HAMMER 56A/Many: Performance tuning - MEDIA STRUCTURES CHANGED!
[dragonfly.git] / sys / vfs / hammer / hammer_vnops.c
1 /*
2  * Copyright (c) 2007-2008 The DragonFly Project.  All rights reserved.
3  * 
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  * 
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  * 
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  * 
34  * $DragonFly: src/sys/vfs/hammer/hammer_vnops.c,v 1.71 2008/06/17 04:02:38 dillon Exp $
35  */
36
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
40 #include <sys/fcntl.h>
41 #include <sys/namecache.h>
42 #include <sys/vnode.h>
43 #include <sys/lockf.h>
44 #include <sys/event.h>
45 #include <sys/stat.h>
46 #include <sys/dirent.h>
47 #include <vm/vm_extern.h>
48 #include <vfs/fifofs/fifo.h>
49 #include "hammer.h"
50
51 /*
52  * USERFS VNOPS
53  */
54 /*static int hammer_vop_vnoperate(struct vop_generic_args *);*/
55 static int hammer_vop_fsync(struct vop_fsync_args *);
56 static int hammer_vop_read(struct vop_read_args *);
57 static int hammer_vop_write(struct vop_write_args *);
58 static int hammer_vop_access(struct vop_access_args *);
59 static int hammer_vop_advlock(struct vop_advlock_args *);
60 static int hammer_vop_close(struct vop_close_args *);
61 static int hammer_vop_ncreate(struct vop_ncreate_args *);
62 static int hammer_vop_getattr(struct vop_getattr_args *);
63 static int hammer_vop_nresolve(struct vop_nresolve_args *);
64 static int hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *);
65 static int hammer_vop_nlink(struct vop_nlink_args *);
66 static int hammer_vop_nmkdir(struct vop_nmkdir_args *);
67 static int hammer_vop_nmknod(struct vop_nmknod_args *);
68 static int hammer_vop_open(struct vop_open_args *);
69 static int hammer_vop_pathconf(struct vop_pathconf_args *);
70 static int hammer_vop_print(struct vop_print_args *);
71 static int hammer_vop_readdir(struct vop_readdir_args *);
72 static int hammer_vop_readlink(struct vop_readlink_args *);
73 static int hammer_vop_nremove(struct vop_nremove_args *);
74 static int hammer_vop_nrename(struct vop_nrename_args *);
75 static int hammer_vop_nrmdir(struct vop_nrmdir_args *);
76 static int hammer_vop_setattr(struct vop_setattr_args *);
77 static int hammer_vop_strategy(struct vop_strategy_args *);
78 static int hammer_vop_bmap(struct vop_bmap_args *ap);
79 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
80 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
81 static int hammer_vop_ioctl(struct vop_ioctl_args *);
82 static int hammer_vop_mountctl(struct vop_mountctl_args *);
83
84 static int hammer_vop_fifoclose (struct vop_close_args *);
85 static int hammer_vop_fiforead (struct vop_read_args *);
86 static int hammer_vop_fifowrite (struct vop_write_args *);
87
88 static int hammer_vop_specclose (struct vop_close_args *);
89 static int hammer_vop_specread (struct vop_read_args *);
90 static int hammer_vop_specwrite (struct vop_write_args *);
91
92 struct vop_ops hammer_vnode_vops = {
93         .vop_default =          vop_defaultop,
94         .vop_fsync =            hammer_vop_fsync,
95         .vop_getpages =         vop_stdgetpages,
96         .vop_putpages =         vop_stdputpages,
97         .vop_read =             hammer_vop_read,
98         .vop_write =            hammer_vop_write,
99         .vop_access =           hammer_vop_access,
100         .vop_advlock =          hammer_vop_advlock,
101         .vop_close =            hammer_vop_close,
102         .vop_ncreate =          hammer_vop_ncreate,
103         .vop_getattr =          hammer_vop_getattr,
104         .vop_inactive =         hammer_vop_inactive,
105         .vop_reclaim =          hammer_vop_reclaim,
106         .vop_nresolve =         hammer_vop_nresolve,
107         .vop_nlookupdotdot =    hammer_vop_nlookupdotdot,
108         .vop_nlink =            hammer_vop_nlink,
109         .vop_nmkdir =           hammer_vop_nmkdir,
110         .vop_nmknod =           hammer_vop_nmknod,
111         .vop_open =             hammer_vop_open,
112         .vop_pathconf =         hammer_vop_pathconf,
113         .vop_print =            hammer_vop_print,
114         .vop_readdir =          hammer_vop_readdir,
115         .vop_readlink =         hammer_vop_readlink,
116         .vop_nremove =          hammer_vop_nremove,
117         .vop_nrename =          hammer_vop_nrename,
118         .vop_nrmdir =           hammer_vop_nrmdir,
119         .vop_setattr =          hammer_vop_setattr,
120         .vop_bmap =             hammer_vop_bmap,
121         .vop_strategy =         hammer_vop_strategy,
122         .vop_nsymlink =         hammer_vop_nsymlink,
123         .vop_nwhiteout =        hammer_vop_nwhiteout,
124         .vop_ioctl =            hammer_vop_ioctl,
125         .vop_mountctl =         hammer_vop_mountctl
126 };
127
128 struct vop_ops hammer_spec_vops = {
129         .vop_default =          spec_vnoperate,
130         .vop_fsync =            hammer_vop_fsync,
131         .vop_read =             hammer_vop_specread,
132         .vop_write =            hammer_vop_specwrite,
133         .vop_access =           hammer_vop_access,
134         .vop_close =            hammer_vop_specclose,
135         .vop_getattr =          hammer_vop_getattr,
136         .vop_inactive =         hammer_vop_inactive,
137         .vop_reclaim =          hammer_vop_reclaim,
138         .vop_setattr =          hammer_vop_setattr
139 };
140
141 struct vop_ops hammer_fifo_vops = {
142         .vop_default =          fifo_vnoperate,
143         .vop_fsync =            hammer_vop_fsync,
144         .vop_read =             hammer_vop_fiforead,
145         .vop_write =            hammer_vop_fifowrite,
146         .vop_access =           hammer_vop_access,
147         .vop_close =            hammer_vop_fifoclose,
148         .vop_getattr =          hammer_vop_getattr,
149         .vop_inactive =         hammer_vop_inactive,
150         .vop_reclaim =          hammer_vop_reclaim,
151         .vop_setattr =          hammer_vop_setattr
152 };
153
154 #ifdef DEBUG_TRUNCATE
155 struct hammer_inode *HammerTruncIp;
156 #endif
157
158 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
159                            struct vnode *dvp, struct ucred *cred, int flags);
160 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
161 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
162 static void hammer_cleanup_write_io(hammer_inode_t ip);
163 static void hammer_update_rsv_databufs(hammer_inode_t ip);
164
165 #if 0
166 static
167 int
168 hammer_vop_vnoperate(struct vop_generic_args *)
169 {
170         return (VOCALL(&hammer_vnode_vops, ap));
171 }
172 #endif
173
174 /*
175  * hammer_vop_fsync { vp, waitfor }
176  */
177 static
178 int
179 hammer_vop_fsync(struct vop_fsync_args *ap)
180 {
181         hammer_inode_t ip = VTOI(ap->a_vp);
182
183         vfsync(ap->a_vp, ap->a_waitfor, 1, NULL, NULL);
184         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
185         if (ap->a_waitfor == MNT_WAIT)
186                 hammer_wait_inode(ip);
187         return (ip->error);
188 }
189
190 /*
191  * hammer_vop_read { vp, uio, ioflag, cred }
192  */
193 static
194 int
195 hammer_vop_read(struct vop_read_args *ap)
196 {
197         struct hammer_transaction trans;
198         hammer_inode_t ip;
199         off_t offset;
200         struct buf *bp;
201         struct uio *uio;
202         int error;
203         int n;
204         int seqcount;
205
206         if (ap->a_vp->v_type != VREG)
207                 return (EINVAL);
208         ip = VTOI(ap->a_vp);
209         error = 0;
210         seqcount = ap->a_ioflag >> 16;
211
212         hammer_start_transaction(&trans, ip->hmp);
213
214         /*
215          * Access the data in HAMMER_BUFSIZE blocks via the buffer cache.
216          */
217         uio = ap->a_uio;
218         while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
219                 offset = uio->uio_offset & HAMMER_BUFMASK;
220                 if (hammer_debug_cluster_enable) {
221                         error = cluster_read(ap->a_vp, ip->ino_data.size,
222                                              uio->uio_offset - offset,
223                                              HAMMER_BUFSIZE,
224                                              MAXBSIZE, seqcount, &bp);
225                 } else {
226                         error = bread(ap->a_vp, uio->uio_offset - offset,
227                                       HAMMER_BUFSIZE, &bp);
228                 }
229                 if (error) {
230                         brelse(bp);
231                         break;
232                 }
233
234                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
235                 n = HAMMER_BUFSIZE - offset;
236                 if (n > uio->uio_resid)
237                         n = uio->uio_resid;
238                 if (n > ip->ino_data.size - uio->uio_offset)
239                         n = (int)(ip->ino_data.size - uio->uio_offset);
240                 error = uiomove((char *)bp->b_data + offset, n, uio);
241
242                 /* data has a lower priority then meta-data */
243                 bp->b_flags |= B_AGE;
244                 bqrelse(bp);
245                 if (error)
246                         break;
247         }
248         if ((ip->flags & HAMMER_INODE_RO) == 0 &&
249             (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
250                 ip->ino_leaf.atime = trans.time;
251                 hammer_modify_inode(ip, HAMMER_INODE_ITIMES);
252         }
253         hammer_done_transaction(&trans);
254         return (error);
255 }
256
257 /*
258  * hammer_vop_write { vp, uio, ioflag, cred }
259  */
260 static
261 int
262 hammer_vop_write(struct vop_write_args *ap)
263 {
264         struct hammer_transaction trans;
265         struct hammer_inode *ip;
266         struct uio *uio;
267         int rel_offset;
268         off_t base_offset;
269         struct buf *bp;
270         int error;
271         int n;
272         int flags;
273         int count;
274         int seqcount;
275
276         if (ap->a_vp->v_type != VREG)
277                 return (EINVAL);
278         ip = VTOI(ap->a_vp);
279         error = 0;
280         seqcount = ap->a_ioflag >> 16;
281
282         if (ip->flags & HAMMER_INODE_RO)
283                 return (EROFS);
284
285         /*
286          * Create a transaction to cover the operations we perform.
287          */
288         hammer_start_transaction(&trans, ip->hmp);
289         uio = ap->a_uio;
290
291         /*
292          * Check append mode
293          */
294         if (ap->a_ioflag & IO_APPEND)
295                 uio->uio_offset = ip->ino_data.size;
296
297         /*
298          * Check for illegal write offsets.  Valid range is 0...2^63-1.
299          *
300          * NOTE: the base_off assignment is required to work around what
301          * I consider to be a GCC-4 optimization bug.
302          */
303         if (uio->uio_offset < 0) {
304                 hammer_done_transaction(&trans);
305                 return (EFBIG);
306         }
307         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
308         if (uio->uio_resid > 0 && base_offset <= 0) {
309                 hammer_done_transaction(&trans);
310                 return (EFBIG);
311         }
312
313         /*
314          * Access the data in HAMMER_BUFSIZE blocks via the buffer cache.
315          */
316         count = 0;
317         while (uio->uio_resid > 0) {
318                 int fixsize = 0;
319
320                 if ((error = hammer_checkspace(trans.hmp)) != 0)
321                         break;
322
323                 /*
324                  * Do not allow HAMMER to blow out the buffer cache.
325                  *
326                  * Do not allow HAMMER to blow out system memory by
327                  * accumulating too many records.   Records are decoupled
328                  * from the buffer cache.
329                  *
330                  * Always check at the beginning so separate writes are
331                  * not able to bypass this code.
332                  *
333                  * WARNING: Cannot unlock vp when doing a NOCOPY write as
334                  * part of a putpages operation.  Doing so could cause us
335                  * to deadlock against the VM system when we try to re-lock.
336                  */
337                 if ((count++ & 15) == 0) {
338                         if (uio->uio_segflg != UIO_NOCOPY) {
339                                 vn_unlock(ap->a_vp);
340                                 if ((ap->a_ioflag & IO_NOBWILL) == 0)
341                                         bwillwrite();
342                         }
343                         if (ip->rsv_recs > hammer_limit_irecs)
344                                 hammer_wait_inode_recs(ip);
345                         if (uio->uio_segflg != UIO_NOCOPY)
346                                 vn_lock(ap->a_vp, LK_EXCLUSIVE|LK_RETRY);
347                 }
348
349                 rel_offset = (int)(uio->uio_offset & HAMMER_BUFMASK);
350                 base_offset = uio->uio_offset & ~HAMMER_BUFMASK64;
351                 n = HAMMER_BUFSIZE - rel_offset;
352                 if (n > uio->uio_resid)
353                         n = uio->uio_resid;
354                 if (uio->uio_offset + n > ip->ino_data.size) {
355                         vnode_pager_setsize(ap->a_vp, uio->uio_offset + n);
356                         fixsize = 1;
357                 }
358
359                 if (uio->uio_segflg == UIO_NOCOPY) {
360                         /*
361                          * Issuing a write with the same data backing the
362                          * buffer.  Instantiate the buffer to collect the
363                          * backing vm pages, then read-in any missing bits.
364                          *
365                          * This case is used by vop_stdputpages().
366                          */
367                         bp = getblk(ap->a_vp, base_offset,
368                                     HAMMER_BUFSIZE, GETBLK_BHEAVY, 0);
369                         if ((bp->b_flags & B_CACHE) == 0) {
370                                 bqrelse(bp);
371                                 error = bread(ap->a_vp, base_offset,
372                                               HAMMER_BUFSIZE, &bp);
373                         }
374                 } else if (rel_offset == 0 && uio->uio_resid >= HAMMER_BUFSIZE) {
375                         /*
376                          * Even though we are entirely overwriting the buffer
377                          * we may still have to zero it out to avoid a 
378                          * mmap/write visibility issue.
379                          */
380                         bp = getblk(ap->a_vp, base_offset,
381                                     HAMMER_BUFSIZE, GETBLK_BHEAVY, 0);
382                         if ((bp->b_flags & B_CACHE) == 0)
383                                 vfs_bio_clrbuf(bp);
384                 } else if (base_offset >= ip->ino_data.size) {
385                         /*
386                          * If the base offset of the buffer is beyond the
387                          * file EOF, we don't have to issue a read.
388                          */
389                         bp = getblk(ap->a_vp, base_offset,
390                                     HAMMER_BUFSIZE, GETBLK_BHEAVY, 0);
391                         vfs_bio_clrbuf(bp);
392                 } else {
393                         /*
394                          * Partial overwrite, read in any missing bits then
395                          * replace the portion being written.
396                          */
397                         error = bread(ap->a_vp, base_offset,
398                                       HAMMER_BUFSIZE, &bp);
399                         if (error == 0)
400                                 bheavy(bp);
401                 }
402                 if (error == 0) {
403                         error = uiomove((char *)bp->b_data + rel_offset,
404                                         n, uio);
405                 }
406
407                 /*
408                  * If we screwed up we have to undo any VM size changes we
409                  * made.
410                  */
411                 if (error) {
412                         brelse(bp);
413                         if (fixsize) {
414                                 vtruncbuf(ap->a_vp, ip->ino_data.size,
415                                           HAMMER_BUFSIZE);
416                         }
417                         break;
418                 }
419                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
420                 if (ip->ino_data.size < uio->uio_offset) {
421                         ip->ino_data.size = uio->uio_offset;
422                         flags = HAMMER_INODE_DDIRTY;
423                         vnode_pager_setsize(ap->a_vp, ip->ino_data.size);
424                 } else {
425                         flags = 0;
426                 }
427                 ip->ino_data.mtime = trans.time;
428                 flags |= HAMMER_INODE_ITIMES | HAMMER_INODE_BUFS;
429                 flags |= HAMMER_INODE_DDIRTY;   /* XXX mtime */
430                 hammer_modify_inode(ip, flags);
431
432                 /*
433                  * Try to keep track of cached dirty data.
434                  */
435                 if ((bp->b_flags & B_DIRTY) == 0) {
436                         ++ip->rsv_databufs;
437                         ++ip->hmp->rsv_databufs;
438                 }
439
440                 /*
441                  * Final buffer disposition.
442                  *
443                  * If write_mode is non-zero we call bawrite()
444                  * unconditionally.  Otherwise we only use bawrite()
445                  * if the writes are clearly sequential.
446                  */
447                 bp->b_flags |= B_AGE;
448                 if (ap->a_ioflag & IO_SYNC) {
449                         bwrite(bp);
450                 } else if (ap->a_ioflag & IO_DIRECT) {
451                         bawrite(bp);
452                 } else if (hammer_write_mode &&
453                            (uio->uio_offset & HAMMER_BUFMASK) == 0) {
454 #if 1
455                         /* strategy write cannot handled clustered writes */
456                         bp->b_flags |= B_CLUSTEROK;
457                         cluster_write(bp, ip->ino_data.size, seqcount);
458 #else
459 #endif
460                         bawrite(bp);
461                 } else if ((ap->a_ioflag >> 16) == IO_SEQMAX &&
462                            (uio->uio_offset & HAMMER_BUFMASK) == 0) {
463                         /*
464                          * If seqcount indicates sequential operation and
465                          * we just finished filling a buffer, push it out
466                          * now to prevent the buffer cache from becoming
467                          * too full, which would trigger non-optimal
468                          * flushes.
469                          */
470                         bawrite(bp);
471                 } else {
472                         bdwrite(bp);
473                 }
474         }
475         hammer_done_transaction(&trans);
476         return (error);
477 }
478
479 /*
480  * hammer_vop_access { vp, mode, cred }
481  */
482 static
483 int
484 hammer_vop_access(struct vop_access_args *ap)
485 {
486         struct hammer_inode *ip = VTOI(ap->a_vp);
487         uid_t uid;
488         gid_t gid;
489         int error;
490
491         uid = hammer_to_unix_xid(&ip->ino_data.uid);
492         gid = hammer_to_unix_xid(&ip->ino_data.gid);
493
494         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
495                                   ip->ino_data.uflags);
496         return (error);
497 }
498
499 /*
500  * hammer_vop_advlock { vp, id, op, fl, flags }
501  */
502 static
503 int
504 hammer_vop_advlock(struct vop_advlock_args *ap)
505 {
506         struct hammer_inode *ip = VTOI(ap->a_vp);
507
508         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
509 }
510
511 /*
512  * hammer_vop_close { vp, fflag }
513  */
514 static
515 int
516 hammer_vop_close(struct vop_close_args *ap)
517 {
518         return (vop_stdclose(ap));
519 }
520
521 /*
522  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
523  *
524  * The operating system has already ensured that the directory entry
525  * does not exist and done all appropriate namespace locking.
526  */
527 static
528 int
529 hammer_vop_ncreate(struct vop_ncreate_args *ap)
530 {
531         struct hammer_transaction trans;
532         struct hammer_inode *dip;
533         struct hammer_inode *nip;
534         struct nchandle *nch;
535         int error;
536
537         nch = ap->a_nch;
538         dip = VTOI(ap->a_dvp);
539
540         if (dip->flags & HAMMER_INODE_RO)
541                 return (EROFS);
542         if ((error = hammer_checkspace(dip->hmp)) != 0)
543                 return (error);
544
545         /*
546          * Create a transaction to cover the operations we perform.
547          */
548         hammer_start_transaction(&trans, dip->hmp);
549
550         /*
551          * Create a new filesystem object of the requested type.  The
552          * returned inode will be referenced and shared-locked to prevent
553          * it from being moved to the flusher.
554          */
555
556         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred, dip, &nip);
557         if (error) {
558                 hkprintf("hammer_create_inode error %d\n", error);
559                 hammer_done_transaction(&trans);
560                 *ap->a_vpp = NULL;
561                 return (error);
562         }
563
564         /*
565          * Add the new filesystem object to the directory.  This will also
566          * bump the inode's link count.
567          */
568         error = hammer_ip_add_directory(&trans, dip, nch->ncp, nip);
569         if (error)
570                 hkprintf("hammer_ip_add_directory error %d\n", error);
571
572         /*
573          * Finish up.
574          */
575         if (error) {
576                 hammer_rel_inode(nip, 0);
577                 hammer_done_transaction(&trans);
578                 *ap->a_vpp = NULL;
579         } else {
580                 error = hammer_get_vnode(nip, ap->a_vpp);
581                 hammer_done_transaction(&trans);
582                 hammer_rel_inode(nip, 0);
583                 if (error == 0) {
584                         cache_setunresolved(ap->a_nch);
585                         cache_setvp(ap->a_nch, *ap->a_vpp);
586                 }
587         }
588         return (error);
589 }
590
591 /*
592  * hammer_vop_getattr { vp, vap }
593  *
594  * Retrieve an inode's attribute information.  When accessing inodes
595  * historically we fake the atime field to ensure consistent results.
596  * The atime field is stored in the B-Tree element and allowed to be
597  * updated without cycling the element.
598  */
599 static
600 int
601 hammer_vop_getattr(struct vop_getattr_args *ap)
602 {
603         struct hammer_inode *ip = VTOI(ap->a_vp);
604         struct vattr *vap = ap->a_vap;
605
606 #if 0
607         if (cache_check_fsmid_vp(ap->a_vp, &ip->fsmid) &&
608             (vp->v_mount->mnt_flag & MNT_RDONLY) == 0 &&
609             ip->obj_asof == XXX
610         ) {
611                 /* LAZYMOD XXX */
612         }
613         hammer_itimes(ap->a_vp);
614 #endif
615
616         vap->va_fsid = ip->hmp->fsid_udev;
617         vap->va_fileid = ip->ino_leaf.base.obj_id;
618         vap->va_mode = ip->ino_data.mode;
619         vap->va_nlink = ip->ino_data.nlinks;
620         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
621         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
622         vap->va_rmajor = 0;
623         vap->va_rminor = 0;
624         vap->va_size = ip->ino_data.size;
625         if (ip->flags & HAMMER_INODE_RO)
626                 hammer_to_timespec(ip->ino_data.mtime, &vap->va_atime);
627         else
628                 hammer_to_timespec(ip->ino_leaf.atime, &vap->va_atime);
629         hammer_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
630         hammer_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
631         vap->va_flags = ip->ino_data.uflags;
632         vap->va_gen = 1;        /* hammer inums are unique for all time */
633         vap->va_blocksize = HAMMER_BUFSIZE;
634         vap->va_bytes = (ip->ino_data.size + 63) & ~63;
635         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
636         vap->va_filerev = 0;    /* XXX */
637         /* mtime uniquely identifies any adjustments made to the file */
638         vap->va_fsmid = ip->ino_data.mtime;
639         vap->va_uid_uuid = ip->ino_data.uid;
640         vap->va_gid_uuid = ip->ino_data.gid;
641         vap->va_fsid_uuid = ip->hmp->fsid;
642         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
643                           VA_FSID_UUID_VALID;
644
645         switch (ip->ino_data.obj_type) {
646         case HAMMER_OBJTYPE_CDEV:
647         case HAMMER_OBJTYPE_BDEV:
648                 vap->va_rmajor = ip->ino_data.rmajor;
649                 vap->va_rminor = ip->ino_data.rminor;
650                 break;
651         default:
652                 break;
653         }
654
655         return(0);
656 }
657
658 /*
659  * hammer_vop_nresolve { nch, dvp, cred }
660  *
661  * Locate the requested directory entry.
662  */
663 static
664 int
665 hammer_vop_nresolve(struct vop_nresolve_args *ap)
666 {
667         struct hammer_transaction trans;
668         struct namecache *ncp;
669         hammer_inode_t dip;
670         hammer_inode_t ip;
671         hammer_tid_t asof;
672         struct hammer_cursor cursor;
673         struct vnode *vp;
674         int64_t namekey;
675         int error;
676         int i;
677         int nlen;
678         int flags;
679         u_int64_t obj_id;
680
681         /*
682          * Misc initialization, plus handle as-of name extensions.  Look for
683          * the '@@' extension.  Note that as-of files and directories cannot
684          * be modified.
685          */
686         dip = VTOI(ap->a_dvp);
687         ncp = ap->a_nch->ncp;
688         asof = dip->obj_asof;
689         nlen = ncp->nc_nlen;
690         flags = dip->flags;
691
692         hammer_simple_transaction(&trans, dip->hmp);
693
694         for (i = 0; i < nlen; ++i) {
695                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
696                         asof = hammer_str_to_tid(ncp->nc_name + i + 2);
697                         flags |= HAMMER_INODE_RO;
698                         break;
699                 }
700         }
701         nlen = i;
702
703         /*
704          * If there is no path component the time extension is relative to
705          * dip.
706          */
707         if (nlen == 0) {
708                 ip = hammer_get_inode(&trans, &dip->cache[1], dip->obj_id,
709                                       asof, flags, &error);
710                 if (error == 0) {
711                         error = hammer_get_vnode(ip, &vp);
712                         hammer_rel_inode(ip, 0);
713                 } else {
714                         vp = NULL;
715                 }
716                 if (error == 0) {
717                         vn_unlock(vp);
718                         cache_setvp(ap->a_nch, vp);
719                         vrele(vp);
720                 }
721                 goto done;
722         }
723
724         /*
725          * Calculate the namekey and setup the key range for the scan.  This
726          * works kinda like a chained hash table where the lower 32 bits
727          * of the namekey synthesize the chain.
728          *
729          * The key range is inclusive of both key_beg and key_end.
730          */
731         namekey = hammer_directory_namekey(ncp->nc_name, nlen);
732
733         error = hammer_init_cursor(&trans, &cursor, &dip->cache[0], dip);
734         cursor.key_beg.localization = HAMMER_LOCALIZE_MISC;
735         cursor.key_beg.obj_id = dip->obj_id;
736         cursor.key_beg.key = namekey;
737         cursor.key_beg.create_tid = 0;
738         cursor.key_beg.delete_tid = 0;
739         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
740         cursor.key_beg.obj_type = 0;
741
742         cursor.key_end = cursor.key_beg;
743         cursor.key_end.key |= 0xFFFFFFFFULL;
744         cursor.asof = asof;
745         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
746
747         /*
748          * Scan all matching records (the chain), locate the one matching
749          * the requested path component.
750          *
751          * The hammer_ip_*() functions merge in-memory records with on-disk
752          * records for the purposes of the search.
753          */
754         obj_id = 0;
755
756         if (error == 0) {
757                 error = hammer_ip_first(&cursor);
758                 while (error == 0) {
759                         error = hammer_ip_resolve_data(&cursor);
760                         if (error)
761                                 break;
762                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
763                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
764                                 obj_id = cursor.data->entry.obj_id;
765                                 break;
766                         }
767                         error = hammer_ip_next(&cursor);
768                 }
769         }
770         hammer_done_cursor(&cursor);
771         if (error == 0) {
772                 ip = hammer_get_inode(&trans, &dip->cache[1],
773                                       obj_id, asof, flags, &error);
774                 if (error == 0) {
775                         error = hammer_get_vnode(ip, &vp);
776                         hammer_rel_inode(ip, 0);
777                 } else {
778                         vp = NULL;
779                 }
780                 if (error == 0) {
781                         vn_unlock(vp);
782                         cache_setvp(ap->a_nch, vp);
783                         vrele(vp);
784                 }
785         } else if (error == ENOENT) {
786                 cache_setvp(ap->a_nch, NULL);
787         }
788 done:
789         hammer_done_transaction(&trans);
790         return (error);
791 }
792
793 /*
794  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
795  *
796  * Locate the parent directory of a directory vnode.
797  *
798  * dvp is referenced but not locked.  *vpp must be returned referenced and
799  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
800  * at the root, instead it could indicate that the directory we were in was
801  * removed.
802  *
803  * NOTE: as-of sequences are not linked into the directory structure.  If
804  * we are at the root with a different asof then the mount point, reload
805  * the same directory with the mount point's asof.   I'm not sure what this
806  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
807  * get confused, but it hasn't been tested.
808  */
809 static
810 int
811 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
812 {
813         struct hammer_transaction trans;
814         struct hammer_inode *dip;
815         struct hammer_inode *ip;
816         int64_t parent_obj_id;
817         hammer_tid_t asof;
818         int error;
819
820         dip = VTOI(ap->a_dvp);
821         asof = dip->obj_asof;
822         parent_obj_id = dip->ino_data.parent_obj_id;
823
824         if (parent_obj_id == 0) {
825                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
826                    asof != dip->hmp->asof) {
827                         parent_obj_id = dip->obj_id;
828                         asof = dip->hmp->asof;
829                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
830                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
831                                    dip->obj_asof);
832                 } else {
833                         *ap->a_vpp = NULL;
834                         return ENOENT;
835                 }
836         }
837
838         hammer_simple_transaction(&trans, dip->hmp);
839
840         ip = hammer_get_inode(&trans, &dip->cache[1], parent_obj_id,
841                               asof, dip->flags, &error);
842         if (ip) {
843                 error = hammer_get_vnode(ip, ap->a_vpp);
844                 hammer_rel_inode(ip, 0);
845         } else {
846                 *ap->a_vpp = NULL;
847         }
848         hammer_done_transaction(&trans);
849         return (error);
850 }
851
852 /*
853  * hammer_vop_nlink { nch, dvp, vp, cred }
854  */
855 static
856 int
857 hammer_vop_nlink(struct vop_nlink_args *ap)
858 {
859         struct hammer_transaction trans;
860         struct hammer_inode *dip;
861         struct hammer_inode *ip;
862         struct nchandle *nch;
863         int error;
864
865         nch = ap->a_nch;
866         dip = VTOI(ap->a_dvp);
867         ip = VTOI(ap->a_vp);
868
869         if (dip->flags & HAMMER_INODE_RO)
870                 return (EROFS);
871         if (ip->flags & HAMMER_INODE_RO)
872                 return (EROFS);
873         if ((error = hammer_checkspace(dip->hmp)) != 0)
874                 return (error);
875
876         /*
877          * Create a transaction to cover the operations we perform.
878          */
879         hammer_start_transaction(&trans, dip->hmp);
880
881         /*
882          * Add the filesystem object to the directory.  Note that neither
883          * dip nor ip are referenced or locked, but their vnodes are
884          * referenced.  This function will bump the inode's link count.
885          */
886         error = hammer_ip_add_directory(&trans, dip, nch->ncp, ip);
887
888         /*
889          * Finish up.
890          */
891         if (error == 0) {
892                 cache_setunresolved(nch);
893                 cache_setvp(nch, ap->a_vp);
894         }
895         hammer_done_transaction(&trans);
896         return (error);
897 }
898
899 /*
900  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
901  *
902  * The operating system has already ensured that the directory entry
903  * does not exist and done all appropriate namespace locking.
904  */
905 static
906 int
907 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
908 {
909         struct hammer_transaction trans;
910         struct hammer_inode *dip;
911         struct hammer_inode *nip;
912         struct nchandle *nch;
913         int error;
914
915         nch = ap->a_nch;
916         dip = VTOI(ap->a_dvp);
917
918         if (dip->flags & HAMMER_INODE_RO)
919                 return (EROFS);
920         if ((error = hammer_checkspace(dip->hmp)) != 0)
921                 return (error);
922
923         /*
924          * Create a transaction to cover the operations we perform.
925          */
926         hammer_start_transaction(&trans, dip->hmp);
927
928         /*
929          * Create a new filesystem object of the requested type.  The
930          * returned inode will be referenced but not locked.
931          */
932         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred, dip, &nip);
933         if (error) {
934                 hkprintf("hammer_mkdir error %d\n", error);
935                 hammer_done_transaction(&trans);
936                 *ap->a_vpp = NULL;
937                 return (error);
938         }
939         /*
940          * Add the new filesystem object to the directory.  This will also
941          * bump the inode's link count.
942          */
943         error = hammer_ip_add_directory(&trans, dip, nch->ncp, nip);
944         if (error)
945                 hkprintf("hammer_mkdir (add) error %d\n", error);
946
947         /*
948          * Finish up.
949          */
950         if (error) {
951                 hammer_rel_inode(nip, 0);
952                 *ap->a_vpp = NULL;
953         } else {
954                 error = hammer_get_vnode(nip, ap->a_vpp);
955                 hammer_rel_inode(nip, 0);
956                 if (error == 0) {
957                         cache_setunresolved(ap->a_nch);
958                         cache_setvp(ap->a_nch, *ap->a_vpp);
959                 }
960         }
961         hammer_done_transaction(&trans);
962         return (error);
963 }
964
965 /*
966  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
967  *
968  * The operating system has already ensured that the directory entry
969  * does not exist and done all appropriate namespace locking.
970  */
971 static
972 int
973 hammer_vop_nmknod(struct vop_nmknod_args *ap)
974 {
975         struct hammer_transaction trans;
976         struct hammer_inode *dip;
977         struct hammer_inode *nip;
978         struct nchandle *nch;
979         int error;
980
981         nch = ap->a_nch;
982         dip = VTOI(ap->a_dvp);
983
984         if (dip->flags & HAMMER_INODE_RO)
985                 return (EROFS);
986         if ((error = hammer_checkspace(dip->hmp)) != 0)
987                 return (error);
988
989         /*
990          * Create a transaction to cover the operations we perform.
991          */
992         hammer_start_transaction(&trans, dip->hmp);
993
994         /*
995          * Create a new filesystem object of the requested type.  The
996          * returned inode will be referenced but not locked.
997          */
998         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred, dip, &nip);
999         if (error) {
1000                 hammer_done_transaction(&trans);
1001                 *ap->a_vpp = NULL;
1002                 return (error);
1003         }
1004
1005         /*
1006          * Add the new filesystem object to the directory.  This will also
1007          * bump the inode's link count.
1008          */
1009         error = hammer_ip_add_directory(&trans, dip, nch->ncp, nip);
1010
1011         /*
1012          * Finish up.
1013          */
1014         if (error) {
1015                 hammer_rel_inode(nip, 0);
1016                 *ap->a_vpp = NULL;
1017         } else {
1018                 error = hammer_get_vnode(nip, ap->a_vpp);
1019                 hammer_rel_inode(nip, 0);
1020                 if (error == 0) {
1021                         cache_setunresolved(ap->a_nch);
1022                         cache_setvp(ap->a_nch, *ap->a_vpp);
1023                 }
1024         }
1025         hammer_done_transaction(&trans);
1026         return (error);
1027 }
1028
1029 /*
1030  * hammer_vop_open { vp, mode, cred, fp }
1031  */
1032 static
1033 int
1034 hammer_vop_open(struct vop_open_args *ap)
1035 {
1036         hammer_inode_t ip;
1037
1038         ip = VTOI(ap->a_vp);
1039
1040         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1041                 return (EROFS);
1042         return(vop_stdopen(ap));
1043 }
1044
1045 /*
1046  * hammer_vop_pathconf { vp, name, retval }
1047  */
1048 static
1049 int
1050 hammer_vop_pathconf(struct vop_pathconf_args *ap)
1051 {
1052         return EOPNOTSUPP;
1053 }
1054
1055 /*
1056  * hammer_vop_print { vp }
1057  */
1058 static
1059 int
1060 hammer_vop_print(struct vop_print_args *ap)
1061 {
1062         return EOPNOTSUPP;
1063 }
1064
1065 /*
1066  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1067  */
1068 static
1069 int
1070 hammer_vop_readdir(struct vop_readdir_args *ap)
1071 {
1072         struct hammer_transaction trans;
1073         struct hammer_cursor cursor;
1074         struct hammer_inode *ip;
1075         struct uio *uio;
1076         hammer_base_elm_t base;
1077         int error;
1078         int cookie_index;
1079         int ncookies;
1080         off_t *cookies;
1081         off_t saveoff;
1082         int r;
1083
1084         ip = VTOI(ap->a_vp);
1085         uio = ap->a_uio;
1086         saveoff = uio->uio_offset;
1087
1088         if (ap->a_ncookies) {
1089                 ncookies = uio->uio_resid / 16 + 1;
1090                 if (ncookies > 1024)
1091                         ncookies = 1024;
1092                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1093                 cookie_index = 0;
1094         } else {
1095                 ncookies = -1;
1096                 cookies = NULL;
1097                 cookie_index = 0;
1098         }
1099
1100         hammer_simple_transaction(&trans, ip->hmp);
1101
1102         /*
1103          * Handle artificial entries
1104          */
1105         error = 0;
1106         if (saveoff == 0) {
1107                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1108                 if (r)
1109                         goto done;
1110                 if (cookies)
1111                         cookies[cookie_index] = saveoff;
1112                 ++saveoff;
1113                 ++cookie_index;
1114                 if (cookie_index == ncookies)
1115                         goto done;
1116         }
1117         if (saveoff == 1) {
1118                 if (ip->ino_data.parent_obj_id) {
1119                         r = vop_write_dirent(&error, uio,
1120                                              ip->ino_data.parent_obj_id,
1121                                              DT_DIR, 2, "..");
1122                 } else {
1123                         r = vop_write_dirent(&error, uio,
1124                                              ip->obj_id, DT_DIR, 2, "..");
1125                 }
1126                 if (r)
1127                         goto done;
1128                 if (cookies)
1129                         cookies[cookie_index] = saveoff;
1130                 ++saveoff;
1131                 ++cookie_index;
1132                 if (cookie_index == ncookies)
1133                         goto done;
1134         }
1135
1136         /*
1137          * Key range (begin and end inclusive) to scan.  Directory keys
1138          * directly translate to a 64 bit 'seek' position.
1139          */
1140         hammer_init_cursor(&trans, &cursor, &ip->cache[0], ip);
1141         cursor.key_beg.localization = HAMMER_LOCALIZE_MISC;
1142         cursor.key_beg.obj_id = ip->obj_id;
1143         cursor.key_beg.create_tid = 0;
1144         cursor.key_beg.delete_tid = 0;
1145         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1146         cursor.key_beg.obj_type = 0;
1147         cursor.key_beg.key = saveoff;
1148
1149         cursor.key_end = cursor.key_beg;
1150         cursor.key_end.key = HAMMER_MAX_KEY;
1151         cursor.asof = ip->obj_asof;
1152         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1153
1154         error = hammer_ip_first(&cursor);
1155
1156         while (error == 0) {
1157                 error = hammer_ip_resolve_data(&cursor);
1158                 if (error)
1159                         break;
1160                 base = &cursor.leaf->base;
1161                 saveoff = base->key;
1162                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1163
1164                 if (base->obj_id != ip->obj_id)
1165                         panic("readdir: bad record at %p", cursor.node);
1166
1167                 r = vop_write_dirent(
1168                              &error, uio, cursor.data->entry.obj_id,
1169                              hammer_get_dtype(cursor.leaf->base.obj_type),
1170                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1171                              (void *)cursor.data->entry.name);
1172                 if (r)
1173                         break;
1174                 ++saveoff;
1175                 if (cookies)
1176                         cookies[cookie_index] = base->key;
1177                 ++cookie_index;
1178                 if (cookie_index == ncookies)
1179                         break;
1180                 error = hammer_ip_next(&cursor);
1181         }
1182         hammer_done_cursor(&cursor);
1183
1184 done:
1185         hammer_done_transaction(&trans);
1186
1187         if (ap->a_eofflag)
1188                 *ap->a_eofflag = (error == ENOENT);
1189         uio->uio_offset = saveoff;
1190         if (error && cookie_index == 0) {
1191                 if (error == ENOENT)
1192                         error = 0;
1193                 if (cookies) {
1194                         kfree(cookies, M_TEMP);
1195                         *ap->a_ncookies = 0;
1196                         *ap->a_cookies = NULL;
1197                 }
1198         } else {
1199                 if (error == ENOENT)
1200                         error = 0;
1201                 if (cookies) {
1202                         *ap->a_ncookies = cookie_index;
1203                         *ap->a_cookies = cookies;
1204                 }
1205         }
1206         return(error);
1207 }
1208
1209 /*
1210  * hammer_vop_readlink { vp, uio, cred }
1211  */
1212 static
1213 int
1214 hammer_vop_readlink(struct vop_readlink_args *ap)
1215 {
1216         struct hammer_transaction trans;
1217         struct hammer_cursor cursor;
1218         struct hammer_inode *ip;
1219         int error;
1220
1221         ip = VTOI(ap->a_vp);
1222
1223         /*
1224          * Shortcut if the symlink data was stuffed into ino_data.
1225          */
1226         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1227                 error = uiomove(ip->ino_data.ext.symlink,
1228                                 ip->ino_data.size, ap->a_uio);
1229                 return(error);
1230         }
1231
1232         /*
1233          * Long version
1234          */
1235         hammer_simple_transaction(&trans, ip->hmp);
1236         hammer_init_cursor(&trans, &cursor, &ip->cache[0], ip);
1237
1238         /*
1239          * Key range (begin and end inclusive) to scan.  Directory keys
1240          * directly translate to a 64 bit 'seek' position.
1241          */
1242         cursor.key_beg.localization = HAMMER_LOCALIZE_MISC; /* XXX */
1243         cursor.key_beg.obj_id = ip->obj_id;
1244         cursor.key_beg.create_tid = 0;
1245         cursor.key_beg.delete_tid = 0;
1246         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1247         cursor.key_beg.obj_type = 0;
1248         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1249         cursor.asof = ip->obj_asof;
1250         cursor.flags |= HAMMER_CURSOR_ASOF;
1251
1252         error = hammer_ip_lookup(&cursor);
1253         if (error == 0) {
1254                 error = hammer_ip_resolve_data(&cursor);
1255                 if (error == 0) {
1256                         KKASSERT(cursor.leaf->data_len >=
1257                                  HAMMER_SYMLINK_NAME_OFF);
1258                         error = uiomove(cursor.data->symlink.name,
1259                                         cursor.leaf->data_len -
1260                                                 HAMMER_SYMLINK_NAME_OFF,
1261                                         ap->a_uio);
1262                 }
1263         }
1264         hammer_done_cursor(&cursor);
1265         hammer_done_transaction(&trans);
1266         return(error);
1267 }
1268
1269 /*
1270  * hammer_vop_nremove { nch, dvp, cred }
1271  */
1272 static
1273 int
1274 hammer_vop_nremove(struct vop_nremove_args *ap)
1275 {
1276         struct hammer_transaction trans;
1277         struct hammer_inode *dip;
1278         int error;
1279
1280         dip = VTOI(ap->a_dvp);
1281
1282         if (hammer_nohistory(dip) == 0 &&
1283             (error = hammer_checkspace(dip->hmp)) != 0) {
1284                 return (error);
1285         }
1286
1287         hammer_start_transaction(&trans, dip->hmp);
1288         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0);
1289         hammer_done_transaction(&trans);
1290
1291         return (error);
1292 }
1293
1294 /*
1295  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1296  */
1297 static
1298 int
1299 hammer_vop_nrename(struct vop_nrename_args *ap)
1300 {
1301         struct hammer_transaction trans;
1302         struct namecache *fncp;
1303         struct namecache *tncp;
1304         struct hammer_inode *fdip;
1305         struct hammer_inode *tdip;
1306         struct hammer_inode *ip;
1307         struct hammer_cursor cursor;
1308         int64_t namekey;
1309         int nlen, error;
1310
1311         fdip = VTOI(ap->a_fdvp);
1312         tdip = VTOI(ap->a_tdvp);
1313         fncp = ap->a_fnch->ncp;
1314         tncp = ap->a_tnch->ncp;
1315         ip = VTOI(fncp->nc_vp);
1316         KKASSERT(ip != NULL);
1317
1318         if (fdip->flags & HAMMER_INODE_RO)
1319                 return (EROFS);
1320         if (tdip->flags & HAMMER_INODE_RO)
1321                 return (EROFS);
1322         if (ip->flags & HAMMER_INODE_RO)
1323                 return (EROFS);
1324         if ((error = hammer_checkspace(fdip->hmp)) != 0)
1325                 return (error);
1326
1327         hammer_start_transaction(&trans, fdip->hmp);
1328
1329         /*
1330          * Remove tncp from the target directory and then link ip as
1331          * tncp. XXX pass trans to dounlink
1332          *
1333          * Force the inode sync-time to match the transaction so it is
1334          * in-sync with the creation of the target directory entry.
1335          */
1336         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp, ap->a_cred, 0);
1337         if (error == 0 || error == ENOENT) {
1338                 error = hammer_ip_add_directory(&trans, tdip, tncp, ip);
1339                 if (error == 0) {
1340                         ip->ino_data.parent_obj_id = tdip->obj_id;
1341                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
1342                 }
1343         }
1344         if (error)
1345                 goto failed; /* XXX */
1346
1347         /*
1348          * Locate the record in the originating directory and remove it.
1349          *
1350          * Calculate the namekey and setup the key range for the scan.  This
1351          * works kinda like a chained hash table where the lower 32 bits
1352          * of the namekey synthesize the chain.
1353          *
1354          * The key range is inclusive of both key_beg and key_end.
1355          */
1356         namekey = hammer_directory_namekey(fncp->nc_name, fncp->nc_nlen);
1357 retry:
1358         hammer_init_cursor(&trans, &cursor, &fdip->cache[0], fdip);
1359         cursor.key_beg.localization = HAMMER_LOCALIZE_MISC;
1360         cursor.key_beg.obj_id = fdip->obj_id;
1361         cursor.key_beg.key = namekey;
1362         cursor.key_beg.create_tid = 0;
1363         cursor.key_beg.delete_tid = 0;
1364         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1365         cursor.key_beg.obj_type = 0;
1366
1367         cursor.key_end = cursor.key_beg;
1368         cursor.key_end.key |= 0xFFFFFFFFULL;
1369         cursor.asof = fdip->obj_asof;
1370         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1371
1372         /*
1373          * Scan all matching records (the chain), locate the one matching
1374          * the requested path component.
1375          *
1376          * The hammer_ip_*() functions merge in-memory records with on-disk
1377          * records for the purposes of the search.
1378          */
1379         error = hammer_ip_first(&cursor);
1380         while (error == 0) {
1381                 if (hammer_ip_resolve_data(&cursor) != 0)
1382                         break;
1383                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
1384                 KKASSERT(nlen > 0);
1385                 if (fncp->nc_nlen == nlen &&
1386                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1387                         break;
1388                 }
1389                 error = hammer_ip_next(&cursor);
1390         }
1391
1392         /*
1393          * If all is ok we have to get the inode so we can adjust nlinks.
1394          *
1395          * WARNING: hammer_ip_del_directory() may have to terminate the
1396          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
1397          * twice.
1398          */
1399         if (error == 0)
1400                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
1401
1402         /*
1403          * XXX A deadlock here will break rename's atomicy for the purposes
1404          * of crash recovery.
1405          */
1406         if (error == EDEADLK) {
1407                 hammer_done_cursor(&cursor);
1408                 goto retry;
1409         }
1410
1411         /*
1412          * Cleanup and tell the kernel that the rename succeeded.
1413          */
1414         hammer_done_cursor(&cursor);
1415         if (error == 0)
1416                 cache_rename(ap->a_fnch, ap->a_tnch);
1417
1418 failed:
1419         hammer_done_transaction(&trans);
1420         return (error);
1421 }
1422
1423 /*
1424  * hammer_vop_nrmdir { nch, dvp, cred }
1425  */
1426 static
1427 int
1428 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
1429 {
1430         struct hammer_transaction trans;
1431         struct hammer_inode *dip;
1432         int error;
1433
1434         dip = VTOI(ap->a_dvp);
1435
1436         if (hammer_nohistory(dip) == 0 &&
1437             (error = hammer_checkspace(dip->hmp)) != 0) {
1438                 return (error);
1439         }
1440
1441         hammer_start_transaction(&trans, dip->hmp);
1442         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0);
1443         hammer_done_transaction(&trans);
1444
1445         return (error);
1446 }
1447
1448 /*
1449  * hammer_vop_setattr { vp, vap, cred }
1450  */
1451 static
1452 int
1453 hammer_vop_setattr(struct vop_setattr_args *ap)
1454 {
1455         struct hammer_transaction trans;
1456         struct vattr *vap;
1457         struct hammer_inode *ip;
1458         int modflags;
1459         int error;
1460         int truncating;
1461         off_t aligned_size;
1462         u_int32_t flags;
1463
1464         vap = ap->a_vap;
1465         ip = ap->a_vp->v_data;
1466         modflags = 0;
1467
1468         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
1469                 return(EROFS);
1470         if (ip->flags & HAMMER_INODE_RO)
1471                 return (EROFS);
1472         if (hammer_nohistory(ip) == 0 &&
1473             (error = hammer_checkspace(ip->hmp)) != 0) {
1474                 return (error);
1475         }
1476
1477         hammer_start_transaction(&trans, ip->hmp);
1478         error = 0;
1479
1480         if (vap->va_flags != VNOVAL) {
1481                 flags = ip->ino_data.uflags;
1482                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
1483                                          hammer_to_unix_xid(&ip->ino_data.uid),
1484                                          ap->a_cred);
1485                 if (error == 0) {
1486                         if (ip->ino_data.uflags != flags) {
1487                                 ip->ino_data.uflags = flags;
1488                                 modflags |= HAMMER_INODE_DDIRTY;
1489                         }
1490                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
1491                                 error = 0;
1492                                 goto done;
1493                         }
1494                 }
1495                 goto done;
1496         }
1497         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
1498                 error = EPERM;
1499                 goto done;
1500         }
1501         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
1502                 mode_t cur_mode = ip->ino_data.mode;
1503                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1504                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1505                 uuid_t uuid_uid;
1506                 uuid_t uuid_gid;
1507
1508                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
1509                                          ap->a_cred,
1510                                          &cur_uid, &cur_gid, &cur_mode);
1511                 if (error == 0) {
1512                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
1513                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
1514                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
1515                                  sizeof(uuid_uid)) ||
1516                             bcmp(&uuid_gid, &ip->ino_data.gid,
1517                                  sizeof(uuid_gid)) ||
1518                             ip->ino_data.mode != cur_mode
1519                         ) {
1520                                 ip->ino_data.uid = uuid_uid;
1521                                 ip->ino_data.gid = uuid_gid;
1522                                 ip->ino_data.mode = cur_mode;
1523                         }
1524                         modflags |= HAMMER_INODE_DDIRTY;
1525                 }
1526         }
1527         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
1528                 switch(ap->a_vp->v_type) {
1529                 case VREG:
1530                         if (vap->va_size == ip->ino_data.size)
1531                                 break;
1532                         /*
1533                          * XXX break atomicy, we can deadlock the backend
1534                          * if we do not release the lock.  Probably not a
1535                          * big deal here.
1536                          */
1537                         if (vap->va_size < ip->ino_data.size) {
1538                                 vtruncbuf(ap->a_vp, vap->va_size,
1539                                           HAMMER_BUFSIZE);
1540                                 truncating = 1;
1541                         } else {
1542                                 vnode_pager_setsize(ap->a_vp, vap->va_size);
1543                                 truncating = 0;
1544                         }
1545                         ip->ino_data.size = vap->va_size;
1546                         modflags |= HAMMER_INODE_DDIRTY;
1547                         aligned_size = (vap->va_size + HAMMER_BUFMASK) &
1548                                        ~HAMMER_BUFMASK64;
1549
1550                         /*
1551                          * on-media truncation is cached in the inode until
1552                          * the inode is synchronized.
1553                          */
1554                         if (truncating) {
1555                                 hammer_ip_frontend_trunc(ip, vap->va_size);
1556                                 hammer_update_rsv_databufs(ip);
1557 #ifdef DEBUG_TRUNCATE
1558                                 if (HammerTruncIp == NULL)
1559                                         HammerTruncIp = ip;
1560 #endif
1561                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
1562                                         ip->flags |= HAMMER_INODE_TRUNCATED;
1563                                         ip->trunc_off = vap->va_size;
1564 #ifdef DEBUG_TRUNCATE
1565                                         if (ip == HammerTruncIp)
1566                                         kprintf("truncate1 %016llx\n", ip->trunc_off);
1567 #endif
1568                                 } else if (ip->trunc_off > vap->va_size) {
1569                                         ip->trunc_off = vap->va_size;
1570 #ifdef DEBUG_TRUNCATE
1571                                         if (ip == HammerTruncIp)
1572                                         kprintf("truncate2 %016llx\n", ip->trunc_off);
1573 #endif
1574                                 } else {
1575 #ifdef DEBUG_TRUNCATE
1576                                         if (ip == HammerTruncIp)
1577                                         kprintf("truncate3 %016llx (ignored)\n", vap->va_size);
1578 #endif
1579                                 }
1580                         }
1581
1582                         /*
1583                          * If truncating we have to clean out a portion of
1584                          * the last block on-disk.  We do this in the
1585                          * front-end buffer cache.
1586                          */
1587                         if (truncating && vap->va_size < aligned_size) {
1588                                 struct buf *bp;
1589                                 int offset;
1590
1591                                 aligned_size -= HAMMER_BUFSIZE;
1592
1593                                 offset = vap->va_size & HAMMER_BUFMASK;
1594                                 error = bread(ap->a_vp, aligned_size,
1595                                               HAMMER_BUFSIZE, &bp);
1596                                 hammer_ip_frontend_trunc(ip, aligned_size);
1597                                 if (error == 0) {
1598                                         bzero(bp->b_data + offset,
1599                                               HAMMER_BUFSIZE - offset);
1600                                         bdwrite(bp);
1601                                 } else {
1602                                         kprintf("ERROR %d\n", error);
1603                                         brelse(bp);
1604                                 }
1605                         }
1606                         break;
1607                 case VDATABASE:
1608                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
1609                                 ip->flags |= HAMMER_INODE_TRUNCATED;
1610                                 ip->trunc_off = vap->va_size;
1611                         } else if (ip->trunc_off > vap->va_size) {
1612                                 ip->trunc_off = vap->va_size;
1613                         }
1614                         hammer_ip_frontend_trunc(ip, vap->va_size);
1615                         ip->ino_data.size = vap->va_size;
1616                         modflags |= HAMMER_INODE_DDIRTY;
1617                         break;
1618                 default:
1619                         error = EINVAL;
1620                         goto done;
1621                 }
1622                 break;
1623         }
1624         if (vap->va_atime.tv_sec != VNOVAL) {
1625                 ip->ino_leaf.atime =
1626                         hammer_timespec_to_transid(&vap->va_atime);
1627                 modflags |= HAMMER_INODE_ITIMES;
1628         }
1629         if (vap->va_mtime.tv_sec != VNOVAL) {
1630                 ip->ino_data.mtime =
1631                         hammer_timespec_to_transid(&vap->va_mtime);
1632                 modflags |= HAMMER_INODE_ITIMES;
1633                 modflags |= HAMMER_INODE_DDIRTY;        /* XXX mtime */
1634         }
1635         if (vap->va_mode != (mode_t)VNOVAL) {
1636                 mode_t   cur_mode = ip->ino_data.mode;
1637                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1638                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1639
1640                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
1641                                          cur_uid, cur_gid, &cur_mode);
1642                 if (error == 0 && ip->ino_data.mode != cur_mode) {
1643                         ip->ino_data.mode = cur_mode;
1644                         modflags |= HAMMER_INODE_DDIRTY;
1645                 }
1646         }
1647 done:
1648         if (error == 0)
1649                 hammer_modify_inode(ip, modflags);
1650         hammer_done_transaction(&trans);
1651         return (error);
1652 }
1653
1654 /*
1655  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
1656  */
1657 static
1658 int
1659 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
1660 {
1661         struct hammer_transaction trans;
1662         struct hammer_inode *dip;
1663         struct hammer_inode *nip;
1664         struct nchandle *nch;
1665         hammer_record_t record;
1666         int error;
1667         int bytes;
1668
1669         ap->a_vap->va_type = VLNK;
1670
1671         nch = ap->a_nch;
1672         dip = VTOI(ap->a_dvp);
1673
1674         if (dip->flags & HAMMER_INODE_RO)
1675                 return (EROFS);
1676         if ((error = hammer_checkspace(dip->hmp)) != 0)
1677                 return (error);
1678
1679         /*
1680          * Create a transaction to cover the operations we perform.
1681          */
1682         hammer_start_transaction(&trans, dip->hmp);
1683
1684         /*
1685          * Create a new filesystem object of the requested type.  The
1686          * returned inode will be referenced but not locked.
1687          */
1688
1689         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred, dip, &nip);
1690         if (error) {
1691                 hammer_done_transaction(&trans);
1692                 *ap->a_vpp = NULL;
1693                 return (error);
1694         }
1695
1696         /*
1697          * Add a record representing the symlink.  symlink stores the link
1698          * as pure data, not a string, and is no \0 terminated.
1699          */
1700         if (error == 0) {
1701                 bytes = strlen(ap->a_target);
1702
1703                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
1704                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
1705                 } else {
1706                         record = hammer_alloc_mem_record(nip, bytes);
1707                         record->type = HAMMER_MEM_RECORD_GENERAL;
1708
1709                         record->leaf.base.localization = HAMMER_LOCALIZE_MISC;
1710                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
1711                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
1712                         record->leaf.data_len = bytes;
1713                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
1714                         bcopy(ap->a_target, record->data->symlink.name, bytes);
1715                         error = hammer_ip_add_record(&trans, record);
1716                 }
1717
1718                 /*
1719                  * Set the file size to the length of the link.
1720                  */
1721                 if (error == 0) {
1722                         nip->ino_data.size = bytes;
1723                         hammer_modify_inode(nip, HAMMER_INODE_DDIRTY);
1724                 }
1725         }
1726         if (error == 0)
1727                 error = hammer_ip_add_directory(&trans, dip, nch->ncp, nip);
1728
1729         /*
1730          * Finish up.
1731          */
1732         if (error) {
1733                 hammer_rel_inode(nip, 0);
1734                 *ap->a_vpp = NULL;
1735         } else {
1736                 error = hammer_get_vnode(nip, ap->a_vpp);
1737                 hammer_rel_inode(nip, 0);
1738                 if (error == 0) {
1739                         cache_setunresolved(ap->a_nch);
1740                         cache_setvp(ap->a_nch, *ap->a_vpp);
1741                 }
1742         }
1743         hammer_done_transaction(&trans);
1744         return (error);
1745 }
1746
1747 /*
1748  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
1749  */
1750 static
1751 int
1752 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
1753 {
1754         struct hammer_transaction trans;
1755         struct hammer_inode *dip;
1756         int error;
1757
1758         dip = VTOI(ap->a_dvp);
1759
1760         if (hammer_nohistory(dip) == 0 &&
1761             (error = hammer_checkspace(dip->hmp)) != 0) {
1762                 return (error);
1763         }
1764
1765         hammer_start_transaction(&trans, dip->hmp);
1766         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
1767                                 ap->a_cred, ap->a_flags);
1768         hammer_done_transaction(&trans);
1769
1770         return (error);
1771 }
1772
1773 /*
1774  * hammer_vop_ioctl { vp, command, data, fflag, cred }
1775  */
1776 static
1777 int
1778 hammer_vop_ioctl(struct vop_ioctl_args *ap)
1779 {
1780         struct hammer_inode *ip = ap->a_vp->v_data;
1781
1782         return(hammer_ioctl(ip, ap->a_command, ap->a_data,
1783                             ap->a_fflag, ap->a_cred));
1784 }
1785
1786 static
1787 int
1788 hammer_vop_mountctl(struct vop_mountctl_args *ap)
1789 {
1790         struct mount *mp;
1791         int error;
1792
1793         mp = ap->a_head.a_ops->head.vv_mount;
1794
1795         switch(ap->a_op) {
1796         case MOUNTCTL_SET_EXPORT:
1797                 if (ap->a_ctllen != sizeof(struct export_args))
1798                         error = EINVAL;
1799                 error = hammer_vfs_export(mp, ap->a_op,
1800                                       (const struct export_args *)ap->a_ctl);
1801                 break;
1802         default:
1803                 error = journal_mountctl(ap);
1804                 break;
1805         }
1806         return(error);
1807 }
1808
1809 /*
1810  * hammer_vop_strategy { vp, bio }
1811  *
1812  * Strategy call, used for regular file read & write only.  Note that the
1813  * bp may represent a cluster.
1814  *
1815  * To simplify operation and allow better optimizations in the future,
1816  * this code does not make any assumptions with regards to buffer alignment
1817  * or size.
1818  */
1819 static
1820 int
1821 hammer_vop_strategy(struct vop_strategy_args *ap)
1822 {
1823         struct buf *bp;
1824         int error;
1825
1826         bp = ap->a_bio->bio_buf;
1827
1828         switch(bp->b_cmd) {
1829         case BUF_CMD_READ:
1830                 error = hammer_vop_strategy_read(ap);
1831                 break;
1832         case BUF_CMD_WRITE:
1833                 error = hammer_vop_strategy_write(ap);
1834                 break;
1835         default:
1836                 bp->b_error = error = EINVAL;
1837                 bp->b_flags |= B_ERROR;
1838                 biodone(ap->a_bio);
1839                 break;
1840         }
1841         return (error);
1842 }
1843
1844 /*
1845  * Read from a regular file.  Iterate the related records and fill in the
1846  * BIO/BUF.  Gaps are zero-filled.
1847  *
1848  * The support code in hammer_object.c should be used to deal with mixed
1849  * in-memory and on-disk records.
1850  *
1851  * XXX atime update
1852  */
1853 static
1854 int
1855 hammer_vop_strategy_read(struct vop_strategy_args *ap)
1856 {
1857         struct hammer_transaction trans;
1858         struct hammer_inode *ip;
1859         struct hammer_cursor cursor;
1860         hammer_base_elm_t base;
1861         struct bio *bio;
1862         struct bio *nbio;
1863         struct buf *bp;
1864         int64_t rec_offset;
1865         int64_t ran_end;
1866         int64_t tmp64;
1867         int error;
1868         int boff;
1869         int roff;
1870         int n;
1871
1872         bio = ap->a_bio;
1873         bp = bio->bio_buf;
1874         ip = ap->a_vp->v_data;
1875
1876         /*
1877          * The zone-2 disk offset may have been set by the cluster code via
1878          * a BMAP operation.  Take care not to confuse it with the bio_offset
1879          * set by hammer_io_direct_write(), which is a device-relative offset.
1880          *
1881          * Checking the high bits should suffice.
1882          */
1883         nbio = push_bio(bio);
1884         if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
1885             HAMMER_ZONE_RAW_BUFFER) {
1886                 error = hammer_io_direct_read(ip->hmp, nbio->bio_offset, bio);
1887                 return (error);
1888         }
1889
1890         /*
1891          * Hard way
1892          */
1893         hammer_simple_transaction(&trans, ip->hmp);
1894         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1895
1896         /*
1897          * Key range (begin and end inclusive) to scan.  Note that the key's
1898          * stored in the actual records represent BASE+LEN, not BASE.  The
1899          * first record containing bio_offset will have a key > bio_offset.
1900          */
1901         cursor.key_beg.localization = HAMMER_LOCALIZE_MISC;
1902         cursor.key_beg.obj_id = ip->obj_id;
1903         cursor.key_beg.create_tid = 0;
1904         cursor.key_beg.delete_tid = 0;
1905         cursor.key_beg.obj_type = 0;
1906         cursor.key_beg.key = bio->bio_offset + 1;
1907         cursor.asof = ip->obj_asof;
1908         cursor.flags |= HAMMER_CURSOR_ASOF;
1909
1910         cursor.key_end = cursor.key_beg;
1911         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
1912 #if 0
1913         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1914                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
1915                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
1916                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
1917         } else
1918 #endif
1919         {
1920                 ran_end = bio->bio_offset + bp->b_bufsize;
1921                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
1922                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
1923                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
1924                 if (tmp64 < ran_end)
1925                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
1926                 else
1927                         cursor.key_end.key = ran_end + MAXPHYS + 1;
1928         }
1929         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
1930
1931         error = hammer_ip_first(&cursor);
1932         boff = 0;
1933
1934         while (error == 0) {
1935                 /*
1936                  * Get the base file offset of the record.  The key for
1937                  * data records is (base + bytes) rather then (base).
1938                  */
1939                 base = &cursor.leaf->base;
1940                 rec_offset = base->key - cursor.leaf->data_len;
1941
1942                 /*
1943                  * Calculate the gap, if any, and zero-fill it.
1944                  *
1945                  * n is the offset of the start of the record verses our
1946                  * current seek offset in the bio.
1947                  */
1948                 n = (int)(rec_offset - (bio->bio_offset + boff));
1949                 if (n > 0) {
1950                         if (n > bp->b_bufsize - boff)
1951                                 n = bp->b_bufsize - boff;
1952                         bzero((char *)bp->b_data + boff, n);
1953                         boff += n;
1954                         n = 0;
1955                 }
1956
1957                 /*
1958                  * Calculate the data offset in the record and the number
1959                  * of bytes we can copy.
1960                  *
1961                  * There are two degenerate cases.  First, boff may already
1962                  * be at bp->b_bufsize.  Secondly, the data offset within
1963                  * the record may exceed the record's size.
1964                  */
1965                 roff = -n;
1966                 rec_offset += roff;
1967                 n = cursor.leaf->data_len - roff;
1968                 if (n <= 0) {
1969                         kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
1970                         n = 0;
1971                 } else if (n > bp->b_bufsize - boff) {
1972                         n = bp->b_bufsize - boff;
1973                 }
1974
1975                 /*
1976                  * Deal with cached truncations.  This cool bit of code
1977                  * allows truncate()/ftruncate() to avoid having to sync
1978                  * the file.
1979                  *
1980                  * If the frontend is truncated then all backend records are
1981                  * subject to the frontend's truncation.
1982                  *
1983                  * If the backend is truncated then backend records on-disk
1984                  * (but not in-memory) are subject to the backend's
1985                  * truncation.  In-memory records owned by the backend
1986                  * represent data written after the truncation point on the
1987                  * backend and must not be truncated.
1988                  *
1989                  * Truncate operations deal with frontend buffer cache
1990                  * buffers and frontend-owned in-memory records synchronously.
1991                  */
1992                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
1993                         if (hammer_cursor_ondisk(&cursor) ||
1994                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
1995                                 if (ip->trunc_off <= rec_offset)
1996                                         n = 0;
1997                                 else if (ip->trunc_off < rec_offset + n)
1998                                         n = (int)(ip->trunc_off - rec_offset);
1999                         }
2000                 }
2001                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2002                         if (hammer_cursor_ondisk(&cursor)) {
2003                                 if (ip->sync_trunc_off <= rec_offset)
2004                                         n = 0;
2005                                 else if (ip->sync_trunc_off < rec_offset + n)
2006                                         n = (int)(ip->sync_trunc_off - rec_offset);
2007                         }
2008                 }
2009
2010                 /*
2011                  * Try to issue a direct read into our bio if possible,
2012                  * otherwise resolve the element data into a hammer_buffer
2013                  * and copy.
2014                  */
2015                 if (n && boff == 0 &&
2016                     ((cursor.leaf->data_offset + roff) & HAMMER_BUFMASK) == 0) {
2017                         error = hammer_io_direct_read(
2018                                         trans.hmp,
2019                                         cursor.leaf->data_offset + roff,
2020                                         bio);
2021                         goto done;
2022                 } else if (n) {
2023                         error = hammer_ip_resolve_data(&cursor);
2024                         if (error == 0) {
2025                                 bcopy((char *)cursor.data + roff,
2026                                       (char *)bp->b_data + boff, n);
2027                         }
2028                 }
2029                 if (error)
2030                         break;
2031
2032                 /*
2033                  * Iterate until we have filled the request.
2034                  */
2035                 boff += n;
2036                 if (boff == bp->b_bufsize)
2037                         break;
2038                 error = hammer_ip_next(&cursor);
2039         }
2040
2041         /*
2042          * There may have been a gap after the last record
2043          */
2044         if (error == ENOENT)
2045                 error = 0;
2046         if (error == 0 && boff != bp->b_bufsize) {
2047                 KKASSERT(boff < bp->b_bufsize);
2048                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2049                 /* boff = bp->b_bufsize; */
2050         }
2051         bp->b_resid = 0;
2052         bp->b_error = error;
2053         if (error)
2054                 bp->b_flags |= B_ERROR;
2055         biodone(ap->a_bio);
2056
2057 done:
2058         if (cursor.node)
2059                 hammer_cache_node(cursor.node, &ip->cache[1]);
2060         hammer_done_cursor(&cursor);
2061         hammer_done_transaction(&trans);
2062         return(error);
2063 }
2064
2065 /*
2066  * BMAP operation - used to support cluster_read() only.
2067  *
2068  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2069  *
2070  * This routine may return EOPNOTSUPP if the opration is not supported for
2071  * the specified offset.  The contents of the pointer arguments do not
2072  * need to be initialized in that case. 
2073  *
2074  * If a disk address is available and properly aligned return 0 with 
2075  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2076  * to the run-length relative to that offset.  Callers may assume that
2077  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2078  * large, so return EOPNOTSUPP if it is not sufficiently large.
2079  */
2080 static
2081 int
2082 hammer_vop_bmap(struct vop_bmap_args *ap)
2083 {
2084         struct hammer_transaction trans;
2085         struct hammer_inode *ip;
2086         struct hammer_cursor cursor;
2087         hammer_base_elm_t base;
2088         int64_t rec_offset;
2089         int64_t ran_end;
2090         int64_t tmp64;
2091         int64_t base_offset;
2092         int64_t base_disk_offset;
2093         int64_t last_offset;
2094         hammer_off_t last_disk_offset;
2095         hammer_off_t disk_offset;
2096         int     rec_len;
2097         int     error;
2098
2099         ip = ap->a_vp->v_data;
2100
2101         /*
2102          * We can only BMAP regular files.  We can't BMAP database files,
2103          * directories, etc.
2104          */
2105         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
2106                 return(EOPNOTSUPP);
2107
2108         /*
2109          * bmap is typically called with runp/runb both NULL when used
2110          * for writing.  We do not support BMAP for writing atm.
2111          */
2112         if (ap->a_runp == NULL && ap->a_runb == NULL)
2113                 return(EOPNOTSUPP);
2114
2115         /*
2116          * Scan the B-Tree to acquire blockmap addresses, then translate
2117          * to raw addresses.
2118          */
2119         hammer_simple_transaction(&trans, ip->hmp);
2120 #if 0
2121         kprintf("bmap_beg %016llx ip->cache %p\n", ap->a_loffset, ip->cache[1]);
2122 #endif
2123         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2124
2125         /*
2126          * Key range (begin and end inclusive) to scan.  Note that the key's
2127          * stored in the actual records represent BASE+LEN, not BASE.  The
2128          * first record containing bio_offset will have a key > bio_offset.
2129          */
2130         cursor.key_beg.localization = HAMMER_LOCALIZE_MISC;
2131         cursor.key_beg.obj_id = ip->obj_id;
2132         cursor.key_beg.create_tid = 0;
2133         cursor.key_beg.delete_tid = 0;
2134         cursor.key_beg.obj_type = 0;
2135         if (ap->a_runb)
2136                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
2137         else
2138                 cursor.key_beg.key = ap->a_loffset + 1;
2139         if (cursor.key_beg.key < 0)
2140                 cursor.key_beg.key = 0;
2141         cursor.asof = ip->obj_asof;
2142         cursor.flags |= HAMMER_CURSOR_ASOF;
2143
2144         cursor.key_end = cursor.key_beg;
2145         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2146
2147         ran_end = ap->a_loffset + MAXPHYS;
2148         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2149         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2150         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2151         if (tmp64 < ran_end)
2152                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2153         else
2154                 cursor.key_end.key = ran_end + MAXPHYS + 1;
2155
2156         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2157
2158         error = hammer_ip_first(&cursor);
2159         base_offset = last_offset = 0;
2160         base_disk_offset = last_disk_offset = 0;
2161
2162         while (error == 0) {
2163                 /*
2164                  * Get the base file offset of the record.  The key for
2165                  * data records is (base + bytes) rather then (base).
2166                  */
2167                 base = &cursor.leaf->base;
2168                 rec_offset = base->key - cursor.leaf->data_len;
2169                 rec_len    = cursor.leaf->data_len;
2170
2171                 /*
2172                  * Incorporate any cached truncation
2173                  */
2174                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2175                         if (hammer_cursor_ondisk(&cursor) ||
2176                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2177                                 if (ip->trunc_off <= rec_offset)
2178                                         rec_len = 0;
2179                                 else if (ip->trunc_off < rec_offset + rec_len)
2180                                         rec_len = (int)(ip->trunc_off - rec_offset);
2181                         }
2182                 }
2183                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2184                         if (hammer_cursor_ondisk(&cursor)) {
2185                                 if (ip->sync_trunc_off <= rec_offset)
2186                                         rec_len = 0;
2187                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
2188                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
2189                         }
2190                 }
2191
2192                 /*
2193                  * Accumulate information.  If we have hit a discontiguous
2194                  * block reset base_offset unless we are already beyond the
2195                  * requested offset.  If we are, that's it, we stop.
2196                  */
2197                 disk_offset = hammer_blockmap_lookup(trans.hmp,
2198                                                      cursor.leaf->data_offset,
2199                                                      &error);
2200                 if (error)
2201                         break;
2202                 if (rec_offset != last_offset ||
2203                     disk_offset != last_disk_offset) {
2204                         if (rec_offset > ap->a_loffset)
2205                                 break;
2206                         base_offset = rec_offset;
2207                         base_disk_offset = disk_offset;
2208                 }
2209                 last_offset = rec_offset + rec_len;
2210                 last_disk_offset = disk_offset + rec_len;
2211
2212                 error = hammer_ip_next(&cursor);
2213         }
2214
2215 #if 0
2216         kprintf("BMAP %016llx:  %016llx - %016llx\n",
2217                 ap->a_loffset, base_offset, last_offset);
2218         kprintf("BMAP %16s:  %016llx - %016llx\n",
2219                 "", base_disk_offset, last_disk_offset);
2220 #endif
2221
2222         if (cursor.node) {
2223                 hammer_cache_node(cursor.node, &ip->cache[1]);
2224 #if 0
2225                 kprintf("bmap_end2 %016llx ip->cache %p\n", ap->a_loffset, ip->cache[1]);
2226 #endif
2227         }
2228         hammer_done_cursor(&cursor);
2229         hammer_done_transaction(&trans);
2230
2231         if (base_offset == 0 || base_offset > ap->a_loffset ||
2232             last_offset < ap->a_loffset) {
2233                 error = EOPNOTSUPP;
2234         } else {
2235                 disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
2236
2237                 /*
2238                  * If doffsetp is not aligned or the forward run size does
2239                  * not cover a whole buffer, disallow the direct I/O.
2240                  */
2241                 if ((disk_offset & HAMMER_BUFMASK) ||
2242                     (last_offset - ap->a_loffset) < HAMMER_BUFSIZE) {
2243                         error = EOPNOTSUPP;
2244                 } else {
2245                         *ap->a_doffsetp = disk_offset;
2246                         if (ap->a_runb)
2247                                 *ap->a_runb = ap->a_loffset - base_offset;
2248                         if (ap->a_runp)
2249                                 *ap->a_runp = last_offset - ap->a_loffset;
2250                         error = 0;
2251                 }
2252         }
2253         return(error);
2254 }
2255
2256 /*
2257  * Write to a regular file.   Because this is a strategy call the OS is
2258  * trying to actually sync data to the media.   HAMMER can only flush
2259  * the entire inode (so the TID remains properly synchronized).
2260  *
2261  * Basically all we do here is place the bio on the inode's flush queue
2262  * and activate the flusher.
2263  */
2264 static
2265 int
2266 hammer_vop_strategy_write(struct vop_strategy_args *ap)
2267 {
2268         hammer_record_t record;
2269         hammer_mount_t hmp;
2270         hammer_inode_t ip;
2271         struct bio *bio;
2272         struct buf *bp;
2273         int bytes;
2274         int error;
2275
2276         bio = ap->a_bio;
2277         bp = bio->bio_buf;
2278         ip = ap->a_vp->v_data;
2279         hmp = ip->hmp;
2280
2281         if (ip->flags & HAMMER_INODE_RO) {
2282                 bp->b_error = EROFS;
2283                 bp->b_flags |= B_ERROR;
2284                 biodone(ap->a_bio);
2285                 hammer_cleanup_write_io(ip);
2286                 return(EROFS);
2287         }
2288
2289         /*
2290          * Interlock with inode destruction (no in-kernel or directory
2291          * topology visibility).  If we queue new IO while trying to
2292          * destroy the inode we can deadlock the vtrunc call in
2293          * hammer_inode_unloadable_check().
2294          */
2295         if (ip->flags & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
2296                 bp->b_resid = 0;
2297                 biodone(ap->a_bio);
2298                 hammer_cleanup_write_io(ip);
2299                 return(0);
2300         }
2301
2302         /*
2303          * Reserve space and issue a direct-write from the front-end. 
2304          * NOTE: The direct_io code will hammer_bread/bcopy smaller
2305          * allocations.
2306          *
2307          * An in-memory record will be installed to reference the storage
2308          * until the flusher can get to it.
2309          *
2310          * Since we own the high level bio the front-end will not try to
2311          * do a direct-read until the write completes.
2312          *
2313          * NOTE: The only time we do not reserve a full-sized buffers
2314          * worth of data is if the file is small.  We do not try to
2315          * allocate a fragment (from the small-data zone) at the end of
2316          * an otherwise large file as this can lead to wildly separated
2317          * data.
2318          */
2319         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
2320         KKASSERT(bio->bio_offset < ip->ino_data.size);
2321         if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
2322                 bytes = (bp->b_bufsize + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
2323         else
2324                 bytes = ((int)ip->ino_data.size + 15) & ~15;
2325
2326         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
2327                                     bytes, &error);
2328         if (record) {
2329                 hammer_io_direct_write(hmp, &record->leaf, bio);
2330                 hammer_rel_mem_record(record);
2331                 if (hmp->rsv_recs > hammer_limit_recs &&
2332                     ip->rsv_recs > hammer_limit_irecs / 10) {
2333                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2334                 } else if (ip->rsv_recs > hammer_limit_irecs / 2) {
2335                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2336                 }
2337         } else {
2338                 bp->b_bio2.bio_offset = NOOFFSET;
2339                 bp->b_error = error;
2340                 bp->b_flags |= B_ERROR;
2341                 biodone(ap->a_bio);
2342         }
2343         hammer_cleanup_write_io(ip);
2344         return(error);
2345 }
2346
2347 /*
2348  * Clean-up after disposing of a dirty frontend buffer's data.
2349  * This is somewhat heuristical so try to be robust.
2350  */
2351 static void
2352 hammer_cleanup_write_io(hammer_inode_t ip)
2353 {
2354         if (ip->rsv_databufs) {
2355                 --ip->rsv_databufs;
2356                 --ip->hmp->rsv_databufs;
2357         }
2358 }
2359
2360 /*
2361  * We can lose track of dirty buffer cache buffers if we truncate, this
2362  * routine will resynchronize the count.
2363  */
2364 static
2365 void
2366 hammer_update_rsv_databufs(hammer_inode_t ip)
2367 {
2368         struct buf *bp;
2369         int delta;
2370         int n;
2371
2372         if (ip->vp) {
2373                 n = 0;
2374                 RB_FOREACH(bp, buf_rb_tree, &ip->vp->v_rbdirty_tree) {
2375                         ++n;
2376                 }
2377         } else {
2378                 n = 0;
2379         }
2380         delta = n - ip->rsv_databufs;
2381         ip->rsv_databufs += delta;
2382         ip->hmp->rsv_databufs += delta;
2383 }
2384
2385 /*
2386  * dounlink - disconnect a directory entry
2387  *
2388  * XXX whiteout support not really in yet
2389  */
2390 static int
2391 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
2392                 struct vnode *dvp, struct ucred *cred, int flags)
2393 {
2394         struct namecache *ncp;
2395         hammer_inode_t dip;
2396         hammer_inode_t ip;
2397         struct hammer_cursor cursor;
2398         int64_t namekey;
2399         int nlen, error;
2400
2401         /*
2402          * Calculate the namekey and setup the key range for the scan.  This
2403          * works kinda like a chained hash table where the lower 32 bits
2404          * of the namekey synthesize the chain.
2405          *
2406          * The key range is inclusive of both key_beg and key_end.
2407          */
2408         dip = VTOI(dvp);
2409         ncp = nch->ncp;
2410
2411         if (dip->flags & HAMMER_INODE_RO)
2412                 return (EROFS);
2413
2414         namekey = hammer_directory_namekey(ncp->nc_name, ncp->nc_nlen);
2415 retry:
2416         hammer_init_cursor(trans, &cursor, &dip->cache[0], dip);
2417         cursor.key_beg.localization = HAMMER_LOCALIZE_MISC;
2418         cursor.key_beg.obj_id = dip->obj_id;
2419         cursor.key_beg.key = namekey;
2420         cursor.key_beg.create_tid = 0;
2421         cursor.key_beg.delete_tid = 0;
2422         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2423         cursor.key_beg.obj_type = 0;
2424
2425         cursor.key_end = cursor.key_beg;
2426         cursor.key_end.key |= 0xFFFFFFFFULL;
2427         cursor.asof = dip->obj_asof;
2428         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2429
2430         /*
2431          * Scan all matching records (the chain), locate the one matching
2432          * the requested path component.  info->last_error contains the
2433          * error code on search termination and could be 0, ENOENT, or
2434          * something else.
2435          *
2436          * The hammer_ip_*() functions merge in-memory records with on-disk
2437          * records for the purposes of the search.
2438          */
2439         error = hammer_ip_first(&cursor);
2440
2441         while (error == 0) {
2442                 error = hammer_ip_resolve_data(&cursor);
2443                 if (error)
2444                         break;
2445                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2446                 KKASSERT(nlen > 0);
2447                 if (ncp->nc_nlen == nlen &&
2448                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2449                         break;
2450                 }
2451                 error = hammer_ip_next(&cursor);
2452         }
2453
2454         /*
2455          * If all is ok we have to get the inode so we can adjust nlinks.
2456          * To avoid a deadlock with the flusher we must release the inode
2457          * lock on the directory when acquiring the inode for the entry.
2458          *
2459          * If the target is a directory, it must be empty.
2460          */
2461         if (error == 0) {
2462                 hammer_unlock(&cursor.ip->lock);
2463                 ip = hammer_get_inode(trans, &dip->cache[1],
2464                                       cursor.data->entry.obj_id,
2465                                       dip->hmp->asof, 0, &error);
2466                 hammer_lock_sh(&cursor.ip->lock);
2467                 if (error == ENOENT) {
2468                         kprintf("obj_id %016llx\n", cursor.data->entry.obj_id);
2469                         Debugger("ENOENT unlinking object that should exist");
2470                 }
2471
2472                 /*
2473                  * If we are trying to remove a directory the directory must
2474                  * be empty.
2475                  *
2476                  * WARNING: hammer_ip_check_directory_empty() may have to
2477                  * terminate the cursor to avoid a deadlock.  It is ok to
2478                  * call hammer_done_cursor() twice.
2479                  */
2480                 if (error == 0 && ip->ino_data.obj_type ==
2481                                   HAMMER_OBJTYPE_DIRECTORY) {
2482                         error = hammer_ip_check_directory_empty(trans, ip);
2483                 }
2484
2485                 /*
2486                  * Delete the directory entry.
2487                  *
2488                  * WARNING: hammer_ip_del_directory() may have to terminate
2489                  * the cursor to avoid a deadlock.  It is ok to call
2490                  * hammer_done_cursor() twice.
2491                  */
2492                 if (error == 0) {
2493                         error = hammer_ip_del_directory(trans, &cursor,
2494                                                         dip, ip);
2495                 }
2496                 hammer_done_cursor(&cursor);
2497                 if (error == 0) {
2498                         cache_setunresolved(nch);
2499                         cache_setvp(nch, NULL);
2500                         /* XXX locking */
2501                         if (ip->vp)
2502                                 cache_inval_vp(ip->vp, CINV_DESTROY);
2503                 }
2504                 if (ip)
2505                         hammer_rel_inode(ip, 0);
2506         } else {
2507                 hammer_done_cursor(&cursor);
2508         }
2509         if (error == EDEADLK)
2510                 goto retry;
2511
2512         return (error);
2513 }
2514
2515 /************************************************************************
2516  *                          FIFO AND SPECFS OPS                         *
2517  ************************************************************************
2518  *
2519  */
2520
2521 static int
2522 hammer_vop_fifoclose (struct vop_close_args *ap)
2523 {
2524         /* XXX update itimes */
2525         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
2526 }
2527
2528 static int
2529 hammer_vop_fiforead (struct vop_read_args *ap)
2530 {
2531         int error;
2532
2533         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2534         /* XXX update access time */
2535         return (error);
2536 }
2537
2538 static int
2539 hammer_vop_fifowrite (struct vop_write_args *ap)
2540 {
2541         int error;
2542
2543         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2544         /* XXX update access time */
2545         return (error);
2546 }
2547
2548 static int
2549 hammer_vop_specclose (struct vop_close_args *ap)
2550 {
2551         /* XXX update itimes */
2552         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2553 }
2554
2555 static int
2556 hammer_vop_specread (struct vop_read_args *ap)
2557 {
2558         /* XXX update access time */
2559         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2560 }
2561
2562 static int
2563 hammer_vop_specwrite (struct vop_write_args *ap)
2564 {
2565         /* XXX update last change time */
2566         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2567 }
2568