOTE: A valid bp has already been acquired, but was not
404                  *       B_CACHE.
405                  */
406                 if (hammer_cluster_enable) {
407                         /*
408                          * Use file_limit to prevent cluster_read() from
409                          * creating buffers of the wrong block size past
410                          * the demarc.
411                          */
412                         file_limit = ip->ino_data.size;
413                         if (base_offset < HAMMER_XDEMARC &&
414                             file_limit > HAMMER_XDEMARC) {
415                                 file_limit = HAMMER_XDEMARC;
416                         }
417                         error = cluster_readx(ap->a_vp,
418                                              file_limit, base_offset,
419                                              blksize, uio->uio_resid,
420                                              seqcount * BKVASIZE, &bp);
421                 } else {
422                         error = breadnx(ap->a_vp, base_offset, blksize,
423                                         NULL, NULL, 0, &bp);
424                 }
425                 if (error) {
426                         brelse(bp);
427                         break;
428                 }
429 skip:
430                 if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IODEBUG)) {
431                         kprintf("doff %016jx read file %016jx@%016jx\n",
432                                 (intmax_t)bp->b_bio2.bio_offset,
433                                 (intmax_t)ip->obj_id,
434                                 (intmax_t)bp->b_loffset);
435                 }
436                 bp->b_flags &= ~B_IODEBUG;
437                 if (blksize == HAMMER_XBUFSIZE)
438                         bp->b_flags |= B_CLUSTEROK;
439
440                 n = blksize - offset;
441                 if (n > uio->uio_resid)
442                         n = uio->uio_resid;
443                 if (n > ip->ino_data.size - uio->uio_offset)
444                         n = (int)(ip->ino_data.size - uio->uio_offset);
445
446                 /*
447                  * Set B_AGE, data has a lower priority than meta-data.
448                  *
449                  * Use a hold/unlock/drop sequence to run the uiomove
450                  * with the buffer unlocked, avoiding deadlocks against
451                  * read()s on mmap()'d spaces.
452                  */
453                 bp->b_flags |= B_AGE;
454                 error = uiomovebp(bp, (char *)bp->b_data + offset, n, uio);
455                 bqrelse(bp);
456
457                 if (error)
458                         break;
459                 hammer_stats_file_read += n;
460         }
461
462 finished:
463
464         /*
465          * Try to update the atime with just the inode lock for maximum
466          * concurrency.  If we can't shortcut it we have to get the full
467          * blown transaction.
468          */
469         if (got_trans == 0 && hammer_update_atime_quick(ip) < 0) {
470                 hammer_start_transaction(&trans, ip->hmp);
471                 got_trans = 1;
472         }
473
474         if (got_trans) {
475                 if ((ip->flags & HAMMER_INODE_RO) == 0 &&
476                     (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
477                         lwkt_gettoken(&hmp->fs_token);
478                         ip->ino_data.atime = trans.time;
479                         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
480                         hammer_done_transaction(&trans);
481                         lwkt_reltoken(&hmp->fs_token);
482                 } else {
483                         hammer_done_transaction(&trans);
484                 }
485         }
486         return (error);
487 }
488
489 /*
490  * hammer_vop_write { vp, uio, ioflag, cred }
491  */
492 static
493 int
494 hammer_vop_write(struct vop_write_args *ap)
495 {
496         struct hammer_transaction trans;
497         struct hammer_inode *ip;
498         hammer_mount_t hmp;
499         thread_t td;
500         struct uio *uio;
501         int offset;
502         off_t base_offset;
503         int64_t cluster_eof;
504         struct buf *bp;
505         int kflags;
506         int error;
507         int n;
508         int flags;
509         int seqcount;
510         int bigwrite;
511
512         if (ap->a_vp->v_type != VREG)
513                 return (EINVAL);
514         ip = VTOI(ap->a_vp);
515         hmp = ip->hmp;
516         error = 0;
517         kflags = 0;
518         seqcount = ap->a_ioflag >> 16;
519
520         if (ip->flags & HAMMER_INODE_RO)
521                 return (EROFS);
522
523         /*
524          * Create a transaction to cover the operations we perform.
525          */
526         hammer_start_transaction(&trans, hmp);
527         uio = ap->a_uio;
528
529         /*
530          * Check append mode
531          */
532         if (ap->a_ioflag & IO_APPEND)
533                 uio->uio_offset = ip->ino_data.size;
534
535         /*
536          * Check for illegal write offsets.  Valid range is 0...2^63-1.
537          *
538          * NOTE: the base_off assignment is required to work around what
539          * I consider to be a GCC-4 optimization bug.
540          */
541         if (uio->uio_offset < 0) {
542                 hammer_done_transaction(&trans);
543                 return (EFBIG);
544         }
545         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
546         if (uio->uio_resid > 0 && base_offset <= uio->uio_offset) {
547                 hammer_done_transaction(&trans);
548                 return (EFBIG);
549         }
550
551         if (uio->uio_resid > 0 && (td = uio->uio_td) != NULL && td->td_proc &&
552             base_offset > td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
553                 hammer_done_transaction(&trans);
554                 lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ);
555                 return (EFBIG);
556         }
557
558         /*
559          * If reading or writing a huge amount of data we have to break
560          * atomicy and allow the operation to be interrupted by a signal
561          * or it can DOS the machine.
562          *
563          * Preset redo_count so we stop generating REDOs earlier if the
564          * limit is exceeded.
565          *
566          * redo_count is heuristical, SMP races are ok
567          */
568         bigwrite = (uio->uio_resid > 100 * 1024 * 1024);
569         if ((ip->flags & HAMMER_INODE_REDO) &&
570             ip->redo_count < hammer_limit_redo) {
571                 ip->redo_count += uio->uio_resid;
572         }
573
574         /*
575          * Access the data typically in HAMMER_BUFSIZE blocks via the
576          * buffer cache, but HAMMER may use a variable block size based
577          * on the offset.
578          */
579         while (uio->uio_resid > 0) {
580                 int fixsize = 0;
581                 int blksize;
582                 int blkmask;
583                 int trivial;
584                 int endofblk;
585                 off_t nsize;
586
587                 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
588                         break;
589                 if (bigwrite && (error = hammer_signal_check(hmp)) != 0)
590                         break;
591
592                 blksize = hammer_blocksize(uio->uio_offset);
593
594                 /*
595                  * Control the number of pending records associated with
596                  * this inode.  If too many have accumulated start a
597                  * flush.  Try to maintain a pipeline with the flusher.
598                  *
599                  * NOTE: It is possible for other sources to grow the
600                  *       records but not necessarily issue another flush,
601                  *       so use a timeout and ensure that a re-flush occurs.
602                  */
603                 if (ip->rsv_recs >= hammer_limit_inode_recs) {
604                         lwkt_gettoken(&hmp->fs_token);
605                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
606                         while (ip->rsv_recs >= hammer_limit_inode_recs * 2) {
607                                 ip->flags |= HAMMER_INODE_RECSW;
608                                 tsleep(&ip->rsv_recs, 0, "hmrwww", hz);
609                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
610                         }
611                         lwkt_reltoken(&hmp->fs_token);
612                 }
613
614                 /*
615                  * Do not allow HAMMER to blow out the buffer cache.  Very
616                  * large UIOs can lockout other processes due to bwillwrite()
617                  * mechanics.
618                  *
619                  * The hammer inode is not locked during these operations.
620                  * The vnode is locked which can interfere with the pageout
621                  * daemon for non-UIO_NOCOPY writes but should not interfere
622                  * with the buffer cache.  Even so, we cannot afford to
623                  * allow the pageout daemon to build up too many dirty buffer
624                  * cache buffers.
625                  *
626                  * Only call this if we aren't being recursively called from
627                  * a virtual disk device (vn), else we may deadlock.
628                  */
629                 if ((ap->a_ioflag & IO_RECURSE) == 0)
630                         bwillwrite(blksize);
631
632                 /*
633                  * Calculate the blocksize at the current offset and figure
634                  * out how much we can actually write.
635                  */
636                 blkmask = blksize - 1;
637                 offset = (int)uio->uio_offset & blkmask;
638                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
639                 n = blksize - offset;
640                 if (n > uio->uio_resid) {
641                         n = uio->uio_resid;
642                         endofblk = 0;
643                 } else {
644                         endofblk = 1;
645                 }
646                 nsize = uio->uio_offset + n;
647                 if (nsize > ip->ino_data.size) {
648                         if (uio->uio_offset > ip->ino_data.size)
649                                 trivial = 0;
650                         else
651                                 trivial = 1;
652                         nvextendbuf(ap->a_vp,
653                                     ip->ino_data.size,
654                                     nsize,
655                                     hammer_blocksize(ip->ino_data.size),
656                                     hammer_blocksize(nsize),
657                                     hammer_blockoff(ip->ino_data.size),
658                                     hammer_blockoff(nsize),
659                                     trivial);
660                         fixsize = 1;
661                         kflags |= NOTE_EXTEND;
662                 }
663
664                 if (uio->uio_segflg == UIO_NOCOPY) {
665                         /*
666                          * Issuing a write with the same data backing the
667                          * buffer.  Instantiate the buffer to collect the
668                          * backing vm pages, then read-in any missing bits.
669                          *
670                          * This case is used by vop_stdputpages().
671                          */
672                         bp = getblk(ap->a_vp, base_offset,
673                                     blksize, GETBLK_BHEAVY, 0);
674                         if ((bp->b_flags & B_CACHE) == 0) {
675                                 bqrelse(bp);
676                                 error = bread(ap->a_vp, base_offset,
677                                               blksize, &bp);
678                         }
679                 } else if (offset == 0 && uio->uio_resid >= blksize) {
680                         /*
681                          * Even though we are entirely overwriting the buffer
682                          * we may still have to zero it out to avoid a
683                          * mmap/write visibility issue.
684                          */
685                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
686                         if ((bp->b_flags & B_CACHE) == 0)
687                                 vfs_bio_clrbuf(bp);
688                 } else if (base_offset >= ip->ino_data.size) {
689                         /*
690                          * If the base offset of the buffer is beyond the
691                          * file EOF, we don't have to issue a read.
692                          */
693                         bp = getblk(ap->a_vp, base_offset,
694                                     blksize, GETBLK_BHEAVY, 0);
695                         vfs_bio_clrbuf(bp);
696                 } else {
697                         /*
698                          * Partial overwrite, read in any missing bits then
699                          * replace the portion being written.
700                          */
701                         error = bread(ap->a_vp, base_offset, blksize, &bp);
702                         if (error == 0)
703                                 bheavy(bp);
704                 }
705                 if (error == 0)
706                         error = uiomovebp(bp, bp->b_data + offset, n, uio);
707
708                 lwkt_gettoken(&hmp->fs_token);
709
710                 /*
711                  * Generate REDO records if enabled and redo_count will not
712                  * exceeded the limit.
713                  *
714                  * If redo_count exceeds the limit we stop generating records
715                  * and clear HAMMER_INODE_REDO.  This will cause the next
716                  * fsync() to do a full meta-data sync instead of just an
717                  * UNDO/REDO fifo update.
718                  *
719                  * When clearing HAMMER_INODE_REDO any pre-existing REDOs
720                  * will still be tracked.  The tracks will be terminated
721                  * when the related meta-data (including possible data
722                  * modifications which are not tracked via REDO) is
723                  * flushed.
724                  */
725                 if ((ip->flags & HAMMER_INODE_REDO) && error == 0) {
726                         if (ip->redo_count < hammer_limit_redo) {
727                                 bp->b_flags |= B_VFSFLAG1;
728                                 error = hammer_generate_redo(&trans, ip,
729                                                      base_offset + offset,
730                                                      HAMMER_REDO_WRITE,
731                                                      bp->b_data + offset,
732                                                      (size_t)n);
733                         } else {
734                                 ip->flags &= ~HAMMER_INODE_REDO;
735                         }
736                 }
737
738                 /*
739                  * If we screwed up we have to undo any VM size changes we
740                  * made.
741                  */
742                 if (error) {
743                         brelse(bp);
744                         if (fixsize) {
745                                 nvtruncbuf(ap->a_vp, ip->ino_data.size,
746                                           hammer_blocksize(ip->ino_data.size),
747                                           hammer_blockoff(ip->ino_data.size),
748                                           0);
749                         }
750                         lwkt_reltoken(&hmp->fs_token);
751                         break;
752                 }
753                 kflags |= NOTE_WRITE;
754                 hammer_stats_file_write += n;
755                 if (blksize == HAMMER_XBUFSIZE)
756                         bp->b_flags |= B_CLUSTEROK;
757                 if (ip->ino_data.size < uio->uio_offset) {
758                         ip->ino_data.size = uio->uio_offset;
759                         flags = HAMMER_INODE_SDIRTY;
760                 } else {
761                         flags = 0;
762                 }
763                 ip->ino_data.mtime = trans.time;
764                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
765                 hammer_modify_inode(&trans, ip, flags);
766
767                 /*
768                  * Once we dirty the buffer any cached zone-X offset
769                  * becomes invalid.  HAMMER NOTE: no-history mode cannot
770                  * allow overwriting over the same data sector unless
771                  * we provide UNDOs for the old data, which we don't.
772                  */
773                 bp->b_bio2.bio_offset = NOOFFSET;
774
775                 lwkt_reltoken(&hmp->fs_token);
776
777                 /*
778                  * Final buffer disposition.
779                  *
780                  * Because meta-data updates are deferred, HAMMER is
781                  * especially sensitive to excessive bdwrite()s because
782                  * the I/O stream is not broken up by disk reads.  So the
783                  * buffer cache simply cannot keep up.
784                  *
785                  * WARNING!  blksize is variable.  cluster_write() is
786                  *           expected to not blow up if it encounters
787                  *           buffers that do not match the passed blksize.
788                  *
789                  * NOTE!  Hammer shouldn't need to bawrite()/cluster_write().
790                  *        The ip->rsv_recs check should burst-flush the data.
791                  *        If we queue it immediately the buf could be left
792                  *        locked on the device queue for a very long time.
793                  *
794                  *        However, failing to flush a dirty buffer out when
795                  *        issued from the pageout daemon can result in a low
796                  *        memory deadlock against bio_page_alloc(), so we
797                  *        have to bawrite() on IO_ASYNC as well.
798                  *
799                  * NOTE!  To avoid degenerate stalls due to mismatched block
800                  *        sizes we only honor IO_DIRECT on the write which
801                  *        abuts the end of the buffer.  However, we must
802                  *        honor IO_SYNC in case someone is silly enough to
803                  *        configure a HAMMER file as swap, or when HAMMER
804                  *        is serving NFS (for commits).  Ick ick.
805                  */
806                 bp->b_flags |= B_AGE;
807                 if (blksize == HAMMER_XBUFSIZE)
808                         bp->b_flags |= B_CLUSTEROK;
809
810                 if (ap->a_ioflag & IO_SYNC) {
811                         bwrite(bp);
812                 } else if ((ap->a_ioflag & IO_DIRECT) && endofblk) {
813                         bawrite(bp);
814                 } else if (ap->a_ioflag & IO_ASYNC) {
815                         bawrite(bp);
816                 } else if (hammer_cluster_enable &&
817                            !(ap->a_vp->v_mount->mnt_flag & MNT_NOCLUSTERW)) {
818                         if (base_offset < HAMMER_XDEMARC)
819                                 cluster_eof = hammer_blockdemarc(base_offset,
820                                                          ip->ino_data.size);
821                         else
822                                 cluster_eof = ip->ino_data.size;
823                         cluster_write(bp, cluster_eof, blksize, seqcount);
824                 } else {
825                         bdwrite(bp);
826                 }
827         }
828         hammer_done_transaction(&trans);
829         hammer_knote(ap->a_vp, kflags);
830
831         return (error);
832 }
833
834 /*
835  * hammer_vop_access { vp, mode, cred }
836  *
837  * MPSAFE - does not require fs_token
838  */
839 static
840 int
841 hammer_vop_access(struct vop_access_args *ap)
842 {
843         struct hammer_inode *ip = VTOI(ap->a_vp);
844         uid_t uid;
845         gid_t gid;
846         int error;
847
848         ++hammer_stats_file_iopsr;
849         uid = hammer_to_unix_xid(&ip->ino_data.uid);
850         gid = hammer_to_unix_xid(&ip->ino_data.gid);
851
852         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
853                                   ip->ino_data.uflags);
854         return (error);
855 }
856
857 /*
858  * hammer_vop_advlock { vp, id, op, fl, flags }
859  *
860  * MPSAFE - does not require fs_token
861  */
862 static
863 int
864 hammer_vop_advlock(struct vop_advlock_args *ap)
865 {
866         hammer_inode_t ip = VTOI(ap->a_vp);
867
868         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
869 }
870
871 /*
872  * hammer_vop_close { vp, fflag }
873  *
874  * We can only sync-on-close for normal closes.  XXX disabled for now.
875  */
876 static
877 int
878 hammer_vop_close(struct vop_close_args *ap)
879 {
880 #if 0
881         struct vnode *vp = ap->a_vp;
882         hammer_inode_t ip = VTOI(vp);
883         int waitfor;
884         if (ip->flags & (HAMMER_INODE_CLOSESYNC|HAMMER_INODE_CLOSEASYNC)) {
885                 if (vn_islocked(vp) == LK_EXCLUSIVE &&
886                     (vp->v_flag & (VINACTIVE|VRECLAIMED)) == 0) {
887                         if (ip->flags & HAMMER_INODE_CLOSESYNC)
888                                 waitfor = MNT_WAIT;
889                         else
890                                 waitfor = MNT_NOWAIT;
891                         ip->flags &= ~(HAMMER_INODE_CLOSESYNC |
892                                        HAMMER_INODE_CLOSEASYNC);
893                         VOP_FSYNC(vp, MNT_NOWAIT, waitfor);
894                 }
895         }
896 #endif
897         return (vop_stdclose(ap));
898 }
899
900 /*
901  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
902  *
903  * The operating system has already ensured that the directory entry
904  * does not exist and done all appropriate namespace locking.
905  */
906 static
907 int
908 hammer_vop_ncreate(struct vop_ncreate_args *ap)
909 {
910         struct hammer_transaction trans;
911         struct hammer_inode *dip;
912         struct hammer_inode *nip;
913         struct nchandle *nch;
914         hammer_mount_t hmp;
915         int error;
916
917         nch = ap->a_nch;
918         dip = VTOI(ap->a_dvp);
919         hmp = dip->hmp;
920
921         if (dip->flags & HAMMER_INODE_RO)
922                 return (EROFS);
923         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
924                 return (error);
925
926         /*
927          * Create a transaction to cover the operations we perform.
928          */
929         lwkt_gettoken(&hmp->fs_token);
930         hammer_start_transaction(&trans, hmp);
931         ++hammer_stats_file_iopsw;
932
933         /*
934          * Create a new filesystem object of the requested type.  The
935          * returned inode will be referenced and shared-locked to prevent
936          * it from being moved to the flusher.
937          */
938         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
939                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
940                                     NULL, &nip);
941         if (error) {
942                 hkprintf("hammer_create_inode error %d\n", error);
943                 hammer_done_transaction(&trans);
944                 *ap->a_vpp = NULL;
945                 lwkt_reltoken(&hmp->fs_token);
946                 return (error);
947         }
948
949         /*
950          * Add the new filesystem object to the directory.  This will also
951          * bump the inode's link count.
952          */
953         error = hammer_ip_add_directory(&trans, dip,
954                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
955                                         nip);
956         if (error)
957                 hkprintf("hammer_ip_add_directory error %d\n", error);
958
959         /*
960          * Finish up.
961          */
962         if (error) {
963                 hammer_rel_inode(nip, 0);
964                 hammer_done_transaction(&trans);
965                 *ap->a_vpp = NULL;
966         } else {
967                 error = hammer_get_vnode(nip, ap->a_vpp);
968                 hammer_done_transaction(&trans);
969                 hammer_rel_inode(nip, 0);
970                 if (error == 0) {
971                         cache_setunresolved(ap->a_nch);
972                         cache_setvp(ap->a_nch, *ap->a_vpp);
973                 }
974                 hammer_knote(ap->a_dvp, NOTE_WRITE);
975         }
976         lwkt_reltoken(&hmp->fs_token);
977         return (error);
978 }
979
980 /*
981  * hammer_vop_getattr { vp, vap }
982  *
983  * Retrieve an inode's attribute information.  When accessing inodes
984  * historically we fake the atime field to ensure consistent results.
985  * The atime field is stored in the B-Tree element and allowed to be
986  * updated without cycling the element.
987  *
988  * MPSAFE - does not require fs_token
989  */
990 static
991 int
992 hammer_vop_getattr(struct vop_getattr_args *ap)
993 {
994         struct hammer_inode *ip = VTOI(ap->a_vp);
995         struct vattr *vap = ap->a_vap;
996
997         /*
998          * We want the fsid to be different when accessing a filesystem
999          * with different as-of's so programs like diff don't think
1000          * the files are the same.
1001          *
1002          * We also want the fsid to be the same when comparing snapshots,
1003          * or when comparing mirrors (which might be backed by different
1004          * physical devices).  HAMMER fsids are based on the PFS's
1005          * shared_uuid field.
1006          *
1007          * XXX there is a chance of collision here.  The va_fsid reported
1008          * by stat is different from the more involved fsid used in the
1009          * mount structure.
1010          */
1011         ++hammer_stats_file_iopsr;
1012         hammer_lock_sh(&ip->lock);
1013         vap->va_fsid = ip->pfsm->fsid_udev ^ (u_int32_t)ip->obj_asof ^
1014                        (u_int32_t)(ip->obj_asof >> 32);
1015
1016         vap->va_fileid = ip->ino_leaf.base.obj_id;
1017         vap->va_mode = ip->ino_data.mode;
1018         vap->va_nlink = ip->ino_data.nlinks;
1019         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1020         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1021         vap->va_rmajor = 0;
1022         vap->va_rminor = 0;
1023         vap->va_size = ip->ino_data.size;
1024
1025         /*
1026          * Special case for @@PFS softlinks.  The actual size of the
1027          * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
1028          * or for MAX_TID is    "@@-1:%05d" == 10 bytes.
1029          *
1030          * Note that userspace hammer command does not allow users to
1031          * create a @@PFS softlink under an existing other PFS (id!=0)
1032          * so the ip localization here for @@PFS softlink is always 0.
1033          */
1034         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
1035             ip->ino_data.size == 10 &&
1036             ip->obj_asof == HAMMER_MAX_TID &&
1037             ip->obj_localization == 0 &&
1038             strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
1039                     if (ip->pfsm->pfsd.mirror_flags & HAMMER_PFSD_SLAVE)
1040                             vap->va_size = 26;
1041                     else
1042                             vap->va_size = 10;
1043         }
1044
1045         /*
1046          * We must provide a consistent atime and mtime for snapshots
1047          * so people can do a 'tar cf - ... | md5' on them and get
1048          * consistent results.
1049          */
1050         if (ip->flags & HAMMER_INODE_RO) {
1051                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
1052                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
1053         } else {
1054                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
1055                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
1056         }
1057         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
1058         vap->va_flags = ip->ino_data.uflags;
1059         vap->va_gen = 1;        /* hammer inums are unique for all time */
1060         vap->va_blocksize = HAMMER_BUFSIZE;
1061         if (ip->ino_data.size >= HAMMER_XDEMARC) {
1062                 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
1063                                 ~HAMMER_XBUFMASK64;
1064         } else if (ip->ino_data.size > HAMMER_HBUFSIZE) {
1065                 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
1066                                 ~HAMMER_BUFMASK64;
1067         } else {
1068                 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
1069         }
1070
1071         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
1072         vap->va_filerev = 0;    /* XXX */
1073         vap->va_uid_uuid = ip->ino_data.uid;
1074         vap->va_gid_uuid = ip->ino_data.gid;
1075         vap->va_fsid_uuid = ip->hmp->fsid;
1076         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
1077                           VA_FSID_UUID_VALID;
1078
1079         switch (ip->ino_data.obj_type) {
1080         case HAMMER_OBJTYPE_CDEV:
1081         case HAMMER_OBJTYPE_BDEV:
1082                 vap->va_rmajor = ip->ino_data.rmajor;
1083                 vap->va_rminor = ip->ino_data.rminor;
1084                 break;
1085         default:
1086                 break;
1087         }
1088         hammer_unlock(&ip->lock);
1089         return(0);
1090 }
1091
1092 /*
1093  * hammer_vop_nresolve { nch, dvp, cred }
1094  *
1095  * Locate the requested directory entry.
1096  */
1097 static
1098 int
1099 hammer_vop_nresolve(struct vop_nresolve_args *ap)
1100 {
1101         struct hammer_transaction trans;
1102         struct namecache *ncp;
1103         hammer_mount_t hmp;
1104         hammer_inode_t dip;
1105         hammer_inode_t ip;
1106         hammer_tid_t asof;
1107         struct hammer_cursor cursor;
1108         struct vnode *vp;
1109         int64_t namekey;
1110         int error;
1111         int i;
1112         int nlen;
1113         int flags;
1114         int ispfs;
1115         int64_t obj_id;
1116         u_int32_t localization;
1117         u_int32_t max_iterations;
1118
1119         /*
1120          * Misc initialization, plus handle as-of name extensions.  Look for
1121          * the '@@' extension.  Note that as-of files and directories cannot
1122          * be modified.
1123          */
1124         dip = VTOI(ap->a_dvp);
1125         ncp = ap->a_nch->ncp;
1126         asof = dip->obj_asof;
1127         localization = dip->obj_localization;   /* for code consistency */
1128         nlen = ncp->nc_nlen;
1129         flags = dip->flags & HAMMER_INODE_RO;
1130         ispfs = 0;
1131         hmp = dip->hmp;
1132
1133         lwkt_gettoken(&hmp->fs_token);
1134         hammer_simple_transaction(&trans, hmp);
1135         ++hammer_stats_file_iopsr;
1136
1137         for (i = 0; i < nlen; ++i) {
1138                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
1139                         error = hammer_str_to_tid(ncp->nc_name + i + 2,
1140                                                   &ispfs, &asof, &localization);
1141                         if (error != 0) {
1142                                 i = nlen;
1143                                 break;
1144                         }
1145                         if (asof != HAMMER_MAX_TID)
1146                                 flags |= HAMMER_INODE_RO;
1147                         break;
1148                 }
1149         }
1150         nlen = i;
1151
1152         /*
1153          * If this is a PFS softlink we dive into the PFS
1154          */
1155         if (ispfs && nlen == 0) {
1156                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1157                                       asof, localization,
1158                                       flags, &error);
1159                 if (error == 0) {
1160                         error = hammer_get_vnode(ip, &vp);
1161                         hammer_rel_inode(ip, 0);
1162                 } else {
1163                         vp = NULL;
1164                 }
1165                 if (error == 0) {
1166                         vn_unlock(vp);
1167                         cache_setvp(ap->a_nch, vp);
1168                         vrele(vp);
1169                 }
1170                 goto done;
1171         }
1172
1173         /*
1174          * If there is no path component the time extension is relative to dip.
1175          * e.g. "fubar/@@<snapshot>"
1176          *
1177          * "." is handled by the kernel, but ".@@<snapshot>" is not.
1178          * e.g. "fubar/.@@<snapshot>"
1179          *
1180          * ".." is handled by the kernel.  We do not currently handle
1181          * "..@<snapshot>".
1182          */
1183         if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1184                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
1185                                       asof, dip->obj_localization,
1186                                       flags, &error);
1187                 if (error == 0) {
1188                         error = hammer_get_vnode(ip, &vp);
1189                         hammer_rel_inode(ip, 0);
1190                 } else {
1191                         vp = NULL;
1192                 }
1193                 if (error == 0) {
1194                         vn_unlock(vp);
1195                         cache_setvp(ap->a_nch, vp);
1196                         vrele(vp);
1197                 }
1198                 goto done;
1199         }
1200
1201         /*
1202          * Calculate the namekey and setup the key range for the scan.  This
1203          * works kinda like a chained hash table where the lower 32 bits
1204          * of the namekey synthesize the chain.
1205          *
1206          * The key range is inclusive of both key_beg and key_end.
1207          */
1208         namekey = hammer_directory_namekey(dip, ncp->nc_name, nlen,
1209                                            &max_iterations);
1210
1211         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1212         cursor.key_beg.localization = dip->obj_localization +
1213                                       hammer_dir_localization(dip);
1214         cursor.key_beg.obj_id = dip->obj_id;
1215         cursor.key_beg.key = namekey;
1216         cursor.key_beg.create_tid = 0;
1217         cursor.key_beg.delete_tid = 0;
1218         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1219         cursor.key_beg.obj_type = 0;
1220
1221         cursor.key_end = cursor.key_beg;
1222         cursor.key_end.key += max_iterations;
1223         cursor.asof = asof;
1224         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1225
1226         /*
1227          * Scan all matching records (the chain), locate the one matching
1228          * the requested path component.
1229          *
1230          * The hammer_ip_*() functions merge in-memory records with on-disk
1231          * records for the purposes of the search.
1232          */
1233         obj_id = 0;
1234         localization = HAMMER_DEF_LOCALIZATION;
1235
1236         if (error == 0) {
1237                 error = hammer_ip_first(&cursor);
1238                 while (error == 0) {
1239                         error = hammer_ip_resolve_data(&cursor);
1240                         if (error)
1241                                 break;
1242                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1243                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1244                                 obj_id = cursor.data->entry.obj_id;
1245                                 localization = cursor.data->entry.localization;
1246                                 break;
1247                         }
1248                         error = hammer_ip_next(&cursor);
1249                 }
1250         }
1251         hammer_done_cursor(&cursor);
1252
1253         /*
1254          * Lookup the obj_id.  This should always succeed.  If it does not
1255          * the filesystem may be damaged and we return a dummy inode.
1256          */
1257         if (error == 0) {
1258                 ip = hammer_get_inode(&trans, dip, obj_id,
1259                                       asof, localization,
1260                                       flags, &error);
1261                 if (error == ENOENT) {
1262                         kprintf("HAMMER: WARNING: Missing "
1263                                 "inode for dirent \"%s\"\n"
1264                                 "\tobj_id = %016llx, asof=%016llx, lo=%08x\n",
1265                                 ncp->nc_name,
1266                                 (long long)obj_id, (long long)asof,
1267                                 localization);
1268                         error = 0;
1269                         ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1270                                                     asof, localization,
1271                                                     flags, &error);
1272                 }
1273                 if (error == 0) {
1274                         error = hammer_get_vnode(ip, &vp);
1275                         hammer_rel_inode(ip, 0);
1276                 } else {
1277                         vp = NULL;
1278                 }
1279                 if (error == 0) {
1280                         vn_unlock(vp);
1281                         cache_setvp(ap->a_nch, vp);
1282                         vrele(vp);
1283                 }
1284         } else if (error == ENOENT) {
1285                 cache_setvp(ap->a_nch, NULL);
1286         }
1287 done:
1288         hammer_done_transaction(&trans);
1289         lwkt_reltoken(&hmp->fs_token);
1290         return (error);
1291 }
1292
1293 /*
1294  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1295  *
1296  * Locate the parent directory of a directory vnode.
1297  *
1298  * dvp is referenced but not locked.  *vpp must be returned referenced and
1299  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
1300  * at the root, instead it could indicate that the directory we were in was
1301  * removed.
1302  *
1303  * NOTE: as-of sequences are not linked into the directory structure.  If
1304  * we are at the root with a different asof then the mount point, reload
1305  * the same directory with the mount point's asof.   I'm not sure what this
1306  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
1307  * get confused, but it hasn't been tested.
1308  */
1309 static
1310 int
1311 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1312 {
1313         struct hammer_transaction trans;
1314         struct hammer_inode *dip;
1315         struct hammer_inode *ip;
1316         hammer_mount_t hmp;
1317         int64_t parent_obj_id;
1318         u_int32_t parent_obj_localization;
1319         hammer_tid_t asof;
1320         int error;
1321
1322         dip = VTOI(ap->a_dvp);
1323         asof = dip->obj_asof;
1324         hmp = dip->hmp;
1325
1326         /*
1327          * Whos are parent?  This could be the root of a pseudo-filesystem
1328          * whos parent is in another localization domain.
1329          */
1330         lwkt_gettoken(&hmp->fs_token);
1331         parent_obj_id = dip->ino_data.parent_obj_id;
1332         if (dip->obj_id == HAMMER_OBJID_ROOT)
1333                 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
1334         else
1335                 parent_obj_localization = dip->obj_localization;
1336
1337         /*
1338          * It's probably a PFS root when dip->ino_data.parent_obj_id is 0.
1339          */
1340         if (parent_obj_id == 0) {
1341                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1342                    asof != hmp->asof) {
1343                         parent_obj_id = dip->obj_id;
1344                         asof = hmp->asof;
1345                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1346                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
1347                                   (long long)dip->obj_asof);
1348                 } else {
1349                         *ap->a_vpp = NULL;
1350                         lwkt_reltoken(&hmp->fs_token);
1351                         return ENOENT;
1352                 }
1353         }
1354
1355         hammer_simple_transaction(&trans, hmp);
1356         ++hammer_stats_file_iopsr;
1357
1358         ip = hammer_get_inode(&trans, dip, parent_obj_id,
1359                               asof, parent_obj_localization,
1360                               dip->flags, &error);
1361         if (ip) {
1362                 error = hammer_get_vnode(ip, ap->a_vpp);
1363                 hammer_rel_inode(ip, 0);
1364         } else {
1365                 *ap->a_vpp = NULL;
1366         }
1367         hammer_done_transaction(&trans);
1368         lwkt_reltoken(&hmp->fs_token);
1369         return (error);
1370 }
1371
1372 /*
1373  * hammer_vop_nlink { nch, dvp, vp, cred }
1374  */
1375 static
1376 int
1377 hammer_vop_nlink(struct vop_nlink_args *ap)
1378 {
1379         struct hammer_transaction trans;
1380         struct hammer_inode *dip;
1381         struct hammer_inode *ip;
1382         struct nchandle *nch;
1383         hammer_mount_t hmp;
1384         int error;
1385
1386         if (ap->a_dvp->v_mount != ap->a_vp->v_mount)
1387                 return(EXDEV);
1388
1389         nch = ap->a_nch;
1390         dip = VTOI(ap->a_dvp);
1391         ip = VTOI(ap->a_vp);
1392         hmp = dip->hmp;
1393
1394         if (dip->obj_localization != ip->obj_localization)
1395                 return(EXDEV);
1396
1397         if (dip->flags & HAMMER_INODE_RO)
1398                 return (EROFS);
1399         if (ip->flags & HAMMER_INODE_RO)
1400                 return (EROFS);
1401         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1402                 return (error);
1403
1404         /*
1405          * Create a transaction to cover the operations we perform.
1406          */
1407         lwkt_gettoken(&hmp->fs_token);
1408         hammer_start_transaction(&trans, hmp);
1409         ++hammer_stats_file_iopsw;
1410
1411         /*
1412          * Add the filesystem object to the directory.  Note that neither
1413          * dip nor ip are referenced or locked, but their vnodes are
1414          * referenced.  This function will bump the inode's link count.
1415          */
1416         error = hammer_ip_add_directory(&trans, dip,
1417                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1418                                         ip);
1419
1420         /*
1421          * Finish up.
1422          */
1423         if (error == 0) {
1424                 cache_setunresolved(nch);
1425                 cache_setvp(nch, ap->a_vp);
1426         }
1427         hammer_done_transaction(&trans);
1428         hammer_knote(ap->a_vp, NOTE_LINK);
1429         hammer_knote(ap->a_dvp, NOTE_WRITE);
1430         lwkt_reltoken(&hmp->fs_token);
1431         return (error);
1432 }
1433
1434 /*
1435  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1436  *
1437  * The operating system has already ensured that the directory entry
1438  * does not exist and done all appropriate namespace locking.
1439  */
1440 static
1441 int
1442 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1443 {
1444         struct hammer_transaction trans;
1445         struct hammer_inode *dip;
1446         struct hammer_inode *nip;
1447         struct nchandle *nch;
1448         hammer_mount_t hmp;
1449         int error;
1450
1451         nch = ap->a_nch;
1452         dip = VTOI(ap->a_dvp);
1453         hmp = dip->hmp;
1454
1455         if (dip->flags & HAMMER_INODE_RO)
1456                 return (EROFS);
1457         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1458                 return (error);
1459
1460         /*
1461          * Create a transaction to cover the operations we perform.
1462          */
1463         lwkt_gettoken(&hmp->fs_token);
1464         hammer_start_transaction(&trans, hmp);
1465         ++hammer_stats_file_iopsw;
1466
1467         /*
1468          * Create a new filesystem object of the requested type.  The
1469          * returned inode will be referenced but not locked.
1470          */
1471         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1472                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1473                                     NULL, &nip);
1474         if (error) {
1475                 hkprintf("hammer_mkdir error %d\n", error);
1476                 hammer_done_transaction(&trans);
1477                 *ap->a_vpp = NULL;
1478                 lwkt_reltoken(&hmp->fs_token);
1479                 return (error);
1480         }
1481         /*
1482          * Add the new filesystem object to the directory.  This will also
1483          * bump the inode's link count.
1484          */
1485         error = hammer_ip_add_directory(&trans, dip,
1486                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1487                                         nip);
1488         if (error)
1489                 hkprintf("hammer_mkdir (add) error %d\n", error);
1490
1491         /*
1492          * Finish up.
1493          */
1494         if (error) {
1495                 hammer_rel_inode(nip, 0);
1496                 *ap->a_vpp = NULL;
1497         } else {
1498                 error = hammer_get_vnode(nip, ap->a_vpp);
1499                 hammer_rel_inode(nip, 0);
1500                 if (error == 0) {
1501                         cache_setunresolved(ap->a_nch);
1502                         cache_setvp(ap->a_nch, *ap->a_vpp);
1503                 }
1504         }
1505         hammer_done_transaction(&trans);
1506         if (error == 0)
1507                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1508         lwkt_reltoken(&hmp->fs_token);
1509         return (error);
1510 }
1511
1512 /*
1513  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1514  *
1515  * The operating system has already ensured that the directory entry
1516  * does not exist and done all appropriate namespace locking.
1517  */
1518 static
1519 int
1520 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1521 {
1522         struct hammer_transaction trans;
1523         struct hammer_inode *dip;
1524         struct hammer_inode *nip;
1525         struct nchandle *nch;
1526         hammer_mount_t hmp;
1527         int error;
1528
1529         nch = ap->a_nch;
1530         dip = VTOI(ap->a_dvp);
1531         hmp = dip->hmp;
1532
1533         if (dip->flags & HAMMER_INODE_RO)
1534                 return (EROFS);
1535         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1536                 return (error);
1537
1538         /*
1539          * Create a transaction to cover the operations we perform.
1540          */
1541         lwkt_gettoken(&hmp->fs_token);
1542         hammer_start_transaction(&trans, hmp);
1543         ++hammer_stats_file_iopsw;
1544
1545         /*
1546          * Create a new filesystem object of the requested type.  The
1547          * returned inode will be referenced but not locked.
1548          *
1549          * If mknod specifies a directory a pseudo-fs is created.
1550          */
1551         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1552                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1553                                     NULL, &nip);
1554         if (error) {
1555                 hammer_done_transaction(&trans);
1556                 *ap->a_vpp = NULL;
1557                 lwkt_reltoken(&hmp->fs_token);
1558                 return (error);
1559         }
1560
1561         /*
1562          * Add the new filesystem object to the directory.  This will also
1563          * bump the inode's link count.
1564          */
1565         error = hammer_ip_add_directory(&trans, dip,
1566                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1567                                         nip);
1568
1569         /*
1570          * Finish up.
1571          */
1572         if (error) {
1573                 hammer_rel_inode(nip, 0);
1574                 *ap->a_vpp = NULL;
1575         } else {
1576                 error = hammer_get_vnode(nip, ap->a_vpp);
1577                 hammer_rel_inode(nip, 0);
1578                 if (error == 0) {
1579                         cache_setunresolved(ap->a_nch);
1580                         cache_setvp(ap->a_nch, *ap->a_vpp);
1581                 }
1582         }
1583         hammer_done_transaction(&trans);
1584         if (error == 0)
1585                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1586         lwkt_reltoken(&hmp->fs_token);
1587         return (error);
1588 }
1589
1590 /*
1591  * hammer_vop_open { vp, mode, cred, fp }
1592  *
1593  * MPSAFE (does not require fs_token)
1594  */
1595 static
1596 int
1597 hammer_vop_open(struct vop_open_args *ap)
1598 {
1599         hammer_inode_t ip;
1600
1601         ++hammer_stats_file_iopsr;
1602         ip = VTOI(ap->a_vp);
1603
1604         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1605                 return (EROFS);
1606         return(vop_stdopen(ap));
1607 }
1608
1609 /*
1610  * hammer_vop_print { vp }
1611  */
1612 static
1613 int
1614 hammer_vop_print(struct vop_print_args *ap)
1615 {
1616         return EOPNOTSUPP;
1617 }
1618
1619 /*
1620  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1621  */
1622 static
1623 int
1624 hammer_vop_readdir(struct vop_readdir_args *ap)
1625 {
1626         struct hammer_transaction trans;
1627         struct hammer_cursor cursor;
1628         struct hammer_inode *ip;
1629         hammer_mount_t hmp;
1630         struct uio *uio;
1631         hammer_base_elm_t base;
1632         int error;
1633         int cookie_index;
1634         int ncookies;
1635         off_t *cookies;
1636         off_t saveoff;
1637         int r;
1638         int dtype;
1639
1640         ++hammer_stats_file_iopsr;
1641         ip = VTOI(ap->a_vp);
1642         uio = ap->a_uio;
1643         saveoff = uio->uio_offset;
1644         hmp = ip->hmp;
1645
1646         if (ap->a_ncookies) {
1647                 ncookies = uio->uio_resid / 16 + 1;
1648                 if (ncookies > 1024)
1649                         ncookies = 1024;
1650                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1651                 cookie_index = 0;
1652         } else {
1653                 ncookies = -1;
1654                 cookies = NULL;
1655                 cookie_index = 0;
1656         }
1657
1658         lwkt_gettoken(&hmp->fs_token);
1659         hammer_simple_transaction(&trans, hmp);
1660
1661         /*
1662          * Handle artificial entries
1663          *
1664          * It should be noted that the minimum value for a directory
1665          * hash key on-media is 0x0000000100000000, so we can use anything
1666          * less then that to represent our 'special' key space.
1667          */
1668         error = 0;
1669         if (saveoff == 0) {
1670                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1671                 if (r)
1672                         goto done;
1673                 if (cookies)
1674                         cookies[cookie_index] = saveoff;
1675                 ++saveoff;
1676                 ++cookie_index;
1677                 if (cookie_index == ncookies)
1678                         goto done;
1679         }
1680         if (saveoff == 1) {
1681                 if (ip->ino_data.parent_obj_id) {
1682                         r = vop_write_dirent(&error, uio,
1683                                              ip->ino_data.parent_obj_id,
1684                                              DT_DIR, 2, "..");
1685                 } else {
1686                         r = vop_write_dirent(&error, uio,
1687                                              ip->obj_id, DT_DIR, 2, "..");
1688                 }
1689                 if (r)
1690                         goto done;
1691                 if (cookies)
1692                         cookies[cookie_index] = saveoff;
1693                 ++saveoff;
1694                 ++cookie_index;
1695                 if (cookie_index == ncookies)
1696                         goto done;
1697         }
1698
1699         /*
1700          * Key range (begin and end inclusive) to scan.  Directory keys
1701          * directly translate to a 64 bit 'seek' position.
1702          */
1703         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1704         cursor.key_beg.localization = ip->obj_localization +
1705                                       hammer_dir_localization(ip);
1706         cursor.key_beg.obj_id = ip->obj_id;
1707         cursor.key_beg.create_tid = 0;
1708         cursor.key_beg.delete_tid = 0;
1709         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1710         cursor.key_beg.obj_type = 0;
1711         cursor.key_beg.key = saveoff;
1712
1713         cursor.key_end = cursor.key_beg;
1714         cursor.key_end.key = HAMMER_MAX_KEY;
1715         cursor.asof = ip->obj_asof;
1716         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1717
1718         error = hammer_ip_first(&cursor);
1719
1720         while (error == 0) {
1721                 error = hammer_ip_resolve_data(&cursor);
1722                 if (error)
1723                         break;
1724                 base = &cursor.leaf->base;
1725                 saveoff = base->key;
1726                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1727
1728                 if (base->obj_id != ip->obj_id)
1729                         panic("readdir: bad record at %p", cursor.node);
1730
1731                 /*
1732                  * Convert pseudo-filesystems into softlinks
1733                  */
1734                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1735                 r = vop_write_dirent(
1736                              &error, uio, cursor.data->entry.obj_id,
1737                              dtype,
1738                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1739                              (void *)cursor.data->entry.name);
1740                 if (r)
1741                         break;
1742                 ++saveoff;
1743                 if (cookies)
1744                         cookies[cookie_index] = base->key;
1745                 ++cookie_index;
1746                 if (cookie_index == ncookies)
1747                         break;
1748                 error = hammer_ip_next(&cursor);
1749         }
1750         hammer_done_cursor(&cursor);
1751
1752 done:
1753         hammer_done_transaction(&trans);
1754
1755         if (ap->a_eofflag)
1756                 *ap->a_eofflag = (error == ENOENT);
1757         uio->uio_offset = saveoff;
1758         if (error && cookie_index == 0) {
1759                 if (error == ENOENT)
1760                         error = 0;
1761                 if (cookies) {
1762                         kfree(cookies, M_TEMP);
1763                         *ap->a_ncookies = 0;
1764                         *ap->a_cookies = NULL;
1765                 }
1766         } else {
1767                 if (error == ENOENT)
1768                         error = 0;
1769                 if (cookies) {
1770                         *ap->a_ncookies = cookie_index;
1771                         *ap->a_cookies = cookies;
1772                 }
1773         }
1774         lwkt_reltoken(&hmp->fs_token);
1775         return(error);
1776 }
1777
1778 /*
1779  * hammer_vop_readlink { vp, uio, cred }
1780  */
1781 static
1782 int
1783 hammer_vop_readlink(struct vop_readlink_args *ap)
1784 {
1785         struct hammer_transaction trans;
1786         struct hammer_cursor cursor;
1787         struct hammer_inode *ip;
1788         hammer_mount_t hmp;
1789         char buf[32];
1790         u_int32_t localization;
1791         hammer_pseudofs_inmem_t pfsm;
1792         int error;
1793
1794         ip = VTOI(ap->a_vp);
1795         hmp = ip->hmp;
1796
1797         lwkt_gettoken(&hmp->fs_token);
1798
1799         /*
1800          * Shortcut if the symlink data was stuffed into ino_data.
1801          *
1802          * Also expand special "@@PFS%05d" softlinks (expansion only
1803          * occurs for non-historical (current) accesses made from the
1804          * primary filesystem).
1805          *
1806          * Note that userspace hammer command does not allow users to
1807          * create a @@PFS softlink under an existing other PFS (id!=0)
1808          * so the ip localization here for @@PFS softlink is always 0.
1809          */
1810         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1811                 char *ptr;
1812                 int bytes;
1813
1814                 ptr = ip->ino_data.ext.symlink;
1815                 bytes = (int)ip->ino_data.size;
1816                 if (bytes == 10 &&
1817                     ip->obj_asof == HAMMER_MAX_TID &&
1818                     ip->obj_localization == 0 &&
1819                     strncmp(ptr, "@@PFS", 5) == 0) {
1820                         hammer_simple_transaction(&trans, hmp);
1821                         bcopy(ptr + 5, buf, 5);
1822                         buf[5] = 0;
1823                         localization = strtoul(buf, NULL, 10) << 16;
1824                         pfsm = hammer_load_pseudofs(&trans, localization,
1825                                                     &error);
1826                         if (error == 0) {
1827                                 if (pfsm->pfsd.mirror_flags &
1828                                     HAMMER_PFSD_SLAVE) {
1829                                         /* vap->va_size == 26 */
1830                                         ksnprintf(buf, sizeof(buf),
1831                                                   "@@0x%016llx:%05d",
1832                                                   (long long)pfsm->pfsd.sync_end_tid,
1833                                                   localization >> 16);
1834                                 } else {
1835                                         /* vap->va_size == 10 */
1836                                         ksnprintf(buf, sizeof(buf),
1837                                                   "@@-1:%05d",
1838                                                   localization >> 16);
1839 #if 0
1840                                         ksnprintf(buf, sizeof(buf),
1841                                                   "@@0x%016llx:%05d",
1842                                                   (long long)HAMMER_MAX_TID,
1843                                                   localization >> 16);
1844 #endif
1845                                 }
1846                                 ptr = buf;
1847                                 bytes = strlen(buf);
1848                         }
1849                         if (pfsm)
1850                                 hammer_rel_pseudofs(hmp, pfsm);
1851                         hammer_done_transaction(&trans);
1852                 }
1853                 error = uiomove(ptr, bytes, ap->a_uio);
1854                 lwkt_reltoken(&hmp->fs_token);
1855                 return(error);
1856         }
1857
1858         /*
1859          * Long version
1860          */
1861         hammer_simple_transaction(&trans, hmp);
1862         ++hammer_stats_file_iopsr;
1863         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1864
1865         /*
1866          * Key range (begin and end inclusive) to scan.  Directory keys
1867          * directly translate to a 64 bit 'seek' position.
1868          */
1869         cursor.key_beg.localization = ip->obj_localization +
1870                                       HAMMER_LOCALIZE_MISC;
1871         cursor.key_beg.obj_id = ip->obj_id;
1872         cursor.key_beg.create_tid = 0;
1873         cursor.key_beg.delete_tid = 0;
1874         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1875         cursor.key_beg.obj_type = 0;
1876         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1877         cursor.asof = ip->obj_asof;
1878         cursor.flags |= HAMMER_CURSOR_ASOF;
1879
1880         error = hammer_ip_lookup(&cursor);
1881         if (error == 0) {
1882                 error = hammer_ip_resolve_data(&cursor);
1883                 if (error == 0) {
1884                         KKASSERT(cursor.leaf->data_len >=
1885                                  HAMMER_SYMLINK_NAME_OFF);
1886                         error = uiomove(cursor.data->symlink.name,
1887                                         cursor.leaf->data_len -
1888                                                 HAMMER_SYMLINK_NAME_OFF,
1889                                         ap->a_uio);
1890                 }
1891         }
1892         hammer_done_cursor(&cursor);
1893         hammer_done_transaction(&trans);
1894         lwkt_reltoken(&hmp->fs_token);
1895         return(error);
1896 }
1897
1898 /*
1899  * hammer_vop_nremove { nch, dvp, cred }
1900  */
1901 static
1902 int
1903 hammer_vop_nremove(struct vop_nremove_args *ap)
1904 {
1905         struct hammer_transaction trans;
1906         struct hammer_inode *dip;
1907         hammer_mount_t hmp;
1908         int error;
1909
1910         dip = VTOI(ap->a_dvp);
1911         hmp = dip->hmp;
1912
1913         if (hammer_nohistory(dip) == 0 &&
1914             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1915                 return (error);
1916         }
1917
1918         lwkt_gettoken(&hmp->fs_token);
1919         hammer_start_transaction(&trans, hmp);
1920         ++hammer_stats_file_iopsw;
1921         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1922         hammer_done_transaction(&trans);
1923         if (error == 0)
1924                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1925         lwkt_reltoken(&hmp->fs_token);
1926         return (error);
1927 }
1928
1929 /*
1930  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1931  */
1932 static
1933 int
1934 hammer_vop_nrename(struct vop_nrename_args *ap)
1935 {
1936         struct hammer_transaction trans;
1937         struct namecache *fncp;
1938         struct namecache *tncp;
1939         struct hammer_inode *fdip;
1940         struct hammer_inode *tdip;
1941         struct hammer_inode *ip;
1942         hammer_mount_t hmp;
1943         struct hammer_cursor cursor;
1944         int64_t namekey;
1945         u_int32_t max_iterations;
1946         int nlen, error;
1947
1948         if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount)
1949                 return(EXDEV);
1950         if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1951                 return(EXDEV);
1952
1953         fdip = VTOI(ap->a_fdvp);
1954         tdip = VTOI(ap->a_tdvp);
1955         fncp = ap->a_fnch->ncp;
1956         tncp = ap->a_tnch->ncp;
1957         ip = VTOI(fncp->nc_vp);
1958         KKASSERT(ip != NULL);
1959
1960         hmp = ip->hmp;
1961
1962         if (fdip->obj_localization != tdip->obj_localization)
1963                 return(EXDEV);
1964         if (fdip->obj_localization != ip->obj_localization)
1965                 return(EXDEV);
1966
1967         if (fdip->flags & HAMMER_INODE_RO)
1968                 return (EROFS);
1969         if (tdip->flags & HAMMER_INODE_RO)
1970                 return (EROFS);
1971         if (ip->flags & HAMMER_INODE_RO)
1972                 return (EROFS);
1973         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1974                 return (error);
1975
1976         lwkt_gettoken(&hmp->fs_token);
1977         hammer_start_transaction(&trans, hmp);
1978         ++hammer_stats_file_iopsw;
1979
1980         /*
1981          * Remove tncp from the target directory and then link ip as
1982          * tncp. XXX pass trans to dounlink
1983          *
1984          * Force the inode sync-time to match the transaction so it is
1985          * in-sync with the creation of the target directory entry.
1986          */
1987         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
1988                                 ap->a_cred, 0, -1);
1989         if (error == 0 || error == ENOENT) {
1990                 error = hammer_ip_add_directory(&trans, tdip,
1991                                                 tncp->nc_name, tncp->nc_nlen,
1992                                                 ip);
1993                 if (error == 0) {
1994                         ip->ino_data.parent_obj_id = tdip->obj_id;
1995                         ip->ino_data.ctime = trans.time;
1996                         hammer_modify_inode(&trans, ip, HAMMER_INODE_DDIRTY);
1997                 }
1998         }
1999         if (error)
2000                 goto failed; /* XXX */
2001
2002         /*
2003          * Locate the record in the originating directory and remove it.
2004          *
2005          * Calculate the namekey and setup the key range for the scan.  This
2006          * works kinda like a chained hash table where the lower 32 bits
2007          * of the namekey synthesize the chain.
2008          *
2009          * The key range is inclusive of both key_beg and key_end.
2010          */
2011         namekey = hammer_directory_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
2012                                            &max_iterations);
2013 retry:
2014         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
2015         cursor.key_beg.localization = fdip->obj_localization +
2016                                       hammer_dir_localization(fdip);
2017         cursor.key_beg.obj_id = fdip->obj_id;
2018         cursor.key_beg.key = namekey;
2019         cursor.key_beg.create_tid = 0;
2020         cursor.key_beg.delete_tid = 0;
2021         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2022         cursor.key_beg.obj_type = 0;
2023
2024         cursor.key_end = cursor.key_beg;
2025         cursor.key_end.key += max_iterations;
2026         cursor.asof = fdip->obj_asof;
2027         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2028
2029         /*
2030          * Scan all matching records (the chain), locate the one matching
2031          * the requested path component.
2032          *
2033          * The hammer_ip_*() functions merge in-memory records with on-disk
2034          * records for the purposes of the search.
2035          */
2036         error = hammer_ip_first(&cursor);
2037         while (error == 0) {
2038                 if (hammer_ip_resolve_data(&cursor) != 0)
2039                         break;
2040                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2041                 KKASSERT(nlen > 0);
2042                 if (fncp->nc_nlen == nlen &&
2043                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2044                         break;
2045                 }
2046                 error = hammer_ip_next(&cursor);
2047         }
2048
2049         /*
2050          * If all is ok we have to get the inode so we can adjust nlinks.
2051          *
2052          * WARNING: hammer_ip_del_directory() may have to terminate the
2053          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
2054          * twice.
2055          */
2056         if (error == 0)
2057                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
2058
2059         /*
2060          * XXX A deadlock here will break rename's atomicy for the purposes
2061          * of crash recovery.
2062          */
2063         if (error == EDEADLK) {
2064                 hammer_done_cursor(&cursor);
2065                 goto retry;
2066         }
2067
2068         /*
2069          * Cleanup and tell the kernel that the rename succeeded.
2070          *
2071          * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2072          *       without formally acquiring the vp since the vp might
2073          *       have zero refs on it, or in the middle of a reclaim,
2074          *       etc.
2075          */
2076         hammer_done_cursor(&cursor);
2077         if (error == 0) {
2078                 cache_rename(ap->a_fnch, ap->a_tnch);
2079                 hammer_knote(ap->a_fdvp, NOTE_WRITE);
2080                 hammer_knote(ap->a_tdvp, NOTE_WRITE);
2081                 while (ip->vp) {
2082                         struct vnode *vp;
2083
2084                         error = hammer_get_vnode(ip, &vp);
2085                         if (error == 0 && vp) {
2086                                 vn_unlock(vp);
2087                                 hammer_knote(ip->vp, NOTE_RENAME);
2088                                 vrele(vp);
2089                                 break;
2090                         }
2091                         kprintf("Debug: HAMMER ip/vp race2 avoided\n");
2092                 }
2093         }
2094
2095 failed:
2096         hammer_done_transaction(&trans);
2097         lwkt_reltoken(&hmp->fs_token);
2098         return (error);
2099 }
2100
2101 /*
2102  * hammer_vop_nrmdir { nch, dvp, cred }
2103  */
2104 static
2105 int
2106 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
2107 {
2108         struct hammer_transaction trans;
2109         struct hammer_inode *dip;
2110         hammer_mount_t hmp;
2111         int error;
2112
2113         dip = VTOI(ap->a_dvp);
2114         hmp = dip->hmp;
2115
2116         if (hammer_nohistory(dip) == 0 &&
2117             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2118                 return (error);
2119         }
2120
2121         lwkt_gettoken(&hmp->fs_token);
2122         hammer_start_transaction(&trans, hmp);
2123         ++hammer_stats_file_iopsw;
2124         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
2125         hammer_done_transaction(&trans);
2126         if (error == 0)
2127                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
2128         lwkt_reltoken(&hmp->fs_token);
2129         return (error);
2130 }
2131
2132 /*
2133  * hammer_vop_markatime { vp, cred }
2134  */
2135 static
2136 int
2137 hammer_vop_markatime(struct vop_markatime_args *ap)
2138 {
2139         struct hammer_transaction trans;
2140         struct hammer_inode *ip;
2141         hammer_mount_t hmp;
2142
2143         ip = VTOI(ap->a_vp);
2144         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2145                 return (EROFS);
2146         if (ip->flags & HAMMER_INODE_RO)
2147                 return (EROFS);
2148         hmp = ip->hmp;
2149         if (hmp->mp->mnt_flag & MNT_NOATIME)
2150                 return (0);
2151         lwkt_gettoken(&hmp->fs_token);
2152         hammer_start_transaction(&trans, hmp);
2153         ++hammer_stats_file_iopsw;
2154
2155         ip->ino_data.atime = trans.time;
2156         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
2157         hammer_done_transaction(&trans);
2158         hammer_knote(ap->a_vp, NOTE_ATTRIB);
2159         lwkt_reltoken(&hmp->fs_token);
2160         return (0);
2161 }
2162
2163 /*
2164  * hammer_vop_setattr { vp, vap, cred }
2165  */
2166 static
2167 int
2168 hammer_vop_setattr(struct vop_setattr_args *ap)
2169 {
2170         struct hammer_transaction trans;
2171         struct hammer_inode *ip;
2172         struct vattr *vap;
2173         hammer_mount_t hmp;
2174         int modflags;
2175         int error;
2176         int truncating;
2177         int blksize;
2178         int kflags;
2179 #if 0
2180         int64_t aligned_size;
2181 #endif
2182         u_int32_t flags;
2183
2184         vap = ap->a_vap;
2185         ip = ap->a_vp->v_data;
2186         modflags = 0;
2187         kflags = 0;
2188         hmp = ip->hmp;
2189
2190         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2191                 return(EROFS);
2192         if (ip->flags & HAMMER_INODE_RO)
2193                 return (EROFS);
2194         if (hammer_nohistory(ip) == 0 &&
2195             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2196                 return (error);
2197         }
2198
2199         lwkt_gettoken(&hmp->fs_token);
2200         hammer_start_transaction(&trans, hmp);
2201         ++hammer_stats_file_iopsw;
2202         error = 0;
2203
2204         if (vap->va_flags != VNOVAL) {
2205                 flags = ip->ino_data.uflags;
2206                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
2207                                          hammer_to_unix_xid(&ip->ino_data.uid),
2208                                          ap->a_cred);
2209                 if (error == 0) {
2210                         if (ip->ino_data.uflags != flags) {
2211                                 ip->ino_data.uflags = flags;
2212                                 ip->ino_data.ctime = trans.time;
2213                                 modflags |= HAMMER_INODE_DDIRTY;
2214                                 kflags |= NOTE_ATTRIB;
2215                         }
2216                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2217                                 error = 0;
2218                                 goto done;
2219                         }
2220                 }
2221                 goto done;
2222         }
2223         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2224                 error = EPERM;
2225                 goto done;
2226         }
2227         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2228                 mode_t cur_mode = ip->ino_data.mode;
2229                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2230                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2231                 uuid_t uuid_uid;
2232                 uuid_t uuid_gid;
2233
2234                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2235                                          ap->a_cred,
2236                                          &cur_uid, &cur_gid, &cur_mode);
2237                 if (error == 0) {
2238                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
2239                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
2240                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
2241                                  sizeof(uuid_uid)) ||
2242                             bcmp(&uuid_gid, &ip->ino_data.gid,
2243                                  sizeof(uuid_gid)) ||
2244                             ip->ino_data.mode != cur_mode) {
2245                                 ip->ino_data.uid = uuid_uid;
2246                                 ip->ino_data.gid = uuid_gid;
2247                                 ip->ino_data.mode = cur_mode;
2248                                 ip->ino_data.ctime = trans.time;
2249                                 modflags |= HAMMER_INODE_DDIRTY;
2250                         }
2251                         kflags |= NOTE_ATTRIB;
2252                 }
2253         }
2254         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2255                 switch(ap->a_vp->v_type) {
2256                 case VREG:
2257                         if (vap->va_size == ip->ino_data.size)
2258                                 break;
2259
2260                         /*
2261                          * Log the operation if in fast-fsync mode or if
2262                          * there are unterminated redo write records present.
2263                          *
2264                          * The second check is needed so the recovery code
2265                          * properly truncates write redos even if nominal
2266                          * REDO operations is turned off due to excessive
2267                          * writes, because the related records might be
2268                          * destroyed and never lay down a TERM_WRITE.
2269                          */
2270                         if ((ip->flags & HAMMER_INODE_REDO) ||
2271                             (ip->flags & HAMMER_INODE_RDIRTY)) {
2272                                 error = hammer_generate_redo(&trans, ip,
2273                                                              vap->va_size,
2274                                                              HAMMER_REDO_TRUNC,
2275                                                              NULL, 0);
2276                         }
2277                         blksize = hammer_blocksize(vap->va_size);
2278
2279                         /*
2280                          * XXX break atomicy, we can deadlock the backend
2281                          * if we do not release the lock.  Probably not a
2282                          * big deal here.
2283                          */
2284                         if (vap->va_size < ip->ino_data.size) {
2285                                 nvtruncbuf(ap->a_vp, vap->va_size,
2286                                            blksize,
2287                                            hammer_blockoff(vap->va_size),
2288                                            0);
2289                                 truncating = 1;
2290                                 kflags |= NOTE_WRITE;
2291                         } else {
2292                                 nvextendbuf(ap->a_vp,
2293                                             ip->ino_data.size,
2294                                             vap->va_size,
2295                                             hammer_blocksize(ip->ino_data.size),
2296                                             hammer_blocksize(vap->va_size),
2297                                             hammer_blockoff(ip->ino_data.size),
2298                                             hammer_blockoff(vap->va_size),
2299                                             0);
2300                                 truncating = 0;
2301                                 kflags |= NOTE_WRITE | NOTE_EXTEND;
2302                         }
2303                         ip->ino_data.size = vap->va_size;
2304                         ip->ino_data.mtime = trans.time;
2305                         /* XXX safe to use SDIRTY instead of DDIRTY here? */
2306                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2307
2308                         /*
2309                          * On-media truncation is cached in the inode until
2310                          * the inode is synchronized.  We must immediately
2311                          * handle any frontend records.
2312                          */
2313                         if (truncating) {
2314                                 hammer_ip_frontend_trunc(ip, vap->va_size);
2315 #ifdef DEBUG_TRUNCATE
2316                                 if (HammerTruncIp == NULL)
2317                                         HammerTruncIp = ip;
2318 #endif
2319                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2320                                         ip->flags |= HAMMER_INODE_TRUNCATED;
2321                                         ip->trunc_off = vap->va_size;
2322                                         hammer_inode_dirty(ip);
2323 #ifdef DEBUG_TRUNCATE
2324                                         if (ip == HammerTruncIp)
2325                                         kprintf("truncate1 %016llx\n",
2326                                                 (long long)ip->trunc_off);
2327 #endif
2328                                 } else if (ip->trunc_off > vap->va_size) {
2329                                         ip->trunc_off = vap->va_size;
2330 #ifdef DEBUG_TRUNCATE
2331                                         if (ip == HammerTruncIp)
2332                                         kprintf("truncate2 %016llx\n",
2333                                                 (long long)ip->trunc_off);
2334 #endif
2335                                 } else {
2336 #ifdef DEBUG_TRUNCATE
2337                                         if (ip == HammerTruncIp)
2338                                         kprintf("truncate3 %016llx (ignored)\n",
2339                                                 (long long)vap->va_size);
2340 #endif
2341                                 }
2342                         }
2343
2344 #if 0
2345                         /*
2346                          * When truncating, nvtruncbuf() may have cleaned out
2347                          * a portion of the last block on-disk in the buffer
2348                          * cache.  We must clean out any frontend records
2349                          * for blocks beyond the new last block.
2350                          */
2351                         aligned_size = (vap->va_size + (blksize - 1)) &
2352                                        ~(int64_t)(blksize - 1);
2353                         if (truncating && vap->va_size < aligned_size) {
2354                                 aligned_size -= blksize;
2355                                 hammer_ip_frontend_trunc(ip, aligned_size);
2356                         }
2357 #endif
2358                         break;
2359                 case VDATABASE:
2360                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2361                                 ip->flags |= HAMMER_INODE_TRUNCATED;
2362                                 ip->trunc_off = vap->va_size;
2363                                 hammer_inode_dirty(ip);
2364                         } else if (ip->trunc_off > vap->va_size) {
2365                                 ip->trunc_off = vap->va_size;
2366                         }
2367                         hammer_ip_frontend_trunc(ip, vap->va_size);
2368                         ip->ino_data.size = vap->va_size;
2369                         ip->ino_data.mtime = trans.time;
2370                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2371                         kflags |= NOTE_ATTRIB;
2372                         break;
2373                 default:
2374                         error = EINVAL;
2375                         goto done;
2376                 }
2377                 break;
2378         }
2379         if (vap->va_atime.tv_sec != VNOVAL) {
2380                 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2381                 modflags |= HAMMER_INODE_ATIME;
2382                 kflags |= NOTE_ATTRIB;
2383         }
2384         if (vap->va_mtime.tv_sec != VNOVAL) {
2385                 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2386                 modflags |= HAMMER_INODE_MTIME;
2387                 kflags |= NOTE_ATTRIB;
2388         }
2389         if (vap->va_mode != (mode_t)VNOVAL) {
2390                 mode_t   cur_mode = ip->ino_data.mode;
2391                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2392                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2393
2394                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2395                                          cur_uid, cur_gid, &cur_mode);
2396                 if (error == 0 && ip->ino_data.mode != cur_mode) {
2397                         ip->ino_data.mode = cur_mode;
2398                         ip->ino_data.ctime = trans.time;
2399                         modflags |= HAMMER_INODE_DDIRTY;
2400                         kflags |= NOTE_ATTRIB;
2401                 }
2402         }
2403 done:
2404         if (error == 0)
2405                 hammer_modify_inode(&trans, ip, modflags);
2406         hammer_done_transaction(&trans);
2407         hammer_knote(ap->a_vp, kflags);
2408         lwkt_reltoken(&hmp->fs_token);
2409         return (error);
2410 }
2411
2412 /*
2413  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2414  */
2415 static
2416 int
2417 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2418 {
2419         struct hammer_transaction trans;
2420         struct hammer_inode *dip;
2421         struct hammer_inode *nip;
2422         hammer_record_t record;
2423         struct nchandle *nch;
2424         hammer_mount_t hmp;
2425         int error;
2426         int bytes;
2427
2428         ap->a_vap->va_type = VLNK;
2429
2430         nch = ap->a_nch;
2431         dip = VTOI(ap->a_dvp);
2432         hmp = dip->hmp;
2433
2434         if (dip->flags & HAMMER_INODE_RO)
2435                 return (EROFS);
2436         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
2437                 return (error);
2438
2439         /*
2440          * Create a transaction to cover the operations we perform.
2441          */
2442         lwkt_gettoken(&hmp->fs_token);
2443         hammer_start_transaction(&trans, hmp);
2444         ++hammer_stats_file_iopsw;
2445
2446         /*
2447          * Create a new filesystem object of the requested type.  The
2448          * returned inode will be referenced but not locked.
2449          */
2450
2451         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2452                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2453                                     NULL, &nip);
2454         if (error) {
2455                 hammer_done_transaction(&trans);
2456                 *ap->a_vpp = NULL;
2457                 lwkt_reltoken(&hmp->fs_token);
2458                 return (error);
2459         }
2460
2461         /*
2462          * Add a record representing the symlink.  symlink stores the link
2463          * as pure data, not a string, and is no \0 terminated.
2464          */
2465         if (error == 0) {
2466                 bytes = strlen(ap->a_target);
2467
2468                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2469                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2470                 } else {
2471                         record = hammer_alloc_mem_record(nip, bytes);
2472                         record->type = HAMMER_MEM_RECORD_GENERAL;
2473
2474                         record->leaf.base.localization = nip->obj_localization +
2475                                                          HAMMER_LOCALIZE_MISC;
2476                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2477                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2478                         record->leaf.data_len = bytes;
2479                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2480                         bcopy(ap->a_target, record->data->symlink.name, bytes);
2481                         error = hammer_ip_add_record(&trans, record);
2482                 }
2483
2484                 /*
2485                  * Set the file size to the length of the link.
2486                  */
2487                 if (error == 0) {
2488                         nip->ino_data.size = bytes;
2489                         hammer_modify_inode(&trans, nip, HAMMER_INODE_DDIRTY);
2490                 }
2491         }
2492         if (error == 0)
2493                 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
2494                                                 nch->ncp->nc_nlen, nip);
2495
2496         /*
2497          * Finish up.
2498          */
2499         if (error) {
2500                 hammer_rel_inode(nip, 0);
2501                 *ap->a_vpp = NULL;
2502         } else {
2503                 error = hammer_get_vnode(nip, ap->a_vpp);
2504                 hammer_rel_inode(nip, 0);
2505                 if (error == 0) {
2506                         cache_setunresolved(ap->a_nch);
2507                         cache_setvp(ap->a_nch, *ap->a_vpp);
2508                         hammer_knote(ap->a_dvp, NOTE_WRITE);
2509                 }
2510         }
2511         hammer_done_transaction(&trans);
2512         lwkt_reltoken(&hmp->fs_token);
2513         return (error);
2514 }
2515
2516 /*
2517  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2518  */
2519 static
2520 int
2521 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2522 {
2523         struct hammer_transaction trans;
2524         struct hammer_inode *dip;
2525         hammer_mount_t hmp;
2526         int error;
2527
2528         dip = VTOI(ap->a_dvp);
2529         hmp = dip->hmp;
2530
2531         if (hammer_nohistory(dip) == 0 &&
2532             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2533                 return (error);
2534         }
2535
2536         lwkt_gettoken(&hmp->fs_token);
2537         hammer_start_transaction(&trans, hmp);
2538         ++hammer_stats_file_iopsw;
2539         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2540                                 ap->a_cred, ap->a_flags, -1);
2541         hammer_done_transaction(&trans);
2542         lwkt_reltoken(&hmp->fs_token);
2543
2544         return (error);
2545 }
2546
2547 /*
2548  * hammer_vop_ioctl { vp, command, data, fflag, cred }
2549  */
2550 static
2551 int
2552 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2553 {
2554         struct hammer_inode *ip = ap->a_vp->v_data;
2555         hammer_mount_t hmp = ip->hmp;
2556         int error;
2557
2558         ++hammer_stats_file_iopsr;
2559         lwkt_gettoken(&hmp->fs_token);
2560         error = hammer_ioctl(ip, ap->a_command, ap->a_data,
2561                              ap->a_fflag, ap->a_cred);
2562         lwkt_reltoken(&hmp->fs_token);
2563         return (error);
2564 }
2565
2566 static
2567 int
2568 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2569 {
2570         static const struct mountctl_opt extraopt[] = {
2571                 { HMNT_NOHISTORY,       "nohistory" },
2572                 { HMNT_MASTERID,        "master" },
2573                 { 0, NULL}
2574
2575         };
2576         struct hammer_mount *hmp;
2577         struct mount *mp;
2578         int usedbytes;
2579         int error;
2580
2581         error = 0;
2582         usedbytes = 0;
2583         mp = ap->a_head.a_ops->head.vv_mount;
2584         KKASSERT(mp->mnt_data != NULL);
2585         hmp = (struct hammer_mount *)mp->mnt_data;
2586
2587         lwkt_gettoken(&hmp->fs_token);
2588
2589         switch(ap->a_op) {
2590         case MOUNTCTL_SET_EXPORT:
2591                 if (ap->a_ctllen != sizeof(struct export_args))
2592                         error = EINVAL;
2593                 else
2594                         error = hammer_vfs_export(mp, ap->a_op,
2595                                       (const struct export_args *)ap->a_ctl);
2596                 break;
2597         case MOUNTCTL_MOUNTFLAGS:
2598         {
2599                 /*
2600                  * Call standard mountctl VOP function
2601                  * so we get user mount flags.
2602                  */
2603                 error = vop_stdmountctl(ap);
2604                 if (error)
2605                         break;
2606
2607                 usedbytes = *ap->a_res;
2608
2609                 if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2610                         usedbytes += vfs_flagstostr(hmp->hflags, extraopt,
2611                                                     ap->a_buf,
2612                                                     ap->a_buflen - usedbytes,
2613                                                     &error);
2614                 }
2615
2616                 *ap->a_res += usedbytes;
2617                 break;
2618         }
2619         default:
2620                 error = vop_stdmountctl(ap);
2621                 break;
2622         }
2623         lwkt_reltoken(&hmp->fs_token);
2624         return(error);
2625 }
2626
2627 /*
2628  * hammer_vop_strategy { vp, bio }
2629  *
2630  * Strategy call, used for regular file read & write only.  Note that the
2631  * bp may represent a cluster.
2632  *
2633  * To simplify operation and allow better optimizations in the future,
2634  * this code does not make any assumptions with regards to buffer alignment
2635  * or size.
2636  */
2637 static
2638 int
2639 hammer_vop_strategy(struct vop_strategy_args *ap)
2640 {
2641         struct buf *bp;
2642         int error;
2643
2644         bp = ap->a_bio->bio_buf;
2645
2646         switch(bp->b_cmd) {
2647         case BUF_CMD_READ:
2648                 error = hammer_vop_strategy_read(ap);
2649                 break;
2650         case BUF_CMD_WRITE:
2651                 error = hammer_vop_strategy_write(ap);
2652                 break;
2653         default:
2654                 bp->b_error = error = EINVAL;
2655                 bp->b_flags |= B_ERROR;
2656                 biodone(ap->a_bio);
2657                 break;
2658         }
2659
2660         /* hammer_dump_dedup_cache(((hammer_inode_t)ap->a_vp->v_data)->hmp); */
2661
2662         return (error);
2663 }
2664
2665 /*
2666  * Read from a regular file.  Iterate the related records and fill in the
2667  * BIO/BUF.  Gaps are zero-filled.
2668  *
2669  * The support code in hammer_object.c should be used to deal with mixed
2670  * in-memory and on-disk records.
2671  *
2672  * NOTE: Can be called from the cluster code with an oversized buf.
2673  *
2674  * XXX atime update
2675  */
2676 static
2677 int
2678 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2679 {
2680         struct hammer_transaction trans;
2681         struct hammer_inode *ip;
2682         struct hammer_inode *dip;
2683         hammer_mount_t hmp;
2684         struct hammer_cursor cursor;
2685         hammer_base_elm_t base;
2686         hammer_off_t disk_offset;
2687         struct bio *bio;
2688         struct bio *nbio;
2689         struct buf *bp;
2690         int64_t rec_offset;
2691         int64_t ran_end;
2692         int64_t tmp64;
2693         int error;
2694         int boff;
2695         int roff;
2696         int n;
2697         int isdedupable;
2698
2699         bio = ap->a_bio;
2700         bp = bio->bio_buf;
2701         ip = ap->a_vp->v_data;
2702         hmp = ip->hmp;
2703
2704         /*
2705          * The zone-2 disk offset may have been set by the cluster code via
2706          * a BMAP operation, or else should be NOOFFSET.
2707          *
2708          * Checking the high bits for a match against zone-2 should suffice.
2709          *
2710          * In cases where a lot of data duplication is present it may be
2711          * more beneficial to drop through and doubule-buffer through the
2712          * device.
2713          */
2714         nbio = push_bio(bio);
2715         if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2716             HAMMER_ZONE_LARGE_DATA) {
2717                 if (hammer_double_buffer == 0) {
2718                         lwkt_gettoken(&hmp->fs_token);
2719                         error = hammer_io_direct_read(hmp, nbio, NULL);
2720                         lwkt_reltoken(&hmp->fs_token);
2721                         return (error);
2722                 }
2723
2724                 /*
2725                  * Try to shortcut requests for double_buffer mode too.
2726                  * Since this mode runs through the device buffer cache
2727                  * only compatible buffer sizes (meaning those generated
2728                  * by normal filesystem buffers) are legal.
2729                  */
2730                 if (hammer_live_dedup == 0 && (bp->b_flags & B_PAGING) == 0) {
2731                         lwkt_gettoken(&hmp->fs_token);
2732                         error = hammer_io_indirect_read(hmp, nbio, NULL);
2733                         lwkt_reltoken(&hmp->fs_token);
2734                         return (error);
2735                 }
2736         }
2737
2738         /*
2739          * Well, that sucked.  Do it the hard way.  If all the stars are
2740          * aligned we may still be able to issue a direct-read.
2741          */
2742         lwkt_gettoken(&hmp->fs_token);
2743         hammer_simple_transaction(&trans, hmp);
2744         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2745
2746         /*
2747          * Key range (begin and end inclusive) to scan.  Note that the key's
2748          * stored in the actual records represent BASE+LEN, not BASE.  The
2749          * first record containing bio_offset will have a key > bio_offset.
2750          */
2751         cursor.key_beg.localization = ip->obj_localization +
2752                                       HAMMER_LOCALIZE_MISC;
2753         cursor.key_beg.obj_id = ip->obj_id;
2754         cursor.key_beg.create_tid = 0;
2755         cursor.key_beg.delete_tid = 0;
2756         cursor.key_beg.obj_type = 0;
2757         cursor.key_beg.key = bio->bio_offset + 1;
2758         cursor.asof = ip->obj_asof;
2759         cursor.flags |= HAMMER_CURSOR_ASOF;
2760
2761         cursor.key_end = cursor.key_beg;
2762         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2763 #if 0
2764         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2765                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2766                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2767                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2768         } else
2769 #endif
2770         {
2771                 ran_end = bio->bio_offset + bp->b_bufsize;
2772                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2773                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2774                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2775                 if (tmp64 < ran_end)
2776                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2777                 else
2778                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2779         }
2780         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2781
2782         /*
2783          * Set NOSWAPCACHE for cursor data extraction if double buffering
2784          * is disabled or (if the file is not marked cacheable via chflags
2785          * and vm.swapcache_use_chflags is enabled).
2786          */
2787         if (hammer_double_buffer == 0 ||
2788             ((ap->a_vp->v_flag & VSWAPCACHE) == 0 &&
2789              vm_swapcache_use_chflags)) {
2790                 cursor.flags |= HAMMER_CURSOR_NOSWAPCACHE;
2791         }
2792
2793         error = hammer_ip_first(&cursor);
2794         boff = 0;
2795
2796         while (error == 0) {
2797                 /*
2798                  * Get the base file offset of the record.  The key for
2799                  * data records is (base + bytes) rather then (base).
2800                  */
2801                 base = &cursor.leaf->base;
2802                 rec_offset = base->key - cursor.leaf->data_len;
2803
2804                 /*
2805                  * Calculate the gap, if any, and zero-fill it.
2806                  *
2807                  * n is the offset of the start of the record verses our
2808                  * current seek offset in the bio.
2809                  */
2810                 n = (int)(rec_offset - (bio->bio_offset + boff));
2811                 if (n > 0) {
2812                         if (n > bp->b_bufsize - boff)
2813                                 n = bp->b_bufsize - boff;
2814                         bzero((char *)bp->b_data + boff, n);
2815                         boff += n;
2816                         n = 0;
2817                 }
2818
2819                 /*
2820                  * Calculate the data offset in the record and the number
2821                  * of bytes we can copy.
2822                  *
2823                  * There are two degenerate cases.  First, boff may already
2824                  * be at bp->b_bufsize.  Secondly, the data offset within
2825                  * the record may exceed the record's size.
2826                  */
2827                 roff = -n;
2828                 rec_offset += roff;
2829                 n = cursor.leaf->data_len - roff;
2830                 if (n <= 0) {
2831                         kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2832                         n = 0;
2833                 } else if (n > bp->b_bufsize - boff) {
2834                         n = bp->b_bufsize - boff;
2835                 }
2836
2837                 /*
2838                  * Deal with cached truncations.  This cool bit of code
2839                  * allows truncate()/ftruncate() to avoid having to sync
2840                  * the file.
2841                  *
2842                  * If the frontend is truncated then all backend records are
2843                  * subject to the frontend's truncation.
2844                  *
2845                  * If the backend is truncated then backend records on-disk
2846                  * (but not in-memory) are subject to the backend's
2847                  * truncation.  In-memory records owned by the backend
2848                  * represent data written after the truncation point on the
2849                  * backend and must not be truncated.
2850                  *
2851                  * Truncate operations deal with frontend buffer cache
2852                  * buffers and frontend-owned in-memory records synchronously.
2853                  */
2854                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2855                         if (hammer_cursor_ondisk(&cursor)/* ||
2856                             cursor.iprec->flush_state == HAMMER_FST_FLUSH*/) {
2857                                 if (ip->trunc_off <= rec_offset)
2858                                         n = 0;
2859                                 else if (ip->trunc_off < rec_offset + n)
2860                                         n = (int)(ip->trunc_off - rec_offset);
2861                         }
2862                 }
2863                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2864                         if (hammer_cursor_ondisk(&cursor)) {
2865                                 if (ip->sync_trunc_off <= rec_offset)
2866                                         n = 0;
2867                                 else if (ip->sync_trunc_off < rec_offset + n)
2868                                         n = (int)(ip->sync_trunc_off - rec_offset);
2869                         }
2870                 }
2871
2872                 /*
2873                  * Try to issue a direct read into our bio if possible,
2874                  * otherwise resolve the element data into a hammer_buffer
2875                  * and copy.
2876                  *
2877                  * The buffer on-disk should be zerod past any real
2878                  * truncation point, but may not be for any synthesized
2879                  * truncation point from above.
2880                  *
2881                  * NOTE: disk_offset is only valid if the cursor data is
2882                  *       on-disk.
2883                  */
2884                 disk_offset = cursor.leaf->data_offset + roff;
2885                 isdedupable = (boff == 0 && n == bp->b_bufsize &&
2886                                hammer_cursor_ondisk(&cursor) &&
2887                                ((int)disk_offset & HAMMER_BUFMASK) == 0);
2888
2889                 if (isdedupable && hammer_double_buffer == 0) {
2890                         /*
2891                          * Direct read case
2892                          */
2893                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2894                                  HAMMER_ZONE_LARGE_DATA);
2895                         nbio->bio_offset = disk_offset;
2896                         error = hammer_io_direct_read(hmp, nbio, cursor.leaf);
2897                         if (hammer_live_dedup && error == 0)
2898                                 hammer_dedup_cache_add(ip, cursor.leaf);
2899                         goto done;
2900                 } else if (isdedupable) {
2901                         /*
2902                          * Async I/O case for reading from backing store
2903                          * and copying the data to the filesystem buffer.
2904                          * live-dedup has to verify the data anyway if it
2905                          * gets a hit later so we can just add the entry
2906                          * now.
2907                          */
2908                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2909                                  HAMMER_ZONE_LARGE_DATA);
2910                         nbio->bio_offset = disk_offset;
2911                         if (hammer_live_dedup)
2912                                 hammer_dedup_cache_add(ip, cursor.leaf);
2913                         error = hammer_io_indirect_read(hmp, nbio, cursor.leaf);
2914                         goto done;
2915                 } else if (n) {
2916                         error = hammer_ip_resolve_data(&cursor);
2917                         if (error == 0) {
2918                                 if (hammer_live_dedup && isdedupable)
2919                                         hammer_dedup_cache_add(ip, cursor.leaf);
2920                                 bcopy((char *)cursor.data + roff,
2921                                       (char *)bp->b_data + boff, n);
2922                         }
2923                 }
2924                 if (error)
2925                         break;
2926
2927                 /*
2928                  * We have to be sure that the only elements added to the
2929                  * dedup cache are those which are already on-media.
2930                  */
2931                 if (hammer_live_dedup && hammer_cursor_ondisk(&cursor))
2932                         hammer_dedup_cache_add(ip, cursor.leaf);
2933
2934                 /*
2935                  * Iterate until we have filled the request.
2936                  */
2937                 boff += n;
2938                 if (boff == bp->b_bufsize)
2939                         break;
2940                 error = hammer_ip_next(&cursor);
2941         }
2942
2943         /*
2944          * There may have been a gap after the last record
2945          */
2946         if (error == ENOENT)
2947                 error = 0;
2948         if (error == 0 && boff != bp->b_bufsize) {
2949                 KKASSERT(boff < bp->b_bufsize);
2950                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2951                 /* boff = bp->b_bufsize; */
2952         }
2953
2954         /*
2955          * Disallow swapcache operation on the vnode buffer if double
2956          * buffering is enabled, the swapcache will get the data via
2957          * the block device buffer.
2958          */
2959         if (hammer_double_buffer)
2960                 bp->b_flags |= B_NOTMETA;
2961
2962         /*
2963          * Cleanup
2964          */
2965         bp->b_resid = 0;
2966         bp->b_error = error;
2967         if (error)
2968                 bp->b_flags |= B_ERROR;
2969         biodone(ap->a_bio);
2970
2971 done:
2972         /*
2973          * Cache the b-tree node for the last data read in cache[1].
2974          *
2975          * If we hit the file EOF then also cache the node in the
2976          * governing directory's cache[3], it will be used to initialize
2977          * the new inode's cache[1] for any inodes looked up via the directory.
2978          *
2979          * This doesn't reduce disk accesses since the B-Tree chain is
2980          * likely cached, but it does reduce cpu overhead when looking
2981          * up file offsets for cpdup/tar/cpio style iterations.
2982          */
2983         if (cursor.node)
2984                 hammer_cache_node(&ip->cache[1], cursor.node);
2985         if (ran_end >= ip->ino_data.size) {
2986                 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
2987                                         ip->obj_asof, ip->obj_localization);
2988                 if (dip) {
2989                         hammer_cache_node(&dip->cache[3], cursor.node);
2990                         hammer_rel_inode(dip, 0);
2991                 }
2992         }
2993         hammer_done_cursor(&cursor);
2994         hammer_done_transaction(&trans);
2995         lwkt_reltoken(&hmp->fs_token);
2996         return(error);
2997 }
2998
2999 /*
3000  * BMAP operation - used to support cluster_read() only.
3001  *
3002  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
3003  *
3004  * This routine may return EOPNOTSUPP if the opration is not supported for
3005  * the specified offset.  The contents of the pointer arguments do not
3006  * need to be initialized in that case.
3007  *
3008  * If a disk address is available and properly aligned return 0 with
3009  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
3010  * to the run-length relative to that offset.  Callers may assume that
3011  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
3012  * large, so return EOPNOTSUPP if it is not sufficiently large.
3013  */
3014 static
3015 int
3016 hammer_vop_bmap(struct vop_bmap_args *ap)
3017 {
3018         struct hammer_transaction trans;
3019         struct hammer_inode *ip;
3020         hammer_mount_t hmp;
3021         struct hammer_cursor cursor;
3022         hammer_base_elm_t base;
3023         int64_t rec_offset;
3024         int64_t ran_end;
3025         int64_t tmp64;
3026         int64_t base_offset;
3027         int64_t base_disk_offset;
3028         int64_t last_offset;
3029         hammer_off_t last_disk_offset;
3030         hammer_off_t disk_offset;
3031         int     rec_len;
3032         int     error;
3033         int     blksize;
3034
3035         ++hammer_stats_file_iopsr;
3036         ip = ap->a_vp->v_data;
3037         hmp = ip->hmp;
3038
3039         /*
3040          * We can only BMAP regular files.  We can't BMAP database files,
3041          * directories, etc.
3042          */
3043         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
3044                 return(EOPNOTSUPP);
3045
3046         /*
3047          * bmap is typically called with runp/runb both NULL when used
3048          * for writing.  We do not support BMAP for writing atm.
3049          */
3050         if (ap->a_cmd != BUF_CMD_READ)
3051                 return(EOPNOTSUPP);
3052
3053         /*
3054          * Scan the B-Tree to acquire blockmap addresses, then translate
3055          * to raw addresses.
3056          */
3057         lwkt_gettoken(&hmp->fs_token);
3058         hammer_simple_transaction(&trans, hmp);
3059 #if 0
3060         kprintf("bmap_beg %016llx ip->cache %p\n",
3061                 (long long)ap->a_loffset, ip->cache[1]);
3062 #endif
3063         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
3064
3065         /*
3066          * Key range (begin and end inclusive) to scan.  Note that the key's
3067          * stored in the actual records represent BASE+LEN, not BASE.  The
3068          * first record containing bio_offset will have a key > bio_offset.
3069          */
3070         cursor.key_beg.localization = ip->obj_localization +
3071                                       HAMMER_LOCALIZE_MISC;
3072         cursor.key_beg.obj_id = ip->obj_id;
3073         cursor.key_beg.create_tid = 0;
3074         cursor.key_beg.delete_tid = 0;
3075         cursor.key_beg.obj_type = 0;
3076         if (ap->a_runb)
3077                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
3078         else
3079                 cursor.key_beg.key = ap->a_loffset + 1;
3080         if (cursor.key_beg.key < 0)
3081                 cursor.key_beg.key = 0;
3082         cursor.asof = ip->obj_asof;
3083         cursor.flags |= HAMMER_CURSOR_ASOF;
3084
3085         cursor.key_end = cursor.key_beg;
3086         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
3087
3088         ran_end = ap->a_loffset + MAXPHYS;
3089         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
3090         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
3091         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
3092         if (tmp64 < ran_end)
3093                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
3094         else
3095                 cursor.key_end.key = ran_end + MAXPHYS + 1;
3096
3097         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
3098
3099         error = hammer_ip_first(&cursor);
3100         base_offset = last_offset = 0;
3101         base_disk_offset = last_disk_offset = 0;
3102
3103         while (error == 0) {
3104                 /*
3105                  * Get the base file offset of the record.  The key for
3106                  * data records is (base + bytes) rather then (base).
3107                  *
3108                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
3109                  * The extra bytes should be zero on-disk and the BMAP op
3110                  * should still be ok.
3111                  */
3112                 base = &cursor.leaf->base;
3113                 rec_offset = base->key - cursor.leaf->data_len;
3114                 rec_len    = cursor.leaf->data_len;
3115
3116                 /*
3117                  * Incorporate any cached truncation.
3118                  *
3119                  * NOTE: Modifications to rec_len based on synthesized
3120                  * truncation points remove the guarantee that any extended
3121                  * data on disk is zero (since the truncations may not have
3122                  * taken place on-media yet).
3123                  */
3124                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
3125                         if (hammer_cursor_ondisk(&cursor) ||
3126                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
3127                                 if (ip->trunc_off <= rec_offset)
3128                                         rec_len = 0;
3129                                 else if (ip->trunc_off < rec_offset + rec_len)
3130                                         rec_len = (int)(ip->trunc_off - rec_offset);
3131                         }
3132                 }
3133                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
3134                         if (hammer_cursor_ondisk(&cursor)) {
3135                                 if (ip->sync_trunc_off <= rec_offset)
3136                                         rec_len = 0;
3137                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
3138                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
3139                         }
3140                 }
3141
3142                 /*
3143                  * Accumulate information.  If we have hit a discontiguous
3144                  * block reset base_offset unless we are already beyond the
3145                  * requested offset.  If we are, that's it, we stop.
3146                  */
3147                 if (error)
3148                         break;
3149                 if (hammer_cursor_ondisk(&cursor)) {
3150                         disk_offset = cursor.leaf->data_offset;
3151                         if (rec_offset != last_offset ||
3152                             disk_offset != last_disk_offset) {
3153                                 if (rec_offset > ap->a_loffset)
3154                                         break;
3155                                 base_offset = rec_offset;
3156                                 base_disk_offset = disk_offset;
3157                         }
3158                         last_offset = rec_offset + rec_len;
3159                         last_disk_offset = disk_offset + rec_len;
3160
3161                         if (hammer_live_dedup)
3162                                 hammer_dedup_cache_add(ip, cursor.leaf);
3163                 }
3164
3165                 error = hammer_ip_next(&cursor);
3166         }
3167
3168 #if 0
3169         kprintf("BMAP %016llx:  %016llx - %016llx\n",
3170                 (long long)ap->a_loffset,
3171                 (long long)base_offset,
3172                 (long long)last_offset);
3173         kprintf("BMAP %16s:  %016llx - %016llx\n", "",
3174                 (long long)base_disk_offset,
3175                 (long long)last_disk_offset);
3176 #endif
3177
3178         if (cursor.node) {
3179                 hammer_cache_node(&ip->cache[1], cursor.node);
3180 #if 0
3181                 kprintf("bmap_end2 %016llx ip->cache %p\n",
3182                         (long long)ap->a_loffset, ip->cache[1]);
3183 #endif
3184         }
3185         hammer_done_cursor(&cursor);
3186         hammer_done_transaction(&trans);
3187         lwkt_reltoken(&hmp->fs_token);
3188
3189         /*
3190          * If we couldn't find any records or the records we did find were
3191          * all behind the requested offset, return failure.  A forward
3192          * truncation can leave a hole w/ no on-disk records.
3193          */
3194         if (last_offset == 0 || last_offset < ap->a_loffset)
3195                 return (EOPNOTSUPP);
3196
3197         /*
3198          * Figure out the block size at the requested offset and adjust
3199          * our limits so the cluster_read() does not create inappropriately
3200          * sized buffer cache buffers.
3201          */
3202         blksize = hammer_blocksize(ap->a_loffset);
3203         if (hammer_blocksize(base_offset) != blksize) {
3204                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
3205         }
3206         if (last_offset != ap->a_loffset &&
3207             hammer_blocksize(last_offset - 1) != blksize) {
3208                 last_offset = hammer_blockdemarc(ap->a_loffset,
3209                                                  last_offset - 1);
3210         }
3211
3212         /*
3213          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3214          * from occuring.
3215          */
3216         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
3217
3218         if ((disk_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_LARGE_DATA) {
3219                 /*
3220                  * Only large-data zones can be direct-IOd
3221                  */
3222                 error = EOPNOTSUPP;
3223         } else if ((disk_offset & HAMMER_BUFMASK) ||
3224                    (last_offset - ap->a_loffset) < blksize) {
3225                 /*
3226                  * doffsetp is not aligned or the forward run size does
3227                  * not cover a whole buffer, disallow the direct I/O.
3228                  */
3229                 error = EOPNOTSUPP;
3230         } else {
3231                 /*
3232                  * We're good.
3233                  */
3234                 *ap->a_doffsetp = disk_offset;
3235                 if (ap->a_runb) {
3236                         *ap->a_runb = ap->a_loffset - base_offset;
3237                         KKASSERT(*ap->a_runb >= 0);
3238                 }
3239                 if (ap->a_runp) {
3240                         *ap->a_runp = last_offset - ap->a_loffset;
3241                         KKASSERT(*ap->a_runp >= 0);
3242                 }
3243                 error = 0;
3244         }
3245         return(error);
3246 }
3247
3248 /*
3249  * Write to a regular file.   Because this is a strategy call the OS is
3250  * trying to actually get data onto the media.
3251  */
3252 static
3253 int
3254 hammer_vop_strategy_write(struct vop_strategy_args *ap)
3255 {
3256         hammer_record_t record;
3257         hammer_mount_t hmp;
3258         hammer_inode_t ip;
3259         struct bio *bio;
3260         struct buf *bp;
3261         int blksize __debugvar;
3262         int bytes;
3263         int error;
3264
3265         bio = ap->a_bio;
3266         bp = bio->bio_buf;
3267         ip = ap->a_vp->v_data;
3268         hmp = ip->hmp;
3269
3270         blksize = hammer_blocksize(bio->bio_offset);
3271         KKASSERT(bp->b_bufsize == blksize);
3272
3273         if (ip->flags & HAMMER_INODE_RO) {
3274                 bp->b_error = EROFS;
3275                 bp->b_flags |= B_ERROR;
3276                 biodone(ap->a_bio);
3277                 return(EROFS);
3278         }
3279
3280         lwkt_gettoken(&hmp->fs_token);
3281
3282         /*
3283          * Disallow swapcache operation on the vnode buffer if double
3284          * buffering is enabled, the swapcache will get the data via
3285          * the block device buffer.
3286          */
3287         if (hammer_double_buffer)
3288                 bp->b_flags |= B_NOTMETA;
3289
3290         /*
3291          * Interlock with inode destruction (no in-kernel or directory
3292          * topology visibility).  If we queue new IO while trying to
3293          * destroy the inode we can deadlock the vtrunc call in
3294          * hammer_inode_unloadable_check().
3295          *
3296          * Besides, there's no point flushing a bp associated with an
3297          * inode that is being destroyed on-media and has no kernel
3298          * references.
3299          */
3300         if ((ip->flags | ip->sync_flags) &
3301             (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
3302                 bp->b_resid = 0;
3303                 biodone(ap->a_bio);
3304                 lwkt_reltoken(&hmp->fs_token);
3305                 return(0);
3306         }
3307
3308         /*
3309          * Reserve space and issue a direct-write from the front-end.
3310          * NOTE: The direct_io code will hammer_bread/bcopy smaller
3311          * allocations.
3312          *
3313          * An in-memory record will be installed to reference the storage
3314          * until the flusher can get to it.
3315          *
3316          * Since we own the high level bio the front-end will not try to
3317          * do a direct-read until the write completes.
3318          *
3319          * NOTE: The only time we do not reserve a full-sized buffers
3320          * worth of data is if the file is small.  We do not try to
3321          * allocate a fragment (from the small-data zone) at the end of
3322          * an otherwise large file as this can lead to wildly separated
3323          * data.
3324          */
3325         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
3326         KKASSERT(bio->bio_offset < ip->ino_data.size);
3327         if (bio->bio_offset || ip->ino_data.size > HAMMER_HBUFSIZE)
3328                 bytes = bp->b_bufsize;
3329         else
3330                 bytes = ((int)ip->ino_data.size + 15) & ~15;
3331
3332         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
3333                                     bytes, &error);
3334
3335         /*
3336          * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3337          * in hammer_vop_write().  We must flag the record so the proper
3338          * REDO_TERM_WRITE entry is generated during the flush.
3339          */
3340         if (record) {
3341                 if (bp->b_flags & B_VFSFLAG1) {
3342                         record->flags |= HAMMER_RECF_REDO;
3343                         bp->b_flags &= ~B_VFSFLAG1;
3344                 }
3345                 if (record->flags & HAMMER_RECF_DEDUPED) {
3346                         bp->b_resid = 0;
3347                         hammer_ip_replace_bulk(hmp, record);
3348                         biodone(ap->a_bio);
3349                 } else {
3350                         hammer_io_direct_write(hmp, bio, record);
3351                 }
3352                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
3353                         hammer_flush_inode(ip, 0);
3354         } else {
3355                 bp->b_bio2.bio_offset = NOOFFSET;
3356                 bp->b_error = error;
3357                 bp->b_flags |= B_ERROR;
3358                 biodone(ap->a_bio);
3359         }
3360         lwkt_reltoken(&hmp->fs_token);
3361         return(error);
3362 }
3363
3364 /*
3365  * dounlink - disconnect a directory entry
3366  *
3367  * XXX whiteout support not really in yet
3368  */
3369 static int
3370 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
3371                 struct vnode *dvp, struct ucred *cred,
3372                 int flags, int isdir)
3373 {
3374         struct namecache *ncp;
3375         hammer_inode_t dip;
3376         hammer_inode_t ip;
3377         hammer_mount_t hmp;
3378         struct hammer_cursor cursor;
3379         int64_t namekey;
3380         u_int32_t max_iterations;
3381         int nlen, error;
3382
3383         /*
3384          * Calculate the namekey and setup the key range for the scan.  This
3385          * works kinda like a chained hash table where the lower 32 bits
3386          * of the namekey synthesize the chain.
3387          *
3388          * The key range is inclusive of both key_beg and key_end.
3389          */
3390         dip = VTOI(dvp);
3391         ncp = nch->ncp;
3392         hmp = dip->hmp;
3393
3394         if (dip->flags & HAMMER_INODE_RO)
3395                 return (EROFS);
3396
3397         namekey = hammer_directory_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3398                                            &max_iterations);
3399 retry:
3400         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3401         cursor.key_beg.localization = dip->obj_localization +
3402                                       hammer_dir_localization(dip);
3403         cursor.key_beg.obj_id = dip->obj_id;
3404         cursor.key_beg.key = namekey;
3405         cursor.key_beg.create_tid = 0;
3406         cursor.key_beg.delete_tid = 0;
3407         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3408         cursor.key_beg.obj_type = 0;
3409
3410         cursor.key_end = cursor.key_beg;
3411         cursor.key_end.key += max_iterations;
3412         cursor.asof = dip->obj_asof;
3413         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3414
3415         /*
3416          * Scan all matching records (the chain), locate the one matching
3417          * the requested path component.  info->last_error contains the
3418          * error code on search termination and could be 0, ENOENT, or
3419          * something else.
3420          *
3421          * The hammer_ip_*() functions merge in-memory records with on-disk
3422          * records for the purposes of the search.
3423          */
3424         error = hammer_ip_first(&cursor);
3425
3426         while (error == 0) {
3427                 error = hammer_ip_resolve_data(&cursor);
3428                 if (error)
3429                         break;
3430                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3431                 KKASSERT(nlen > 0);
3432                 if (ncp->nc_nlen == nlen &&
3433                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3434                         break;
3435                 }
3436                 error = hammer_ip_next(&cursor);
3437         }
3438
3439         /*
3440          * If all is ok we have to get the inode so we can adjust nlinks.
3441          * To avoid a deadlock with the flusher we must release the inode
3442          * lock on the directory when acquiring the inode for the entry.
3443          *
3444          * If the target is a directory, it must be empty.
3445          */
3446         if (error == 0) {
3447                 hammer_unlock(&cursor.ip->lock);
3448                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3449                                       hmp->asof,
3450                                       cursor.data->entry.localization,
3451                                       0, &error);
3452                 hammer_lock_sh(&cursor.ip->lock);
3453                 if (error == ENOENT) {
3454                         kprintf("HAMMER: WARNING: Removing "
3455                                 "dirent w/missing inode \"%s\"\n"
3456                                 "\tobj_id = %016llx\n",
3457                                 ncp->nc_name,
3458                                 (long long)cursor.data->entry.obj_id);
3459                         error = 0;
3460                 }
3461
3462                 /*
3463                  * If isdir >= 0 we validate that the entry is or is not a
3464                  * directory.  If isdir < 0 we don't care.
3465                  */
3466                 if (error == 0 && isdir >= 0 && ip) {
3467                         if (isdir &&
3468                             ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3469                                 error = ENOTDIR;
3470                         } else if (isdir == 0 &&
3471                             ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3472                                 error = EISDIR;
3473                         }
3474                 }
3475
3476                 /*
3477                  * If we are trying to remove a directory the directory must
3478                  * be empty.
3479                  *
3480                  * The check directory code can loop and deadlock/retry.  Our
3481                  * own cursor's node locks must be released to avoid a 3-way
3482                  * deadlock with the flusher if the check directory code
3483                  * blocks.
3484                  *
3485                  * If any changes whatsoever have been made to the cursor
3486                  * set EDEADLK and retry.
3487                  *
3488                  * WARNING: See warnings in hammer_unlock_cursor()
3489                  *          function.
3490                  */
3491                 if (error == 0 && ip && ip->ino_data.obj_type ==
3492                                         HAMMER_OBJTYPE_DIRECTORY) {
3493                         hammer_unlock_cursor(&cursor);
3494                         error = hammer_ip_check_directory_empty(trans, ip);
3495                         hammer_lock_cursor(&cursor);
3496                         if (cursor.flags & HAMMER_CURSOR_RETEST) {
3497                                 kprintf("HAMMER: Warning: avoided deadlock "
3498                                         "on rmdir '%s'\n",
3499                                         ncp->nc_name);
3500                                 error = EDEADLK;
3501                         }
3502                 }
3503
3504                 /*
3505                  * Delete the directory entry.
3506                  *
3507                  * WARNING: hammer_ip_del_directory() may have to terminate
3508                  * the cursor to avoid a deadlock.  It is ok to call
3509                  * hammer_done_cursor() twice.
3510                  */
3511                 if (error == 0) {
3512                         error = hammer_ip_del_directory(trans, &cursor,
3513                                                         dip, ip);
3514                 }
3515                 hammer_done_cursor(&cursor);
3516                 if (error == 0) {
3517                         /*
3518                          * Tell the namecache that we are now unlinked.
3519                          */
3520                         cache_unlink(nch);
3521
3522                         /*
3523                          * NOTE: ip->vp, if non-NULL, cannot be directly
3524                          *       referenced without formally acquiring the
3525                          *       vp since the vp might have zero refs on it,
3526                          *       or in the middle of a reclaim, etc.
3527                          *
3528                          * NOTE: The cache_setunresolved() can rip the vp
3529                          *       out from under us since the vp may not have
3530                          *       any refs, in which case ip->vp will be NULL
3531                          *       from the outset.
3532                          */
3533                         while (ip && ip->vp) {
3534                                 struct vnode *vp;
3535
3536                                 error = hammer_get_vnode(ip, &vp);
3537                                 if (error == 0 && vp) {
3538                                         vn_unlock(vp);
3539                                         hammer_knote(ip->vp, NOTE_DELETE);
3540 #if 0
3541                                         /*
3542                                          * Don't do this, it can deadlock
3543                                          * on concurrent rm's of hardlinks.
3544                                          * Shouldn't be needed any more.
3545                                          */
3546                                         cache_inval_vp(ip->vp, CINV_DESTROY);
3547 #endif
3548                                         vrele(vp);
3549                                         break;
3550                                 }
3551                                 kprintf("Debug: HAMMER ip/vp race1 avoided\n");
3552                         }
3553                 }
3554                 if (ip)
3555                         hammer_rel_inode(ip, 0);
3556         } else {
3557                 hammer_done_cursor(&cursor);
3558         }
3559         if (error == EDEADLK)
3560                 goto retry;
3561
3562         return (error);
3563 }
3564
3565 /************************************************************************
3566  *                          FIFO AND SPECFS OPS                         *
3567  ************************************************************************
3568  *
3569  */
3570 static int
3571 hammer_vop_fifoclose (struct vop_close_args *ap)
3572 {
3573         /* XXX update itimes */
3574         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3575 }
3576
3577 static int
3578 hammer_vop_fiforead (struct vop_read_args *ap)
3579 {
3580         int error;
3581
3582         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3583         /* XXX update access time */
3584         return (error);
3585 }
3586
3587 static int
3588 hammer_vop_fifowrite (struct vop_write_args *ap)
3589 {
3590         int error;
3591
3592         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3593         /* XXX update access time */
3594         return (error);
3595 }
3596
3597 static
3598 int
3599 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3600 {
3601         int error;
3602
3603         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3604         if (error)
3605                 error = hammer_vop_kqfilter(ap);
3606         return(error);
3607 }
3608
3609 /************************************************************************
3610  *                          KQFILTER OPS                                *
3611  ************************************************************************
3612  *
3613  */
3614 static void filt_hammerdetach(struct knote *kn);
3615 static int filt_hammerread(struct knote *kn, long hint);
3616 static int filt_hammerwrite(struct knote *kn, long hint);
3617 static int filt_hammervnode(struct knote *kn, long hint);
3618
3619 static struct filterops hammerread_filtops =
3620         { FILTEROP_ISFD | FILTEROP_MPSAFE,
3621           NULL, filt_hammerdetach, filt_hammerread };
3622 static struct filterops hammerwrite_filtops =
3623         { FILTEROP_ISFD | FILTEROP_MPSAFE,
3624           NULL, filt_hammerdetach, filt_hammerwrite };
3625 static struct filterops hammervnode_filtops =
3626         { FILTEROP_ISFD | FILTEROP_MPSAFE,
3627           NULL, filt_hammerdetach, filt_hammervnode };
3628
3629 static
3630 int
3631 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3632 {
3633         struct vnode *vp = ap->a_vp;
3634         struct knote *kn = ap->a_kn;
3635
3636         switch (kn->kn_filter) {
3637         case EVFILT_READ:
3638                 kn->kn_fop = &hammerread_filtops;
3639                 break;
3640         case EVFILT_WRITE:
3641                 kn->kn_fop = &hammerwrite_filtops;
3642                 break;
3643         case EVFILT_VNODE:
3644                 kn->kn_fop = &hammervnode_filtops;
3645                 break;
3646         default:
3647                 return (EOPNOTSUPP);
3648         }
3649
3650         kn->kn_hook = (caddr_t)vp;
3651
3652         knote_insert(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3653
3654         return(0);
3655 }
3656
3657 static void
3658 filt_hammerdetach(struct knote *kn)
3659 {
3660         struct vnode *vp = (void *)kn->kn_hook;
3661
3662         knote_remove(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3663 }
3664
3665 static int
3666 filt_hammerread(struct knote *kn, long hint)
3667 {
3668         struct vnode *vp = (void *)kn->kn_hook;
3669         hammer_inode_t ip = VTOI(vp);
3670         hammer_mount_t hmp = ip->hmp;
3671         off_t off;
3672
3673         if (hint == NOTE_REVOKE) {
3674                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3675                 return(1);
3676         }
3677         lwkt_gettoken(&hmp->fs_token);  /* XXX use per-ip-token */
3678         off = ip->ino_data.size - kn->kn_fp->f_offset;
3679         kn->kn_data = (off < INTPTR_MAX) ? off : INTPTR_MAX;
3680         lwkt_reltoken(&hmp->fs_token);
3681         if (kn->kn_sfflags & NOTE_OLDAPI)
3682                 return(1);
3683         return (kn->kn_data != 0);
3684 }
3685
3686 static int
3687 filt_hammerwrite(struct knote *kn, long hint)
3688 {
3689         if (hint == NOTE_REVOKE)
3690                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3691         kn->kn_data = 0;
3692         return (1);
3693 }
3694
3695 static int
3696 filt_hammervnode(struct knote *kn, long hint)
3697 {
3698         if (kn->kn_sfflags & hint)
3699                 kn->kn_fflags |= hint;
3700         if (hint == NOTE_REVOKE) {
3701                 kn->kn_flags |= (EV_EOF | EV_NODATA);
3702                 return (1);
3703         }
3704         return (kn->kn_fflags != 0);
3705 }
3706