xref: /dragonfly/sys/vfs/hammer/hammer_vnops.c (revision 0212bfce)
1 /*
2  * Copyright (c) 2007-2008 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  */
34 
35 #include <sys/mountctl.h>
36 #include <sys/namecache.h>
37 #include <sys/buf2.h>
38 #include <vfs/fifofs/fifo.h>
39 
40 #include "hammer.h"
41 
42 /*
43  * USERFS VNOPS
44  */
45 static int hammer_vop_fsync(struct vop_fsync_args *);
46 static int hammer_vop_read(struct vop_read_args *);
47 static int hammer_vop_write(struct vop_write_args *);
48 static int hammer_vop_access(struct vop_access_args *);
49 static int hammer_vop_advlock(struct vop_advlock_args *);
50 static int hammer_vop_close(struct vop_close_args *);
51 static int hammer_vop_ncreate(struct vop_ncreate_args *);
52 static int hammer_vop_getattr(struct vop_getattr_args *);
53 static int hammer_vop_nresolve(struct vop_nresolve_args *);
54 static int hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *);
55 static int hammer_vop_nlink(struct vop_nlink_args *);
56 static int hammer_vop_nmkdir(struct vop_nmkdir_args *);
57 static int hammer_vop_nmknod(struct vop_nmknod_args *);
58 static int hammer_vop_open(struct vop_open_args *);
59 static int hammer_vop_print(struct vop_print_args *);
60 static int hammer_vop_readdir(struct vop_readdir_args *);
61 static int hammer_vop_readlink(struct vop_readlink_args *);
62 static int hammer_vop_nremove(struct vop_nremove_args *);
63 static int hammer_vop_nrename(struct vop_nrename_args *);
64 static int hammer_vop_nrmdir(struct vop_nrmdir_args *);
65 static int hammer_vop_markatime(struct vop_markatime_args *);
66 static int hammer_vop_setattr(struct vop_setattr_args *);
67 static int hammer_vop_strategy(struct vop_strategy_args *);
68 static int hammer_vop_bmap(struct vop_bmap_args *ap);
69 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
70 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
71 static int hammer_vop_ioctl(struct vop_ioctl_args *);
72 static int hammer_vop_mountctl(struct vop_mountctl_args *);
73 static int hammer_vop_kqfilter (struct vop_kqfilter_args *);
74 
75 static int hammer_vop_fifoclose (struct vop_close_args *);
76 static int hammer_vop_fiforead (struct vop_read_args *);
77 static int hammer_vop_fifowrite (struct vop_write_args *);
78 static int hammer_vop_fifokqfilter (struct vop_kqfilter_args *);
79 
80 struct vop_ops hammer_vnode_vops = {
81 	.vop_default =		vop_defaultop,
82 	.vop_fsync =		hammer_vop_fsync,
83 	.vop_getpages =		vop_stdgetpages,
84 	.vop_putpages =		vop_stdputpages,
85 	.vop_read =		hammer_vop_read,
86 	.vop_write =		hammer_vop_write,
87 	.vop_access =		hammer_vop_access,
88 	.vop_advlock =		hammer_vop_advlock,
89 	.vop_close =		hammer_vop_close,
90 	.vop_ncreate =		hammer_vop_ncreate,
91 	.vop_getattr =		hammer_vop_getattr,
92 	.vop_inactive =		hammer_vop_inactive,
93 	.vop_reclaim =		hammer_vop_reclaim,
94 	.vop_nresolve =		hammer_vop_nresolve,
95 	.vop_nlookupdotdot =	hammer_vop_nlookupdotdot,
96 	.vop_nlink =		hammer_vop_nlink,
97 	.vop_nmkdir =		hammer_vop_nmkdir,
98 	.vop_nmknod =		hammer_vop_nmknod,
99 	.vop_open =		hammer_vop_open,
100 	.vop_pathconf =		vop_stdpathconf,
101 	.vop_print =		hammer_vop_print,
102 	.vop_readdir =		hammer_vop_readdir,
103 	.vop_readlink =		hammer_vop_readlink,
104 	.vop_nremove =		hammer_vop_nremove,
105 	.vop_nrename =		hammer_vop_nrename,
106 	.vop_nrmdir =		hammer_vop_nrmdir,
107 	.vop_markatime =	hammer_vop_markatime,
108 	.vop_setattr =		hammer_vop_setattr,
109 	.vop_bmap =		hammer_vop_bmap,
110 	.vop_strategy =		hammer_vop_strategy,
111 	.vop_nsymlink =		hammer_vop_nsymlink,
112 	.vop_nwhiteout =	hammer_vop_nwhiteout,
113 	.vop_ioctl =		hammer_vop_ioctl,
114 	.vop_mountctl =		hammer_vop_mountctl,
115 	.vop_kqfilter =		hammer_vop_kqfilter
116 };
117 
118 struct vop_ops hammer_spec_vops = {
119 	.vop_default =		vop_defaultop,
120 	.vop_fsync =		hammer_vop_fsync,
121 	.vop_read =		vop_stdnoread,
122 	.vop_write =		vop_stdnowrite,
123 	.vop_access =		hammer_vop_access,
124 	.vop_close =		hammer_vop_close,
125 	.vop_markatime =	hammer_vop_markatime,
126 	.vop_getattr =		hammer_vop_getattr,
127 	.vop_inactive =		hammer_vop_inactive,
128 	.vop_reclaim =		hammer_vop_reclaim,
129 	.vop_setattr =		hammer_vop_setattr
130 };
131 
132 struct vop_ops hammer_fifo_vops = {
133 	.vop_default =		fifo_vnoperate,
134 	.vop_fsync =		hammer_vop_fsync,
135 	.vop_read =		hammer_vop_fiforead,
136 	.vop_write =		hammer_vop_fifowrite,
137 	.vop_access =		hammer_vop_access,
138 	.vop_close =		hammer_vop_fifoclose,
139 	.vop_markatime =	hammer_vop_markatime,
140 	.vop_getattr =		hammer_vop_getattr,
141 	.vop_inactive =		hammer_vop_inactive,
142 	.vop_reclaim =		hammer_vop_reclaim,
143 	.vop_setattr =		hammer_vop_setattr,
144 	.vop_kqfilter =		hammer_vop_fifokqfilter
145 };
146 
147 static __inline
148 void
149 hammer_knote(struct vnode *vp, int flags)
150 {
151 	if (flags)
152 		KNOTE(&vp->v_pollinfo.vpi_kqinfo.ki_note, flags);
153 }
154 
155 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
156 			   struct vnode *dvp, struct ucred *cred,
157 			   int flags, int isdir);
158 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
159 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
160 
161 /*
162  * hammer_vop_fsync { vp, waitfor }
163  *
164  * fsync() an inode to disk and wait for it to be completely committed
165  * such that the information would not be undone if a crash occured after
166  * return.
167  *
168  * NOTE: HAMMER's fsync()'s are going to remain expensive until we implement
169  *	 a REDO log.  A sysctl is provided to relax HAMMER's fsync()
170  *	 operation.
171  *
172  *	 Ultimately the combination of a REDO log and use of fast storage
173  *	 to front-end cluster caches will make fsync fast, but it aint
174  *	 here yet.  And, in anycase, we need real transactional
175  *	 all-or-nothing features which are not restricted to a single file.
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 	hammer_mount_t hmp = ip->hmp;
183 	int waitfor = ap->a_waitfor;
184 	int mode;
185 
186 	lwkt_gettoken(&hmp->fs_token);
187 
188 	/*
189 	 * Fsync rule relaxation (default is either full synchronous flush
190 	 * or REDO semantics with synchronous flush).
191 	 */
192 	if (ap->a_flags & VOP_FSYNC_SYSCALL) {
193 		switch(hammer_fsync_mode) {
194 		case 0:
195 mode0:
196 			/* no REDO, full synchronous flush */
197 			goto skip;
198 		case 1:
199 mode1:
200 			/* no REDO, full asynchronous flush */
201 			if (waitfor == MNT_WAIT)
202 				waitfor = MNT_NOWAIT;
203 			goto skip;
204 		case 2:
205 			/* REDO semantics, synchronous flush */
206 			if (hmp->version < HAMMER_VOL_VERSION_FOUR)
207 				goto mode0;
208 			mode = HAMMER_FLUSH_UNDOS_AUTO;
209 			break;
210 		case 3:
211 			/* REDO semantics, relaxed asynchronous flush */
212 			if (hmp->version < HAMMER_VOL_VERSION_FOUR)
213 				goto mode1;
214 			mode = HAMMER_FLUSH_UNDOS_RELAXED;
215 			if (waitfor == MNT_WAIT)
216 				waitfor = MNT_NOWAIT;
217 			break;
218 		case 4:
219 			/* ignore the fsync() system call */
220 			lwkt_reltoken(&hmp->fs_token);
221 			return(0);
222 		default:
223 			/* we have to do something */
224 			mode = HAMMER_FLUSH_UNDOS_RELAXED;
225 			if (waitfor == MNT_WAIT)
226 				waitfor = MNT_NOWAIT;
227 			break;
228 		}
229 
230 		/*
231 		 * Fast fsync only needs to flush the UNDO/REDO fifo if
232 		 * HAMMER_INODE_REDO is non-zero and the only modifications
233 		 * made to the file are write or write-extends.
234 		 */
235 		if ((ip->flags & HAMMER_INODE_REDO) &&
236 		    (ip->flags & HAMMER_INODE_MODMASK_NOREDO) == 0) {
237 			++hammer_count_fsyncs;
238 			hammer_flusher_flush_undos(hmp, mode);
239 			ip->redo_count = 0;
240 			if (ip->vp && (ip->flags & HAMMER_INODE_MODMASK) == 0)
241 				vclrisdirty(ip->vp);
242 			lwkt_reltoken(&hmp->fs_token);
243 			return(0);
244 		}
245 
246 		/*
247 		 * REDO is enabled by fsync(), the idea being we really only
248 		 * want to lay down REDO records when programs are using
249 		 * fsync() heavily.  The first fsync() on the file starts
250 		 * the gravy train going and later fsync()s keep it hot by
251 		 * resetting the redo_count.
252 		 *
253 		 * We weren't running REDOs before now so we have to fall
254 		 * through and do a full fsync of what we have.
255 		 */
256 		if (hmp->version >= HAMMER_VOL_VERSION_FOUR &&
257 		    (hmp->flags & HAMMER_MOUNT_REDO_RECOVERY_RUN) == 0) {
258 			ip->flags |= HAMMER_INODE_REDO;
259 			ip->redo_count = 0;
260 		}
261 	}
262 skip:
263 
264 	/*
265 	 * Do a full flush sequence.
266 	 *
267 	 * Attempt to release the vnode while waiting for the inode to
268 	 * finish flushing.  This can really mess up inactive->reclaim
269 	 * sequences so only do it if the vnode is active.
270 	 *
271 	 * WARNING! The VX lock functions must be used.  vn_lock() will
272 	 *	    fail when this is part of a VOP_RECLAIM sequence.
273 	 */
274 	++hammer_count_fsyncs;
275 	vfsync(ap->a_vp, waitfor, 1, NULL, NULL);
276 	hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
277 	if (waitfor == MNT_WAIT) {
278 		int dorelock;
279 
280 		if ((ap->a_vp->v_flag & VRECLAIMED) == 0) {
281 			vx_unlock(ap->a_vp);
282 			dorelock = 1;
283 		} else {
284 			dorelock = 0;
285 		}
286 		hammer_wait_inode(ip);
287 		if (dorelock)
288 			vx_lock(ap->a_vp);
289 	}
290 	if (ip->vp && (ip->flags & HAMMER_INODE_MODMASK) == 0)
291 		vclrisdirty(ip->vp);
292 	lwkt_reltoken(&hmp->fs_token);
293 	return (ip->error);
294 }
295 
296 /*
297  * hammer_vop_read { vp, uio, ioflag, cred }
298  *
299  * MPSAFE (for the cache safe does not require fs_token)
300  */
301 static
302 int
303 hammer_vop_read(struct vop_read_args *ap)
304 {
305 	struct hammer_transaction trans;
306 	hammer_inode_t ip;
307 	hammer_mount_t hmp;
308 	off_t offset;
309 	struct buf *bp;
310 	struct uio *uio;
311 	int error;
312 	int n;
313 	int seqcount;
314 	int ioseqcount;
315 	int blksize;
316 	int bigread;
317 	int got_trans;
318 	size_t resid;
319 
320 	if (ap->a_vp->v_type != VREG)
321 		return (EINVAL);
322 	ip = VTOI(ap->a_vp);
323 	hmp = ip->hmp;
324 	error = 0;
325 	got_trans = 0;
326 	uio = ap->a_uio;
327 
328 	/*
329 	 * Attempt to shortcut directly to the VM object using lwbufs.
330 	 * This is much faster than instantiating buffer cache buffers.
331 	 */
332 	resid = uio->uio_resid;
333 	error = vop_helper_read_shortcut(ap);
334 	hammer_stats_file_read += resid - uio->uio_resid;
335 	if (error)
336 		return (error);
337 	if (uio->uio_resid == 0)
338 		goto finished;
339 
340 	/*
341 	 * Allow the UIO's size to override the sequential heuristic.
342 	 */
343 	blksize = hammer_blocksize(uio->uio_offset);
344 	seqcount = howmany(uio->uio_resid, MAXBSIZE);
345 	ioseqcount = ap->a_ioflag >> IO_SEQSHIFT;
346 	if (seqcount < ioseqcount)
347 		seqcount = ioseqcount;
348 
349 	/*
350 	 * If reading or writing a huge amount of data we have to break
351 	 * atomicy and allow the operation to be interrupted by a signal
352 	 * or it can DOS the machine.
353 	 */
354 	bigread = (uio->uio_resid > 100 * 1024 * 1024);
355 
356 	/*
357 	 * Access the data typically in HAMMER_BUFSIZE blocks via the
358 	 * buffer cache, but HAMMER may use a variable block size based
359 	 * on the offset.
360 	 *
361 	 * XXX Temporary hack, delay the start transaction while we remain
362 	 *     MPSAFE.  NOTE: ino_data.size cannot change while vnode is
363 	 *     locked-shared.
364 	 */
365 	while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
366 		int64_t base_offset;
367 		int64_t file_limit;
368 
369 		blksize = hammer_blocksize(uio->uio_offset);
370 		offset = (int)uio->uio_offset & (blksize - 1);
371 		base_offset = uio->uio_offset - offset;
372 
373 		if (bigread && (error = hammer_signal_check(ip->hmp)) != 0)
374 			break;
375 
376 		/*
377 		 * MPSAFE
378 		 */
379 		bp = getblk(ap->a_vp, base_offset, blksize, 0, 0);
380 		if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == B_CACHE) {
381 			bp->b_flags &= ~B_AGE;
382 			error = 0;
383 			goto skip;
384 		}
385 		if (ap->a_ioflag & IO_NRDELAY) {
386 			bqrelse(bp);
387 			return (EWOULDBLOCK);
388 		}
389 
390 		/*
391 		 * MPUNSAFE
392 		 */
393 		if (got_trans == 0) {
394 			hammer_start_transaction(&trans, ip->hmp);
395 			got_trans = 1;
396 		}
397 
398 		/*
399 		 * NOTE: A valid bp has already been acquired, but was not
400 		 *	 B_CACHE.
401 		 */
402 		if (hammer_cluster_enable) {
403 			/*
404 			 * Use file_limit to prevent cluster_read() from
405 			 * creating buffers of the wrong block size past
406 			 * the demarc.
407 			 */
408 			file_limit = ip->ino_data.size;
409 			if (base_offset < HAMMER_XDEMARC &&
410 			    file_limit > HAMMER_XDEMARC) {
411 				file_limit = HAMMER_XDEMARC;
412 			}
413 			error = cluster_readx(ap->a_vp,
414 					     file_limit, base_offset,
415 					     blksize, B_NOTMETA,
416 					     uio->uio_resid,
417 					     seqcount * MAXBSIZE,
418 					     &bp);
419 		} else {
420 			error = breadnx(ap->a_vp, base_offset,
421 					blksize, B_NOTMETA,
422 					NULL, NULL, 0, &bp);
423 		}
424 		if (error) {
425 			brelse(bp);
426 			break;
427 		}
428 skip:
429 		if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IOISSUED)) {
430 			hdkprintf("zone2_offset %016jx read file %016jx@%016jx\n",
431 				(intmax_t)bp->b_bio2.bio_offset,
432 				(intmax_t)ip->obj_id,
433 				(intmax_t)bp->b_loffset);
434 		}
435 		bp->b_flags &= ~B_IOISSUED;
436 		if (blksize == HAMMER_XBUFSIZE)
437 			bp->b_flags |= B_CLUSTEROK;
438 
439 		n = blksize - offset;
440 		if (n > uio->uio_resid)
441 			n = uio->uio_resid;
442 		if (n > ip->ino_data.size - uio->uio_offset)
443 			n = (int)(ip->ino_data.size - uio->uio_offset);
444 
445 		/*
446 		 * Set B_AGE, data has a lower priority than meta-data.
447 		 *
448 		 * Use a hold/unlock/drop sequence to run the uiomove
449 		 * with the buffer unlocked, avoiding deadlocks against
450 		 * read()s on mmap()'d spaces.
451 		 */
452 		bp->b_flags |= B_AGE;
453 		error = uiomovebp(bp, (char *)bp->b_data + offset, n, uio);
454 		bqrelse(bp);
455 
456 		if (error)
457 			break;
458 		hammer_stats_file_read += n;
459 	}
460 
461 finished:
462 
463 	/*
464 	 * Try to update the atime with just the inode lock for maximum
465 	 * concurrency.  If we can't shortcut it we have to get the full
466 	 * blown transaction.
467 	 */
468 	if (got_trans == 0 && hammer_update_atime_quick(ip) < 0) {
469 		hammer_start_transaction(&trans, ip->hmp);
470 		got_trans = 1;
471 	}
472 
473 	if (got_trans) {
474 		if ((ip->flags & HAMMER_INODE_RO) == 0 &&
475 		    (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
476 			lwkt_gettoken(&hmp->fs_token);
477 			ip->ino_data.atime = trans.time;
478 			hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
479 			hammer_done_transaction(&trans);
480 			lwkt_reltoken(&hmp->fs_token);
481 		} else {
482 			hammer_done_transaction(&trans);
483 		}
484 	}
485 	return (error);
486 }
487 
488 /*
489  * hammer_vop_write { vp, uio, ioflag, cred }
490  */
491 static
492 int
493 hammer_vop_write(struct vop_write_args *ap)
494 {
495 	struct hammer_transaction trans;
496 	hammer_inode_t ip;
497 	hammer_mount_t hmp;
498 	thread_t td;
499 	struct vnode *vp;
500 	struct uio *uio;
501 	int offset;
502 	off_t base_offset;
503 	int64_t cluster_eof;
504 	struct buf *bp;
505 	int kflags;
506 	int error;
507 	int n;
508 	int flags;
509 	int seqcount;
510 	int bigwrite;
511 
512 	vp = ap->a_vp;
513 	if (vp->v_type != VREG)
514 		return (EINVAL);
515 	ip = VTOI(ap->a_vp);
516 	hmp = ip->hmp;
517 	error = 0;
518 	kflags = 0;
519 	seqcount = ap->a_ioflag >> IO_SEQSHIFT;
520 
521 	if (ip->flags & HAMMER_INODE_RO)
522 		return (EROFS);
523 
524 	/*
525 	 * Create a transaction to cover the operations we perform.
526 	 */
527 	hammer_start_transaction(&trans, hmp);
528 	uio = ap->a_uio;
529 
530 	/*
531 	 * Use v_lastwrite_ts if file not open for writing
532 	 * (i.e. a late msync)
533 	 */
534 	if (uio->uio_segflg == UIO_NOCOPY) {
535 		if (vp->v_flag & VLASTWRITETS) {
536 			trans.time = vp->v_lastwrite_ts.tv_sec * 1000000 +
537 				     vp->v_lastwrite_ts.tv_nsec / 1000;
538 		} else {
539 			trans.time = ip->ino_data.mtime;
540 		}
541 	} else {
542 		vclrflags(vp, VLASTWRITETS);
543 	}
544 
545 	/*
546 	 * Check append mode
547 	 */
548 	if (ap->a_ioflag & IO_APPEND)
549 		uio->uio_offset = ip->ino_data.size;
550 
551 	/*
552 	 * Check for illegal write offsets.  Valid range is 0...2^63-1.
553 	 *
554 	 * NOTE: the base_off assignment is required to work around what
555 	 * I consider to be a GCC-4 optimization bug.
556 	 */
557 	if (uio->uio_offset < 0) {
558 		hammer_done_transaction(&trans);
559 		return (EFBIG);
560 	}
561 	base_offset = uio->uio_offset + uio->uio_resid;	/* work around gcc-4 */
562 	if (uio->uio_resid > 0 && base_offset <= uio->uio_offset) {
563 		hammer_done_transaction(&trans);
564 		return (EFBIG);
565 	}
566 
567 	if (uio->uio_resid > 0 && (td = uio->uio_td) != NULL && td->td_proc &&
568 	    base_offset > td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
569 		hammer_done_transaction(&trans);
570 		lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ);
571 		return (EFBIG);
572 	}
573 
574 	/*
575 	 * If reading or writing a huge amount of data we have to break
576 	 * atomicy and allow the operation to be interrupted by a signal
577 	 * or it can DOS the machine.
578 	 *
579 	 * Preset redo_count so we stop generating REDOs earlier if the
580 	 * limit is exceeded.
581 	 *
582 	 * redo_count is heuristical, SMP races are ok
583 	 */
584 	bigwrite = (uio->uio_resid > 100 * 1024 * 1024);
585 	if ((ip->flags & HAMMER_INODE_REDO) &&
586 	    ip->redo_count < hammer_limit_redo) {
587 		ip->redo_count += uio->uio_resid;
588 	}
589 
590 	/*
591 	 * Access the data typically in HAMMER_BUFSIZE blocks via the
592 	 * buffer cache, but HAMMER may use a variable block size based
593 	 * on the offset.
594 	 */
595 	while (uio->uio_resid > 0) {
596 		int fixsize = 0;
597 		int blksize;
598 		int blkmask;
599 		int trivial;
600 		int endofblk;
601 		off_t nsize;
602 
603 		if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
604 			break;
605 		if (bigwrite && (error = hammer_signal_check(hmp)) != 0)
606 			break;
607 
608 		blksize = hammer_blocksize(uio->uio_offset);
609 
610 		/*
611 		 * Control the number of pending records associated with
612 		 * this inode.  If too many have accumulated start a
613 		 * flush.  Try to maintain a pipeline with the flusher.
614 		 *
615 		 * NOTE: It is possible for other sources to grow the
616 		 *	 records but not necessarily issue another flush,
617 		 *	 so use a timeout and ensure that a re-flush occurs.
618 		 */
619 		if (ip->rsv_recs >= hammer_limit_inode_recs) {
620 			lwkt_gettoken(&hmp->fs_token);
621 			hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
622 			while (ip->rsv_recs >= hammer_limit_inode_recs * 2) {
623 				ip->flags |= HAMMER_INODE_RECSW;
624 				tsleep(&ip->rsv_recs, 0, "hmrwww", hz);
625 				hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
626 			}
627 			lwkt_reltoken(&hmp->fs_token);
628 		}
629 
630 		/*
631 		 * Do not allow HAMMER to blow out the buffer cache.  Very
632 		 * large UIOs can lockout other processes due to bwillwrite()
633 		 * mechanics.
634 		 *
635 		 * The hammer inode is not locked during these operations.
636 		 * The vnode is locked which can interfere with the pageout
637 		 * daemon for non-UIO_NOCOPY writes but should not interfere
638 		 * with the buffer cache.  Even so, we cannot afford to
639 		 * allow the pageout daemon to build up too many dirty buffer
640 		 * cache buffers.
641 		 *
642 		 * Only call this if we aren't being recursively called from
643 		 * a virtual disk device (vn), else we may deadlock.
644 		 */
645 		if ((ap->a_ioflag & IO_RECURSE) == 0)
646 			bwillwrite(blksize);
647 
648 		/*
649 		 * Calculate the blocksize at the current offset and figure
650 		 * out how much we can actually write.
651 		 */
652 		blkmask = blksize - 1;
653 		offset = (int)uio->uio_offset & blkmask;
654 		base_offset = uio->uio_offset & ~(int64_t)blkmask;
655 		n = blksize - offset;
656 		if (n > uio->uio_resid) {
657 			n = uio->uio_resid;
658 			endofblk = 0;
659 		} else {
660 			endofblk = 1;
661 		}
662 		nsize = uio->uio_offset + n;
663 		if (nsize > ip->ino_data.size) {
664 			if (uio->uio_offset > ip->ino_data.size)
665 				trivial = 0;
666 			else
667 				trivial = 1;
668 			nvextendbuf(ap->a_vp,
669 				    ip->ino_data.size,
670 				    nsize,
671 				    hammer_blocksize(ip->ino_data.size),
672 				    hammer_blocksize(nsize),
673 				    hammer_blockoff(ip->ino_data.size),
674 				    hammer_blockoff(nsize),
675 				    trivial);
676 			fixsize = 1;
677 			kflags |= NOTE_EXTEND;
678 		}
679 
680 		if (uio->uio_segflg == UIO_NOCOPY) {
681 			/*
682 			 * Issuing a write with the same data backing the
683 			 * buffer.  Instantiate the buffer to collect the
684 			 * backing vm pages, then read-in any missing bits.
685 			 *
686 			 * This case is used by vop_stdputpages().
687 			 */
688 			bp = getblk(ap->a_vp, base_offset,
689 				    blksize, GETBLK_BHEAVY, 0);
690 			if ((bp->b_flags & B_CACHE) == 0) {
691 				bqrelse(bp);
692 				error = bread(ap->a_vp, base_offset,
693 					      blksize, &bp);
694 			}
695 		} else if (offset == 0 && uio->uio_resid >= blksize) {
696 			/*
697 			 * Even though we are entirely overwriting the buffer
698 			 * we may still have to zero it out to avoid a
699 			 * mmap/write visibility issue.
700 			 */
701 			bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
702 			if ((bp->b_flags & B_CACHE) == 0)
703 				vfs_bio_clrbuf(bp);
704 		} else if (base_offset >= ip->ino_data.size) {
705 			/*
706 			 * If the base offset of the buffer is beyond the
707 			 * file EOF, we don't have to issue a read.
708 			 */
709 			bp = getblk(ap->a_vp, base_offset,
710 				    blksize, GETBLK_BHEAVY, 0);
711 			vfs_bio_clrbuf(bp);
712 		} else {
713 			/*
714 			 * Partial overwrite, read in any missing bits then
715 			 * replace the portion being written.
716 			 */
717 			error = bread(ap->a_vp, base_offset, blksize, &bp);
718 			if (error == 0)
719 				bheavy(bp);
720 		}
721 		if (error == 0)
722 			error = uiomovebp(bp, bp->b_data + offset, n, uio);
723 
724 		lwkt_gettoken(&hmp->fs_token);
725 
726 		/*
727 		 * Generate REDO records if enabled and redo_count will not
728 		 * exceeded the limit.
729 		 *
730 		 * If redo_count exceeds the limit we stop generating records
731 		 * and clear HAMMER_INODE_REDO.  This will cause the next
732 		 * fsync() to do a full meta-data sync instead of just an
733 		 * UNDO/REDO fifo update.
734 		 *
735 		 * When clearing HAMMER_INODE_REDO any pre-existing REDOs
736 		 * will still be tracked.  The tracks will be terminated
737 		 * when the related meta-data (including possible data
738 		 * modifications which are not tracked via REDO) is
739 		 * flushed.
740 		 */
741 		if ((ip->flags & HAMMER_INODE_REDO) && error == 0) {
742 			if (ip->redo_count < hammer_limit_redo) {
743 				bp->b_flags |= B_VFSFLAG1;
744 				error = hammer_generate_redo(&trans, ip,
745 						     base_offset + offset,
746 						     HAMMER_REDO_WRITE,
747 						     bp->b_data + offset,
748 						     (size_t)n);
749 			} else {
750 				ip->flags &= ~HAMMER_INODE_REDO;
751 			}
752 		}
753 
754 		/*
755 		 * If we screwed up we have to undo any VM size changes we
756 		 * made.
757 		 */
758 		if (error) {
759 			brelse(bp);
760 			if (fixsize) {
761 				nvtruncbuf(ap->a_vp, ip->ino_data.size,
762 					  hammer_blocksize(ip->ino_data.size),
763 					  hammer_blockoff(ip->ino_data.size),
764 					  0);
765 			}
766 			lwkt_reltoken(&hmp->fs_token);
767 			break;
768 		}
769 		kflags |= NOTE_WRITE;
770 		hammer_stats_file_write += n;
771 		if (blksize == HAMMER_XBUFSIZE)
772 			bp->b_flags |= B_CLUSTEROK;
773 		if (ip->ino_data.size < uio->uio_offset) {
774 			ip->ino_data.size = uio->uio_offset;
775 			flags = HAMMER_INODE_SDIRTY;
776 		} else {
777 			flags = 0;
778 		}
779 		ip->ino_data.mtime = trans.time;
780 		flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
781 		hammer_modify_inode(&trans, ip, flags);
782 
783 		/*
784 		 * Once we dirty the buffer any cached zone-X offset
785 		 * becomes invalid.  HAMMER NOTE: no-history mode cannot
786 		 * allow overwriting over the same data sector unless
787 		 * we provide UNDOs for the old data, which we don't.
788 		 */
789 		bp->b_bio2.bio_offset = NOOFFSET;
790 
791 		lwkt_reltoken(&hmp->fs_token);
792 
793 		/*
794 		 * Final buffer disposition.
795 		 *
796 		 * Because meta-data updates are deferred, HAMMER is
797 		 * especially sensitive to excessive bdwrite()s because
798 		 * the I/O stream is not broken up by disk reads.  So the
799 		 * buffer cache simply cannot keep up.
800 		 *
801 		 * WARNING!  blksize is variable.  cluster_write() is
802 		 *	     expected to not blow up if it encounters
803 		 *	     buffers that do not match the passed blksize.
804 		 *
805 		 * NOTE!  Hammer shouldn't need to bawrite()/cluster_write().
806 		 *	  The ip->rsv_recs check should burst-flush the data.
807 		 *	  If we queue it immediately the buf could be left
808 		 *	  locked on the device queue for a very long time.
809 		 *
810 		 *	  However, failing to flush a dirty buffer out when
811 		 *        issued from the pageout daemon can result in a low
812 		 *        memory deadlock against bio_page_alloc(), so we
813 		 *	  have to bawrite() on IO_ASYNC as well.
814 		 *
815 		 * NOTE!  To avoid degenerate stalls due to mismatched block
816 		 *	  sizes we only honor IO_DIRECT on the write which
817 		 *	  abuts the end of the buffer.  However, we must
818 		 *	  honor IO_SYNC in case someone is silly enough to
819 		 *	  configure a HAMMER file as swap, or when HAMMER
820 		 *	  is serving NFS (for commits).  Ick ick.
821 		 */
822 		bp->b_flags |= B_AGE;
823 		if (blksize == HAMMER_XBUFSIZE)
824 			bp->b_flags |= B_CLUSTEROK;
825 
826 		if (ap->a_ioflag & IO_SYNC) {
827 			bwrite(bp);
828 		} else if ((ap->a_ioflag & IO_DIRECT) && endofblk) {
829 			bawrite(bp);
830 		} else if (ap->a_ioflag & IO_ASYNC) {
831 			bawrite(bp);
832 		} else if (hammer_cluster_enable &&
833 			   !(ap->a_vp->v_mount->mnt_flag & MNT_NOCLUSTERW)) {
834 			if (base_offset < HAMMER_XDEMARC)
835 				cluster_eof = hammer_blockdemarc(base_offset,
836 							 ip->ino_data.size);
837 			else
838 				cluster_eof = ip->ino_data.size;
839 			cluster_write(bp, cluster_eof, blksize, seqcount);
840 		} else {
841 			bdwrite(bp);
842 		}
843 	}
844 	hammer_done_transaction(&trans);
845 	hammer_knote(ap->a_vp, kflags);
846 
847 	return (error);
848 }
849 
850 /*
851  * hammer_vop_access { vp, mode, cred }
852  *
853  * MPSAFE - does not require fs_token
854  */
855 static
856 int
857 hammer_vop_access(struct vop_access_args *ap)
858 {
859 	hammer_inode_t ip = VTOI(ap->a_vp);
860 	uid_t uid;
861 	gid_t gid;
862 	int error;
863 
864 	uid = hammer_to_unix_xid(&ip->ino_data.uid);
865 	gid = hammer_to_unix_xid(&ip->ino_data.gid);
866 
867 	error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
868 				  ip->ino_data.uflags);
869 	return (error);
870 }
871 
872 /*
873  * hammer_vop_advlock { vp, id, op, fl, flags }
874  *
875  * MPSAFE - does not require fs_token
876  */
877 static
878 int
879 hammer_vop_advlock(struct vop_advlock_args *ap)
880 {
881 	hammer_inode_t ip = VTOI(ap->a_vp);
882 
883 	return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
884 }
885 
886 /*
887  * hammer_vop_close { vp, fflag }
888  *
889  * We can only sync-on-close for normal closes.  XXX disabled for now.
890  */
891 static
892 int
893 hammer_vop_close(struct vop_close_args *ap)
894 {
895 #if 0
896 	struct vnode *vp = ap->a_vp;
897 	hammer_inode_t ip = VTOI(vp);
898 	int waitfor;
899 	if (ip->flags & (HAMMER_INODE_CLOSESYNC|HAMMER_INODE_CLOSEASYNC)) {
900 		if (vn_islocked(vp) == LK_EXCLUSIVE &&
901 		    (vp->v_flag & (VINACTIVE|VRECLAIMED)) == 0) {
902 			if (ip->flags & HAMMER_INODE_CLOSESYNC)
903 				waitfor = MNT_WAIT;
904 			else
905 				waitfor = MNT_NOWAIT;
906 			ip->flags &= ~(HAMMER_INODE_CLOSESYNC |
907 				       HAMMER_INODE_CLOSEASYNC);
908 			VOP_FSYNC(vp, MNT_NOWAIT, waitfor);
909 		}
910 	}
911 #endif
912 	return (vop_stdclose(ap));
913 }
914 
915 /*
916  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
917  *
918  * The operating system has already ensured that the directory entry
919  * does not exist and done all appropriate namespace locking.
920  */
921 static
922 int
923 hammer_vop_ncreate(struct vop_ncreate_args *ap)
924 {
925 	struct hammer_transaction trans;
926 	hammer_inode_t dip;
927 	hammer_inode_t nip;
928 	struct nchandle *nch;
929 	hammer_mount_t hmp;
930 	int error;
931 
932 	nch = ap->a_nch;
933 	dip = VTOI(ap->a_dvp);
934 	hmp = dip->hmp;
935 
936 	if (dip->flags & HAMMER_INODE_RO)
937 		return (EROFS);
938 	if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
939 		return (error);
940 
941 	/*
942 	 * Create a transaction to cover the operations we perform.
943 	 */
944 	lwkt_gettoken(&hmp->fs_token);
945 	hammer_start_transaction(&trans, hmp);
946 
947 	/*
948 	 * Create a new filesystem object of the requested type.  The
949 	 * returned inode will be referenced and shared-locked to prevent
950 	 * it from being moved to the flusher.
951 	 */
952 	error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
953 				    dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
954 				    NULL, &nip);
955 	if (error) {
956 		hkprintf("hammer_create_inode error %d\n", error);
957 		hammer_done_transaction(&trans);
958 		*ap->a_vpp = NULL;
959 		lwkt_reltoken(&hmp->fs_token);
960 		return (error);
961 	}
962 
963 	/*
964 	 * Add the new filesystem object to the directory.  This will also
965 	 * bump the inode's link count.
966 	 */
967 	error = hammer_ip_add_direntry(&trans, dip,
968 					nch->ncp->nc_name, nch->ncp->nc_nlen,
969 					nip);
970 	if (error)
971 		hkprintf("hammer_ip_add_direntry error %d\n", error);
972 
973 	/*
974 	 * Finish up.
975 	 */
976 	if (error) {
977 		hammer_rel_inode(nip, 0);
978 		hammer_done_transaction(&trans);
979 		*ap->a_vpp = NULL;
980 	} else {
981 		error = hammer_get_vnode(nip, ap->a_vpp);
982 		hammer_done_transaction(&trans);
983 		hammer_rel_inode(nip, 0);
984 		if (error == 0) {
985 			cache_setunresolved(ap->a_nch);
986 			cache_setvp(ap->a_nch, *ap->a_vpp);
987 		}
988 		hammer_knote(ap->a_dvp, NOTE_WRITE);
989 	}
990 	lwkt_reltoken(&hmp->fs_token);
991 	return (error);
992 }
993 
994 /*
995  * hammer_vop_getattr { vp, vap }
996  *
997  * Retrieve an inode's attribute information.  When accessing inodes
998  * historically we fake the atime field to ensure consistent results.
999  * The atime field is stored in the B-Tree element and allowed to be
1000  * updated without cycling the element.
1001  *
1002  * MPSAFE - does not require fs_token
1003  */
1004 static
1005 int
1006 hammer_vop_getattr(struct vop_getattr_args *ap)
1007 {
1008 	hammer_inode_t ip = VTOI(ap->a_vp);
1009 	struct vattr *vap = ap->a_vap;
1010 
1011 	/*
1012 	 * We want the fsid to be different when accessing a filesystem
1013 	 * with different as-of's so programs like diff don't think
1014 	 * the files are the same.
1015 	 *
1016 	 * We also want the fsid to be the same when comparing snapshots,
1017 	 * or when comparing mirrors (which might be backed by different
1018 	 * physical devices).  HAMMER fsids are based on the PFS's
1019 	 * shared_uuid field.
1020 	 *
1021 	 * XXX there is a chance of collision here.  The va_fsid reported
1022 	 * by stat is different from the more involved fsid used in the
1023 	 * mount structure.
1024 	 */
1025 	hammer_lock_sh(&ip->lock);
1026 	vap->va_fsid = ip->pfsm->fsid_udev ^ (uint32_t)ip->obj_asof ^
1027 		       (uint32_t)(ip->obj_asof >> 32);
1028 
1029 	vap->va_fileid = ip->ino_leaf.base.obj_id;
1030 	vap->va_mode = ip->ino_data.mode;
1031 	vap->va_nlink = ip->ino_data.nlinks;
1032 	vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1033 	vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1034 	vap->va_rmajor = 0;
1035 	vap->va_rminor = 0;
1036 	vap->va_size = ip->ino_data.size;
1037 
1038 	/*
1039 	 * Special case for @@PFS softlinks.  The actual size of the
1040 	 * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
1041 	 * or for MAX_TID is    "@@-1:%05d" == 10 bytes.
1042 	 *
1043 	 * Note that userspace hammer command does not allow users to
1044 	 * create a @@PFS softlink under an existing other PFS (id!=0)
1045 	 * so the ip localization here for @@PFS softlink is always 0.
1046 	 */
1047 	if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
1048 	    ip->ino_data.size == 10 &&
1049 	    ip->obj_asof == HAMMER_MAX_TID &&
1050 	    ip->obj_localization == HAMMER_DEF_LOCALIZATION &&
1051 	    strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
1052 		if (hammer_is_pfs_slave(&ip->pfsm->pfsd))
1053 			vap->va_size = 26;
1054 		else
1055 			vap->va_size = 10;
1056 	}
1057 
1058 	/*
1059 	 * We must provide a consistent atime and mtime for snapshots
1060 	 * so people can do a 'tar cf - ... | md5' on them and get
1061 	 * consistent results.
1062 	 */
1063 	if (ip->flags & HAMMER_INODE_RO) {
1064 		hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
1065 		hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
1066 	} else {
1067 		hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
1068 		hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
1069 	}
1070 	hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
1071 	vap->va_flags = ip->ino_data.uflags;
1072 	vap->va_gen = 1;	/* hammer inums are unique for all time */
1073 	vap->va_blocksize = HAMMER_BUFSIZE;
1074 	if (ip->ino_data.size >= HAMMER_XDEMARC) {
1075 		vap->va_bytes = HAMMER_XBUFSIZE64_DOALIGN(ip->ino_data.size);
1076 	} else if (ip->ino_data.size > HAMMER_HBUFSIZE) {
1077 		vap->va_bytes = HAMMER_BUFSIZE64_DOALIGN(ip->ino_data.size);
1078 	} else {
1079 		vap->va_bytes = HAMMER_DATA_DOALIGN(ip->ino_data.size);
1080 	}
1081 
1082 	vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
1083 	vap->va_filerev = 0;	/* XXX */
1084 	vap->va_uid_uuid = ip->ino_data.uid;
1085 	vap->va_gid_uuid = ip->ino_data.gid;
1086 	vap->va_fsid_uuid = ip->hmp->fsid;
1087 	vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
1088 			  VA_FSID_UUID_VALID;
1089 
1090 	switch (ip->ino_data.obj_type) {
1091 	case HAMMER_OBJTYPE_CDEV:
1092 	case HAMMER_OBJTYPE_BDEV:
1093 		vap->va_rmajor = ip->ino_data.rmajor;
1094 		vap->va_rminor = ip->ino_data.rminor;
1095 		break;
1096 	default:
1097 		break;
1098 	}
1099 	hammer_unlock(&ip->lock);
1100 	return(0);
1101 }
1102 
1103 /*
1104  * hammer_vop_nresolve { nch, dvp, cred }
1105  *
1106  * Locate the requested directory entry.
1107  */
1108 static
1109 int
1110 hammer_vop_nresolve(struct vop_nresolve_args *ap)
1111 {
1112 	struct hammer_transaction trans;
1113 	struct namecache *ncp;
1114 	hammer_mount_t hmp;
1115 	hammer_inode_t dip;
1116 	hammer_inode_t ip;
1117 	hammer_tid_t asof;
1118 	struct hammer_cursor cursor;
1119 	struct vnode *vp;
1120 	int64_t namekey;
1121 	int error;
1122 	int i;
1123 	int nlen;
1124 	int flags;
1125 	int ispfs;
1126 	int64_t obj_id;
1127 	uint32_t localization;
1128 	uint32_t max_iterations;
1129 
1130 	/*
1131 	 * Misc initialization, plus handle as-of name extensions.  Look for
1132 	 * the '@@' extension.  Note that as-of files and directories cannot
1133 	 * be modified.
1134 	 */
1135 	dip = VTOI(ap->a_dvp);
1136 	ncp = ap->a_nch->ncp;
1137 	asof = dip->obj_asof;
1138 	localization = dip->obj_localization;	/* for code consistency */
1139 	nlen = ncp->nc_nlen;
1140 	flags = dip->flags & HAMMER_INODE_RO;
1141 	ispfs = 0;
1142 	hmp = dip->hmp;
1143 
1144 	lwkt_gettoken(&hmp->fs_token);
1145 	hammer_simple_transaction(&trans, hmp);
1146 
1147 	for (i = 0; i < nlen; ++i) {
1148 		if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
1149 			error = hammer_str_to_tid(ncp->nc_name + i + 2,
1150 						  &ispfs, &asof, &localization);
1151 			if (error != 0) {
1152 				i = nlen;
1153 				break;
1154 			}
1155 			if (asof != HAMMER_MAX_TID)
1156 				flags |= HAMMER_INODE_RO;
1157 			break;
1158 		}
1159 	}
1160 	nlen = i;
1161 
1162 	/*
1163 	 * If this is a PFS we dive into the PFS root inode
1164 	 */
1165 	if (ispfs && nlen == 0) {
1166 		ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1167 				      asof, localization,
1168 				      flags, &error);
1169 		if (error == 0) {
1170 			error = hammer_get_vnode(ip, &vp);
1171 			hammer_rel_inode(ip, 0);
1172 		} else {
1173 			vp = NULL;
1174 		}
1175 		if (error == 0) {
1176 			vn_unlock(vp);
1177 			cache_setvp(ap->a_nch, vp);
1178 			vrele(vp);
1179 		}
1180 		goto done;
1181 	}
1182 
1183 	/*
1184 	 * If there is no path component the time extension is relative to dip.
1185 	 * e.g. "fubar/@@<snapshot>"
1186 	 *
1187 	 * "." is handled by the kernel, but ".@@<snapshot>" is not.
1188 	 * e.g. "fubar/.@@<snapshot>"
1189 	 *
1190 	 * ".." is handled by the kernel.  We do not currently handle
1191 	 * "..@<snapshot>".
1192 	 */
1193 	if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1194 		ip = hammer_get_inode(&trans, dip, dip->obj_id,
1195 				      asof, dip->obj_localization,
1196 				      flags, &error);
1197 		if (error == 0) {
1198 			error = hammer_get_vnode(ip, &vp);
1199 			hammer_rel_inode(ip, 0);
1200 		} else {
1201 			vp = NULL;
1202 		}
1203 		if (error == 0) {
1204 			vn_unlock(vp);
1205 			cache_setvp(ap->a_nch, vp);
1206 			vrele(vp);
1207 		}
1208 		goto done;
1209 	}
1210 
1211 	/*
1212 	 * Calculate the namekey and setup the key range for the scan.  This
1213 	 * works kinda like a chained hash table where the lower 32 bits
1214 	 * of the namekey synthesize the chain.
1215 	 *
1216 	 * The key range is inclusive of both key_beg and key_end.
1217 	 */
1218 	namekey = hammer_direntry_namekey(dip, ncp->nc_name, nlen,
1219 					   &max_iterations);
1220 
1221 	error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1222 	cursor.key_beg.localization = dip->obj_localization |
1223 				      hammer_dir_localization(dip);
1224         cursor.key_beg.obj_id = dip->obj_id;
1225 	cursor.key_beg.key = namekey;
1226         cursor.key_beg.create_tid = 0;
1227         cursor.key_beg.delete_tid = 0;
1228         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1229         cursor.key_beg.obj_type = 0;
1230 
1231 	cursor.key_end = cursor.key_beg;
1232 	cursor.key_end.key += max_iterations;
1233 	cursor.asof = asof;
1234 	cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1235 
1236 	/*
1237 	 * Scan all matching records (the chain), locate the one matching
1238 	 * the requested path component.
1239 	 *
1240 	 * The hammer_ip_*() functions merge in-memory records with on-disk
1241 	 * records for the purposes of the search.
1242 	 */
1243 	obj_id = 0;
1244 	localization = HAMMER_DEF_LOCALIZATION;
1245 
1246 	if (error == 0) {
1247 		error = hammer_ip_first(&cursor);
1248 		while (error == 0) {
1249 			error = hammer_ip_resolve_data(&cursor);
1250 			if (error)
1251 				break;
1252 			if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1253 			    bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1254 				obj_id = cursor.data->entry.obj_id;
1255 				localization = cursor.data->entry.localization;
1256 				break;
1257 			}
1258 			error = hammer_ip_next(&cursor);
1259 		}
1260 	}
1261 	hammer_done_cursor(&cursor);
1262 
1263 	/*
1264 	 * Lookup the obj_id.  This should always succeed.  If it does not
1265 	 * the filesystem may be damaged and we return a dummy inode.
1266 	 */
1267 	if (error == 0) {
1268 		ip = hammer_get_inode(&trans, dip, obj_id,
1269 				      asof, localization,
1270 				      flags, &error);
1271 		if (error == ENOENT) {
1272 			hkprintf("WARNING: Missing inode for dirent \"%s\"\n"
1273 				"\tobj_id = %016jx, asof=%016jx, lo=%08x\n",
1274 				ncp->nc_name,
1275 				(intmax_t)obj_id, (intmax_t)asof,
1276 				localization);
1277 			error = 0;
1278 			ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1279 						    asof, localization,
1280 						    flags, &error);
1281 		}
1282 		if (error == 0) {
1283 			error = hammer_get_vnode(ip, &vp);
1284 			hammer_rel_inode(ip, 0);
1285 		} else {
1286 			vp = NULL;
1287 		}
1288 		if (error == 0) {
1289 			vn_unlock(vp);
1290 			cache_setvp(ap->a_nch, vp);
1291 			vrele(vp);
1292 		}
1293 	} else if (error == ENOENT) {
1294 		cache_setvp(ap->a_nch, NULL);
1295 	}
1296 done:
1297 	hammer_done_transaction(&trans);
1298 	lwkt_reltoken(&hmp->fs_token);
1299 	return (error);
1300 }
1301 
1302 /*
1303  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1304  *
1305  * Locate the parent directory of a directory vnode.
1306  *
1307  * dvp is referenced but not locked.  *vpp must be returned referenced and
1308  * locked.  A parent_obj_id of 0 indicates that we are at the root.
1309  *
1310  * NOTE: as-of sequences are not linked into the directory structure.  If
1311  * we are at the root with a different asof then the mount point, reload
1312  * the same directory with the mount point's asof.   I'm not sure what this
1313  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
1314  * get confused, but it hasn't been tested.
1315  */
1316 static
1317 int
1318 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1319 {
1320 	struct hammer_transaction trans;
1321 	hammer_inode_t dip;
1322 	hammer_inode_t ip;
1323 	hammer_mount_t hmp;
1324 	int64_t parent_obj_id;
1325 	uint32_t parent_obj_localization;
1326 	hammer_tid_t asof;
1327 	int error;
1328 
1329 	dip = VTOI(ap->a_dvp);
1330 	asof = dip->obj_asof;
1331 	hmp = dip->hmp;
1332 
1333 	/*
1334 	 * Whos are parent?  This could be the root of a pseudo-filesystem
1335 	 * whos parent is in another localization domain.
1336 	 */
1337 	lwkt_gettoken(&hmp->fs_token);
1338 	parent_obj_id = dip->ino_data.parent_obj_id;
1339 	if (dip->obj_id == HAMMER_OBJID_ROOT)
1340 		parent_obj_localization = HAMMER_DEF_LOCALIZATION;
1341 	else
1342 		parent_obj_localization = dip->obj_localization;
1343 
1344 	/*
1345 	 * It's probably a PFS root when dip->ino_data.parent_obj_id is 0.
1346 	 */
1347 	if (parent_obj_id == 0) {
1348 		if (dip->obj_id == HAMMER_OBJID_ROOT &&
1349 		   asof != hmp->asof) {
1350 			parent_obj_id = dip->obj_id;
1351 			asof = hmp->asof;
1352 			*ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1353 			ksnprintf(*ap->a_fakename, 19, "0x%016jx",
1354 				  (intmax_t)dip->obj_asof);
1355 		} else {
1356 			*ap->a_vpp = NULL;
1357 			lwkt_reltoken(&hmp->fs_token);
1358 			return ENOENT;
1359 		}
1360 	}
1361 
1362 	hammer_simple_transaction(&trans, hmp);
1363 
1364 	ip = hammer_get_inode(&trans, dip, parent_obj_id,
1365 			      asof, parent_obj_localization,
1366 			      dip->flags, &error);
1367 	if (ip) {
1368 		error = hammer_get_vnode(ip, ap->a_vpp);
1369 		hammer_rel_inode(ip, 0);
1370 	} else {
1371 		*ap->a_vpp = NULL;
1372 	}
1373 	hammer_done_transaction(&trans);
1374 	lwkt_reltoken(&hmp->fs_token);
1375 	return (error);
1376 }
1377 
1378 /*
1379  * hammer_vop_nlink { nch, dvp, vp, cred }
1380  */
1381 static
1382 int
1383 hammer_vop_nlink(struct vop_nlink_args *ap)
1384 {
1385 	struct hammer_transaction trans;
1386 	hammer_inode_t dip;
1387 	hammer_inode_t ip;
1388 	struct nchandle *nch;
1389 	hammer_mount_t hmp;
1390 	int error;
1391 
1392 	if (ap->a_dvp->v_mount != ap->a_vp->v_mount)
1393 		return(EXDEV);
1394 
1395 	nch = ap->a_nch;
1396 	dip = VTOI(ap->a_dvp);
1397 	ip = VTOI(ap->a_vp);
1398 	hmp = dip->hmp;
1399 
1400 	if (dip->obj_localization != ip->obj_localization)
1401 		return(EXDEV);
1402 
1403 	if (dip->flags & HAMMER_INODE_RO)
1404 		return (EROFS);
1405 	if (ip->flags & HAMMER_INODE_RO)
1406 		return (EROFS);
1407 	if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1408 		return (error);
1409 
1410 	/*
1411 	 * Create a transaction to cover the operations we perform.
1412 	 */
1413 	lwkt_gettoken(&hmp->fs_token);
1414 	hammer_start_transaction(&trans, hmp);
1415 
1416 	/*
1417 	 * Add the filesystem object to the directory.  Note that neither
1418 	 * dip nor ip are referenced or locked, but their vnodes are
1419 	 * referenced.  This function will bump the inode's link count.
1420 	 */
1421 	error = hammer_ip_add_direntry(&trans, dip,
1422 					nch->ncp->nc_name, nch->ncp->nc_nlen,
1423 					ip);
1424 
1425 	/*
1426 	 * Finish up.
1427 	 */
1428 	if (error == 0) {
1429 		cache_setunresolved(nch);
1430 		cache_setvp(nch, ap->a_vp);
1431 	}
1432 	hammer_done_transaction(&trans);
1433 	hammer_knote(ap->a_vp, NOTE_LINK);
1434 	hammer_knote(ap->a_dvp, NOTE_WRITE);
1435 	lwkt_reltoken(&hmp->fs_token);
1436 	return (error);
1437 }
1438 
1439 /*
1440  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1441  *
1442  * The operating system has already ensured that the directory entry
1443  * does not exist and done all appropriate namespace locking.
1444  */
1445 static
1446 int
1447 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1448 {
1449 	struct hammer_transaction trans;
1450 	hammer_inode_t dip;
1451 	hammer_inode_t nip;
1452 	struct nchandle *nch;
1453 	hammer_mount_t hmp;
1454 	int error;
1455 
1456 	nch = ap->a_nch;
1457 	dip = VTOI(ap->a_dvp);
1458 	hmp = dip->hmp;
1459 
1460 	if (dip->flags & HAMMER_INODE_RO)
1461 		return (EROFS);
1462 	if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1463 		return (error);
1464 
1465 	/*
1466 	 * Create a transaction to cover the operations we perform.
1467 	 */
1468 	lwkt_gettoken(&hmp->fs_token);
1469 	hammer_start_transaction(&trans, hmp);
1470 
1471 	/*
1472 	 * Create a new filesystem object of the requested type.  The
1473 	 * returned inode will be referenced but not locked.
1474 	 */
1475 	error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1476 				    dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1477 				    NULL, &nip);
1478 	if (error) {
1479 		hammer_done_transaction(&trans);
1480 		*ap->a_vpp = NULL;
1481 		lwkt_reltoken(&hmp->fs_token);
1482 		return (error);
1483 	}
1484 	/*
1485 	 * Add the new filesystem object to the directory.  This will also
1486 	 * bump the inode's link count.
1487 	 */
1488 	error = hammer_ip_add_direntry(&trans, dip,
1489 					nch->ncp->nc_name, nch->ncp->nc_nlen,
1490 					nip);
1491 	if (error)
1492 		hkprintf("hammer_mkdir (add) error %d\n", error);
1493 
1494 	/*
1495 	 * Finish up.
1496 	 */
1497 	if (error) {
1498 		hammer_rel_inode(nip, 0);
1499 		*ap->a_vpp = NULL;
1500 	} else {
1501 		error = hammer_get_vnode(nip, ap->a_vpp);
1502 		hammer_rel_inode(nip, 0);
1503 		if (error == 0) {
1504 			cache_setunresolved(ap->a_nch);
1505 			cache_setvp(ap->a_nch, *ap->a_vpp);
1506 		}
1507 	}
1508 	hammer_done_transaction(&trans);
1509 	if (error == 0)
1510 		hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1511 	lwkt_reltoken(&hmp->fs_token);
1512 	return (error);
1513 }
1514 
1515 /*
1516  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1517  *
1518  * The operating system has already ensured that the directory entry
1519  * does not exist and done all appropriate namespace locking.
1520  */
1521 static
1522 int
1523 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1524 {
1525 	struct hammer_transaction trans;
1526 	hammer_inode_t dip;
1527 	hammer_inode_t nip;
1528 	struct nchandle *nch;
1529 	hammer_mount_t hmp;
1530 	int error;
1531 
1532 	nch = ap->a_nch;
1533 	dip = VTOI(ap->a_dvp);
1534 	hmp = dip->hmp;
1535 
1536 	if (dip->flags & HAMMER_INODE_RO)
1537 		return (EROFS);
1538 	if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1539 		return (error);
1540 
1541 	/*
1542 	 * Create a transaction to cover the operations we perform.
1543 	 */
1544 	lwkt_gettoken(&hmp->fs_token);
1545 	hammer_start_transaction(&trans, hmp);
1546 
1547 	/*
1548 	 * Create a new filesystem object of the requested type.  The
1549 	 * returned inode will be referenced but not locked.
1550 	 *
1551 	 * If mknod specifies a directory a pseudo-fs is created.
1552 	 */
1553 	error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1554 				    dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1555 				    NULL, &nip);
1556 	if (error) {
1557 		hammer_done_transaction(&trans);
1558 		*ap->a_vpp = NULL;
1559 		lwkt_reltoken(&hmp->fs_token);
1560 		return (error);
1561 	}
1562 
1563 	/*
1564 	 * Add the new filesystem object to the directory.  This will also
1565 	 * bump the inode's link count.
1566 	 */
1567 	error = hammer_ip_add_direntry(&trans, dip,
1568 					nch->ncp->nc_name, nch->ncp->nc_nlen,
1569 					nip);
1570 
1571 	/*
1572 	 * Finish up.
1573 	 */
1574 	if (error) {
1575 		hammer_rel_inode(nip, 0);
1576 		*ap->a_vpp = NULL;
1577 	} else {
1578 		error = hammer_get_vnode(nip, ap->a_vpp);
1579 		hammer_rel_inode(nip, 0);
1580 		if (error == 0) {
1581 			cache_setunresolved(ap->a_nch);
1582 			cache_setvp(ap->a_nch, *ap->a_vpp);
1583 		}
1584 	}
1585 	hammer_done_transaction(&trans);
1586 	if (error == 0)
1587 		hammer_knote(ap->a_dvp, NOTE_WRITE);
1588 	lwkt_reltoken(&hmp->fs_token);
1589 	return (error);
1590 }
1591 
1592 /*
1593  * hammer_vop_open { vp, mode, cred, fp }
1594  *
1595  * MPSAFE (does not require fs_token)
1596  */
1597 static
1598 int
1599 hammer_vop_open(struct vop_open_args *ap)
1600 {
1601 	hammer_inode_t ip;
1602 
1603 	ip = VTOI(ap->a_vp);
1604 
1605 	if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1606 		return (EROFS);
1607 	return(vop_stdopen(ap));
1608 }
1609 
1610 /*
1611  * hammer_vop_print { vp }
1612  */
1613 static
1614 int
1615 hammer_vop_print(struct vop_print_args *ap)
1616 {
1617 	return EOPNOTSUPP;
1618 }
1619 
1620 /*
1621  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1622  */
1623 static
1624 int
1625 hammer_vop_readdir(struct vop_readdir_args *ap)
1626 {
1627 	struct hammer_transaction trans;
1628 	struct hammer_cursor cursor;
1629 	hammer_inode_t ip;
1630 	hammer_mount_t hmp;
1631 	struct uio *uio;
1632 	hammer_base_elm_t base;
1633 	int error;
1634 	int cookie_index;
1635 	int ncookies;
1636 	off_t *cookies;
1637 	off_t saveoff;
1638 	int r;
1639 	int dtype;
1640 
1641 	ip = VTOI(ap->a_vp);
1642 	uio = ap->a_uio;
1643 	saveoff = uio->uio_offset;
1644 	hmp = ip->hmp;
1645 
1646 	if (ap->a_ncookies) {
1647 		ncookies = uio->uio_resid / 16 + 1;
1648 		if (ncookies > 1024)
1649 			ncookies = 1024;
1650 		cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1651 		cookie_index = 0;
1652 	} else {
1653 		ncookies = -1;
1654 		cookies = NULL;
1655 		cookie_index = 0;
1656 	}
1657 
1658 	lwkt_gettoken(&hmp->fs_token);
1659 	hammer_simple_transaction(&trans, hmp);
1660 
1661 	/*
1662 	 * Handle artificial entries
1663 	 *
1664 	 * It should be noted that the minimum value for a directory
1665 	 * hash key on-media is 0x0000000100000000, so we can use anything
1666 	 * less then that to represent our 'special' key space.
1667 	 */
1668 	error = 0;
1669 	if (saveoff == 0) {
1670 		r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1671 		if (r)
1672 			goto done;
1673 		if (cookies)
1674 			cookies[cookie_index] = saveoff;
1675 		++saveoff;
1676 		++cookie_index;
1677 		if (cookie_index == ncookies)
1678 			goto done;
1679 	}
1680 	if (saveoff == 1) {
1681 		if (ip->ino_data.parent_obj_id) {
1682 			r = vop_write_dirent(&error, uio,
1683 					     ip->ino_data.parent_obj_id,
1684 					     DT_DIR, 2, "..");
1685 		} else {
1686 			r = vop_write_dirent(&error, uio,
1687 					     ip->obj_id, DT_DIR, 2, "..");
1688 		}
1689 		if (r)
1690 			goto done;
1691 		if (cookies)
1692 			cookies[cookie_index] = saveoff;
1693 		++saveoff;
1694 		++cookie_index;
1695 		if (cookie_index == ncookies)
1696 			goto done;
1697 	}
1698 
1699 	/*
1700 	 * Key range (begin and end inclusive) to scan.  Directory keys
1701 	 * directly translate to a 64 bit 'seek' position.
1702 	 */
1703 	hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1704 	cursor.key_beg.localization = ip->obj_localization |
1705 				      hammer_dir_localization(ip);
1706 	cursor.key_beg.obj_id = ip->obj_id;
1707 	cursor.key_beg.create_tid = 0;
1708 	cursor.key_beg.delete_tid = 0;
1709         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1710 	cursor.key_beg.obj_type = 0;
1711 	cursor.key_beg.key = saveoff;
1712 
1713 	cursor.key_end = cursor.key_beg;
1714 	cursor.key_end.key = HAMMER_MAX_KEY;
1715 	cursor.asof = ip->obj_asof;
1716 	cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1717 
1718 	error = hammer_ip_first(&cursor);
1719 
1720 	while (error == 0) {
1721 		error = hammer_ip_resolve_data(&cursor);
1722 		if (error)
1723 			break;
1724 		base = &cursor.leaf->base;
1725 		saveoff = base->key;
1726 		KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1727 
1728 		if (base->obj_id != ip->obj_id)
1729 			hpanic("bad record at %p", cursor.node);
1730 
1731 		dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1732 		r = vop_write_dirent(
1733 			     &error, uio, cursor.data->entry.obj_id,
1734 			     dtype,
1735 			     cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1736 			     (void *)cursor.data->entry.name);
1737 		if (r)
1738 			break;
1739 		++saveoff;
1740 		if (cookies)
1741 			cookies[cookie_index] = base->key;
1742 		++cookie_index;
1743 		if (cookie_index == ncookies)
1744 			break;
1745 		error = hammer_ip_next(&cursor);
1746 	}
1747 	hammer_done_cursor(&cursor);
1748 
1749 done:
1750 	hammer_done_transaction(&trans);
1751 
1752 	if (ap->a_eofflag)
1753 		*ap->a_eofflag = (error == ENOENT);
1754 	uio->uio_offset = saveoff;
1755 	if (error && cookie_index == 0) {
1756 		if (error == ENOENT)
1757 			error = 0;
1758 		if (cookies) {
1759 			kfree(cookies, M_TEMP);
1760 			*ap->a_ncookies = 0;
1761 			*ap->a_cookies = NULL;
1762 		}
1763 	} else {
1764 		if (error == ENOENT)
1765 			error = 0;
1766 		if (cookies) {
1767 			*ap->a_ncookies = cookie_index;
1768 			*ap->a_cookies = cookies;
1769 		}
1770 	}
1771 	lwkt_reltoken(&hmp->fs_token);
1772 	return(error);
1773 }
1774 
1775 /*
1776  * hammer_vop_readlink { vp, uio, cred }
1777  */
1778 static
1779 int
1780 hammer_vop_readlink(struct vop_readlink_args *ap)
1781 {
1782 	struct hammer_transaction trans;
1783 	struct hammer_cursor cursor;
1784 	hammer_inode_t ip;
1785 	hammer_mount_t hmp;
1786 	char buf[32];
1787 	uint32_t localization;
1788 	hammer_pseudofs_inmem_t pfsm;
1789 	int error;
1790 
1791 	ip = VTOI(ap->a_vp);
1792 	hmp = ip->hmp;
1793 
1794 	lwkt_gettoken(&hmp->fs_token);
1795 
1796 	/*
1797 	 * Shortcut if the symlink data was stuffed into ino_data.
1798 	 *
1799 	 * Also expand special "@@PFS%05d" softlinks (expansion only
1800 	 * occurs for non-historical (current) accesses made from the
1801 	 * primary filesystem).
1802 	 *
1803 	 * Note that userspace hammer command does not allow users to
1804 	 * create a @@PFS softlink under an existing other PFS (id!=0)
1805 	 * so the ip localization here for @@PFS softlink is always 0.
1806 	 */
1807 	if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1808 		char *ptr;
1809 		int bytes;
1810 
1811 		ptr = ip->ino_data.ext.symlink;
1812 		bytes = (int)ip->ino_data.size;
1813 		if (bytes == 10 &&
1814 		    ip->obj_asof == HAMMER_MAX_TID &&
1815 		    ip->obj_localization == HAMMER_DEF_LOCALIZATION &&
1816 		    strncmp(ptr, "@@PFS", 5) == 0) {
1817 			hammer_simple_transaction(&trans, hmp);
1818 			bcopy(ptr + 5, buf, 5);
1819 			buf[5] = 0;
1820 			localization = pfs_to_lo(strtoul(buf, NULL, 10));
1821 			pfsm = hammer_load_pseudofs(&trans, localization,
1822 						    &error);
1823 			if (error == 0) {
1824 				if (hammer_is_pfs_slave(&pfsm->pfsd)) {
1825 					/* vap->va_size == 26 */
1826 					ksnprintf(buf, sizeof(buf),
1827 						  "@@0x%016jx:%05d",
1828 						  (intmax_t)pfsm->pfsd.sync_end_tid,
1829 						  lo_to_pfs(localization));
1830 				} else {
1831 					/* vap->va_size == 10 */
1832 					ksnprintf(buf, sizeof(buf),
1833 						  "@@-1:%05d",
1834 						  lo_to_pfs(localization));
1835 				}
1836 				ptr = buf;
1837 				bytes = strlen(buf);
1838 			}
1839 			if (pfsm)
1840 				hammer_rel_pseudofs(hmp, pfsm);
1841 			hammer_done_transaction(&trans);
1842 		}
1843 		error = uiomove(ptr, bytes, ap->a_uio);
1844 		lwkt_reltoken(&hmp->fs_token);
1845 		return(error);
1846 	}
1847 
1848 	/*
1849 	 * Long version
1850 	 */
1851 	hammer_simple_transaction(&trans, hmp);
1852 	hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1853 
1854 	/*
1855 	 * Key range (begin and end inclusive) to scan.  Directory keys
1856 	 * directly translate to a 64 bit 'seek' position.
1857 	 */
1858 	cursor.key_beg.localization = ip->obj_localization |
1859 				      HAMMER_LOCALIZE_MISC;
1860 	cursor.key_beg.obj_id = ip->obj_id;
1861 	cursor.key_beg.create_tid = 0;
1862 	cursor.key_beg.delete_tid = 0;
1863         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1864 	cursor.key_beg.obj_type = 0;
1865 	cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1866 	cursor.asof = ip->obj_asof;
1867 	cursor.flags |= HAMMER_CURSOR_ASOF;
1868 
1869 	error = hammer_ip_lookup(&cursor);
1870 	if (error == 0) {
1871 		error = hammer_ip_resolve_data(&cursor);
1872 		if (error == 0) {
1873 			KKASSERT(cursor.leaf->data_len >=
1874 				 HAMMER_SYMLINK_NAME_OFF);
1875 			error = uiomove(cursor.data->symlink.name,
1876 					cursor.leaf->data_len -
1877 						HAMMER_SYMLINK_NAME_OFF,
1878 					ap->a_uio);
1879 		}
1880 	}
1881 	hammer_done_cursor(&cursor);
1882 	hammer_done_transaction(&trans);
1883 	lwkt_reltoken(&hmp->fs_token);
1884 	return(error);
1885 }
1886 
1887 /*
1888  * hammer_vop_nremove { nch, dvp, cred }
1889  */
1890 static
1891 int
1892 hammer_vop_nremove(struct vop_nremove_args *ap)
1893 {
1894 	struct hammer_transaction trans;
1895 	hammer_inode_t dip;
1896 	hammer_mount_t hmp;
1897 	int error;
1898 
1899 	dip = VTOI(ap->a_dvp);
1900 	hmp = dip->hmp;
1901 
1902 	if (hammer_nohistory(dip) == 0 &&
1903 	    (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1904 		return (error);
1905 	}
1906 
1907 	lwkt_gettoken(&hmp->fs_token);
1908 	hammer_start_transaction(&trans, hmp);
1909 	error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1910 	hammer_done_transaction(&trans);
1911 	if (error == 0)
1912 		hammer_knote(ap->a_dvp, NOTE_WRITE);
1913 	lwkt_reltoken(&hmp->fs_token);
1914 	return (error);
1915 }
1916 
1917 /*
1918  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1919  */
1920 static
1921 int
1922 hammer_vop_nrename(struct vop_nrename_args *ap)
1923 {
1924 	struct hammer_transaction trans;
1925 	struct namecache *fncp;
1926 	struct namecache *tncp;
1927 	hammer_inode_t fdip;
1928 	hammer_inode_t tdip;
1929 	hammer_inode_t ip;
1930 	hammer_mount_t hmp;
1931 	struct hammer_cursor cursor;
1932 	int64_t namekey;
1933 	uint32_t max_iterations;
1934 	int nlen, error;
1935 
1936 	if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount)
1937 		return(EXDEV);
1938 	if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1939 		return(EXDEV);
1940 
1941 	fdip = VTOI(ap->a_fdvp);
1942 	tdip = VTOI(ap->a_tdvp);
1943 	fncp = ap->a_fnch->ncp;
1944 	tncp = ap->a_tnch->ncp;
1945 	ip = VTOI(fncp->nc_vp);
1946 	KKASSERT(ip != NULL);
1947 
1948 	hmp = ip->hmp;
1949 
1950 	if (fdip->obj_localization != tdip->obj_localization)
1951 		return(EXDEV);
1952 	if (fdip->obj_localization != ip->obj_localization)
1953 		return(EXDEV);
1954 
1955 	if (fdip->flags & HAMMER_INODE_RO)
1956 		return (EROFS);
1957 	if (tdip->flags & HAMMER_INODE_RO)
1958 		return (EROFS);
1959 	if (ip->flags & HAMMER_INODE_RO)
1960 		return (EROFS);
1961 	if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1962 		return (error);
1963 
1964 	lwkt_gettoken(&hmp->fs_token);
1965 	hammer_start_transaction(&trans, hmp);
1966 
1967 	/*
1968 	 * Remove tncp from the target directory and then link ip as
1969 	 * tncp. XXX pass trans to dounlink
1970 	 *
1971 	 * Force the inode sync-time to match the transaction so it is
1972 	 * in-sync with the creation of the target directory entry.
1973 	 */
1974 	error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
1975 				ap->a_cred, 0, -1);
1976 	if (error == 0 || error == ENOENT) {
1977 		error = hammer_ip_add_direntry(&trans, tdip,
1978 						tncp->nc_name, tncp->nc_nlen,
1979 						ip);
1980 		if (error == 0) {
1981 			ip->ino_data.parent_obj_id = tdip->obj_id;
1982 			ip->ino_data.ctime = trans.time;
1983 			hammer_modify_inode(&trans, ip, HAMMER_INODE_DDIRTY);
1984 		}
1985 	}
1986 	if (error)
1987 		goto failed; /* XXX */
1988 
1989 	/*
1990 	 * Locate the record in the originating directory and remove it.
1991 	 *
1992 	 * Calculate the namekey and setup the key range for the scan.  This
1993 	 * works kinda like a chained hash table where the lower 32 bits
1994 	 * of the namekey synthesize the chain.
1995 	 *
1996 	 * The key range is inclusive of both key_beg and key_end.
1997 	 */
1998 	namekey = hammer_direntry_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
1999 					   &max_iterations);
2000 retry:
2001 	hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
2002 	cursor.key_beg.localization = fdip->obj_localization |
2003 				      hammer_dir_localization(fdip);
2004         cursor.key_beg.obj_id = fdip->obj_id;
2005 	cursor.key_beg.key = namekey;
2006         cursor.key_beg.create_tid = 0;
2007         cursor.key_beg.delete_tid = 0;
2008         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2009         cursor.key_beg.obj_type = 0;
2010 
2011 	cursor.key_end = cursor.key_beg;
2012 	cursor.key_end.key += max_iterations;
2013 	cursor.asof = fdip->obj_asof;
2014 	cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2015 
2016 	/*
2017 	 * Scan all matching records (the chain), locate the one matching
2018 	 * the requested path component.
2019 	 *
2020 	 * The hammer_ip_*() functions merge in-memory records with on-disk
2021 	 * records for the purposes of the search.
2022 	 */
2023 	error = hammer_ip_first(&cursor);
2024 	while (error == 0) {
2025 		if (hammer_ip_resolve_data(&cursor) != 0)
2026 			break;
2027 		nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2028 		KKASSERT(nlen > 0);
2029 		if (fncp->nc_nlen == nlen &&
2030 		    bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2031 			break;
2032 		}
2033 		error = hammer_ip_next(&cursor);
2034 	}
2035 
2036 	/*
2037 	 * If all is ok we have to get the inode so we can adjust nlinks.
2038 	 *
2039 	 * WARNING: hammer_ip_del_direntry() may have to terminate the
2040 	 * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
2041 	 * twice.
2042 	 */
2043 	if (error == 0)
2044 		error = hammer_ip_del_direntry(&trans, &cursor, fdip, ip);
2045 
2046 	/*
2047 	 * XXX A deadlock here will break rename's atomicy for the purposes
2048 	 * of crash recovery.
2049 	 */
2050 	if (error == EDEADLK) {
2051 		hammer_done_cursor(&cursor);
2052 		goto retry;
2053 	}
2054 
2055 	/*
2056 	 * Cleanup and tell the kernel that the rename succeeded.
2057 	 *
2058 	 * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2059 	 *	 without formally acquiring the vp since the vp might
2060 	 *	 have zero refs on it, or in the middle of a reclaim,
2061 	 *	 etc.
2062 	 */
2063         hammer_done_cursor(&cursor);
2064 	if (error == 0) {
2065 		cache_rename(ap->a_fnch, ap->a_tnch);
2066 		hammer_knote(ap->a_fdvp, NOTE_WRITE);
2067 		hammer_knote(ap->a_tdvp, NOTE_WRITE);
2068 		while (ip->vp) {
2069 			struct vnode *vp;
2070 
2071 			error = hammer_get_vnode(ip, &vp);
2072 			if (error == 0 && vp) {
2073 				vn_unlock(vp);
2074 				hammer_knote(ip->vp, NOTE_RENAME);
2075 				vrele(vp);
2076 				break;
2077 			}
2078 			hdkprintf("ip/vp race2 avoided\n");
2079 		}
2080 	}
2081 
2082 failed:
2083 	hammer_done_transaction(&trans);
2084 	lwkt_reltoken(&hmp->fs_token);
2085 	return (error);
2086 }
2087 
2088 /*
2089  * hammer_vop_nrmdir { nch, dvp, cred }
2090  */
2091 static
2092 int
2093 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
2094 {
2095 	struct hammer_transaction trans;
2096 	hammer_inode_t dip;
2097 	hammer_mount_t hmp;
2098 	int error;
2099 
2100 	dip = VTOI(ap->a_dvp);
2101 	hmp = dip->hmp;
2102 
2103 	if (hammer_nohistory(dip) == 0 &&
2104 	    (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2105 		return (error);
2106 	}
2107 
2108 	lwkt_gettoken(&hmp->fs_token);
2109 	hammer_start_transaction(&trans, hmp);
2110 	error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
2111 	hammer_done_transaction(&trans);
2112 	if (error == 0)
2113 		hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
2114 	lwkt_reltoken(&hmp->fs_token);
2115 	return (error);
2116 }
2117 
2118 /*
2119  * hammer_vop_markatime { vp, cred }
2120  */
2121 static
2122 int
2123 hammer_vop_markatime(struct vop_markatime_args *ap)
2124 {
2125 	struct hammer_transaction trans;
2126 	hammer_inode_t ip;
2127 	hammer_mount_t hmp;
2128 
2129 	ip = VTOI(ap->a_vp);
2130 	if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2131 		return (EROFS);
2132 	if (ip->flags & HAMMER_INODE_RO)
2133 		return (EROFS);
2134 	hmp = ip->hmp;
2135 	if (hmp->mp->mnt_flag & MNT_NOATIME)
2136 		return (0);
2137 	lwkt_gettoken(&hmp->fs_token);
2138 	hammer_start_transaction(&trans, hmp);
2139 
2140 	ip->ino_data.atime = trans.time;
2141 	hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
2142 	hammer_done_transaction(&trans);
2143 	hammer_knote(ap->a_vp, NOTE_ATTRIB);
2144 	lwkt_reltoken(&hmp->fs_token);
2145 	return (0);
2146 }
2147 
2148 /*
2149  * hammer_vop_setattr { vp, vap, cred }
2150  */
2151 static
2152 int
2153 hammer_vop_setattr(struct vop_setattr_args *ap)
2154 {
2155 	struct hammer_transaction trans;
2156 	hammer_inode_t ip;
2157 	struct vattr *vap;
2158 	hammer_mount_t hmp;
2159 	int modflags;
2160 	int error;
2161 	int truncating;
2162 	int blksize;
2163 	int kflags;
2164 #if 0
2165 	int64_t aligned_size;
2166 #endif
2167 	uint32_t flags;
2168 
2169 	vap = ap->a_vap;
2170 	ip = ap->a_vp->v_data;
2171 	modflags = 0;
2172 	kflags = 0;
2173 	hmp = ip->hmp;
2174 
2175 	if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2176 		return(EROFS);
2177 	if (ip->flags & HAMMER_INODE_RO)
2178 		return (EROFS);
2179 	if (hammer_nohistory(ip) == 0 &&
2180 	    (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2181 		return (error);
2182 	}
2183 
2184 	lwkt_gettoken(&hmp->fs_token);
2185 	hammer_start_transaction(&trans, hmp);
2186 	error = 0;
2187 
2188 	if (vap->va_flags != VNOVAL) {
2189 		flags = ip->ino_data.uflags;
2190 		error = vop_helper_setattr_flags(&flags, vap->va_flags,
2191 					 hammer_to_unix_xid(&ip->ino_data.uid),
2192 					 ap->a_cred);
2193 		if (error == 0) {
2194 			if (ip->ino_data.uflags != flags) {
2195 				ip->ino_data.uflags = flags;
2196 				ip->ino_data.ctime = trans.time;
2197 				modflags |= HAMMER_INODE_DDIRTY;
2198 				kflags |= NOTE_ATTRIB;
2199 			}
2200 			if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2201 				error = 0;
2202 				goto done;
2203 			}
2204 		}
2205 		goto done;
2206 	}
2207 	if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2208 		error = EPERM;
2209 		goto done;
2210 	}
2211 	if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2212 		mode_t cur_mode = ip->ino_data.mode;
2213 		uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2214 		gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2215 		hammer_uuid_t uuid_uid;
2216 		hammer_uuid_t uuid_gid;
2217 
2218 		error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2219 					 ap->a_cred,
2220 					 &cur_uid, &cur_gid, &cur_mode);
2221 		if (error == 0) {
2222 			hammer_guid_to_uuid(&uuid_uid, cur_uid);
2223 			hammer_guid_to_uuid(&uuid_gid, cur_gid);
2224 			if (kuuid_compare(&uuid_uid, &ip->ino_data.uid) ||
2225 			    kuuid_compare(&uuid_gid, &ip->ino_data.gid) ||
2226 			    ip->ino_data.mode != cur_mode) {
2227 				ip->ino_data.uid = uuid_uid;
2228 				ip->ino_data.gid = uuid_gid;
2229 				ip->ino_data.mode = cur_mode;
2230 				ip->ino_data.ctime = trans.time;
2231 				modflags |= HAMMER_INODE_DDIRTY;
2232 			}
2233 			kflags |= NOTE_ATTRIB;
2234 		}
2235 	}
2236 	while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2237 		switch(ap->a_vp->v_type) {
2238 		case VREG:
2239 			if (vap->va_size == ip->ino_data.size)
2240 				break;
2241 
2242 			/*
2243 			 * Log the operation if in fast-fsync mode or if
2244 			 * there are unterminated redo write records present.
2245 			 *
2246 			 * The second check is needed so the recovery code
2247 			 * properly truncates write redos even if nominal
2248 			 * REDO operations is turned off due to excessive
2249 			 * writes, because the related records might be
2250 			 * destroyed and never lay down a TERM_WRITE.
2251 			 */
2252 			if ((ip->flags & HAMMER_INODE_REDO) ||
2253 			    (ip->flags & HAMMER_INODE_RDIRTY)) {
2254 				error = hammer_generate_redo(&trans, ip,
2255 							     vap->va_size,
2256 							     HAMMER_REDO_TRUNC,
2257 							     NULL, 0);
2258 			}
2259 			blksize = hammer_blocksize(vap->va_size);
2260 
2261 			/*
2262 			 * XXX break atomicy, we can deadlock the backend
2263 			 * if we do not release the lock.  Probably not a
2264 			 * big deal here.
2265 			 */
2266 			if (vap->va_size < ip->ino_data.size) {
2267 				nvtruncbuf(ap->a_vp, vap->va_size,
2268 					   blksize,
2269 					   hammer_blockoff(vap->va_size),
2270 					   0);
2271 				truncating = 1;
2272 				kflags |= NOTE_WRITE;
2273 			} else {
2274 				nvextendbuf(ap->a_vp,
2275 					    ip->ino_data.size,
2276 					    vap->va_size,
2277 					    hammer_blocksize(ip->ino_data.size),
2278 					    hammer_blocksize(vap->va_size),
2279 					    hammer_blockoff(ip->ino_data.size),
2280 					    hammer_blockoff(vap->va_size),
2281 					    0);
2282 				truncating = 0;
2283 				kflags |= NOTE_WRITE | NOTE_EXTEND;
2284 			}
2285 			ip->ino_data.size = vap->va_size;
2286 			ip->ino_data.mtime = trans.time;
2287 			/* XXX safe to use SDIRTY instead of DDIRTY here? */
2288 			modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2289 			vclrflags(ap->a_vp, VLASTWRITETS);
2290 
2291 			/*
2292 			 * On-media truncation is cached in the inode until
2293 			 * the inode is synchronized.  We must immediately
2294 			 * handle any frontend records.
2295 			 */
2296 			if (truncating) {
2297 				hammer_ip_frontend_trunc(ip, vap->va_size);
2298 				if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2299 					ip->flags |= HAMMER_INODE_TRUNCATED;
2300 					ip->trunc_off = vap->va_size;
2301 					hammer_inode_dirty(ip);
2302 				} else if (ip->trunc_off > vap->va_size) {
2303 					ip->trunc_off = vap->va_size;
2304 				}
2305 			}
2306 
2307 #if 0
2308 			/*
2309 			 * When truncating, nvtruncbuf() may have cleaned out
2310 			 * a portion of the last block on-disk in the buffer
2311 			 * cache.  We must clean out any frontend records
2312 			 * for blocks beyond the new last block.
2313 			 */
2314 			aligned_size = (vap->va_size + (blksize - 1)) &
2315 				       ~(int64_t)(blksize - 1);
2316 			if (truncating && vap->va_size < aligned_size) {
2317 				aligned_size -= blksize;
2318 				hammer_ip_frontend_trunc(ip, aligned_size);
2319 			}
2320 #endif
2321 			break;
2322 		case VDATABASE:
2323 			if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2324 				ip->flags |= HAMMER_INODE_TRUNCATED;
2325 				ip->trunc_off = vap->va_size;
2326 				hammer_inode_dirty(ip);
2327 			} else if (ip->trunc_off > vap->va_size) {
2328 				ip->trunc_off = vap->va_size;
2329 			}
2330 			hammer_ip_frontend_trunc(ip, vap->va_size);
2331 			ip->ino_data.size = vap->va_size;
2332 			ip->ino_data.mtime = trans.time;
2333 			modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2334 			vclrflags(ap->a_vp, VLASTWRITETS);
2335 			kflags |= NOTE_ATTRIB;
2336 			break;
2337 		default:
2338 			error = EINVAL;
2339 			goto done;
2340 		}
2341 		break;
2342 	}
2343 	if (vap->va_atime.tv_sec != VNOVAL) {
2344 		ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2345 		modflags |= HAMMER_INODE_ATIME;
2346 		kflags |= NOTE_ATTRIB;
2347 	}
2348 	if (vap->va_mtime.tv_sec != VNOVAL) {
2349 		ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2350 		modflags |= HAMMER_INODE_MTIME;
2351 		kflags |= NOTE_ATTRIB;
2352 		vclrflags(ap->a_vp, VLASTWRITETS);
2353 	}
2354 	if (vap->va_mode != (mode_t)VNOVAL) {
2355 		mode_t   cur_mode = ip->ino_data.mode;
2356 		uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2357 		gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2358 
2359 		error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2360 					 cur_uid, cur_gid, &cur_mode);
2361 		if (error == 0 && ip->ino_data.mode != cur_mode) {
2362 			ip->ino_data.mode = cur_mode;
2363 			ip->ino_data.ctime = trans.time;
2364 			modflags |= HAMMER_INODE_DDIRTY;
2365 			kflags |= NOTE_ATTRIB;
2366 		}
2367 	}
2368 done:
2369 	if (error == 0)
2370 		hammer_modify_inode(&trans, ip, modflags);
2371 	hammer_done_transaction(&trans);
2372 	hammer_knote(ap->a_vp, kflags);
2373 	lwkt_reltoken(&hmp->fs_token);
2374 	return (error);
2375 }
2376 
2377 /*
2378  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2379  */
2380 static
2381 int
2382 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2383 {
2384 	struct hammer_transaction trans;
2385 	hammer_inode_t dip;
2386 	hammer_inode_t nip;
2387 	hammer_record_t record;
2388 	struct nchandle *nch;
2389 	hammer_mount_t hmp;
2390 	int error;
2391 	int bytes;
2392 
2393 	ap->a_vap->va_type = VLNK;
2394 
2395 	nch = ap->a_nch;
2396 	dip = VTOI(ap->a_dvp);
2397 	hmp = dip->hmp;
2398 
2399 	if (dip->flags & HAMMER_INODE_RO)
2400 		return (EROFS);
2401 	if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
2402 		return (error);
2403 
2404 	/*
2405 	 * Create a transaction to cover the operations we perform.
2406 	 */
2407 	lwkt_gettoken(&hmp->fs_token);
2408 	hammer_start_transaction(&trans, hmp);
2409 
2410 	/*
2411 	 * Create a new filesystem object of the requested type.  The
2412 	 * returned inode will be referenced but not locked.
2413 	 */
2414 
2415 	error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2416 				    dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2417 				    NULL, &nip);
2418 	if (error) {
2419 		hammer_done_transaction(&trans);
2420 		*ap->a_vpp = NULL;
2421 		lwkt_reltoken(&hmp->fs_token);
2422 		return (error);
2423 	}
2424 
2425 	/*
2426 	 * Add a record representing the symlink.  symlink stores the link
2427 	 * as pure data, not a string, and is no \0 terminated.
2428 	 */
2429 	if (error == 0) {
2430 		bytes = strlen(ap->a_target);
2431 
2432 		if (bytes <= HAMMER_INODE_BASESYMLEN) {
2433 			bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2434 		} else {
2435 			record = hammer_alloc_mem_record(nip, bytes);
2436 			record->type = HAMMER_MEM_RECORD_GENERAL;
2437 
2438 			record->leaf.base.localization = nip->obj_localization |
2439 							 HAMMER_LOCALIZE_MISC;
2440 			record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2441 			record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2442 			record->leaf.data_len = bytes;
2443 			KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2444 			bcopy(ap->a_target, record->data->symlink.name, bytes);
2445 			error = hammer_ip_add_record(&trans, record);
2446 		}
2447 
2448 		/*
2449 		 * Set the file size to the length of the link.
2450 		 */
2451 		if (error == 0) {
2452 			nip->ino_data.size = bytes;
2453 			hammer_modify_inode(&trans, nip, HAMMER_INODE_DDIRTY);
2454 		}
2455 	}
2456 	if (error == 0)
2457 		error = hammer_ip_add_direntry(&trans, dip, nch->ncp->nc_name,
2458 						nch->ncp->nc_nlen, nip);
2459 
2460 	/*
2461 	 * Finish up.
2462 	 */
2463 	if (error) {
2464 		hammer_rel_inode(nip, 0);
2465 		*ap->a_vpp = NULL;
2466 	} else {
2467 		error = hammer_get_vnode(nip, ap->a_vpp);
2468 		hammer_rel_inode(nip, 0);
2469 		if (error == 0) {
2470 			cache_setunresolved(ap->a_nch);
2471 			cache_setvp(ap->a_nch, *ap->a_vpp);
2472 			hammer_knote(ap->a_dvp, NOTE_WRITE);
2473 		}
2474 	}
2475 	hammer_done_transaction(&trans);
2476 	lwkt_reltoken(&hmp->fs_token);
2477 	return (error);
2478 }
2479 
2480 /*
2481  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2482  */
2483 static
2484 int
2485 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2486 {
2487 	struct hammer_transaction trans;
2488 	hammer_inode_t dip;
2489 	hammer_mount_t hmp;
2490 	int error;
2491 
2492 	dip = VTOI(ap->a_dvp);
2493 	hmp = dip->hmp;
2494 
2495 	if (hammer_nohistory(dip) == 0 &&
2496 	    (error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2497 		return (error);
2498 	}
2499 
2500 	lwkt_gettoken(&hmp->fs_token);
2501 	hammer_start_transaction(&trans, hmp);
2502 	error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2503 				ap->a_cred, ap->a_flags, -1);
2504 	hammer_done_transaction(&trans);
2505 	lwkt_reltoken(&hmp->fs_token);
2506 
2507 	return (error);
2508 }
2509 
2510 /*
2511  * hammer_vop_ioctl { vp, command, data, fflag, cred }
2512  */
2513 static
2514 int
2515 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2516 {
2517 	hammer_inode_t ip = ap->a_vp->v_data;
2518 	hammer_mount_t hmp = ip->hmp;
2519 	int error;
2520 
2521 	lwkt_gettoken(&hmp->fs_token);
2522 	error = hammer_ioctl(ip, ap->a_command, ap->a_data,
2523 			     ap->a_fflag, ap->a_cred);
2524 	lwkt_reltoken(&hmp->fs_token);
2525 	return (error);
2526 }
2527 
2528 static
2529 int
2530 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2531 {
2532 	static const struct mountctl_opt extraopt[] = {
2533 		{ HMNT_NOHISTORY,	"nohistory" },
2534 		{ HMNT_MASTERID,	"master" },
2535 		{ HMNT_NOMIRROR,	"nomirror" },
2536 		{ 0, NULL}
2537 
2538 	};
2539 	hammer_mount_t hmp;
2540 	struct mount *mp;
2541 	int usedbytes;
2542 	int error;
2543 
2544 	error = 0;
2545 	usedbytes = 0;
2546 	mp = ap->a_head.a_ops->head.vv_mount;
2547 	KKASSERT(mp->mnt_data != NULL);
2548 	hmp = (hammer_mount_t)mp->mnt_data;
2549 
2550 	lwkt_gettoken(&hmp->fs_token);
2551 
2552 	switch(ap->a_op) {
2553 	case MOUNTCTL_SET_EXPORT:
2554 		if (ap->a_ctllen != sizeof(struct export_args))
2555 			error = EINVAL;
2556 		else
2557 			error = hammer_vfs_export(mp, ap->a_op,
2558 				      (const struct export_args *)ap->a_ctl);
2559 		break;
2560 	case MOUNTCTL_MOUNTFLAGS:
2561 		/*
2562 		 * Call standard mountctl VOP function
2563 		 * so we get user mount flags.
2564 		 */
2565 		error = vop_stdmountctl(ap);
2566 		if (error)
2567 			break;
2568 
2569 		usedbytes = *ap->a_res;
2570 
2571 		if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2572 			usedbytes += vfs_flagstostr(hmp->hflags, extraopt,
2573 						    ap->a_buf,
2574 						    ap->a_buflen - usedbytes,
2575 						    &error);
2576 		}
2577 
2578 		*ap->a_res += usedbytes;
2579 		break;
2580 	default:
2581 		error = vop_stdmountctl(ap);
2582 		break;
2583 	}
2584 	lwkt_reltoken(&hmp->fs_token);
2585 	return(error);
2586 }
2587 
2588 /*
2589  * hammer_vop_strategy { vp, bio }
2590  *
2591  * Strategy call, used for regular file read & write only.  Note that the
2592  * bp may represent a cluster.
2593  *
2594  * To simplify operation and allow better optimizations in the future,
2595  * this code does not make any assumptions with regards to buffer alignment
2596  * or size.
2597  */
2598 static
2599 int
2600 hammer_vop_strategy(struct vop_strategy_args *ap)
2601 {
2602 	struct buf *bp;
2603 	int error;
2604 
2605 	bp = ap->a_bio->bio_buf;
2606 
2607 	switch(bp->b_cmd) {
2608 	case BUF_CMD_READ:
2609 		error = hammer_vop_strategy_read(ap);
2610 		break;
2611 	case BUF_CMD_WRITE:
2612 		error = hammer_vop_strategy_write(ap);
2613 		break;
2614 	default:
2615 		bp->b_error = error = EINVAL;
2616 		bp->b_flags |= B_ERROR;
2617 		biodone(ap->a_bio);
2618 		break;
2619 	}
2620 	return (error);
2621 }
2622 
2623 /*
2624  * Read from a regular file.  Iterate the related records and fill in the
2625  * BIO/BUF.  Gaps are zero-filled.
2626  *
2627  * The support code in hammer_object.c should be used to deal with mixed
2628  * in-memory and on-disk records.
2629  *
2630  * NOTE: Can be called from the cluster code with an oversized buf.
2631  *
2632  * XXX atime update
2633  */
2634 static
2635 int
2636 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2637 {
2638 	struct hammer_transaction trans;
2639 	hammer_inode_t ip;
2640 	hammer_inode_t dip;
2641 	hammer_mount_t hmp;
2642 	struct hammer_cursor cursor;
2643 	hammer_base_elm_t base;
2644 	hammer_off_t disk_offset;
2645 	struct bio *bio;
2646 	struct bio *nbio;
2647 	struct buf *bp;
2648 	int64_t rec_offset;
2649 	int64_t ran_end;
2650 	int64_t tmp64;
2651 	int error;
2652 	int boff;
2653 	int roff;
2654 	int n;
2655 	int isdedupable;
2656 
2657 	bio = ap->a_bio;
2658 	bp = bio->bio_buf;
2659 	ip = ap->a_vp->v_data;
2660 	hmp = ip->hmp;
2661 
2662 	/*
2663 	 * The zone-2 disk offset may have been set by the cluster code via
2664 	 * a BMAP operation, or else should be NOOFFSET.
2665 	 *
2666 	 * Checking the high bits for a match against zone-2 should suffice.
2667 	 *
2668 	 * In cases where a lot of data duplication is present it may be
2669 	 * more beneficial to drop through and doubule-buffer through the
2670 	 * device.
2671 	 */
2672 	nbio = push_bio(bio);
2673 	if (hammer_is_zone_large_data(nbio->bio_offset)) {
2674 		if (hammer_double_buffer == 0) {
2675 			lwkt_gettoken(&hmp->fs_token);
2676 			error = hammer_io_direct_read(hmp, nbio, NULL);
2677 			lwkt_reltoken(&hmp->fs_token);
2678 			return (error);
2679 		}
2680 
2681 		/*
2682 		 * Try to shortcut requests for double_buffer mode too.
2683 		 * Since this mode runs through the device buffer cache
2684 		 * only compatible buffer sizes (meaning those generated
2685 		 * by normal filesystem buffers) are legal.
2686 		 */
2687 		if ((bp->b_flags & B_PAGING) == 0) {
2688 			lwkt_gettoken(&hmp->fs_token);
2689 			error = hammer_io_indirect_read(hmp, nbio, NULL);
2690 			lwkt_reltoken(&hmp->fs_token);
2691 			return (error);
2692 		}
2693 	}
2694 
2695 	/*
2696 	 * Well, that sucked.  Do it the hard way.  If all the stars are
2697 	 * aligned we may still be able to issue a direct-read.
2698 	 */
2699 	lwkt_gettoken(&hmp->fs_token);
2700 	hammer_simple_transaction(&trans, hmp);
2701 	hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2702 
2703 	/*
2704 	 * Key range (begin and end inclusive) to scan.  Note that the key's
2705 	 * stored in the actual records represent BASE+LEN, not BASE.  The
2706 	 * first record containing bio_offset will have a key > bio_offset.
2707 	 */
2708 	cursor.key_beg.localization = ip->obj_localization |
2709 				      HAMMER_LOCALIZE_MISC;
2710 	cursor.key_beg.obj_id = ip->obj_id;
2711 	cursor.key_beg.create_tid = 0;
2712 	cursor.key_beg.delete_tid = 0;
2713 	cursor.key_beg.obj_type = 0;
2714 	cursor.key_beg.key = bio->bio_offset + 1;
2715 	cursor.asof = ip->obj_asof;
2716 	cursor.flags |= HAMMER_CURSOR_ASOF;
2717 
2718 	cursor.key_end = cursor.key_beg;
2719 	KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2720 #if 0
2721 	if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2722 		cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2723 		cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2724 		cursor.key_end.key = HAMMER_MAX_KEY;
2725 	} else
2726 #endif
2727 	{
2728 		ran_end = bio->bio_offset + bp->b_bufsize;
2729 		cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2730 		cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2731 		tmp64 = ran_end + MAXPHYS + 1;	/* work-around GCC-4 bug */
2732 		if (tmp64 < ran_end)
2733 			cursor.key_end.key = HAMMER_MAX_KEY;
2734 		else
2735 			cursor.key_end.key = ran_end + MAXPHYS + 1;
2736 	}
2737 	cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2738 
2739 	/*
2740 	 * Set NOSWAPCACHE for cursor data extraction if double buffering
2741 	 * is disabled or (if the file is not marked cacheable via chflags
2742 	 * and vm.swapcache_use_chflags is enabled).
2743 	 */
2744 	if (hammer_double_buffer == 0 ||
2745 	    ((ap->a_vp->v_flag & VSWAPCACHE) == 0 &&
2746 	     vm_swapcache_use_chflags)) {
2747 		cursor.flags |= HAMMER_CURSOR_NOSWAPCACHE;
2748 	}
2749 
2750 	error = hammer_ip_first(&cursor);
2751 	boff = 0;
2752 
2753 	while (error == 0) {
2754 		/*
2755 		 * Get the base file offset of the record.  The key for
2756 		 * data records is (base + bytes) rather then (base).
2757 		 */
2758 		base = &cursor.leaf->base;
2759 		rec_offset = base->key - cursor.leaf->data_len;
2760 
2761 		/*
2762 		 * Calculate the gap, if any, and zero-fill it.
2763 		 *
2764 		 * n is the offset of the start of the record verses our
2765 		 * current seek offset in the bio.
2766 		 */
2767 		n = (int)(rec_offset - (bio->bio_offset + boff));
2768 		if (n > 0) {
2769 			if (n > bp->b_bufsize - boff)
2770 				n = bp->b_bufsize - boff;
2771 			bzero((char *)bp->b_data + boff, n);
2772 			boff += n;
2773 			n = 0;
2774 		}
2775 
2776 		/*
2777 		 * Calculate the data offset in the record and the number
2778 		 * of bytes we can copy.
2779 		 *
2780 		 * There are two degenerate cases.  First, boff may already
2781 		 * be at bp->b_bufsize.  Secondly, the data offset within
2782 		 * the record may exceed the record's size.
2783 		 */
2784 		roff = -n;
2785 		rec_offset += roff;
2786 		n = cursor.leaf->data_len - roff;
2787 		if (n <= 0) {
2788 			hdkprintf("bad n=%d roff=%d\n", n, roff);
2789 			n = 0;
2790 		} else if (n > bp->b_bufsize - boff) {
2791 			n = bp->b_bufsize - boff;
2792 		}
2793 
2794 		/*
2795 		 * Deal with cached truncations.  This cool bit of code
2796 		 * allows truncate()/ftruncate() to avoid having to sync
2797 		 * the file.
2798 		 *
2799 		 * If the frontend is truncated then all backend records are
2800 		 * subject to the frontend's truncation.
2801 		 *
2802 		 * If the backend is truncated then backend records on-disk
2803 		 * (but not in-memory) are subject to the backend's
2804 		 * truncation.  In-memory records owned by the backend
2805 		 * represent data written after the truncation point on the
2806 		 * backend and must not be truncated.
2807 		 *
2808 		 * Truncate operations deal with frontend buffer cache
2809 		 * buffers and frontend-owned in-memory records synchronously.
2810 		 */
2811 		if (ip->flags & HAMMER_INODE_TRUNCATED) {
2812 			if (hammer_cursor_ondisk(&cursor)/* ||
2813 			    cursor.iprec->flush_state == HAMMER_FST_FLUSH*/) {
2814 				if (ip->trunc_off <= rec_offset)
2815 					n = 0;
2816 				else if (ip->trunc_off < rec_offset + n)
2817 					n = (int)(ip->trunc_off - rec_offset);
2818 			}
2819 		}
2820 		if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2821 			if (hammer_cursor_ondisk(&cursor)) {
2822 				if (ip->sync_trunc_off <= rec_offset)
2823 					n = 0;
2824 				else if (ip->sync_trunc_off < rec_offset + n)
2825 					n = (int)(ip->sync_trunc_off - rec_offset);
2826 			}
2827 		}
2828 
2829 		/*
2830 		 * Try to issue a direct read into our bio if possible,
2831 		 * otherwise resolve the element data into a hammer_buffer
2832 		 * and copy.
2833 		 *
2834 		 * The buffer on-disk should be zerod past any real
2835 		 * truncation point, but may not be for any synthesized
2836 		 * truncation point from above.
2837 		 *
2838 		 * NOTE: disk_offset is only valid if the cursor data is
2839 		 *	 on-disk.
2840 		 */
2841 		disk_offset = cursor.leaf->data_offset + roff;
2842 		isdedupable = (boff == 0 && n == bp->b_bufsize &&
2843 			       hammer_cursor_ondisk(&cursor) &&
2844 			       ((int)disk_offset & HAMMER_BUFMASK) == 0);
2845 
2846 		if (isdedupable && hammer_double_buffer == 0) {
2847 			/*
2848 			 * Direct read case
2849 			 */
2850 			KKASSERT(hammer_is_zone_large_data(disk_offset));
2851 			nbio->bio_offset = disk_offset;
2852 			error = hammer_io_direct_read(hmp, nbio, cursor.leaf);
2853 			goto done;
2854 		} else if (isdedupable) {
2855 			/*
2856 			 * Async I/O case for reading from backing store
2857 			 * and copying the data to the filesystem buffer.
2858 			 */
2859 			KKASSERT(hammer_is_zone_large_data(disk_offset));
2860 			nbio->bio_offset = disk_offset;
2861 			error = hammer_io_indirect_read(hmp, nbio, cursor.leaf);
2862 			goto done;
2863 		} else if (n) {
2864 			error = hammer_ip_resolve_data(&cursor);
2865 			if (error == 0) {
2866 				bcopy((char *)cursor.data + roff,
2867 				      (char *)bp->b_data + boff, n);
2868 			}
2869 		}
2870 		if (error)
2871 			break;
2872 
2873 		/*
2874 		 * Iterate until we have filled the request.
2875 		 */
2876 		boff += n;
2877 		if (boff == bp->b_bufsize)
2878 			break;
2879 		error = hammer_ip_next(&cursor);
2880 	}
2881 
2882 	/*
2883 	 * There may have been a gap after the last record
2884 	 */
2885 	if (error == ENOENT)
2886 		error = 0;
2887 	if (error == 0 && boff != bp->b_bufsize) {
2888 		KKASSERT(boff < bp->b_bufsize);
2889 		bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2890 		/* boff = bp->b_bufsize; */
2891 	}
2892 
2893 	/*
2894 	 * Disallow swapcache operation on the vnode buffer if double
2895 	 * buffering is enabled, the swapcache will get the data via
2896 	 * the block device buffer.
2897 	 */
2898 	if (hammer_double_buffer)
2899 		bp->b_flags |= B_NOTMETA;
2900 
2901 	/*
2902 	 * Cleanup
2903 	 */
2904 	bp->b_resid = 0;
2905 	bp->b_error = error;
2906 	if (error)
2907 		bp->b_flags |= B_ERROR;
2908 	biodone(ap->a_bio);
2909 
2910 done:
2911 	/*
2912 	 * Cache the b-tree node for the last data read in cache[1].
2913 	 *
2914 	 * If we hit the file EOF then also cache the node in the
2915 	 * governing directory's cache[3], it will be used to initialize
2916 	 * the new inode's cache[1] for any inodes looked up via the directory.
2917 	 *
2918 	 * This doesn't reduce disk accesses since the B-Tree chain is
2919 	 * likely cached, but it does reduce cpu overhead when looking
2920 	 * up file offsets for cpdup/tar/cpio style iterations.
2921 	 */
2922 	if (cursor.node)
2923 		hammer_cache_node(&ip->cache[1], cursor.node);
2924 	if (ran_end >= ip->ino_data.size) {
2925 		dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
2926 					ip->obj_asof, ip->obj_localization);
2927 		if (dip) {
2928 			hammer_cache_node(&dip->cache[3], cursor.node);
2929 			hammer_rel_inode(dip, 0);
2930 		}
2931 	}
2932 	hammer_done_cursor(&cursor);
2933 	hammer_done_transaction(&trans);
2934 	lwkt_reltoken(&hmp->fs_token);
2935 	return(error);
2936 }
2937 
2938 /*
2939  * BMAP operation - used to support cluster_read() only.
2940  *
2941  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2942  *
2943  * This routine may return EOPNOTSUPP if the opration is not supported for
2944  * the specified offset.  The contents of the pointer arguments do not
2945  * need to be initialized in that case.
2946  *
2947  * If a disk address is available and properly aligned return 0 with
2948  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2949  * to the run-length relative to that offset.  Callers may assume that
2950  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2951  * large, so return EOPNOTSUPP if it is not sufficiently large.
2952  */
2953 static
2954 int
2955 hammer_vop_bmap(struct vop_bmap_args *ap)
2956 {
2957 	struct hammer_transaction trans;
2958 	hammer_inode_t ip;
2959 	hammer_mount_t hmp;
2960 	struct hammer_cursor cursor;
2961 	hammer_base_elm_t base;
2962 	int64_t rec_offset;
2963 	int64_t ran_end;
2964 	int64_t tmp64;
2965 	int64_t base_offset;
2966 	int64_t base_disk_offset;
2967 	int64_t last_offset;
2968 	hammer_off_t last_disk_offset;
2969 	hammer_off_t disk_offset;
2970 	int	rec_len;
2971 	int	error;
2972 	int	blksize;
2973 
2974 	ip = ap->a_vp->v_data;
2975 	hmp = ip->hmp;
2976 
2977 	/*
2978 	 * We can only BMAP regular files.  We can't BMAP database files,
2979 	 * directories, etc.
2980 	 */
2981 	if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
2982 		return(EOPNOTSUPP);
2983 
2984 	/*
2985 	 * bmap is typically called with runp/runb both NULL when used
2986 	 * for writing.  We do not support BMAP for writing atm.
2987 	 */
2988 	if (ap->a_cmd != BUF_CMD_READ)
2989 		return(EOPNOTSUPP);
2990 
2991 	/*
2992 	 * Scan the B-Tree to acquire blockmap addresses, then translate
2993 	 * to raw addresses.
2994 	 */
2995 	lwkt_gettoken(&hmp->fs_token);
2996 	hammer_simple_transaction(&trans, hmp);
2997 
2998 	hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2999 
3000 	/*
3001 	 * Key range (begin and end inclusive) to scan.  Note that the key's
3002 	 * stored in the actual records represent BASE+LEN, not BASE.  The
3003 	 * first record containing bio_offset will have a key > bio_offset.
3004 	 */
3005 	cursor.key_beg.localization = ip->obj_localization |
3006 				      HAMMER_LOCALIZE_MISC;
3007 	cursor.key_beg.obj_id = ip->obj_id;
3008 	cursor.key_beg.create_tid = 0;
3009 	cursor.key_beg.delete_tid = 0;
3010 	cursor.key_beg.obj_type = 0;
3011 	if (ap->a_runb)
3012 		cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
3013 	else
3014 		cursor.key_beg.key = ap->a_loffset + 1;
3015 	if (cursor.key_beg.key < 0)
3016 		cursor.key_beg.key = 0;
3017 	cursor.asof = ip->obj_asof;
3018 	cursor.flags |= HAMMER_CURSOR_ASOF;
3019 
3020 	cursor.key_end = cursor.key_beg;
3021 	KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
3022 
3023 	ran_end = ap->a_loffset + MAXPHYS;
3024 	cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
3025 	cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
3026 	tmp64 = ran_end + MAXPHYS + 1;	/* work-around GCC-4 bug */
3027 	if (tmp64 < ran_end)
3028 		cursor.key_end.key = HAMMER_MAX_KEY;
3029 	else
3030 		cursor.key_end.key = ran_end + MAXPHYS + 1;
3031 
3032 	cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
3033 
3034 	error = hammer_ip_first(&cursor);
3035 	base_offset = last_offset = 0;
3036 	base_disk_offset = last_disk_offset = 0;
3037 
3038 	while (error == 0) {
3039 		/*
3040 		 * Get the base file offset of the record.  The key for
3041 		 * data records is (base + bytes) rather then (base).
3042 		 *
3043 		 * NOTE: rec_offset + rec_len may exceed the end-of-file.
3044 		 * The extra bytes should be zero on-disk and the BMAP op
3045 		 * should still be ok.
3046 		 */
3047 		base = &cursor.leaf->base;
3048 		rec_offset = base->key - cursor.leaf->data_len;
3049 		rec_len    = cursor.leaf->data_len;
3050 
3051 		/*
3052 		 * Incorporate any cached truncation.
3053 		 *
3054 		 * NOTE: Modifications to rec_len based on synthesized
3055 		 * truncation points remove the guarantee that any extended
3056 		 * data on disk is zero (since the truncations may not have
3057 		 * taken place on-media yet).
3058 		 */
3059 		if (ip->flags & HAMMER_INODE_TRUNCATED) {
3060 			if (hammer_cursor_ondisk(&cursor) ||
3061 			    cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
3062 				if (ip->trunc_off <= rec_offset)
3063 					rec_len = 0;
3064 				else if (ip->trunc_off < rec_offset + rec_len)
3065 					rec_len = (int)(ip->trunc_off - rec_offset);
3066 			}
3067 		}
3068 		if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
3069 			if (hammer_cursor_ondisk(&cursor)) {
3070 				if (ip->sync_trunc_off <= rec_offset)
3071 					rec_len = 0;
3072 				else if (ip->sync_trunc_off < rec_offset + rec_len)
3073 					rec_len = (int)(ip->sync_trunc_off - rec_offset);
3074 			}
3075 		}
3076 
3077 		/*
3078 		 * Accumulate information.  If we have hit a discontiguous
3079 		 * block reset base_offset unless we are already beyond the
3080 		 * requested offset.  If we are, that's it, we stop.
3081 		 */
3082 		if (error)
3083 			break;
3084 		if (hammer_cursor_ondisk(&cursor)) {
3085 			disk_offset = cursor.leaf->data_offset;
3086 			if (rec_offset != last_offset ||
3087 			    disk_offset != last_disk_offset) {
3088 				if (rec_offset > ap->a_loffset)
3089 					break;
3090 				base_offset = rec_offset;
3091 				base_disk_offset = disk_offset;
3092 			}
3093 			last_offset = rec_offset + rec_len;
3094 			last_disk_offset = disk_offset + rec_len;
3095 		}
3096 		error = hammer_ip_next(&cursor);
3097 	}
3098 
3099 	if (cursor.node)
3100 		hammer_cache_node(&ip->cache[1], cursor.node);
3101 
3102 	hammer_done_cursor(&cursor);
3103 	hammer_done_transaction(&trans);
3104 	lwkt_reltoken(&hmp->fs_token);
3105 
3106 	/*
3107 	 * If we couldn't find any records or the records we did find were
3108 	 * all behind the requested offset, return failure.  A forward
3109 	 * truncation can leave a hole w/ no on-disk records.
3110 	 */
3111 	if (last_offset == 0 || last_offset < ap->a_loffset)
3112 		return (EOPNOTSUPP);
3113 
3114 	/*
3115 	 * Figure out the block size at the requested offset and adjust
3116 	 * our limits so the cluster_read() does not create inappropriately
3117 	 * sized buffer cache buffers.
3118 	 */
3119 	blksize = hammer_blocksize(ap->a_loffset);
3120 	if (hammer_blocksize(base_offset) != blksize) {
3121 		base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
3122 	}
3123 	if (last_offset != ap->a_loffset &&
3124 	    hammer_blocksize(last_offset - 1) != blksize) {
3125 		last_offset = hammer_blockdemarc(ap->a_loffset,
3126 						 last_offset - 1);
3127 	}
3128 
3129 	/*
3130 	 * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3131 	 * from occuring.
3132 	 */
3133 	disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
3134 
3135 	if (!hammer_is_zone_large_data(disk_offset)) {
3136 		/*
3137 		 * Only large-data zones can be direct-IOd
3138 		 */
3139 		error = EOPNOTSUPP;
3140 	} else if ((disk_offset & HAMMER_BUFMASK) ||
3141 		   (last_offset - ap->a_loffset) < blksize) {
3142 		/*
3143 		 * doffsetp is not aligned or the forward run size does
3144 		 * not cover a whole buffer, disallow the direct I/O.
3145 		 */
3146 		error = EOPNOTSUPP;
3147 	} else {
3148 		/*
3149 		 * We're good.
3150 		 */
3151 		*ap->a_doffsetp = disk_offset;
3152 		if (ap->a_runb) {
3153 			*ap->a_runb = ap->a_loffset - base_offset;
3154 			KKASSERT(*ap->a_runb >= 0);
3155 		}
3156 		if (ap->a_runp) {
3157 			*ap->a_runp = last_offset - ap->a_loffset;
3158 			KKASSERT(*ap->a_runp >= 0);
3159 		}
3160 		error = 0;
3161 	}
3162 	return(error);
3163 }
3164 
3165 /*
3166  * Write to a regular file.   Because this is a strategy call the OS is
3167  * trying to actually get data onto the media.
3168  */
3169 static
3170 int
3171 hammer_vop_strategy_write(struct vop_strategy_args *ap)
3172 {
3173 	hammer_record_t record;
3174 	hammer_mount_t hmp;
3175 	hammer_inode_t ip;
3176 	struct bio *bio;
3177 	struct buf *bp;
3178 	int blksize __debugvar;
3179 	int bytes;
3180 	int error;
3181 
3182 	bio = ap->a_bio;
3183 	bp = bio->bio_buf;
3184 	ip = ap->a_vp->v_data;
3185 	hmp = ip->hmp;
3186 
3187 	blksize = hammer_blocksize(bio->bio_offset);
3188 	KKASSERT(bp->b_bufsize == blksize);
3189 
3190 	if (ip->flags & HAMMER_INODE_RO) {
3191 		bp->b_error = EROFS;
3192 		bp->b_flags |= B_ERROR;
3193 		biodone(ap->a_bio);
3194 		return(EROFS);
3195 	}
3196 
3197 	lwkt_gettoken(&hmp->fs_token);
3198 
3199 	/*
3200 	 * Disallow swapcache operation on the vnode buffer if double
3201 	 * buffering is enabled, the swapcache will get the data via
3202 	 * the block device buffer.
3203 	 */
3204 	if (hammer_double_buffer)
3205 		bp->b_flags |= B_NOTMETA;
3206 
3207 	/*
3208 	 * Interlock with inode destruction (no in-kernel or directory
3209 	 * topology visibility).  If we queue new IO while trying to
3210 	 * destroy the inode we can deadlock the vtrunc call in
3211 	 * hammer_inode_unloadable_check().
3212 	 *
3213 	 * Besides, there's no point flushing a bp associated with an
3214 	 * inode that is being destroyed on-media and has no kernel
3215 	 * references.
3216 	 */
3217 	if ((ip->flags | ip->sync_flags) &
3218 	    (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
3219 		bp->b_resid = 0;
3220 		biodone(ap->a_bio);
3221 		lwkt_reltoken(&hmp->fs_token);
3222 		return(0);
3223 	}
3224 
3225 	/*
3226 	 * Reserve space and issue a direct-write from the front-end.
3227 	 * NOTE: The direct_io code will hammer_bread/bcopy smaller
3228 	 * allocations.
3229 	 *
3230 	 * An in-memory record will be installed to reference the storage
3231 	 * until the flusher can get to it.
3232 	 *
3233 	 * Since we own the high level bio the front-end will not try to
3234 	 * do a direct-read until the write completes.
3235 	 *
3236 	 * NOTE: The only time we do not reserve a full-sized buffers
3237 	 * worth of data is if the file is small.  We do not try to
3238 	 * allocate a fragment (from the small-data zone) at the end of
3239 	 * an otherwise large file as this can lead to wildly separated
3240 	 * data.
3241 	 */
3242 	KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
3243 	KKASSERT(bio->bio_offset < ip->ino_data.size);
3244 	if (bio->bio_offset || ip->ino_data.size > HAMMER_HBUFSIZE)
3245 		bytes = bp->b_bufsize;
3246 	else
3247 		bytes = HAMMER_DATA_DOALIGN_WITH(int, ip->ino_data.size);
3248 
3249 	record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
3250 				    bytes, &error);
3251 
3252 	/*
3253 	 * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3254 	 * in hammer_vop_write().  We must flag the record so the proper
3255 	 * REDO_TERM_WRITE entry is generated during the flush.
3256 	 */
3257 	if (record) {
3258 		if (bp->b_flags & B_VFSFLAG1) {
3259 			record->flags |= HAMMER_RECF_REDO;
3260 			bp->b_flags &= ~B_VFSFLAG1;
3261 		}
3262 		hammer_io_direct_write(hmp, bio, record);
3263 		if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
3264 			hammer_flush_inode(ip, 0);
3265 	} else {
3266 		bp->b_bio2.bio_offset = NOOFFSET;
3267 		bp->b_error = error;
3268 		bp->b_flags |= B_ERROR;
3269 		biodone(ap->a_bio);
3270 	}
3271 	lwkt_reltoken(&hmp->fs_token);
3272 	return(error);
3273 }
3274 
3275 /*
3276  * dounlink - disconnect a directory entry
3277  *
3278  * XXX whiteout support not really in yet
3279  */
3280 static int
3281 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
3282 		struct vnode *dvp, struct ucred *cred,
3283 		int flags, int isdir)
3284 {
3285 	struct namecache *ncp;
3286 	hammer_inode_t dip;
3287 	hammer_inode_t ip;
3288 	hammer_mount_t hmp;
3289 	struct hammer_cursor cursor;
3290 	int64_t namekey;
3291 	uint32_t max_iterations;
3292 	int nlen, error;
3293 
3294 	/*
3295 	 * Calculate the namekey and setup the key range for the scan.  This
3296 	 * works kinda like a chained hash table where the lower 32 bits
3297 	 * of the namekey synthesize the chain.
3298 	 *
3299 	 * The key range is inclusive of both key_beg and key_end.
3300 	 */
3301 	dip = VTOI(dvp);
3302 	ncp = nch->ncp;
3303 	hmp = dip->hmp;
3304 
3305 	if (dip->flags & HAMMER_INODE_RO)
3306 		return (EROFS);
3307 
3308 	namekey = hammer_direntry_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3309 					   &max_iterations);
3310 retry:
3311 	hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3312 	cursor.key_beg.localization = dip->obj_localization |
3313 				      hammer_dir_localization(dip);
3314         cursor.key_beg.obj_id = dip->obj_id;
3315 	cursor.key_beg.key = namekey;
3316         cursor.key_beg.create_tid = 0;
3317         cursor.key_beg.delete_tid = 0;
3318         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3319         cursor.key_beg.obj_type = 0;
3320 
3321 	cursor.key_end = cursor.key_beg;
3322 	cursor.key_end.key += max_iterations;
3323 	cursor.asof = dip->obj_asof;
3324 	cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3325 
3326 	/*
3327 	 * Scan all matching records (the chain), locate the one matching
3328 	 * the requested path component.  info->last_error contains the
3329 	 * error code on search termination and could be 0, ENOENT, or
3330 	 * something else.
3331 	 *
3332 	 * The hammer_ip_*() functions merge in-memory records with on-disk
3333 	 * records for the purposes of the search.
3334 	 */
3335 	error = hammer_ip_first(&cursor);
3336 
3337 	while (error == 0) {
3338 		error = hammer_ip_resolve_data(&cursor);
3339 		if (error)
3340 			break;
3341 		nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3342 		KKASSERT(nlen > 0);
3343 		if (ncp->nc_nlen == nlen &&
3344 		    bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3345 			break;
3346 		}
3347 		error = hammer_ip_next(&cursor);
3348 	}
3349 
3350 	/*
3351 	 * If all is ok we have to get the inode so we can adjust nlinks.
3352 	 * To avoid a deadlock with the flusher we must release the inode
3353 	 * lock on the directory when acquiring the inode for the entry.
3354 	 *
3355 	 * If the target is a directory, it must be empty.
3356 	 */
3357 	if (error == 0) {
3358 		hammer_unlock(&cursor.ip->lock);
3359 		ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3360 				      hmp->asof,
3361 				      cursor.data->entry.localization,
3362 				      0, &error);
3363 		hammer_lock_sh(&cursor.ip->lock);
3364 		if (error == ENOENT) {
3365 			hkprintf("WARNING: Removing dirent w/missing inode "
3366 				"\"%s\"\n"
3367 				"\tobj_id = %016jx\n",
3368 				ncp->nc_name,
3369 				(intmax_t)cursor.data->entry.obj_id);
3370 			error = 0;
3371 		}
3372 
3373 		/*
3374 		 * If isdir >= 0 we validate that the entry is or is not a
3375 		 * directory.  If isdir < 0 we don't care.
3376 		 */
3377 		if (error == 0 && isdir >= 0 && ip) {
3378 			if (isdir &&
3379 			    ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3380 				error = ENOTDIR;
3381 			} else if (isdir == 0 &&
3382 			    ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3383 				error = EISDIR;
3384 			}
3385 		}
3386 
3387 		/*
3388 		 * If we are trying to remove a directory the directory must
3389 		 * be empty.
3390 		 *
3391 		 * The check directory code can loop and deadlock/retry.  Our
3392 		 * own cursor's node locks must be released to avoid a 3-way
3393 		 * deadlock with the flusher if the check directory code
3394 		 * blocks.
3395 		 *
3396 		 * If any changes whatsoever have been made to the cursor
3397 		 * set EDEADLK and retry.
3398 		 *
3399 		 * WARNING: See warnings in hammer_unlock_cursor()
3400 		 *	    function.
3401 		 */
3402 		if (error == 0 && ip && ip->ino_data.obj_type ==
3403 				        HAMMER_OBJTYPE_DIRECTORY) {
3404 			hammer_unlock_cursor(&cursor);
3405 			error = hammer_ip_check_directory_empty(trans, ip);
3406 			hammer_lock_cursor(&cursor);
3407 			if (cursor.flags & HAMMER_CURSOR_RETEST) {
3408 				hkprintf("Warning: avoided deadlock "
3409 					"on rmdir '%s'\n",
3410 					ncp->nc_name);
3411 				error = EDEADLK;
3412 			}
3413 		}
3414 
3415 		/*
3416 		 * Delete the directory entry.
3417 		 *
3418 		 * WARNING: hammer_ip_del_direntry() may have to terminate
3419 		 * the cursor to avoid a deadlock.  It is ok to call
3420 		 * hammer_done_cursor() twice.
3421 		 */
3422 		if (error == 0) {
3423 			error = hammer_ip_del_direntry(trans, &cursor,
3424 							dip, ip);
3425 		}
3426 		hammer_done_cursor(&cursor);
3427 		if (error == 0) {
3428 			/*
3429 			 * Tell the namecache that we are now unlinked.
3430 			 */
3431 			cache_unlink(nch);
3432 
3433 			/*
3434 			 * NOTE: ip->vp, if non-NULL, cannot be directly
3435 			 *	 referenced without formally acquiring the
3436 			 *	 vp since the vp might have zero refs on it,
3437 			 *	 or in the middle of a reclaim, etc.
3438 			 *
3439 			 * NOTE: The cache_setunresolved() can rip the vp
3440 			 *	 out from under us since the vp may not have
3441 			 *	 any refs, in which case ip->vp will be NULL
3442 			 *	 from the outset.
3443 			 */
3444 			while (ip && ip->vp) {
3445 				struct vnode *vp;
3446 
3447 				error = hammer_get_vnode(ip, &vp);
3448 				if (error == 0 && vp) {
3449 					vn_unlock(vp);
3450 					hammer_knote(ip->vp, NOTE_DELETE);
3451 #if 0
3452 					/*
3453 					 * Don't do this, it can deadlock
3454 					 * on concurrent rm's of hardlinks.
3455 					 * Shouldn't be needed any more.
3456 					 */
3457 					cache_inval_vp(ip->vp, CINV_DESTROY);
3458 #endif
3459 					vrele(vp);
3460 					break;
3461 				}
3462 				hdkprintf("ip/vp race1 avoided\n");
3463 			}
3464 		}
3465 		if (ip)
3466 			hammer_rel_inode(ip, 0);
3467 	} else {
3468 		hammer_done_cursor(&cursor);
3469 	}
3470 	if (error == EDEADLK)
3471 		goto retry;
3472 
3473 	return (error);
3474 }
3475 
3476 /************************************************************************
3477  *			    FIFO AND SPECFS OPS				*
3478  ************************************************************************
3479  *
3480  */
3481 static int
3482 hammer_vop_fifoclose (struct vop_close_args *ap)
3483 {
3484 	/* XXX update itimes */
3485 	return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3486 }
3487 
3488 static int
3489 hammer_vop_fiforead (struct vop_read_args *ap)
3490 {
3491 	int error;
3492 
3493 	error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3494 	/* XXX update access time */
3495 	return (error);
3496 }
3497 
3498 static int
3499 hammer_vop_fifowrite (struct vop_write_args *ap)
3500 {
3501 	int error;
3502 
3503 	error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3504 	/* XXX update access time */
3505 	return (error);
3506 }
3507 
3508 static
3509 int
3510 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3511 {
3512 	int error;
3513 
3514 	error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3515 	if (error)
3516 		error = hammer_vop_kqfilter(ap);
3517 	return(error);
3518 }
3519 
3520 /************************************************************************
3521  *			    KQFILTER OPS				*
3522  ************************************************************************
3523  *
3524  */
3525 static void filt_hammerdetach(struct knote *kn);
3526 static int filt_hammerread(struct knote *kn, long hint);
3527 static int filt_hammerwrite(struct knote *kn, long hint);
3528 static int filt_hammervnode(struct knote *kn, long hint);
3529 
3530 static struct filterops hammerread_filtops =
3531 	{ FILTEROP_ISFD | FILTEROP_MPSAFE,
3532 	  NULL, filt_hammerdetach, filt_hammerread };
3533 static struct filterops hammerwrite_filtops =
3534 	{ FILTEROP_ISFD | FILTEROP_MPSAFE,
3535 	  NULL, filt_hammerdetach, filt_hammerwrite };
3536 static struct filterops hammervnode_filtops =
3537 	{ FILTEROP_ISFD | FILTEROP_MPSAFE,
3538 	  NULL, filt_hammerdetach, filt_hammervnode };
3539 
3540 static
3541 int
3542 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3543 {
3544 	struct vnode *vp = ap->a_vp;
3545 	struct knote *kn = ap->a_kn;
3546 
3547 	switch (kn->kn_filter) {
3548 	case EVFILT_READ:
3549 		kn->kn_fop = &hammerread_filtops;
3550 		break;
3551 	case EVFILT_WRITE:
3552 		kn->kn_fop = &hammerwrite_filtops;
3553 		break;
3554 	case EVFILT_VNODE:
3555 		kn->kn_fop = &hammervnode_filtops;
3556 		break;
3557 	default:
3558 		return (EOPNOTSUPP);
3559 	}
3560 
3561 	kn->kn_hook = (caddr_t)vp;
3562 
3563 	knote_insert(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3564 
3565 	return(0);
3566 }
3567 
3568 static void
3569 filt_hammerdetach(struct knote *kn)
3570 {
3571 	struct vnode *vp = (void *)kn->kn_hook;
3572 
3573 	knote_remove(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3574 }
3575 
3576 static int
3577 filt_hammerread(struct knote *kn, long hint)
3578 {
3579 	struct vnode *vp = (void *)kn->kn_hook;
3580 	hammer_inode_t ip = VTOI(vp);
3581 	hammer_mount_t hmp = ip->hmp;
3582 	off_t off;
3583 
3584 	if (hint == NOTE_REVOKE) {
3585 		kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3586 		return(1);
3587 	}
3588 	lwkt_gettoken(&hmp->fs_token);	/* XXX use per-ip-token */
3589 	off = ip->ino_data.size - kn->kn_fp->f_offset;
3590 	kn->kn_data = (off < INTPTR_MAX) ? off : INTPTR_MAX;
3591 	lwkt_reltoken(&hmp->fs_token);
3592 	if (kn->kn_sfflags & NOTE_OLDAPI)
3593 		return(1);
3594 	return (kn->kn_data != 0);
3595 }
3596 
3597 static int
3598 filt_hammerwrite(struct knote *kn, long hint)
3599 {
3600 	if (hint == NOTE_REVOKE)
3601 		kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3602 	kn->kn_data = 0;
3603 	return (1);
3604 }
3605 
3606 static int
3607 filt_hammervnode(struct knote *kn, long hint)
3608 {
3609 	if (kn->kn_sfflags & hint)
3610 		kn->kn_fflags |= hint;
3611 	if (hint == NOTE_REVOKE) {
3612 		kn->kn_flags |= (EV_EOF | EV_NODATA);
3613 		return (1);
3614 	}
3615 	return (kn->kn_fflags != 0);
3616 }
3617 
3618