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/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/fcntl.h>
39 #include <sys/namecache.h>
40 #include <sys/vnode.h>
41 #include <sys/lockf.h>
42 #include <sys/event.h>
43 #include <sys/stat.h>
44 #include <sys/dirent.h>
45 #include <sys/file.h>
46 #include <vm/vm_extern.h>
47 #include <vm/swap_pager.h>
48 #include <vfs/fifofs/fifo.h>
49
50 #include "hammer.h"
51
52 /*
53 * USERFS VNOPS
54 */
55 /*static int hammer_vop_vnoperate(struct vop_generic_args *);*/
56 static int hammer_vop_fsync(struct vop_fsync_args *);
57 static int hammer_vop_read(struct vop_read_args *);
58 static int hammer_vop_write(struct vop_write_args *);
59 static int hammer_vop_access(struct vop_access_args *);
60 static int hammer_vop_advlock(struct vop_advlock_args *);
61 static int hammer_vop_close(struct vop_close_args *);
62 static int hammer_vop_ncreate(struct vop_ncreate_args *);
63 static int hammer_vop_getattr(struct vop_getattr_args *);
64 static int hammer_vop_nresolve(struct vop_nresolve_args *);
65 static int hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *);
66 static int hammer_vop_nlink(struct vop_nlink_args *);
67 static int hammer_vop_nmkdir(struct vop_nmkdir_args *);
68 static int hammer_vop_nmknod(struct vop_nmknod_args *);
69 static int hammer_vop_open(struct vop_open_args *);
70 static int hammer_vop_print(struct vop_print_args *);
71 static int hammer_vop_readdir(struct vop_readdir_args *);
72 static int hammer_vop_readlink(struct vop_readlink_args *);
73 static int hammer_vop_nremove(struct vop_nremove_args *);
74 static int hammer_vop_nrename(struct vop_nrename_args *);
75 static int hammer_vop_nrmdir(struct vop_nrmdir_args *);
76 static int hammer_vop_markatime(struct vop_markatime_args *);
77 static int hammer_vop_setattr(struct vop_setattr_args *);
78 static int hammer_vop_strategy(struct vop_strategy_args *);
79 static int hammer_vop_bmap(struct vop_bmap_args *ap);
80 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
81 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
82 static int hammer_vop_ioctl(struct vop_ioctl_args *);
83 static int hammer_vop_mountctl(struct vop_mountctl_args *);
84 static int hammer_vop_kqfilter (struct vop_kqfilter_args *);
85
86 static int hammer_vop_fifoclose (struct vop_close_args *);
87 static int hammer_vop_fiforead (struct vop_read_args *);
88 static int hammer_vop_fifowrite (struct vop_write_args *);
89 static int hammer_vop_fifokqfilter (struct vop_kqfilter_args *);
90
91 struct vop_ops hammer_vnode_vops = {
92 .vop_default = vop_defaultop,
93 .vop_fsync = hammer_vop_fsync,
94 .vop_getpages = vop_stdgetpages,
95 .vop_putpages = vop_stdputpages,
96 .vop_read = hammer_vop_read,
97 .vop_write = hammer_vop_write,
98 .vop_access = hammer_vop_access,
99 .vop_advlock = hammer_vop_advlock,
100 .vop_close = hammer_vop_close,
101 .vop_ncreate = hammer_vop_ncreate,
102 .vop_getattr = hammer_vop_getattr,
103 .vop_inactive = hammer_vop_inactive,
104 .vop_reclaim = hammer_vop_reclaim,
105 .vop_nresolve = hammer_vop_nresolve,
106 .vop_nlookupdotdot = hammer_vop_nlookupdotdot,
107 .vop_nlink = hammer_vop_nlink,
108 .vop_nmkdir = hammer_vop_nmkdir,
109 .vop_nmknod = hammer_vop_nmknod,
110 .vop_open = hammer_vop_open,
111 .vop_pathconf = vop_stdpathconf,
112 .vop_print = hammer_vop_print,
113 .vop_readdir = hammer_vop_readdir,
114 .vop_readlink = hammer_vop_readlink,
115 .vop_nremove = hammer_vop_nremove,
116 .vop_nrename = hammer_vop_nrename,
117 .vop_nrmdir = hammer_vop_nrmdir,
118 .vop_markatime = hammer_vop_markatime,
119 .vop_setattr = hammer_vop_setattr,
120 .vop_bmap = hammer_vop_bmap,
121 .vop_strategy = hammer_vop_strategy,
122 .vop_nsymlink = hammer_vop_nsymlink,
123 .vop_nwhiteout = hammer_vop_nwhiteout,
124 .vop_ioctl = hammer_vop_ioctl,
125 .vop_mountctl = hammer_vop_mountctl,
126 .vop_kqfilter = hammer_vop_kqfilter
127 };
128
129 struct vop_ops hammer_spec_vops = {
130 .vop_default = vop_defaultop,
131 .vop_fsync = hammer_vop_fsync,
132 .vop_read = vop_stdnoread,
133 .vop_write = vop_stdnowrite,
134 .vop_access = hammer_vop_access,
135 .vop_close = hammer_vop_close,
136 .vop_markatime = hammer_vop_markatime,
137 .vop_getattr = hammer_vop_getattr,
138 .vop_inactive = hammer_vop_inactive,
139 .vop_reclaim = hammer_vop_reclaim,
140 .vop_setattr = hammer_vop_setattr
141 };
142
143 struct vop_ops hammer_fifo_vops = {
144 .vop_default = fifo_vnoperate,
145 .vop_fsync = hammer_vop_fsync,
146 .vop_read = hammer_vop_fiforead,
147 .vop_write = hammer_vop_fifowrite,
148 .vop_access = hammer_vop_access,
149 .vop_close = hammer_vop_fifoclose,
150 .vop_markatime = hammer_vop_markatime,
151 .vop_getattr = hammer_vop_getattr,
152 .vop_inactive = hammer_vop_inactive,
153 .vop_reclaim = hammer_vop_reclaim,
154 .vop_setattr = hammer_vop_setattr,
155 .vop_kqfilter = hammer_vop_fifokqfilter
156 };
157
158 static __inline
159 void
160 hammer_knote(struct vnode *vp, int flags)
161 {
162 if (flags)
163 KNOTE(&vp->v_pollinfo.vpi_kqinfo.ki_note, flags);
164 }
165
166 #ifdef DEBUG_TRUNCATE
167 struct hammer_inode *HammerTruncIp;
168 #endif
169
170 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
171 struct vnode *dvp, struct ucred *cred,
172 int flags, int isdir);
173 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
174 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
175
176 #if 0
177 static
178 int
179 hammer_vop_vnoperate(struct vop_generic_args *)
180 {
181 return (VOCALL(&hammer_vnode_vops, ap));
182 }
183 #endif
184
185 /*
186 * hammer_vop_fsync { vp, waitfor }
187 *
188 * fsync() an inode to disk and wait for it to be completely committed
189 * such that the information would not be undone if a crash occured after
190 * return.
191 *
192 * NOTE: HAMMER's fsync()'s are going to remain expensive until we implement
193 * a REDO log. A sysctl is provided to relax HAMMER's fsync()
194 * operation.
195 *
196 * Ultimately the combination of a REDO log and use of fast storage
197 * to front-end cluster caches will make fsync fast, but it aint
198 * here yet. And, in anycase, we need real transactional
199 * all-or-nothing features which are not restricted to a single file.
200 */
201 static
202 int
203 hammer_vop_fsync(struct vop_fsync_args *ap)
204 {
205 hammer_inode_t ip = VTOI(ap->a_vp);
206 hammer_mount_t hmp = ip->hmp;
207 int waitfor = ap->a_waitfor;
208 int mode;
209
210 lwkt_gettoken(&hmp->fs_token);
211
212 /*
213 * Fsync rule relaxation (default is either full synchronous flush
214 * or REDO semantics with synchronous flush).
215 */
216 if (ap->a_flags & VOP_FSYNC_SYSCALL) {
217 switch(hammer_fsync_mode) {
218 case 0:
219 mode0:
220 /* no REDO, full synchronous flush */
221 goto skip;
222 case 1:
223 mode1:
224 /* no REDO, full asynchronous flush */
225 if (waitfor == MNT_WAIT)
226 waitfor = MNT_NOWAIT;
227 goto skip;
228 case 2:
229 /* REDO semantics, synchronous flush */
230 if (hmp->version < HAMMER_VOL_VERSION_FOUR)
231 goto mode0;
232 mode = HAMMER_FLUSH_UNDOS_AUTO;
233 break;
234 case 3:
235 /* REDO semantics, relaxed asynchronous flush */
236 if (hmp->version < HAMMER_VOL_VERSION_FOUR)
237 goto mode1;
238 mode = HAMMER_FLUSH_UNDOS_RELAXED;
239 if (waitfor == MNT_WAIT)
240 waitfor = MNT_NOWAIT;
241 break;
242 case 4:
243 /* ignore the fsync() system call */
244 lwkt_reltoken(&hmp->fs_token);
245 return(0);
246 default:
247 /* we have to do something */
248 mode = HAMMER_FLUSH_UNDOS_RELAXED;
249 if (waitfor == MNT_WAIT)
250 waitfor = MNT_NOWAIT;
251 break;
252 }
253
254 /*
255 * Fast fsync only needs to flush the UNDO/REDO fifo if
256 * HAMMER_INODE_REDO is non-zero and the only modifications
257 * made to the file are write or write-extends.
258 */
259 if ((ip->flags & HAMMER_INODE_REDO) &&
260 (ip->flags & HAMMER_INODE_MODMASK_NOREDO) == 0
261 ) {
262 ++hammer_count_fsyncs;
263 hammer_flusher_flush_undos(hmp, mode);
264 ip->redo_count = 0;
265 if (ip->vp && (ip->flags & HAMMER_INODE_MODMASK) == 0)
266 vclrisdirty(ip->vp);
267 lwkt_reltoken(&hmp->fs_token);
268 return(0);
269 }
270
271 /*
272 * REDO is enabled by fsync(), the idea being we really only
273 * want to lay down REDO records when programs are using
274 * fsync() heavily. The first fsync() on the file starts
275 * the gravy train going and later fsync()s keep it hot by
276 * resetting the redo_count.
277 *
278 * We weren't running REDOs before now so we have to fall
279 * through and do a full fsync of what we have.
280 */
281 if (hmp->version >= HAMMER_VOL_VERSION_FOUR &&
282 (hmp->flags & HAMMER_MOUNT_REDO_RECOVERY_RUN) == 0) {
283 ip->flags |= HAMMER_INODE_REDO;
284 ip->redo_count = 0;
285 }
286 }
287 skip:
288
289 /*
290 * Do a full flush sequence.
291 *
292 * Attempt to release the vnode while waiting for the inode to
293 * finish flushing. This can really mess up inactive->reclaim
294 * sequences so only do it if the vnode is active.
295 *
296 * WARNING! The VX lock functions must be used. vn_lock() will
297 * fail when this is part of a VOP_RECLAIM sequence.
298 */
299 ++hammer_count_fsyncs;
300 vfsync(ap->a_vp, waitfor, 1, NULL, NULL);
301 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
302 if (waitfor == MNT_WAIT) {
303 int dorelock;
304
305 if ((ap->a_vp->v_flag & VRECLAIMED) == 0) {
306 vx_unlock(ap->a_vp);
307 dorelock = 1;
308 } else {
309 dorelock = 0;
310 }
311 hammer_wait_inode(ip);
312 if (dorelock)
313 vx_lock(ap->a_vp);
314 }
315 if (ip->vp && (ip->flags & HAMMER_INODE_MODMASK) == 0)
316 vclrisdirty(ip->vp);
317 lwkt_reltoken(&hmp->fs_token);
318 return (ip->error);
319 }
320
321 /*
322 * hammer_vop_read { vp, uio, ioflag, cred }
323 *
324 * MPSAFE (for the cache safe does not require fs_token)
325 */
326 static
327 int
328 hammer_vop_read(struct vop_read_args *ap)
329 {
330 struct hammer_transaction trans;
331 hammer_inode_t ip;
332 hammer_mount_t hmp;
333 off_t offset;
334 struct buf *bp;
335 struct uio *uio;
336 int error;
337 int n;
338 int seqcount;
339 int ioseqcount;
340 int blksize;
341 int bigread;
342 int got_trans;
343 size_t resid;
344
345 if (ap->a_vp->v_type != VREG)
346 return (EINVAL);
347 ip = VTOI(ap->a_vp);
348 hmp = ip->hmp;
349 error = 0;
350 got_trans = 0;
351 uio = ap->a_uio;
352
353 /*
354 * Attempt to shortcut directly to the VM object using lwbufs.
355 * This is much faster than instantiating buffer cache buffers.
356 */
357 resid = uio->uio_resid;
358 error = vop_helper_read_shortcut(ap);
359 hammer_stats_file_read += resid - uio->uio_resid;
360 if (error)
361 return (error);
362 if (uio->uio_resid == 0)
363 goto finished;
364
365 /*
366 * Allow the UIO's size to override the sequential heuristic.
367 */
368 blksize = hammer_blocksize(uio->uio_offset);
369 seqcount = (uio->uio_resid + (BKVASIZE - 1)) / BKVASIZE;
370 ioseqcount = (ap->a_ioflag >> 16);
371 if (seqcount < ioseqcount)
372 seqcount = ioseqcount;
373
374 /*
375 * If reading or writing a huge amount of data we have to break
376 * atomicy and allow the operation to be interrupted by a signal
377 * or it can DOS the machine.
378 */
379 bigread = (uio->uio_resid > 100 * 1024 * 1024);
380
381 /*
382 * Access the data typically in HAMMER_BUFSIZE blocks via the
383 * buffer cache, but HAMMER may use a variable block size based
384 * on the offset.
385 *
386 * XXX Temporary hack, delay the start transaction while we remain
387 * MPSAFE. NOTE: ino_data.size cannot change while vnode is
388 * locked-shared.
389 */
390 while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
391 int64_t base_offset;
392 int64_t file_limit;
393
394 blksize = hammer_blocksize(uio->uio_offset);
395 offset = (int)uio->uio_offset & (blksize - 1);
396 base_offset = uio->uio_offset - offset;
397
398 if (bigread && (error = hammer_signal_check(ip->hmp)) != 0)
399 break;
400
401 /*
402 * MPSAFE
403 */
404 bp = getblk(ap->a_vp, base_offset, blksize, 0, 0);
405 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == B_CACHE) {
406 bp->b_flags &= ~B_AGE;
407 error = 0;
408 goto skip;
409 }
410 if (ap->a_ioflag & IO_NRDELAY) {
411 bqrelse(bp);
412 return (EWOULDBLOCK);
413 }
414
415 /*
416 * MPUNSAFE
417 */
418 if (got_trans == 0) {
419 hammer_start_transaction(&trans, ip->hmp);
420 got_trans = 1;
421 }
422
423 /*
424 * NOTE: A valid bp has already been acquired, but was not
425 * B_CACHE.
426 */
427 if (hammer_cluster_enable) {
428 /*
429 * Use file_limit to prevent cluster_read() from
430 * creating buffers of the wrong block size past
431 * the demarc.
432 */
433 file_limit = ip->ino_data.size;
434 if (base_offset < HAMMER_XDEMARC &&
435 file_limit > HAMMER_XDEMARC) {
436 file_limit = HAMMER_XDEMARC;
437 }
438 error = cluster_readx(ap->a_vp,
439 file_limit, base_offset,
440 blksize, uio->uio_resid,
441 seqcount * BKVASIZE, &bp);
442 } else {
443 error = breadnx(ap->a_vp, base_offset, blksize,
444 NULL, NULL, 0, &bp);
445 }
446 if (error) {
447 brelse(bp);
448 break;
449 }
450 skip:
451 if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IODEBUG)) {
452 kprintf("doff %016jx read file %016jx@%016jx\n",
453 (intmax_t)bp->b_bio2.bio_offset,
454 (intmax_t)ip->obj_id,
455 (intmax_t)bp->b_loffset);
456 }
457 bp->b_flags &= ~B_IODEBUG;
458 if (blksize == HAMMER_XBUFSIZE)
459 bp->b_flags |= B_CLUSTEROK;
460
461 n = blksize - offset;
462 if (n > uio->uio_resid)
463 n = uio->uio_resid;
464 if (n > ip->ino_data.size - uio->uio_offset)
465 n = (int)(ip->ino_data.size - uio->uio_offset);
466
467 /*
468 * Set B_AGE, data has a lower priority than meta-data.
469 *
470 * Use a hold/unlock/drop sequence to run the uiomove
471 * with the buffer unlocked, avoiding deadlocks against
472 * read()s on mmap()'d spaces.
473 */
474 bp->b_flags |= B_AGE;
475 error = uiomovebp(bp, (char *)bp->b_data + offset, n, uio);
476 bqrelse(bp);
477
478 if (error)
479 break;
480 hammer_stats_file_read += n;
481 }
482
483 finished:
484
485 /*
486 * Try to update the atime with just the inode lock for maximum
487 * concurrency. If we can't shortcut it we have to get the full
488 * blown transaction.
489 */
490 if (got_trans == 0 && hammer_update_atime_quick(ip) < 0) {
491 hammer_start_transaction(&trans, ip->hmp);
492 got_trans = 1;
493 }
494
495 if (got_trans) {
496 if ((ip->flags & HAMMER_INODE_RO) == 0 &&
497 (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
498 lwkt_gettoken(&hmp->fs_token);
499 ip->ino_data.atime = trans.time;
500 hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
501 hammer_done_transaction(&trans);
502 lwkt_reltoken(&hmp->fs_token);
503 } else {
504 hammer_done_transaction(&trans);
505 }
506 }
507 return (error);
508 }
509
510 /*
511 * hammer_vop_write { vp, uio, ioflag, cred }
512 */
513 static
514 int
515 hammer_vop_write(struct vop_write_args *ap)
516 {
517 struct hammer_transaction trans;
518 struct hammer_inode *ip;
519 hammer_mount_t hmp;
520 thread_t td;
521 struct uio *uio;
522 int offset;
523 off_t base_offset;
524 int64_t cluster_eof;
525 struct buf *bp;
526 int kflags;
527 int error;
528 int n;
529 int flags;
530 int seqcount;
531 int bigwrite;
532
533 if (ap->a_vp->v_type != VREG)
534 return (EINVAL);
535 ip = VTOI(ap->a_vp);
536 hmp = ip->hmp;
537 error = 0;
538 kflags = 0;
539 seqcount = ap->a_ioflag >> 16;
540
541 if (ip->flags & HAMMER_INODE_RO)
542 return (EROFS);
543
544 /*
545 * Create a transaction to cover the operations we perform.
546 */
547 hammer_start_transaction(&trans, hmp);
548 uio = ap->a_uio;
549
550 /*
551 * Check append mode
552 */
553 if (ap->a_ioflag & IO_APPEND)
554 uio->uio_offset = ip->ino_data.size;
555
556 /*
557 * Check for illegal write offsets. Valid range is 0...2^63-1.
558 *
559 * NOTE: the base_off assignment is required to work around what
560 * I consider to be a GCC-4 optimization bug.
561 */
562 if (uio->uio_offset < 0) {
563 hammer_done_transaction(&trans);
564 return (EFBIG);
565 }
566 base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
567 if (uio->uio_resid > 0 && base_offset <= uio->uio_offset) {
568 hammer_done_transaction(&trans);
569 return (EFBIG);
570 }
571
572 if (uio->uio_resid > 0 && (td = uio->uio_td) != NULL && td->td_proc &&
573 base_offset > td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
574 hammer_done_transaction(&trans);
575 lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ);
576 return (EFBIG);
577 }
578
579 /*
580 * If reading or writing a huge amount of data we have to break
581 * atomicy and allow the operation to be interrupted by a signal
582 * or it can DOS the machine.
583 *
584 * Preset redo_count so we stop generating REDOs earlier if the
585 * limit is exceeded.
586 *
587 * redo_count is heuristical, SMP races are ok
588 */
589 bigwrite = (uio->uio_resid > 100 * 1024 * 1024);
590 if ((ip->flags & HAMMER_INODE_REDO) &&
591 ip->redo_count < hammer_limit_redo) {
592 ip->redo_count += uio->uio_resid;
593 }
594
595 /*
596 * Access the data typically in HAMMER_BUFSIZE blocks via the
597 * buffer cache, but HAMMER may use a variable block size based
598 * on the offset.
599 */
600 while (uio->uio_resid > 0) {
601 int fixsize = 0;
602 int blksize;
603 int blkmask;
604 int trivial;
605 int endofblk;
606 off_t nsize;
607
608 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
609 break;
610 if (bigwrite && (error = hammer_signal_check(hmp)) != 0)
611 break;
612
613 blksize = hammer_blocksize(uio->uio_offset);
614
615 /*
616 * Control the number of pending records associated with
617 * this inode. If too many have accumulated start a
618 * flush. Try to maintain a pipeline with the flusher.
619 *
620 * NOTE: It is possible for other sources to grow the
621 * records but not necessarily issue another flush,
622 * so use a timeout and ensure that a re-flush occurs.
623 */
624 if (ip->rsv_recs >= hammer_limit_inode_recs) {
625 lwkt_gettoken(&hmp->fs_token);
626 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
627 while (ip->rsv_recs >= hammer_limit_inode_recs * 2) {
628 ip->flags |= HAMMER_INODE_RECSW;
629 tsleep(&ip->rsv_recs, 0, "hmrwww", hz);
630 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
631 }
632 lwkt_reltoken(&hmp->fs_token);
633 }
634
635 /*
636 * Do not allow HAMMER to blow out the buffer cache. Very
637 * large UIOs can lockout other processes due to bwillwrite()
638 * mechanics.
639 *
640 * The hammer inode is not locked during these operations.
641 * The vnode is locked which can interfere with the pageout
642 * daemon for non-UIO_NOCOPY writes but should not interfere
643 * with the buffer cache. Even so, we cannot afford to
644 * allow the pageout daemon to build up too many dirty buffer
645 * cache buffers.
646 *
647 * Only call this if we aren't being recursively called from
648 * a virtual disk device (vn), else we may deadlock.
649 */
650 if ((ap->a_ioflag & IO_RECURSE) == 0)
651 bwillwrite(blksize);
652
653 /*
654 * Calculate the blocksize at the current offset and figure
655 * out how much we can actually write.
656 */
657 blkmask = blksize - 1;
658 offset = (int)uio->uio_offset & blkmask;
659 base_offset = uio->uio_offset & ~(int64_t)blkmask;
660 n = blksize - offset;
661 if (n > uio->uio_resid) {
662 n = uio->uio_resid;
663 endofblk = 0;
664 } else {
665 endofblk = 1;
666 }
667 nsize = uio->uio_offset + n;
668 if (nsize > ip->ino_data.size) {
669 if (uio->uio_offset > ip->ino_data.size)
670 trivial = 0;
671 else
672 trivial = 1;
673 nvextendbuf(ap->a_vp,
674 ip->ino_data.size,
675 nsize,
676 hammer_blocksize(ip->ino_data.size),
677 hammer_blocksize(nsize),
678 hammer_blockoff(ip->ino_data.size),
679 hammer_blockoff(nsize),
680 trivial);
681 fixsize = 1;
682 kflags |= NOTE_EXTEND;
683 }
684
685 if (uio->uio_segflg == UIO_NOCOPY) {
686 /*
687 * Issuing a write with the same data backing the
688 * buffer. Instantiate the buffer to collect the
689 * backing vm pages, then read-in any missing bits.
690 *
691 * This case is used by vop_stdputpages().
692 */
693 bp = getblk(ap->a_vp, base_offset,
694 blksize, GETBLK_BHEAVY, 0);
695 if ((bp->b_flags & B_CACHE) == 0) {
696 bqrelse(bp);
697 error = bread(ap->a_vp, base_offset,
698 blksize, &bp);
699 }
700 } else if (offset == 0 && uio->uio_resid >= blksize) {
701 /*
702 * Even though we are entirely overwriting the buffer
703 * we may still have to zero it out to avoid a
704 * mmap/write visibility issue.
705 */
706 bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
707 if ((bp->b_flags & B_CACHE) == 0)
708 vfs_bio_clrbuf(bp);
709 } else if (base_offset >= ip->ino_data.size) {
710 /*
711 * If the base offset of the buffer is beyond the
712 * file EOF, we don't have to issue a read.
713 */
714 bp = getblk(ap->a_vp, base_offset,
715 blksize, GETBLK_BHEAVY, 0);
716 vfs_bio_clrbuf(bp);
717 } else {
718 /*
719 * Partial overwrite, read in any missing bits then
720 * replace the portion being written.
721 */
722 error = bread(ap->a_vp, base_offset, blksize, &bp);
723 if (error == 0)
724 bheavy(bp);
725 }
726 if (error == 0)
727 error = uiomovebp(bp, bp->b_data + offset, n, uio);
728
729 lwkt_gettoken(&hmp->fs_token);
730
731 /*
732 * Generate REDO records if enabled and redo_count will not
733 * exceeded the limit.
734 *
735 * If redo_count exceeds the limit we stop generating records
736 * and clear HAMMER_INODE_REDO. This will cause the next
737 * fsync() to do a full meta-data sync instead of just an
738 * UNDO/REDO fifo update.
739 *
740 * When clearing HAMMER_INODE_REDO any pre-existing REDOs
741 * will still be tracked. The tracks will be terminated
742 * when the related meta-data (including possible data
743 * modifications which are not tracked via REDO) is
744 * flushed.
745 */
746 if ((ip->flags & HAMMER_INODE_REDO) && error == 0) {
747 if (ip->redo_count < hammer_limit_redo) {
748 bp->b_flags |= B_VFSFLAG1;
749 error = hammer_generate_redo(&trans, ip,
750 base_offset + offset,
751 HAMMER_REDO_WRITE,
752 bp->b_data + offset,
753 (size_t)n);
754 } else {
755 ip->flags &= ~HAMMER_INODE_REDO;
756 }
757 }
758
759 /*
760 * If we screwed up we have to undo any VM size changes we
761 * made.
762 */
763 if (error) {
764 brelse(bp);
765 if (fixsize) {
766 nvtruncbuf(ap->a_vp, ip->ino_data.size,
767 hammer_blocksize(ip->ino_data.size),
768 hammer_blockoff(ip->ino_data.size),
769 0);
770 }
771 break;
772 }
773 kflags |= NOTE_WRITE;
774 hammer_stats_file_write += n;
775 if (blksize == HAMMER_XBUFSIZE)
776 bp->b_flags |= B_CLUSTEROK;
777 if (ip->ino_data.size < uio->uio_offset) {
778 ip->ino_data.size = uio->uio_offset;
779 flags = HAMMER_INODE_SDIRTY;
780 } else {
781 flags = 0;
782 }
783 ip->ino_data.mtime = trans.time;
784 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
785 hammer_modify_inode(&trans, ip, flags);
786
787 /*
788 * Once we dirty the buffer any cached zone-X offset
789 * becomes invalid. HAMMER NOTE: no-history mode cannot
790 * allow overwriting over the same data sector unless
791 * we provide UNDOs for the old data, which we don't.
792 */
793 bp->b_bio2.bio_offset = NOOFFSET;
794
795 lwkt_reltoken(&hmp->fs_token);
796
797 /*
798 * Final buffer disposition.
799 *
800 * Because meta-data updates are deferred, HAMMER is
801 * especially sensitive to excessive bdwrite()s because
802 * the I/O stream is not broken up by disk reads. So the
803 * buffer cache simply cannot keep up.
804 *
805 * WARNING! blksize is variable. cluster_write() is
806 * expected to not blow up if it encounters
807 * buffers that do not match the passed blksize.
808 *
809 * NOTE! Hammer shouldn't need to bawrite()/cluster_write().
810 * The ip->rsv_recs check should burst-flush the data.
811 * If we queue it immediately the buf could be left
812 * locked on the device queue for a very long time.
813 *
814 * However, failing to flush a dirty buffer out when
815 * issued from the pageout daemon can result in a low
816 * memory deadlock against bio_page_alloc(), so we
817 * have to bawrite() on IO_ASYNC as well.
818 *
819 * NOTE! To avoid degenerate stalls due to mismatched block
820 * sizes we only honor IO_DIRECT on the write which
821 * abuts the end of the buffer. However, we must
822 * honor IO_SYNC in case someone is silly enough to
823 * configure a HAMMER file as swap, or when HAMMER
824 * is serving NFS (for commits). Ick ick.
825 */
826 bp->b_flags |= B_AGE;
827 if (blksize == HAMMER_XBUFSIZE)
828 bp->b_flags |= B_CLUSTEROK;
829
830 if (ap->a_ioflag & IO_SYNC) {
831 bwrite(bp);
832 } else if ((ap->a_ioflag & IO_DIRECT) && endofblk) {
833 bawrite(bp);
834 } else if (ap->a_ioflag & IO_ASYNC) {
835 bawrite(bp);
836 } else if (hammer_cluster_enable &&
837 !(ap->a_vp->v_mount->mnt_flag & MNT_NOCLUSTERW)) {
838 if (base_offset < HAMMER_XDEMARC)
839 cluster_eof = hammer_blockdemarc(base_offset,
840 ip->ino_data.size);
841 else
842 cluster_eof = ip->ino_data.size;
843 cluster_write(bp, cluster_eof, blksize, seqcount);
844 } else {
845 bdwrite(bp);
846 }
847 }
848 hammer_done_transaction(&trans);
849 hammer_knote(ap->a_vp, kflags);
850
851 return (error);
852 }
853
854 /*
855 * hammer_vop_access { vp, mode, cred }
856 *
857 * MPSAFE - does not require fs_token
858 */
859 static
860 int
861 hammer_vop_access(struct vop_access_args *ap)
862 {
863 struct hammer_inode *ip = VTOI(ap->a_vp);
864 uid_t uid;
865 gid_t gid;
866 int error;
867
868 ++hammer_stats_file_iopsr;
869 uid = hammer_to_unix_xid(&ip->ino_data.uid);
870 gid = hammer_to_unix_xid(&ip->ino_data.gid);
871
872 error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
873 ip->ino_data.uflags);
874 return (error);
875 }
876
877 /*
878 * hammer_vop_advlock { vp, id, op, fl, flags }
879 *
880 * MPSAFE - does not require fs_token
881 */
882 static
883 int
884 hammer_vop_advlock(struct vop_advlock_args *ap)
885 {
886 hammer_inode_t ip = VTOI(ap->a_vp);
887
888 return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
889 }
890
891 /*
892 * hammer_vop_close { vp, fflag }
893 *
894 * We can only sync-on-close for normal closes. XXX disabled for now.
895 */
896 static
897 int
898 hammer_vop_close(struct vop_close_args *ap)
899 {
900 #if 0
901 struct vnode *vp = ap->a_vp;
902 hammer_inode_t ip = VTOI(vp);
903 int waitfor;
904 if (ip->flags & (HAMMER_INODE_CLOSESYNC|HAMMER_INODE_CLOSEASYNC)) {
905 if (vn_islocked(vp) == LK_EXCLUSIVE &&
906 (vp->v_flag & (VINACTIVE|VRECLAIMED)) == 0) {
907 if (ip->flags & HAMMER_INODE_CLOSESYNC)
908 waitfor = MNT_WAIT;
909 else
910 waitfor = MNT_NOWAIT;
911 ip->flags &= ~(HAMMER_INODE_CLOSESYNC |
912 HAMMER_INODE_CLOSEASYNC);
913 VOP_FSYNC(vp, MNT_NOWAIT, waitfor);
914 }
915 }
916 #endif
917 return (vop_stdclose(ap));
918 }
919
920 /*
921 * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
922 *
923 * The operating system has already ensured that the directory entry
924 * does not exist and done all appropriate namespace locking.
925 */
926 static
927 int
928 hammer_vop_ncreate(struct vop_ncreate_args *ap)
929 {
930 struct hammer_transaction trans;
931 struct hammer_inode *dip;
932 struct hammer_inode *nip;
933 struct nchandle *nch;
934 hammer_mount_t hmp;
935 int error;
936
937 nch = ap->a_nch;
938 dip = VTOI(ap->a_dvp);
939 hmp = dip->hmp;
940
941 if (dip->flags & HAMMER_INODE_RO)
942 return (EROFS);
943 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
944 return (error);
945
946 /*
947 * Create a transaction to cover the operations we perform.
948 */
949 lwkt_gettoken(&hmp->fs_token);
950 hammer_start_transaction(&trans, hmp);
951 ++hammer_stats_file_iopsw;
952
953 /*
954 * Create a new filesystem object of the requested type. The
955 * returned inode will be referenced and shared-locked to prevent
956 * it from being moved to the flusher.
957 */
958 error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
959 dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
960 NULL, &nip);
961 if (error) {
962 hkprintf("hammer_create_inode error %d\n", error);
963 hammer_done_transaction(&trans);
964 *ap->a_vpp = NULL;
965 lwkt_reltoken(&hmp->fs_token);
966 return (error);
967 }
968
969 /*
970 * Add the new filesystem object to the directory. This will also
971 * bump the inode's link count.
972 */
973 error = hammer_ip_add_directory(&trans, dip,
974 nch->ncp->nc_name, nch->ncp->nc_nlen,
975 nip);
976 if (error)
977 hkprintf("hammer_ip_add_directory error %d\n", error);
978
979 /*
980 * Finish up.
981 */
982 if (error) {
983 hammer_rel_inode(nip, 0);
984 hammer_done_transaction(&trans);
985 *ap->a_vpp = NULL;
986 } else {
987 error = hammer_get_vnode(nip, ap->a_vpp);
988 hammer_done_transaction(&trans);
989 hammer_rel_inode(nip, 0);
990 if (error == 0) {
991 cache_setunresolved(ap->a_nch);
992 cache_setvp(ap->a_nch, *ap->a_vpp);
993 }
994 hammer_knote(ap->a_dvp, NOTE_WRITE);
995 }
996 lwkt_reltoken(&hmp->fs_token);
997 return (error);
998 }
999
1000 /*
1001 * hammer_vop_getattr { vp, vap }
1002 *
1003 * Retrieve an inode's attribute information. When accessing inodes
1004 * historically we fake the atime field to ensure consistent results.
1005 * The atime field is stored in the B-Tree element and allowed to be
1006 * updated without cycling the element.
1007 *
1008 * MPSAFE - does not require fs_token
1009 */
1010 static
1011 int
1012 hammer_vop_getattr(struct vop_getattr_args *ap)
1013 {
1014 struct hammer_inode *ip = VTOI(ap->a_vp);
1015 struct vattr *vap = ap->a_vap;
1016
1017 /*
1018 * We want the fsid to be different when accessing a filesystem
1019 * with different as-of's so programs like diff don't think
1020 * the files are the same.
1021 *
1022 * We also want the fsid to be the same when comparing snapshots,
1023 * or when comparing mirrors (which might be backed by different
1024 * physical devices). HAMMER fsids are based on the PFS's
1025 * shared_uuid field.
1026 *
1027 * XXX there is a chance of collision here. The va_fsid reported
1028 * by stat is different from the more involved fsid used in the
1029 * mount structure.
1030 */
1031 ++hammer_stats_file_iopsr;
1032 hammer_lock_sh(&ip->lock);
1033 vap->va_fsid = ip->pfsm->fsid_udev ^ (u_int32_t)ip->obj_asof ^
1034 (u_int32_t)(ip->obj_asof >> 32);
1035
1036 vap->va_fileid = ip->ino_leaf.base.obj_id;
1037 vap->va_mode = ip->ino_data.mode;
1038 vap->va_nlink = ip->ino_data.nlinks;
1039 vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1040 vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1041 vap->va_rmajor = 0;
1042 vap->va_rminor = 0;
1043 vap->va_size = ip->ino_data.size;
1044
1045 /*
1046 * Special case for @@PFS softlinks. The actual size of the
1047 * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
1048 * or for MAX_TID is "@@-1:%05d" == 10 bytes.
1049 */
1050 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
1051 ip->ino_data.size == 10 &&
1052 ip->obj_asof == HAMMER_MAX_TID &&
1053 ip->obj_localization == 0 &&
1054 strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
1055 if (ip->pfsm->pfsd.mirror_flags & HAMMER_PFSD_SLAVE)
1056 vap->va_size = 26;
1057 else
1058 vap->va_size = 10;
1059 }
1060
1061 /*
1062 * We must provide a consistent atime and mtime for snapshots
1063 * so people can do a 'tar cf - ... | md5' on them and get
1064 * consistent results.
1065 */
1066 if (ip->flags & HAMMER_INODE_RO) {
1067 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
1068 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
1069 } else {
1070 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
1071 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
1072 }
1073 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
1074 vap->va_flags = ip->ino_data.uflags;
1075 vap->va_gen = 1; /* hammer inums are unique for all time */
1076 vap->va_blocksize = HAMMER_BUFSIZE;
1077 if (ip->ino_data.size >= HAMMER_XDEMARC) {
1078 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
1079 ~HAMMER_XBUFMASK64;
1080 } else if (ip->ino_data.size > HAMMER_BUFSIZE / 2) {
1081 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
1082 ~HAMMER_BUFMASK64;
1083 } else {
1084 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
1085 }
1086
1087 vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
1088 vap->va_filerev = 0; /* XXX */
1089 vap->va_uid_uuid = ip->ino_data.uid;
1090 vap->va_gid_uuid = ip->ino_data.gid;
1091 vap->va_fsid_uuid = ip->hmp->fsid;
1092 vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
1093 VA_FSID_UUID_VALID;
1094
1095 switch (ip->ino_data.obj_type) {
1096 case HAMMER_OBJTYPE_CDEV:
1097 case HAMMER_OBJTYPE_BDEV:
1098 vap->va_rmajor = ip->ino_data.rmajor;
1099 vap->va_rminor = ip->ino_data.rminor;
1100 break;
1101 default:
1102 break;
1103 }
1104 hammer_unlock(&ip->lock);
1105 return(0);
1106 }
1107
1108 /*
1109 * hammer_vop_nresolve { nch, dvp, cred }
1110 *
1111 * Locate the requested directory entry.
1112 */
1113 static
1114 int
1115 hammer_vop_nresolve(struct vop_nresolve_args *ap)
1116 {
1117 struct hammer_transaction trans;
1118 struct namecache *ncp;
1119 hammer_mount_t hmp;
1120 hammer_inode_t dip;
1121 hammer_inode_t ip;
1122 hammer_tid_t asof;
1123 struct hammer_cursor cursor;
1124 struct vnode *vp;
1125 int64_t namekey;
1126 int error;
1127 int i;
1128 int nlen;
1129 int flags;
1130 int ispfs;
1131 int64_t obj_id;
1132 u_int32_t localization;
1133 u_int32_t max_iterations;
1134
1135 /*
1136 * Misc initialization, plus handle as-of name extensions. Look for
1137 * the '@@' extension. Note that as-of files and directories cannot
1138 * be modified.
1139 */
1140 dip = VTOI(ap->a_dvp);
1141 ncp = ap->a_nch->ncp;
1142 asof = dip->obj_asof;
1143 localization = dip->obj_localization; /* for code consistency */
1144 nlen = ncp->nc_nlen;
1145 flags = dip->flags & HAMMER_INODE_RO;
1146 ispfs = 0;
1147 hmp = dip->hmp;
1148
1149 lwkt_gettoken(&hmp->fs_token);
1150 hammer_simple_transaction(&trans, hmp);
1151 ++hammer_stats_file_iopsr;
1152
1153 for (i = 0; i < nlen; ++i) {
1154 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
1155 error = hammer_str_to_tid(ncp->nc_name + i + 2,
1156 &ispfs, &asof, &localization);
1157 if (error != 0) {
1158 i = nlen;
1159 break;
1160 }
1161 if (asof != HAMMER_MAX_TID)
1162 flags |= HAMMER_INODE_RO;
1163 break;
1164 }
1165 }
1166 nlen = i;
1167
1168 /*
1169 * If this is a PFS softlink we dive into the PFS
1170 */
1171 if (ispfs && nlen == 0) {
1172 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1173 asof, localization,
1174 flags, &error);
1175 if (error == 0) {
1176 error = hammer_get_vnode(ip, &vp);
1177 hammer_rel_inode(ip, 0);
1178 } else {
1179 vp = NULL;
1180 }
1181 if (error == 0) {
1182 vn_unlock(vp);
1183 cache_setvp(ap->a_nch, vp);
1184 vrele(vp);
1185 }
1186 goto done;
1187 }
1188
1189 /*
1190 * If there is no path component the time extension is relative to dip.
1191 * e.g. "fubar/@@<snapshot>"
1192 *
1193 * "." is handled by the kernel, but ".@@<snapshot>" is not.
1194 * e.g. "fubar/.@@<snapshot>"
1195 *
1196 * ".." is handled by the kernel. We do not currently handle
1197 * "..@<snapshot>".
1198 */
1199 if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1200 ip = hammer_get_inode(&trans, dip, dip->obj_id,
1201 asof, dip->obj_localization,
1202 flags, &error);
1203 if (error == 0) {
1204 error = hammer_get_vnode(ip, &vp);
1205 hammer_rel_inode(ip, 0);
1206 } else {
1207 vp = NULL;
1208 }
1209 if (error == 0) {
1210 vn_unlock(vp);
1211 cache_setvp(ap->a_nch, vp);
1212 vrele(vp);
1213 }
1214 goto done;
1215 }
1216
1217 /*
1218 * Calculate the namekey and setup the key range for the scan. This
1219 * works kinda like a chained hash table where the lower 32 bits
1220 * of the namekey synthesize the chain.
1221 *
1222 * The key range is inclusive of both key_beg and key_end.
1223 */
1224 namekey = hammer_directory_namekey(dip, ncp->nc_name, nlen,
1225 &max_iterations);
1226
1227 error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1228 cursor.key_beg.localization = dip->obj_localization +
1229 hammer_dir_localization(dip);
1230 cursor.key_beg.obj_id = dip->obj_id;
1231 cursor.key_beg.key = namekey;
1232 cursor.key_beg.create_tid = 0;
1233 cursor.key_beg.delete_tid = 0;
1234 cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1235 cursor.key_beg.obj_type = 0;
1236
1237 cursor.key_end = cursor.key_beg;
1238 cursor.key_end.key += max_iterations;
1239 cursor.asof = asof;
1240 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1241
1242 /*
1243 * Scan all matching records (the chain), locate the one matching
1244 * the requested path component.
1245 *
1246 * The hammer_ip_*() functions merge in-memory records with on-disk
1247 * records for the purposes of the search.
1248 */
1249 obj_id = 0;
1250 localization = HAMMER_DEF_LOCALIZATION;
1251
1252 if (error == 0) {
1253 error = hammer_ip_first(&cursor);
1254 while (error == 0) {
1255 error = hammer_ip_resolve_data(&cursor);
1256 if (error)
1257 break;
1258 if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1259 bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1260 obj_id = cursor.data->entry.obj_id;
1261 localization = cursor.data->entry.localization;
1262 break;
1263 }
1264 error = hammer_ip_next(&cursor);
1265 }
1266 }
1267 hammer_done_cursor(&cursor);
1268
1269 /*
1270 * Lookup the obj_id. This should always succeed. If it does not
1271 * the filesystem may be damaged and we return a dummy inode.
1272 */
1273 if (error == 0) {
1274 ip = hammer_get_inode(&trans, dip, obj_id,
1275 asof, localization,
1276 flags, &error);
1277 if (error == ENOENT) {
1278 kprintf("HAMMER: WARNING: Missing "
1279 "inode for dirent \"%s\"\n"
1280 "\tobj_id = %016llx, asof=%016llx, lo=%08x\n",
1281 ncp->nc_name,
1282 (long long)obj_id, (long long)asof,
1283 localization);
1284 error = 0;
1285 ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1286 asof, localization,
1287 flags, &error);
1288 }
1289 if (error == 0) {
1290 error = hammer_get_vnode(ip, &vp);
1291 hammer_rel_inode(ip, 0);
1292 } else {
1293 vp = NULL;
1294 }
1295 if (error == 0) {
1296 vn_unlock(vp);
1297 cache_setvp(ap->a_nch, vp);
1298 vrele(vp);
1299 }
1300 } else if (error == ENOENT) {
1301 cache_setvp(ap->a_nch, NULL);
1302 }
1303 done:
1304 hammer_done_transaction(&trans);
1305 lwkt_reltoken(&hmp->fs_token);
1306 return (error);
1307 }
1308
1309 /*
1310 * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1311 *
1312 * Locate the parent directory of a directory vnode.
1313 *
1314 * dvp is referenced but not locked. *vpp must be returned referenced and
1315 * locked. A parent_obj_id of 0 does not necessarily indicate that we are
1316 * at the root, instead it could indicate that the directory we were in was
1317 * removed.
1318 *
1319 * NOTE: as-of sequences are not linked into the directory structure. If
1320 * we are at the root with a different asof then the mount point, reload
1321 * the same directory with the mount point's asof. I'm not sure what this
1322 * will do to NFS. We encode ASOF stamps in NFS file handles so it might not
1323 * get confused, but it hasn't been tested.
1324 */
1325 static
1326 int
1327 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1328 {
1329 struct hammer_transaction trans;
1330 struct hammer_inode *dip;
1331 struct hammer_inode *ip;
1332 hammer_mount_t hmp;
1333 int64_t parent_obj_id;
1334 u_int32_t parent_obj_localization;
1335 hammer_tid_t asof;
1336 int error;
1337
1338 dip = VTOI(ap->a_dvp);
1339 asof = dip->obj_asof;
1340 hmp = dip->hmp;
1341
1342 /*
1343 * Whos are parent? This could be the root of a pseudo-filesystem
1344 * whos parent is in another localization domain.
1345 */
1346 lwkt_gettoken(&hmp->fs_token);
1347 parent_obj_id = dip->ino_data.parent_obj_id;
1348 if (dip->obj_id == HAMMER_OBJID_ROOT)
1349 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
1350 else
1351 parent_obj_localization = dip->obj_localization;
1352
1353 if (parent_obj_id == 0) {
1354 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1355 asof != hmp->asof) {
1356 parent_obj_id = dip->obj_id;
1357 asof = hmp->asof;
1358 *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1359 ksnprintf(*ap->a_fakename, 19, "0x%016llx",
1360 (long long)dip->obj_asof);
1361 } else {
1362 *ap->a_vpp = NULL;
1363 lwkt_reltoken(&hmp->fs_token);
1364 return ENOENT;
1365 }
1366 }
1367
1368 hammer_simple_transaction(&trans, hmp);
1369 ++hammer_stats_file_iopsr;
1370
1371 ip = hammer_get_inode(&trans, dip, parent_obj_id,
1372 asof, parent_obj_localization,
1373 dip->flags, &error);
1374 if (ip) {
1375 error = hammer_get_vnode(ip, ap->a_vpp);
1376 hammer_rel_inode(ip, 0);
1377 } else {
1378 *ap->a_vpp = NULL;
1379 }
1380 hammer_done_transaction(&trans);
1381 lwkt_reltoken(&hmp->fs_token);
1382 return (error);
1383 }
1384
1385 /*
1386 * hammer_vop_nlink { nch, dvp, vp, cred }
1387 */
1388 static
1389 int
1390 hammer_vop_nlink(struct vop_nlink_args *ap)
1391 {
1392 struct hammer_transaction trans;
1393 struct hammer_inode *dip;
1394 struct hammer_inode *ip;
1395 struct nchandle *nch;
1396 hammer_mount_t hmp;
1397 int error;
1398
1399 if (ap->a_dvp->v_mount != ap->a_vp->v_mount)
1400 return(EXDEV);
1401
1402 nch = ap->a_nch;
1403 dip = VTOI(ap->a_dvp);
1404 ip = VTOI(ap->a_vp);
1405 hmp = dip->hmp;
1406
1407 if (dip->obj_localization != ip->obj_localization)
1408 return(EXDEV);
1409
1410 if (dip->flags & HAMMER_INODE_RO)
1411 return (EROFS);
1412 if (ip->flags & HAMMER_INODE_RO)
1413 return (EROFS);
1414 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1415 return (error);
1416
1417 /*
1418 * Create a transaction to cover the operations we perform.
1419 */
1420 lwkt_gettoken(&hmp->fs_token);
1421 hammer_start_transaction(&trans, hmp);
1422 ++hammer_stats_file_iopsw;
1423
1424 /*
1425 * Add the filesystem object to the directory. Note that neither
1426 * dip nor ip are referenced or locked, but their vnodes are
1427 * referenced. This function will bump the inode's link count.
1428 */
1429 error = hammer_ip_add_directory(&trans, dip,
1430 nch->ncp->nc_name, nch->ncp->nc_nlen,
1431 ip);
1432
1433 /*
1434 * Finish up.
1435 */
1436 if (error == 0) {
1437 cache_setunresolved(nch);
1438 cache_setvp(nch, ap->a_vp);
1439 }
1440 hammer_done_transaction(&trans);
1441 hammer_knote(ap->a_vp, NOTE_LINK);
1442 hammer_knote(ap->a_dvp, NOTE_WRITE);
1443 lwkt_reltoken(&hmp->fs_token);
1444 return (error);
1445 }
1446
1447 /*
1448 * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1449 *
1450 * The operating system has already ensured that the directory entry
1451 * does not exist and done all appropriate namespace locking.
1452 */
1453 static
1454 int
1455 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1456 {
1457 struct hammer_transaction trans;
1458 struct hammer_inode *dip;
1459 struct hammer_inode *nip;
1460 struct nchandle *nch;
1461 hammer_mount_t hmp;
1462 int error;
1463
1464 nch = ap->a_nch;
1465 dip = VTOI(ap->a_dvp);
1466 hmp = dip->hmp;
1467
1468 if (dip->flags & HAMMER_INODE_RO)
1469 return (EROFS);
1470 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1471 return (error);
1472
1473 /*
1474 * Create a transaction to cover the operations we perform.
1475 */
1476 lwkt_gettoken(&hmp->fs_token);
1477 hammer_start_transaction(&trans, hmp);
1478 ++hammer_stats_file_iopsw;
1479
1480 /*
1481 * Create a new filesystem object of the requested type. The
1482 * returned inode will be referenced but not locked.
1483 */
1484 error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1485 dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1486 NULL, &nip);
1487 if (error) {
1488 hkprintf("hammer_mkdir error %d\n", error);
1489 hammer_done_transaction(&trans);
1490 *ap->a_vpp = NULL;
1491 lwkt_reltoken(&hmp->fs_token);
1492 return (error);
1493 }
1494 /*
1495 * Add the new filesystem object to the directory. This will also
1496 * bump the inode's link count.
1497 */
1498 error = hammer_ip_add_directory(&trans, dip,
1499 nch->ncp->nc_name, nch->ncp->nc_nlen,
1500 nip);
1501 if (error)
1502 hkprintf("hammer_mkdir (add) error %d\n", error);
1503
1504 /*
1505 * Finish up.
1506 */
1507 if (error) {
1508 hammer_rel_inode(nip, 0);
1509 *ap->a_vpp = NULL;
1510 } else {
1511 error = hammer_get_vnode(nip, ap->a_vpp);
1512 hammer_rel_inode(nip, 0);
1513 if (error == 0) {
1514 cache_setunresolved(ap->a_nch);
1515 cache_setvp(ap->a_nch, *ap->a_vpp);
1516 }
1517 }
1518 hammer_done_transaction(&trans);
1519 if (error == 0)
1520 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1521 lwkt_reltoken(&hmp->fs_token);
1522 return (error);
1523 }
1524
1525 /*
1526 * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1527 *
1528 * The operating system has already ensured that the directory entry
1529 * does not exist and done all appropriate namespace locking.
1530 */
1531 static
1532 int
1533 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1534 {
1535 struct hammer_transaction trans;
1536 struct hammer_inode *dip;
1537 struct hammer_inode *nip;
1538 struct nchandle *nch;
1539 hammer_mount_t hmp;
1540 int error;
1541
1542 nch = ap->a_nch;
1543 dip = VTOI(ap->a_dvp);
1544 hmp = dip->hmp;
1545
1546 if (dip->flags & HAMMER_INODE_RO)
1547 return (EROFS);
1548 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1549 return (error);
1550
1551 /*
1552 * Create a transaction to cover the operations we perform.
1553 */
1554 lwkt_gettoken(&hmp->fs_token);
1555 hammer_start_transaction(&trans, hmp);
1556 ++hammer_stats_file_iopsw;
1557
1558 /*
1559 * Create a new filesystem object of the requested type. The
1560 * returned inode will be referenced but not locked.
1561 *
1562 * If mknod specifies a directory a pseudo-fs is created.
1563 */
1564 error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1565 dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1566 NULL, &nip);
1567 if (error) {
1568 hammer_done_transaction(&trans);
1569 *ap->a_vpp = NULL;
1570 lwkt_reltoken(&hmp->fs_token);
1571 return (error);
1572 }
1573
1574 /*
1575 * Add the new filesystem object to the directory. This will also
1576 * bump the inode's link count.
1577 */
1578 error = hammer_ip_add_directory(&trans, dip,
1579 nch->ncp->nc_name, nch->ncp->nc_nlen,
1580 nip);
1581
1582 /*
1583 * Finish up.
1584 */
1585 if (error) {
1586 hammer_rel_inode(nip, 0);
1587 *ap->a_vpp = NULL;
1588 } else {
1589 error = hammer_get_vnode(nip, ap->a_vpp);
1590 hammer_rel_inode(nip, 0);
1591 if (error == 0) {
1592 cache_setunresolved(ap->a_nch);
1593 cache_setvp(ap->a_nch, *ap->a_vpp);
1594 }
1595 }
1596 hammer_done_transaction(&trans);
1597 if (error == 0)
1598 hammer_knote(ap->a_dvp, NOTE_WRITE);
1599 lwkt_reltoken(&hmp->fs_token);
1600 return (error);
1601 }
1602
1603 /*
1604 * hammer_vop_open { vp, mode, cred, fp }
1605 *
1606 * MPSAFE (does not require fs_token)
1607 */
1608 static
1609 int
1610 hammer_vop_open(struct vop_open_args *ap)
1611 {
1612 hammer_inode_t ip;
1613
1614 ++hammer_stats_file_iopsr;
1615 ip = VTOI(ap->a_vp);
1616
1617 if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1618 return (EROFS);
1619 return(vop_stdopen(ap));
1620 }
1621
1622 /*
1623 * hammer_vop_print { vp }
1624 */
1625 static
1626 int
1627 hammer_vop_print(struct vop_print_args *ap)
1628 {
1629 return EOPNOTSUPP;
1630 }
1631
1632 /*
1633 * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1634 */
1635 static
1636 int
1637 hammer_vop_readdir(struct vop_readdir_args *ap)
1638 {
1639 struct hammer_transaction trans;
1640 struct hammer_cursor cursor;
1641 struct hammer_inode *ip;
1642 hammer_mount_t hmp;
1643 struct uio *uio;
1644 hammer_base_elm_t base;
1645 int error;
1646 int cookie_index;
1647 int ncookies;
1648 off_t *cookies;
1649 off_t saveoff;
1650 int r;
1651 int dtype;
1652
1653 ++hammer_stats_file_iopsr;
1654 ip = VTOI(ap->a_vp);
1655 uio = ap->a_uio;
1656 saveoff = uio->uio_offset;
1657 hmp = ip->hmp;
1658
1659 if (ap->a_ncookies) {
1660 ncookies = uio->uio_resid / 16 + 1;
1661 if (ncookies > 1024)
1662 ncookies = 1024;
1663 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1664 cookie_index = 0;
1665 } else {
1666 ncookies = -1;
1667 cookies = NULL;
1668 cookie_index = 0;
1669 }
1670
1671 lwkt_gettoken(&hmp->fs_token);
1672 hammer_simple_transaction(&trans, hmp);
1673
1674 /*
1675 * Handle artificial entries
1676 *
1677 * It should be noted that the minimum value for a directory
1678 * hash key on-media is 0x0000000100000000, so we can use anything
1679 * less then that to represent our 'special' key space.
1680 */
1681 error = 0;
1682 if (saveoff == 0) {
1683 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1684 if (r)
1685 goto done;
1686 if (cookies)
1687 cookies[cookie_index] = saveoff;
1688 ++saveoff;
1689 ++cookie_index;
1690 if (cookie_index == ncookies)
1691 goto done;
1692 }
1693 if (saveoff == 1) {
1694 if (ip->ino_data.parent_obj_id) {
1695 r = vop_write_dirent(&error, uio,
1696 ip->ino_data.parent_obj_id,
1697 DT_DIR, 2, "..");
1698 } else {
1699 r = vop_write_dirent(&error, uio,
1700 ip->obj_id, DT_DIR, 2, "..");
1701 }
1702 if (r)
1703 goto done;
1704 if (cookies)
1705 cookies[cookie_index] = saveoff;
1706 ++saveoff;
1707 ++cookie_index;
1708 if (cookie_index == ncookies)
1709 goto done;
1710 }
1711
1712 /*
1713 * Key range (begin and end inclusive) to scan. Directory keys
1714 * directly translate to a 64 bit 'seek' position.
1715 */
1716 hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1717 cursor.key_beg.localization = ip->obj_localization +
1718 hammer_dir_localization(ip);
1719 cursor.key_beg.obj_id = ip->obj_id;
1720 cursor.key_beg.create_tid = 0;
1721 cursor.key_beg.delete_tid = 0;
1722 cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1723 cursor.key_beg.obj_type = 0;
1724 cursor.key_beg.key = saveoff;
1725
1726 cursor.key_end = cursor.key_beg;
1727 cursor.key_end.key = HAMMER_MAX_KEY;
1728 cursor.asof = ip->obj_asof;
1729 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1730
1731 error = hammer_ip_first(&cursor);
1732
1733 while (error == 0) {
1734 error = hammer_ip_resolve_data(&cursor);
1735 if (error)
1736 break;
1737 base = &cursor.leaf->base;
1738 saveoff = base->key;
1739 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1740
1741 if (base->obj_id != ip->obj_id)
1742 panic("readdir: bad record at %p", cursor.node);
1743
1744 /*
1745 * Convert pseudo-filesystems into softlinks
1746 */
1747 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1748 r = vop_write_dirent(
1749 &error, uio, cursor.data->entry.obj_id,
1750 dtype,
1751 cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1752 (void *)cursor.data->entry.name);
1753 if (r)
1754 break;
1755 ++saveoff;
1756 if (cookies)
1757 cookies[cookie_index] = base->key;
1758 ++cookie_index;
1759 if (cookie_index == ncookies)
1760 break;
1761 error = hammer_ip_next(&cursor);
1762 }
1763 hammer_done_cursor(&cursor);
1764
1765 done:
1766 hammer_done_transaction(&trans);
1767
1768 if (ap->a_eofflag)
1769 *ap->a_eofflag = (error == ENOENT);
1770 uio->uio_offset = saveoff;
1771 if (error && cookie_index == 0) {
1772 if (error == ENOENT)
1773 error = 0;
1774 if (cookies) {
1775 kfree(cookies, M_TEMP);
1776 *ap->a_ncookies = 0;
1777 *ap->a_cookies = NULL;
1778 }
1779 } else {
1780 if (error == ENOENT)
1781 error = 0;
1782 if (cookies) {
1783 *ap->a_ncookies = cookie_index;
1784 *ap->a_cookies = cookies;
1785 }
1786 }
1787 lwkt_reltoken(&hmp->fs_token);
1788 return(error);
1789 }
1790
1791 /*
1792 * hammer_vop_readlink { vp, uio, cred }
1793 */
1794 static
1795 int
1796 hammer_vop_readlink(struct vop_readlink_args *ap)
1797 {
1798 struct hammer_transaction trans;
1799 struct hammer_cursor cursor;
1800 struct hammer_inode *ip;
1801 hammer_mount_t hmp;
1802 char buf[32];
1803 u_int32_t localization;
1804 hammer_pseudofs_inmem_t pfsm;
1805 int error;
1806
1807 ip = VTOI(ap->a_vp);
1808 hmp = ip->hmp;
1809
1810 lwkt_gettoken(&hmp->fs_token);
1811
1812 /*
1813 * Shortcut if the symlink data was stuffed into ino_data.
1814 *
1815 * Also expand special "@@PFS%05d" softlinks (expansion only
1816 * occurs for non-historical (current) accesses made from the
1817 * primary filesystem).
1818 */
1819 if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1820 char *ptr;
1821 int bytes;
1822
1823 ptr = ip->ino_data.ext.symlink;
1824 bytes = (int)ip->ino_data.size;
1825 if (bytes == 10 &&
1826 ip->obj_asof == HAMMER_MAX_TID &&
1827 ip->obj_localization == 0 &&
1828 strncmp(ptr, "@@PFS", 5) == 0) {
1829 hammer_simple_transaction(&trans, hmp);
1830 bcopy(ptr + 5, buf, 5);
1831 buf[5] = 0;
1832 localization = strtoul(buf, NULL, 10) << 16;
1833 pfsm = hammer_load_pseudofs(&trans, localization,
1834 &error);
1835 if (error == 0) {
1836 if (pfsm->pfsd.mirror_flags &
1837 HAMMER_PFSD_SLAVE) {
1838 /* vap->va_size == 26 */
1839 ksnprintf(buf, sizeof(buf),
1840 "@@0x%016llx:%05d",
1841 (long long)pfsm->pfsd.sync_end_tid,
1842 localization >> 16);
1843 } else {
1844 /* vap->va_size == 10 */
1845 ksnprintf(buf, sizeof(buf),
1846 "@@-1:%05d",
1847 localization >> 16);
1848 #if 0
1849 ksnprintf(buf, sizeof(buf),
1850 "@@0x%016llx:%05d",
1851 (long long)HAMMER_MAX_TID,
1852 localization >> 16);
1853 #endif
1854 }
1855 ptr = buf;
1856 bytes = strlen(buf);
1857 }
1858 if (pfsm)
1859 hammer_rel_pseudofs(hmp, pfsm);
1860 hammer_done_transaction(&trans);
1861 }
1862 error = uiomove(ptr, bytes, ap->a_uio);
1863 lwkt_reltoken(&hmp->fs_token);
1864 return(error);
1865 }
1866
1867 /*
1868 * Long version
1869 */
1870 hammer_simple_transaction(&trans, hmp);
1871 ++hammer_stats_file_iopsr;
1872 hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1873
1874 /*
1875 * Key range (begin and end inclusive) to scan. Directory keys
1876 * directly translate to a 64 bit 'seek' position.
1877 */
1878 cursor.key_beg.localization = ip->obj_localization +
1879 HAMMER_LOCALIZE_MISC;
1880 cursor.key_beg.obj_id = ip->obj_id;
1881 cursor.key_beg.create_tid = 0;
1882 cursor.key_beg.delete_tid = 0;
1883 cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1884 cursor.key_beg.obj_type = 0;
1885 cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1886 cursor.asof = ip->obj_asof;
1887 cursor.flags |= HAMMER_CURSOR_ASOF;
1888
1889 error = hammer_ip_lookup(&cursor);
1890 if (error == 0) {
1891 error = hammer_ip_resolve_data(&cursor);
1892 if (error == 0) {
1893 KKASSERT(cursor.leaf->data_len >=
1894 HAMMER_SYMLINK_NAME_OFF);
1895 error = uiomove(cursor.data->symlink.name,
1896 cursor.leaf->data_len -
1897 HAMMER_SYMLINK_NAME_OFF,
1898 ap->a_uio);
1899 }
1900 }
1901 hammer_done_cursor(&cursor);
1902 hammer_done_transaction(&trans);
1903 lwkt_reltoken(&hmp->fs_token);
1904 return(error);
1905 }
1906
1907 /*
1908 * hammer_vop_nremove { nch, dvp, cred }
1909 */
1910 static
1911 int
1912 hammer_vop_nremove(struct vop_nremove_args *ap)
1913 {
1914 struct hammer_transaction trans;
1915 struct hammer_inode *dip;
1916 hammer_mount_t hmp;
1917 int error;
1918
1919 dip = VTOI(ap->a_dvp);
1920 hmp = dip->hmp;
1921
1922 if (hammer_nohistory(dip) == 0 &&
1923 (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1924 return (error);
1925 }
1926
1927 lwkt_gettoken(&hmp->fs_token);
1928 hammer_start_transaction(&trans, hmp);
1929 ++hammer_stats_file_iopsw;
1930 error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1931 hammer_done_transaction(&trans);
1932 if (error == 0)
1933 hammer_knote(ap->a_dvp, NOTE_WRITE);
1934 lwkt_reltoken(&hmp->fs_token);
1935 return (error);
1936 }
1937
1938 /*
1939 * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1940 */
1941 static
1942 int
1943 hammer_vop_nrename(struct vop_nrename_args *ap)
1944 {
1945 struct hammer_transaction trans;
1946 struct namecache *fncp;
1947 struct namecache *tncp;
1948 struct hammer_inode *fdip;
1949 struct hammer_inode *tdip;
1950 struct hammer_inode *ip;
1951 hammer_mount_t hmp;
1952 struct hammer_cursor cursor;
1953 int64_t namekey;
1954 u_int32_t max_iterations;
1955 int nlen, error;
1956
1957 if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount)
1958 return(EXDEV);
1959 if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1960 return(EXDEV);
1961
1962 fdip = VTOI(ap->a_fdvp);
1963 tdip = VTOI(ap->a_tdvp);
1964 fncp = ap->a_fnch->ncp;
1965 tncp = ap->a_tnch->ncp;
1966 ip = VTOI(fncp->nc_vp);
1967 KKASSERT(ip != NULL);
1968
1969 hmp = ip->hmp;
1970
1971 if (fdip->obj_localization != tdip->obj_localization)
1972 return(EXDEV);
1973 if (fdip->obj_localization != ip->obj_localization)
1974 return(EXDEV);
1975
1976 if (fdip->flags & HAMMER_INODE_RO)
1977 return (EROFS);
1978 if (tdip->flags & HAMMER_INODE_RO)
1979 return (EROFS);
1980 if (ip->flags & HAMMER_INODE_RO)
1981 return (EROFS);
1982 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1983 return (error);
1984
1985 lwkt_gettoken(&hmp->fs_token);
1986 hammer_start_transaction(&trans, hmp);
1987 ++hammer_stats_file_iopsw;
1988
1989 /*
1990 * Remove tncp from the target directory and then link ip as
1991 * tncp. XXX pass trans to dounlink
1992 *
1993 * Force the inode sync-time to match the transaction so it is
1994 * in-sync with the creation of the target directory entry.
1995 */
1996 error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
1997 ap->a_cred, 0, -1);
1998 if (error == 0 || error == ENOENT) {
1999 error = hammer_ip_add_directory(&trans, tdip,
2000 tncp->nc_name, tncp->nc_nlen,
2001 ip);
2002 if (error == 0) {
2003 ip->ino_data.parent_obj_id = tdip->obj_id;
2004 ip->ino_data.ctime = trans.time;
2005 hammer_modify_inode(&trans, ip, HAMMER_INODE_DDIRTY);
2006 }
2007 }
2008 if (error)
2009 goto failed; /* XXX */
2010
2011 /*
2012 * Locate the record in the originating directory and remove it.
2013 *
2014 * Calculate the namekey and setup the key range for the scan. This
2015 * works kinda like a chained hash table where the lower 32 bits
2016 * of the namekey synthesize the chain.
2017 *
2018 * The key range is inclusive of both key_beg and key_end.
2019 */
2020 namekey = hammer_directory_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
2021 &max_iterations);
2022 retry:
2023 hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
2024 cursor.key_beg.localization = fdip->obj_localization +
2025 hammer_dir_localization(fdip);
2026 cursor.key_beg.obj_id = fdip->obj_id;
2027 cursor.key_beg.key = namekey;
2028 cursor.key_beg.create_tid = 0;
2029 cursor.key_beg.delete_tid = 0;
2030 cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2031 cursor.key_beg.obj_type = 0;
2032
2033 cursor.key_end = cursor.key_beg;
2034 cursor.key_end.key += max_iterations;
2035 cursor.asof = fdip->obj_asof;
2036 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2037
2038 /*
2039 * Scan all matching records (the chain), locate the one matching
2040 * the requested path component.
2041 *
2042 * The hammer_ip_*() functions merge in-memory records with on-disk
2043 * records for the purposes of the search.
2044 */
2045 error = hammer_ip_first(&cursor);
2046 while (error == 0) {
2047 if (hammer_ip_resolve_data(&cursor) != 0)
2048 break;
2049 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2050 KKASSERT(nlen > 0);
2051 if (fncp->nc_nlen == nlen &&
2052 bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2053 break;
2054 }
2055 error = hammer_ip_next(&cursor);
2056 }
2057
2058 /*
2059 * If all is ok we have to get the inode so we can adjust nlinks.
2060 *
2061 * WARNING: hammer_ip_del_directory() may have to terminate the
2062 * cursor to avoid a recursion. It's ok to call hammer_done_cursor()
2063 * twice.
2064 */
2065 if (error == 0)
2066 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
2067
2068 /*
2069 * XXX A deadlock here will break rename's atomicy for the purposes
2070 * of crash recovery.
2071 */
2072 if (error == EDEADLK) {
2073 hammer_done_cursor(&cursor);
2074 goto retry;
2075 }
2076
2077 /*
2078 * Cleanup and tell the kernel that the rename succeeded.
2079 *
2080 * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2081 * without formally acquiring the vp since the vp might
2082 * have zero refs on it, or in the middle of a reclaim,
2083 * etc.
2084 */
2085 hammer_done_cursor(&cursor);
2086 if (error == 0) {
2087 cache_rename(ap->a_fnch, ap->a_tnch);
2088 hammer_knote(ap->a_fdvp, NOTE_WRITE);
2089 hammer_knote(ap->a_tdvp, NOTE_WRITE);
2090 while (ip->vp) {
2091 struct vnode *vp;
2092
2093 error = hammer_get_vnode(ip, &vp);
2094 if (error == 0 && vp) {
2095 vn_unlock(vp);
2096 hammer_knote(ip->vp, NOTE_RENAME);
2097 vrele(vp);
2098 break;
2099 }
2100 kprintf("Debug: HAMMER ip/vp race2 avoided\n");
2101 }
2102 }
2103
2104 failed:
2105 hammer_done_transaction(&trans);
2106 lwkt_reltoken(&hmp->fs_token);
2107 return (error);
2108 }
2109
2110 /*
2111 * hammer_vop_nrmdir { nch, dvp, cred }
2112 */
2113 static
2114 int
2115 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
2116 {
2117 struct hammer_transaction trans;
2118 struct hammer_inode *dip;
2119 hammer_mount_t hmp;
2120 int error;
2121
2122 dip = VTOI(ap->a_dvp);
2123 hmp = dip->hmp;
2124
2125 if (hammer_nohistory(dip) == 0 &&
2126 (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2127 return (error);
2128 }
2129
2130 lwkt_gettoken(&hmp->fs_token);
2131 hammer_start_transaction(&trans, hmp);
2132 ++hammer_stats_file_iopsw;
2133 error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
2134 hammer_done_transaction(&trans);
2135 if (error == 0)
2136 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
2137 lwkt_reltoken(&hmp->fs_token);
2138 return (error);
2139 }
2140
2141 /*
2142 * hammer_vop_markatime { vp, cred }
2143 */
2144 static
2145 int
2146 hammer_vop_markatime(struct vop_markatime_args *ap)
2147 {
2148 struct hammer_transaction trans;
2149 struct hammer_inode *ip;
2150 hammer_mount_t hmp;
2151
2152 ip = VTOI(ap->a_vp);
2153 if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2154 return (EROFS);
2155 if (ip->flags & HAMMER_INODE_RO)
2156 return (EROFS);
2157 hmp = ip->hmp;
2158 if (hmp->mp->mnt_flag & MNT_NOATIME)
2159 return (0);
2160 lwkt_gettoken(&hmp->fs_token);
2161 hammer_start_transaction(&trans, hmp);
2162 ++hammer_stats_file_iopsw;
2163
2164 ip->ino_data.atime = trans.time;
2165 hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
2166 hammer_done_transaction(&trans);
2167 hammer_knote(ap->a_vp, NOTE_ATTRIB);
2168 lwkt_reltoken(&hmp->fs_token);
2169 return (0);
2170 }
2171
2172 /*
2173 * hammer_vop_setattr { vp, vap, cred }
2174 */
2175 static
2176 int
2177 hammer_vop_setattr(struct vop_setattr_args *ap)
2178 {
2179 struct hammer_transaction trans;
2180 struct hammer_inode *ip;
2181 struct vattr *vap;
2182 hammer_mount_t hmp;
2183 int modflags;
2184 int error;
2185 int truncating;
2186 int blksize;
2187 int kflags;
2188 #if 0
2189 int64_t aligned_size;
2190 #endif
2191 u_int32_t flags;
2192
2193 vap = ap->a_vap;
2194 ip = ap->a_vp->v_data;
2195 modflags = 0;
2196 kflags = 0;
2197 hmp = ip->hmp;
2198
2199 if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2200 return(EROFS);
2201 if (ip->flags & HAMMER_INODE_RO)
2202 return (EROFS);
2203 if (hammer_nohistory(ip) == 0 &&
2204 (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2205 return (error);
2206 }
2207
2208 lwkt_gettoken(&hmp->fs_token);
2209 hammer_start_transaction(&trans, hmp);
2210 ++hammer_stats_file_iopsw;
2211 error = 0;
2212
2213 if (vap->va_flags != VNOVAL) {
2214 flags = ip->ino_data.uflags;
2215 error = vop_helper_setattr_flags(&flags, vap->va_flags,
2216 hammer_to_unix_xid(&ip->ino_data.uid),
2217 ap->a_cred);
2218 if (error == 0) {
2219 if (ip->ino_data.uflags != flags) {
2220 ip->ino_data.uflags = flags;
2221 ip->ino_data.ctime = trans.time;
2222 modflags |= HAMMER_INODE_DDIRTY;
2223 kflags |= NOTE_ATTRIB;
2224 }
2225 if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2226 error = 0;
2227 goto done;
2228 }
2229 }
2230 goto done;
2231 }
2232 if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2233 error = EPERM;
2234 goto done;
2235 }
2236 if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2237 mode_t cur_mode = ip->ino_data.mode;
2238 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2239 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2240 uuid_t uuid_uid;
2241 uuid_t uuid_gid;
2242
2243 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2244 ap->a_cred,
2245 &cur_uid, &cur_gid, &cur_mode);
2246 if (error == 0) {
2247 hammer_guid_to_uuid(&uuid_uid, cur_uid);
2248 hammer_guid_to_uuid(&uuid_gid, cur_gid);
2249 if (bcmp(&uuid_uid, &ip->ino_data.uid,
2250 sizeof(uuid_uid)) ||
2251 bcmp(&uuid_gid, &ip->ino_data.gid,
2252 sizeof(uuid_gid)) ||
2253 ip->ino_data.mode != cur_mode
2254 ) {
2255 ip->ino_data.uid = uuid_uid;
2256 ip->ino_data.gid = uuid_gid;
2257 ip->ino_data.mode = cur_mode;
2258 ip->ino_data.ctime = trans.time;
2259 modflags |= HAMMER_INODE_DDIRTY;
2260 }
2261 kflags |= NOTE_ATTRIB;
2262 }
2263 }
2264 while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2265 switch(ap->a_vp->v_type) {
2266 case VREG:
2267 if (vap->va_size == ip->ino_data.size)
2268 break;
2269
2270 /*
2271 * Log the operation if in fast-fsync mode or if
2272 * there are unterminated redo write records present.
2273 *
2274 * The second check is needed so the recovery code
2275 * properly truncates write redos even if nominal
2276 * REDO operations is turned off due to excessive
2277 * writes, because the related records might be
2278 * destroyed and never lay down a TERM_WRITE.
2279 */
2280 if ((ip->flags & HAMMER_INODE_REDO) ||
2281 (ip->flags & HAMMER_INODE_RDIRTY)) {
2282 error = hammer_generate_redo(&trans, ip,
2283 vap->va_size,
2284 HAMMER_REDO_TRUNC,
2285 NULL, 0);
2286 }
2287 blksize = hammer_blocksize(vap->va_size);
2288
2289 /*
2290 * XXX break atomicy, we can deadlock the backend
2291 * if we do not release the lock. Probably not a
2292 * big deal here.
2293 */
2294 if (vap->va_size < ip->ino_data.size) {
2295 nvtruncbuf(ap->a_vp, vap->va_size,
2296 blksize,
2297 hammer_blockoff(vap->va_size),
2298 0);
2299 truncating = 1;
2300 kflags |= NOTE_WRITE;
2301 } else {
2302 nvextendbuf(ap->a_vp,
2303 ip->ino_data.size,
2304 vap->va_size,
2305 hammer_blocksize(ip->ino_data.size),
2306 hammer_blocksize(vap->va_size),
2307 hammer_blockoff(ip->ino_data.size),
2308 hammer_blockoff(vap->va_size),
2309 0);
2310 truncating = 0;
2311 kflags |= NOTE_WRITE | NOTE_EXTEND;
2312 }
2313 ip->ino_data.size = vap->va_size;
2314 ip->ino_data.mtime = trans.time;
2315 /* XXX safe to use SDIRTY instead of DDIRTY here? */
2316 modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2317
2318 /*
2319 * On-media truncation is cached in the inode until
2320 * the inode is synchronized. We must immediately
2321 * handle any frontend records.
2322 */
2323 if (truncating) {
2324 hammer_ip_frontend_trunc(ip, vap->va_size);
2325 #ifdef DEBUG_TRUNCATE
2326 if (HammerTruncIp == NULL)
2327 HammerTruncIp = ip;
2328 #endif
2329 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2330 ip->flags |= HAMMER_INODE_TRUNCATED;
2331 ip->trunc_off = vap->va_size;
2332 hammer_inode_dirty(ip);
2333 #ifdef DEBUG_TRUNCATE
2334 if (ip == HammerTruncIp)
2335 kprintf("truncate1 %016llx\n",
2336 (long long)ip->trunc_off);
2337 #endif
2338 } else if (ip->trunc_off > vap->va_size) {
2339 ip->trunc_off = vap->va_size;
2340 #ifdef DEBUG_TRUNCATE
2341 if (ip == HammerTruncIp)
2342 kprintf("truncate2 %016llx\n",
2343 (long long)ip->trunc_off);
2344 #endif
2345 } else {
2346 #ifdef DEBUG_TRUNCATE
2347 if (ip == HammerTruncIp)
2348 kprintf("truncate3 %016llx (ignored)\n",
2349 (long long)vap->va_size);
2350 #endif
2351 }
2352 }
2353
2354 #if 0
2355 /*
2356 * When truncating, nvtruncbuf() may have cleaned out
2357 * a portion of the last block on-disk in the buffer
2358 * cache. We must clean out any frontend records
2359 * for blocks beyond the new last block.
2360 */
2361 aligned_size = (vap->va_size + (blksize - 1)) &
2362 ~(int64_t)(blksize - 1);
2363 if (truncating && vap->va_size < aligned_size) {
2364 aligned_size -= blksize;
2365 hammer_ip_frontend_trunc(ip, aligned_size);
2366 }
2367 #endif
2368 break;
2369 case VDATABASE:
2370 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2371 ip->flags |= HAMMER_INODE_TRUNCATED;
2372 ip->trunc_off = vap->va_size;
2373 hammer_inode_dirty(ip);
2374 } else if (ip->trunc_off > vap->va_size) {
2375 ip->trunc_off = vap->va_size;
2376 }
2377 hammer_ip_frontend_trunc(ip, vap->va_size);
2378 ip->ino_data.size = vap->va_size;
2379 ip->ino_data.mtime = trans.time;
2380 modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2381 kflags |= NOTE_ATTRIB;
2382 break;
2383 default:
2384 error = EINVAL;
2385 goto done;
2386 }
2387 break;
2388 }
2389 if (vap->va_atime.tv_sec != VNOVAL) {
2390 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2391 modflags |= HAMMER_INODE_ATIME;
2392 kflags |= NOTE_ATTRIB;
2393 }
2394 if (vap->va_mtime.tv_sec != VNOVAL) {
2395 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2396 modflags |= HAMMER_INODE_MTIME;
2397 kflags |= NOTE_ATTRIB;
2398 }
2399 if (vap->va_mode != (mode_t)VNOVAL) {
2400 mode_t cur_mode = ip->ino_data.mode;
2401 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2402 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2403
2404 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2405 cur_uid, cur_gid, &cur_mode);
2406 if (error == 0 && ip->ino_data.mode != cur_mode) {
2407 ip->ino_data.mode = cur_mode;
2408 ip->ino_data.ctime = trans.time;
2409 modflags |= HAMMER_INODE_DDIRTY;
2410 kflags |= NOTE_ATTRIB;
2411 }
2412 }
2413 done:
2414 if (error == 0)
2415 hammer_modify_inode(&trans, ip, modflags);
2416 hammer_done_transaction(&trans);
2417 hammer_knote(ap->a_vp, kflags);
2418 lwkt_reltoken(&hmp->fs_token);
2419 return (error);
2420 }
2421
2422 /*
2423 * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2424 */
2425 static
2426 int
2427 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2428 {
2429 struct hammer_transaction trans;
2430 struct hammer_inode *dip;
2431 struct hammer_inode *nip;
2432 hammer_record_t record;
2433 struct nchandle *nch;
2434 hammer_mount_t hmp;
2435 int error;
2436 int bytes;
2437
2438 ap->a_vap->va_type = VLNK;
2439
2440 nch = ap->a_nch;
2441 dip = VTOI(ap->a_dvp);
2442 hmp = dip->hmp;
2443
2444 if (dip->flags & HAMMER_INODE_RO)
2445 return (EROFS);
2446 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
2447 return (error);
2448
2449 /*
2450 * Create a transaction to cover the operations we perform.
2451 */
2452 lwkt_gettoken(&hmp->fs_token);
2453 hammer_start_transaction(&trans, hmp);
2454 ++hammer_stats_file_iopsw;
2455
2456 /*
2457 * Create a new filesystem object of the requested type. The
2458 * returned inode will be referenced but not locked.
2459 */
2460
2461 error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2462 dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2463 NULL, &nip);
2464 if (error) {
2465 hammer_done_transaction(&trans);
2466 *ap->a_vpp = NULL;
2467 lwkt_reltoken(&hmp->fs_token);
2468 return (error);
2469 }
2470
2471 /*
2472 * Add a record representing the symlink. symlink stores the link
2473 * as pure data, not a string, and is no \0 terminated.
2474 */
2475 if (error == 0) {
2476 bytes = strlen(ap->a_target);
2477
2478 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2479 bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2480 } else {
2481 record = hammer_alloc_mem_record(nip, bytes);
2482 record->type = HAMMER_MEM_RECORD_GENERAL;
2483
2484 record->leaf.base.localization = nip->obj_localization +
2485 HAMMER_LOCALIZE_MISC;
2486 record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2487 record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2488 record->leaf.data_len = bytes;
2489 KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2490 bcopy(ap->a_target, record->data->symlink.name, bytes);
2491 error = hammer_ip_add_record(&trans, record);
2492 }
2493
2494 /*
2495 * Set the file size to the length of the link.
2496 */
2497 if (error == 0) {
2498 nip->ino_data.size = bytes;
2499 hammer_modify_inode(&trans, nip, HAMMER_INODE_DDIRTY);
2500 }
2501 }
2502 if (error == 0)
2503 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
2504 nch->ncp->nc_nlen, nip);
2505
2506 /*
2507 * Finish up.
2508 */
2509 if (error) {
2510 hammer_rel_inode(nip, 0);
2511 *ap->a_vpp = NULL;
2512 } else {
2513 error = hammer_get_vnode(nip, ap->a_vpp);
2514 hammer_rel_inode(nip, 0);
2515 if (error == 0) {
2516 cache_setunresolved(ap->a_nch);
2517 cache_setvp(ap->a_nch, *ap->a_vpp);
2518 hammer_knote(ap->a_dvp, NOTE_WRITE);
2519 }
2520 }
2521 hammer_done_transaction(&trans);
2522 lwkt_reltoken(&hmp->fs_token);
2523 return (error);
2524 }
2525
2526 /*
2527 * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2528 */
2529 static
2530 int
2531 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2532 {
2533 struct hammer_transaction trans;
2534 struct hammer_inode *dip;
2535 hammer_mount_t hmp;
2536 int error;
2537
2538 dip = VTOI(ap->a_dvp);
2539 hmp = dip->hmp;
2540
2541 if (hammer_nohistory(dip) == 0 &&
2542 (error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2543 return (error);
2544 }
2545
2546 lwkt_gettoken(&hmp->fs_token);
2547 hammer_start_transaction(&trans, hmp);
2548 ++hammer_stats_file_iopsw;
2549 error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2550 ap->a_cred, ap->a_flags, -1);
2551 hammer_done_transaction(&trans);
2552 lwkt_reltoken(&hmp->fs_token);
2553
2554 return (error);
2555 }
2556
2557 /*
2558 * hammer_vop_ioctl { vp, command, data, fflag, cred }
2559 */
2560 static
2561 int
2562 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2563 {
2564 struct hammer_inode *ip = ap->a_vp->v_data;
2565 hammer_mount_t hmp = ip->hmp;
2566 int error;
2567
2568 ++hammer_stats_file_iopsr;
2569 lwkt_gettoken(&hmp->fs_token);
2570 error = hammer_ioctl(ip, ap->a_command, ap->a_data,
2571 ap->a_fflag, ap->a_cred);
2572 lwkt_reltoken(&hmp->fs_token);
2573 return (error);
2574 }
2575
2576 static
2577 int
2578 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2579 {
2580 static const struct mountctl_opt extraopt[] = {
2581 { HMNT_NOHISTORY, "nohistory" },
2582 { HMNT_MASTERID, "master" },
2583 { 0, NULL}
2584
2585 };
2586 struct hammer_mount *hmp;
2587 struct mount *mp;
2588 int usedbytes;
2589 int error;
2590
2591 error = 0;
2592 usedbytes = 0;
2593 mp = ap->a_head.a_ops->head.vv_mount;
2594 KKASSERT(mp->mnt_data != NULL);
2595 hmp = (struct hammer_mount *)mp->mnt_data;
2596
2597 lwkt_gettoken(&hmp->fs_token);
2598
2599 switch(ap->a_op) {
2600 case MOUNTCTL_SET_EXPORT:
2601 if (ap->a_ctllen != sizeof(struct export_args))
2602 error = EINVAL;
2603 else
2604 error = hammer_vfs_export(mp, ap->a_op,
2605 (const struct export_args *)ap->a_ctl);
2606 break;
2607 case MOUNTCTL_MOUNTFLAGS:
2608 {
2609 /*
2610 * Call standard mountctl VOP function
2611 * so we get user mount flags.
2612 */
2613 error = vop_stdmountctl(ap);
2614 if (error)
2615 break;
2616
2617 usedbytes = *ap->a_res;
2618
2619 if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2620 usedbytes += vfs_flagstostr(hmp->hflags, extraopt,
2621 ap->a_buf,
2622 ap->a_buflen - usedbytes,
2623 &error);
2624 }
2625
2626 *ap->a_res += usedbytes;
2627 break;
2628 }
2629 default:
2630 error = vop_stdmountctl(ap);
2631 break;
2632 }
2633 lwkt_reltoken(&hmp->fs_token);
2634 return(error);
2635 }
2636
2637 /*
2638 * hammer_vop_strategy { vp, bio }
2639 *
2640 * Strategy call, used for regular file read & write only. Note that the
2641 * bp may represent a cluster.
2642 *
2643 * To simplify operation and allow better optimizations in the future,
2644 * this code does not make any assumptions with regards to buffer alignment
2645 * or size.
2646 */
2647 static
2648 int
2649 hammer_vop_strategy(struct vop_strategy_args *ap)
2650 {
2651 struct buf *bp;
2652 int error;
2653
2654 bp = ap->a_bio->bio_buf;
2655
2656 switch(bp->b_cmd) {
2657 case BUF_CMD_READ:
2658 error = hammer_vop_strategy_read(ap);
2659 break;
2660 case BUF_CMD_WRITE:
2661 error = hammer_vop_strategy_write(ap);
2662 break;
2663 default:
2664 bp->b_error = error = EINVAL;
2665 bp->b_flags |= B_ERROR;
2666 biodone(ap->a_bio);
2667 break;
2668 }
2669
2670 /* hammer_dump_dedup_cache(((hammer_inode_t)ap->a_vp->v_data)->hmp); */
2671
2672 return (error);
2673 }
2674
2675 /*
2676 * Read from a regular file. Iterate the related records and fill in the
2677 * BIO/BUF. Gaps are zero-filled.
2678 *
2679 * The support code in hammer_object.c should be used to deal with mixed
2680 * in-memory and on-disk records.
2681 *
2682 * NOTE: Can be called from the cluster code with an oversized buf.
2683 *
2684 * XXX atime update
2685 */
2686 static
2687 int
2688 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2689 {
2690 struct hammer_transaction trans;
2691 struct hammer_inode *ip;
2692 struct hammer_inode *dip;
2693 hammer_mount_t hmp;
2694 struct hammer_cursor cursor;
2695 hammer_base_elm_t base;
2696 hammer_off_t disk_offset;
2697 struct bio *bio;
2698 struct bio *nbio;
2699 struct buf *bp;
2700 int64_t rec_offset;
2701 int64_t ran_end;
2702 int64_t tmp64;
2703 int error;
2704 int boff;
2705 int roff;
2706 int n;
2707 int isdedupable;
2708
2709 bio = ap->a_bio;
2710 bp = bio->bio_buf;
2711 ip = ap->a_vp->v_data;
2712 hmp = ip->hmp;
2713
2714 /*
2715 * The zone-2 disk offset may have been set by the cluster code via
2716 * a BMAP operation, or else should be NOOFFSET.
2717 *
2718 * Checking the high bits for a match against zone-2 should suffice.
2719 *
2720 * In cases where a lot of data duplication is present it may be
2721 * more beneficial to drop through and doubule-buffer through the
2722 * device.
2723 */
2724 nbio = push_bio(bio);
2725 if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2726 HAMMER_ZONE_LARGE_DATA) {
2727 if (hammer_double_buffer == 0) {
2728 lwkt_gettoken(&hmp->fs_token);
2729 error = hammer_io_direct_read(hmp, nbio, NULL);
2730 lwkt_reltoken(&hmp->fs_token);
2731 return (error);
2732 }
2733
2734 /*
2735 * Try to shortcut requests for double_buffer mode too.
2736 * Since this mode runs through the device buffer cache
2737 * only compatible buffer sizes (meaning those generated
2738 * by normal filesystem buffers) are legal.
2739 */
2740 if (hammer_live_dedup == 0 && (bp->b_flags & B_PAGING) == 0) {
2741 lwkt_gettoken(&hmp->fs_token);
2742 error = hammer_io_indirect_read(hmp, nbio, NULL);
2743 lwkt_reltoken(&hmp->fs_token);
2744 return (error);
2745 }
2746 }
2747
2748 /*
2749 * Well, that sucked. Do it the hard way. If all the stars are
2750 * aligned we may still be able to issue a direct-read.
2751 */
2752 lwkt_gettoken(&hmp->fs_token);
2753 hammer_simple_transaction(&trans, hmp);
2754 hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2755
2756 /*
2757 * Key range (begin and end inclusive) to scan. Note that the key's
2758 * stored in the actual records represent BASE+LEN, not BASE. The
2759 * first record containing bio_offset will have a key > bio_offset.
2760 */
2761 cursor.key_beg.localization = ip->obj_localization +
2762 HAMMER_LOCALIZE_MISC;
2763 cursor.key_beg.obj_id = ip->obj_id;
2764 cursor.key_beg.create_tid = 0;
2765 cursor.key_beg.delete_tid = 0;
2766 cursor.key_beg.obj_type = 0;
2767 cursor.key_beg.key = bio->bio_offset + 1;
2768 cursor.asof = ip->obj_asof;
2769 cursor.flags |= HAMMER_CURSOR_ASOF;
2770
2771 cursor.key_end = cursor.key_beg;
2772 KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2773 #if 0
2774 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2775 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2776 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2777 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2778 } else
2779 #endif
2780 {
2781 ran_end = bio->bio_offset + bp->b_bufsize;
2782 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2783 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2784 tmp64 = ran_end + MAXPHYS + 1; /* work-around GCC-4 bug */
2785 if (tmp64 < ran_end)
2786 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2787 else
2788 cursor.key_end.key = ran_end + MAXPHYS + 1;
2789 }
2790 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2791
2792 /*
2793 * Set NOSWAPCACHE for cursor data extraction if double buffering
2794 * is disabled or (if the file is not marked cacheable via chflags
2795 * and vm.swapcache_use_chflags is enabled).
2796 */
2797 if (hammer_double_buffer == 0 ||
2798 ((ap->a_vp->v_flag & VSWAPCACHE) == 0 &&
2799 vm_swapcache_use_chflags)) {
2800 cursor.flags |= HAMMER_CURSOR_NOSWAPCACHE;
2801 }
2802
2803 error = hammer_ip_first(&cursor);
2804 boff = 0;
2805
2806 while (error == 0) {
2807 /*
2808 * Get the base file offset of the record. The key for
2809 * data records is (base + bytes) rather then (base).
2810 */
2811 base = &cursor.leaf->base;
2812 rec_offset = base->key - cursor.leaf->data_len;
2813
2814 /*
2815 * Calculate the gap, if any, and zero-fill it.
2816 *
2817 * n is the offset of the start of the record verses our
2818 * current seek offset in the bio.
2819 */
2820 n = (int)(rec_offset - (bio->bio_offset + boff));
2821 if (n > 0) {
2822 if (n > bp->b_bufsize - boff)
2823 n = bp->b_bufsize - boff;
2824 bzero((char *)bp->b_data + boff, n);
2825 boff += n;
2826 n = 0;
2827 }
2828
2829 /*
2830 * Calculate the data offset in the record and the number
2831 * of bytes we can copy.
2832 *
2833 * There are two degenerate cases. First, boff may already
2834 * be at bp->b_bufsize. Secondly, the data offset within
2835 * the record may exceed the record's size.
2836 */
2837 roff = -n;
2838 rec_offset += roff;
2839 n = cursor.leaf->data_len - roff;
2840 if (n <= 0) {
2841 kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2842 n = 0;
2843 } else if (n > bp->b_bufsize - boff) {
2844 n = bp->b_bufsize - boff;
2845 }
2846
2847 /*
2848 * Deal with cached truncations. This cool bit of code
2849 * allows truncate()/ftruncate() to avoid having to sync
2850 * the file.
2851 *
2852 * If the frontend is truncated then all backend records are
2853 * subject to the frontend's truncation.
2854 *
2855 * If the backend is truncated then backend records on-disk
2856 * (but not in-memory) are subject to the backend's
2857 * truncation. In-memory records owned by the backend
2858 * represent data written after the truncation point on the
2859 * backend and must not be truncated.
2860 *
2861 * Truncate operations deal with frontend buffer cache
2862 * buffers and frontend-owned in-memory records synchronously.
2863 */
2864 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2865 if (hammer_cursor_ondisk(&cursor)/* ||
2866 cursor.iprec->flush_state == HAMMER_FST_FLUSH*/) {
2867 if (ip->trunc_off <= rec_offset)
2868 n = 0;
2869 else if (ip->trunc_off < rec_offset + n)
2870 n = (int)(ip->trunc_off - rec_offset);
2871 }
2872 }
2873 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2874 if (hammer_cursor_ondisk(&cursor)) {
2875 if (ip->sync_trunc_off <= rec_offset)
2876 n = 0;
2877 else if (ip->sync_trunc_off < rec_offset + n)
2878 n = (int)(ip->sync_trunc_off - rec_offset);
2879 }
2880 }
2881
2882 /*
2883 * Try to issue a direct read into our bio if possible,
2884 * otherwise resolve the element data into a hammer_buffer
2885 * and copy.
2886 *
2887 * The buffer on-disk should be zerod past any real
2888 * truncation point, but may not be for any synthesized
2889 * truncation point from above.
2890 *
2891 * NOTE: disk_offset is only valid if the cursor data is
2892 * on-disk.
2893 */
2894 disk_offset = cursor.leaf->data_offset + roff;
2895 isdedupable = (boff == 0 && n == bp->b_bufsize &&
2896 hammer_cursor_ondisk(&cursor) &&
2897 ((int)disk_offset & HAMMER_BUFMASK) == 0);
2898
2899 if (isdedupable && hammer_double_buffer == 0) {
2900 /*
2901 * Direct read case
2902 */
2903 KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2904 HAMMER_ZONE_LARGE_DATA);
2905 nbio->bio_offset = disk_offset;
2906 error = hammer_io_direct_read(hmp, nbio, cursor.leaf);
2907 if (hammer_live_dedup && error == 0)
2908 hammer_dedup_cache_add(ip, cursor.leaf);
2909 goto done;
2910 } else if (isdedupable) {
2911 /*
2912 * Async I/O case for reading from backing store
2913 * and copying the data to the filesystem buffer.
2914 * live-dedup has to verify the data anyway if it
2915 * gets a hit later so we can just add the entry
2916 * now.
2917 */
2918 KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2919 HAMMER_ZONE_LARGE_DATA);
2920 nbio->bio_offset = disk_offset;
2921 if (hammer_live_dedup)
2922 hammer_dedup_cache_add(ip, cursor.leaf);
2923 error = hammer_io_indirect_read(hmp, nbio, cursor.leaf);
2924 goto done;
2925 } else if (n) {
2926 error = hammer_ip_resolve_data(&cursor);
2927 if (error == 0) {
2928 if (hammer_live_dedup && isdedupable)
2929 hammer_dedup_cache_add(ip, cursor.leaf);
2930 bcopy((char *)cursor.data + roff,
2931 (char *)bp->b_data + boff, n);
2932 }
2933 }
2934 if (error)
2935 break;
2936
2937 /*
2938 * We have to be sure that the only elements added to the
2939 * dedup cache are those which are already on-media.
2940 */
2941 if (hammer_live_dedup && hammer_cursor_ondisk(&cursor))
2942 hammer_dedup_cache_add(ip, cursor.leaf);
2943
2944 /*
2945 * Iterate until we have filled the request.
2946 */
2947 boff += n;
2948 if (boff == bp->b_bufsize)
2949 break;
2950 error = hammer_ip_next(&cursor);
2951 }
2952
2953 /*
2954 * There may have been a gap after the last record
2955 */
2956 if (error == ENOENT)
2957 error = 0;
2958 if (error == 0 && boff != bp->b_bufsize) {
2959 KKASSERT(boff < bp->b_bufsize);
2960 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2961 /* boff = bp->b_bufsize; */
2962 }
2963
2964 /*
2965 * Disallow swapcache operation on the vnode buffer if double
2966 * buffering is enabled, the swapcache will get the data via
2967 * the block device buffer.
2968 */
2969 if (hammer_double_buffer)
2970 bp->b_flags |= B_NOTMETA;
2971
2972 /*
2973 * Cleanup
2974 */
2975 bp->b_resid = 0;
2976 bp->b_error = error;
2977 if (error)
2978 bp->b_flags |= B_ERROR;
2979 biodone(ap->a_bio);
2980
2981 done:
2982 /*
2983 * Cache the b-tree node for the last data read in cache[1].
2984 *
2985 * If we hit the file EOF then also cache the node in the
2986 * governing director's cache[3], it will be used to initialize
2987 * the inode's cache[1] for any inodes looked up via the directory.
2988 *
2989 * This doesn't reduce disk accesses since the B-Tree chain is
2990 * likely cached, but it does reduce cpu overhead when looking
2991 * up file offsets for cpdup/tar/cpio style iterations.
2992 */
2993 if (cursor.node)
2994 hammer_cache_node(&ip->cache[1], cursor.node);
2995 if (ran_end >= ip->ino_data.size) {
2996 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
2997 ip->obj_asof, ip->obj_localization);
2998 if (dip) {
2999 hammer_cache_node(&dip->cache[3], cursor.node);
3000 hammer_rel_inode(dip, 0);
3001 }
3002 }
3003 hammer_done_cursor(&cursor);
3004 hammer_done_transaction(&trans);
3005 lwkt_reltoken(&hmp->fs_token);
3006 return(error);
3007 }
3008
3009 /*
3010 * BMAP operation - used to support cluster_read() only.
3011 *
3012 * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
3013 *
3014 * This routine may return EOPNOTSUPP if the opration is not supported for
3015 * the specified offset. The contents of the pointer arguments do not
3016 * need to be initialized in that case.
3017 *
3018 * If a disk address is available and properly aligned return 0 with
3019 * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
3020 * to the run-length relative to that offset. Callers may assume that
3021 * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
3022 * large, so return EOPNOTSUPP if it is not sufficiently large.
3023 */
3024 static
3025 int
3026 hammer_vop_bmap(struct vop_bmap_args *ap)
3027 {
3028 struct hammer_transaction trans;
3029 struct hammer_inode *ip;
3030 hammer_mount_t hmp;
3031 struct hammer_cursor cursor;
3032 hammer_base_elm_t base;
3033 int64_t rec_offset;
3034 int64_t ran_end;
3035 int64_t tmp64;
3036 int64_t base_offset;
3037 int64_t base_disk_offset;
3038 int64_t last_offset;
3039 hammer_off_t last_disk_offset;
3040 hammer_off_t disk_offset;
3041 int rec_len;
3042 int error;
3043 int blksize;
3044
3045 ++hammer_stats_file_iopsr;
3046 ip = ap->a_vp->v_data;
3047 hmp = ip->hmp;
3048
3049 /*
3050 * We can only BMAP regular files. We can't BMAP database files,
3051 * directories, etc.
3052 */
3053 if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
3054 return(EOPNOTSUPP);
3055
3056 /*
3057 * bmap is typically called with runp/runb both NULL when used
3058 * for writing. We do not support BMAP for writing atm.
3059 */
3060 if (ap->a_cmd != BUF_CMD_READ)
3061 return(EOPNOTSUPP);
3062
3063 /*
3064 * Scan the B-Tree to acquire blockmap addresses, then translate
3065 * to raw addresses.
3066 */
3067 lwkt_gettoken(&hmp->fs_token);
3068 hammer_simple_transaction(&trans, hmp);
3069 #if 0
3070 kprintf("bmap_beg %016llx ip->cache %p\n",
3071 (long long)ap->a_loffset, ip->cache[1]);
3072 #endif
3073 hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
3074
3075 /*
3076 * Key range (begin and end inclusive) to scan. Note that the key's
3077 * stored in the actual records represent BASE+LEN, not BASE. The
3078 * first record containing bio_offset will have a key > bio_offset.
3079 */
3080 cursor.key_beg.localization = ip->obj_localization +
3081 HAMMER_LOCALIZE_MISC;
3082 cursor.key_beg.obj_id = ip->obj_id;
3083 cursor.key_beg.create_tid = 0;
3084 cursor.key_beg.delete_tid = 0;
3085 cursor.key_beg.obj_type = 0;
3086 if (ap->a_runb)
3087 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
3088 else
3089 cursor.key_beg.key = ap->a_loffset + 1;
3090 if (cursor.key_beg.key < 0)
3091 cursor.key_beg.key = 0;
3092 cursor.asof = ip->obj_asof;
3093 cursor.flags |= HAMMER_CURSOR_ASOF;
3094
3095 cursor.key_end = cursor.key_beg;
3096 KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
3097
3098 ran_end = ap->a_loffset + MAXPHYS;
3099 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
3100 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
3101 tmp64 = ran_end + MAXPHYS + 1; /* work-around GCC-4 bug */
3102 if (tmp64 < ran_end)
3103 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
3104 else
3105 cursor.key_end.key = ran_end + MAXPHYS + 1;
3106
3107 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
3108
3109 error = hammer_ip_first(&cursor);
3110 base_offset = last_offset = 0;
3111 base_disk_offset = last_disk_offset = 0;
3112
3113 while (error == 0) {
3114 /*
3115 * Get the base file offset of the record. The key for
3116 * data records is (base + bytes) rather then (base).
3117 *
3118 * NOTE: rec_offset + rec_len may exceed the end-of-file.
3119 * The extra bytes should be zero on-disk and the BMAP op
3120 * should still be ok.
3121 */
3122 base = &cursor.leaf->base;
3123 rec_offset = base->key - cursor.leaf->data_len;
3124 rec_len = cursor.leaf->data_len;
3125
3126 /*
3127 * Incorporate any cached truncation.
3128 *
3129 * NOTE: Modifications to rec_len based on synthesized
3130 * truncation points remove the guarantee that any extended
3131 * data on disk is zero (since the truncations may not have
3132 * taken place on-media yet).
3133 */
3134 if (ip->flags & HAMMER_INODE_TRUNCATED) {
3135 if (hammer_cursor_ondisk(&cursor) ||
3136 cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
3137 if (ip->trunc_off <= rec_offset)
3138 rec_len = 0;
3139 else if (ip->trunc_off < rec_offset + rec_len)
3140 rec_len = (int)(ip->trunc_off - rec_offset);
3141 }
3142 }
3143 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
3144 if (hammer_cursor_ondisk(&cursor)) {
3145 if (ip->sync_trunc_off <= rec_offset)
3146 rec_len = 0;
3147 else if (ip->sync_trunc_off < rec_offset + rec_len)
3148 rec_len = (int)(ip->sync_trunc_off - rec_offset);
3149 }
3150 }
3151
3152 /*
3153 * Accumulate information. If we have hit a discontiguous
3154 * block reset base_offset unless we are already beyond the
3155 * requested offset. If we are, that's it, we stop.
3156 */
3157 if (error)
3158 break;
3159 if (hammer_cursor_ondisk(&cursor)) {
3160 disk_offset = cursor.leaf->data_offset;
3161 if (rec_offset != last_offset ||
3162 disk_offset != last_disk_offset) {
3163 if (rec_offset > ap->a_loffset)
3164 break;
3165 base_offset = rec_offset;
3166 base_disk_offset = disk_offset;
3167 }
3168 last_offset = rec_offset + rec_len;
3169 last_disk_offset = disk_offset + rec_len;
3170
3171 if (hammer_live_dedup)
3172 hammer_dedup_cache_add(ip, cursor.leaf);
3173 }
3174
3175 error = hammer_ip_next(&cursor);
3176 }
3177
3178 #if 0
3179 kprintf("BMAP %016llx: %016llx - %016llx\n",
3180 (long long)ap->a_loffset,
3181 (long long)base_offset,
3182 (long long)last_offset);
3183 kprintf("BMAP %16s: %016llx - %016llx\n", "",
3184 (long long)base_disk_offset,
3185 (long long)last_disk_offset);
3186 #endif
3187
3188 if (cursor.node) {
3189 hammer_cache_node(&ip->cache[1], cursor.node);
3190 #if 0
3191 kprintf("bmap_end2 %016llx ip->cache %p\n",
3192 (long long)ap->a_loffset, ip->cache[1]);
3193 #endif
3194 }
3195 hammer_done_cursor(&cursor);
3196 hammer_done_transaction(&trans);
3197 lwkt_reltoken(&hmp->fs_token);
3198
3199 /*
3200 * If we couldn't find any records or the records we did find were
3201 * all behind the requested offset, return failure. A forward
3202 * truncation can leave a hole w/ no on-disk records.
3203 */
3204 if (last_offset == 0 || last_offset < ap->a_loffset)
3205 return (EOPNOTSUPP);
3206
3207 /*
3208 * Figure out the block size at the requested offset and adjust
3209 * our limits so the cluster_read() does not create inappropriately
3210 * sized buffer cache buffers.
3211 */
3212 blksize = hammer_blocksize(ap->a_loffset);
3213 if (hammer_blocksize(base_offset) != blksize) {
3214 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
3215 }
3216 if (last_offset != ap->a_loffset &&
3217 hammer_blocksize(last_offset - 1) != blksize) {
3218 last_offset = hammer_blockdemarc(ap->a_loffset,
3219 last_offset - 1);
3220 }
3221
3222 /*
3223 * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3224 * from occuring.
3225 */
3226 disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
3227
3228 if ((disk_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_LARGE_DATA) {
3229 /*
3230 * Only large-data zones can be direct-IOd
3231 */
3232 error = EOPNOTSUPP;
3233 } else if ((disk_offset & HAMMER_BUFMASK) ||
3234 (last_offset - ap->a_loffset) < blksize) {
3235 /*
3236 * doffsetp is not aligned or the forward run size does
3237 * not cover a whole buffer, disallow the direct I/O.
3238 */
3239 error = EOPNOTSUPP;
3240 } else {
3241 /*
3242 * We're good.
3243 */
3244 *ap->a_doffsetp = disk_offset;
3245 if (ap->a_runb) {
3246 *ap->a_runb = ap->a_loffset - base_offset;
3247 KKASSERT(*ap->a_runb >= 0);
3248 }
3249 if (ap->a_runp) {
3250 *ap->a_runp = last_offset - ap->a_loffset;
3251 KKASSERT(*ap->a_runp >= 0);
3252 }
3253 error = 0;
3254 }
3255 return(error);
3256 }
3257
3258 /*
3259 * Write to a regular file. Because this is a strategy call the OS is
3260 * trying to actually get data onto the media.
3261 */
3262 static
3263 int
3264 hammer_vop_strategy_write(struct vop_strategy_args *ap)
3265 {
3266 hammer_record_t record;
3267 hammer_mount_t hmp;
3268 hammer_inode_t ip;
3269 struct bio *bio;
3270 struct buf *bp;
3271 int blksize __debugvar;
3272 int bytes;
3273 int error;
3274
3275 bio = ap->a_bio;
3276 bp = bio->bio_buf;
3277 ip = ap->a_vp->v_data;
3278 hmp = ip->hmp;
3279
3280 blksize = hammer_blocksize(bio->bio_offset);
3281 KKASSERT(bp->b_bufsize == blksize);
3282
3283 if (ip->flags & HAMMER_INODE_RO) {
3284 bp->b_error = EROFS;
3285 bp->b_flags |= B_ERROR;
3286 biodone(ap->a_bio);
3287 return(EROFS);
3288 }
3289
3290 lwkt_gettoken(&hmp->fs_token);
3291
3292 /*
3293 * Disallow swapcache operation on the vnode buffer if double
3294 * buffering is enabled, the swapcache will get the data via
3295 * the block device buffer.
3296 */
3297 if (hammer_double_buffer)
3298 bp->b_flags |= B_NOTMETA;
3299
3300 /*
3301 * Interlock with inode destruction (no in-kernel or directory
3302 * topology visibility). If we queue new IO while trying to
3303 * destroy the inode we can deadlock the vtrunc call in
3304 * hammer_inode_unloadable_check().
3305 *
3306 * Besides, there's no point flushing a bp associated with an
3307 * inode that is being destroyed on-media and has no kernel
3308 * references.
3309 */
3310 if ((ip->flags | ip->sync_flags) &
3311 (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
3312 bp->b_resid = 0;
3313 biodone(ap->a_bio);
3314 lwkt_reltoken(&hmp->fs_token);
3315 return(0);
3316 }
3317
3318 /*
3319 * Reserve space and issue a direct-write from the front-end.
3320 * NOTE: The direct_io code will hammer_bread/bcopy smaller
3321 * allocations.
3322 *
3323 * An in-memory record will be installed to reference the storage
3324 * until the flusher can get to it.
3325 *
3326 * Since we own the high level bio the front-end will not try to
3327 * do a direct-read until the write completes.
3328 *
3329 * NOTE: The only time we do not reserve a full-sized buffers
3330 * worth of data is if the file is small. We do not try to
3331 * allocate a fragment (from the small-data zone) at the end of
3332 * an otherwise large file as this can lead to wildly separated
3333 * data.
3334 */
3335 KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
3336 KKASSERT(bio->bio_offset < ip->ino_data.size);
3337 if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
3338 bytes = bp->b_bufsize;
3339 else
3340 bytes = ((int)ip->ino_data.size + 15) & ~15;
3341
3342 record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
3343 bytes, &error);
3344
3345 /*
3346 * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3347 * in hammer_vop_write(). We must flag the record so the proper
3348 * REDO_TERM_WRITE entry is generated during the flush.
3349 */
3350 if (record) {
3351 if (bp->b_flags & B_VFSFLAG1) {
3352 record->flags |= HAMMER_RECF_REDO;
3353 bp->b_flags &= ~B_VFSFLAG1;
3354 }
3355 if (record->flags & HAMMER_RECF_DEDUPED) {
3356 bp->b_resid = 0;
3357 hammer_ip_replace_bulk(hmp, record);
3358 biodone(ap->a_bio);
3359 } else {
3360 hammer_io_direct_write(hmp, bio, record);
3361 }
3362 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
3363 hammer_flush_inode(ip, 0);
3364 } else {
3365 bp->b_bio2.bio_offset = NOOFFSET;
3366 bp->b_error = error;
3367 bp->b_flags |= B_ERROR;
3368 biodone(ap->a_bio);
3369 }
3370 lwkt_reltoken(&hmp->fs_token);
3371 return(error);
3372 }
3373
3374 /*
3375 * dounlink - disconnect a directory entry
3376 *
3377 * XXX whiteout support not really in yet
3378 */
3379 static int
3380 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
3381 struct vnode *dvp, struct ucred *cred,
3382 int flags, int isdir)
3383 {
3384 struct namecache *ncp;
3385 hammer_inode_t dip;
3386 hammer_inode_t ip;
3387 hammer_mount_t hmp;
3388 struct hammer_cursor cursor;
3389 int64_t namekey;
3390 u_int32_t max_iterations;
3391 int nlen, error;
3392
3393 /*
3394 * Calculate the namekey and setup the key range for the scan. This
3395 * works kinda like a chained hash table where the lower 32 bits
3396 * of the namekey synthesize the chain.
3397 *
3398 * The key range is inclusive of both key_beg and key_end.
3399 */
3400 dip = VTOI(dvp);
3401 ncp = nch->ncp;
3402 hmp = dip->hmp;
3403
3404 if (dip->flags & HAMMER_INODE_RO)
3405 return (EROFS);
3406
3407 namekey = hammer_directory_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3408 &max_iterations);
3409 retry:
3410 hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3411 cursor.key_beg.localization = dip->obj_localization +
3412 hammer_dir_localization(dip);
3413 cursor.key_beg.obj_id = dip->obj_id;
3414 cursor.key_beg.key = namekey;
3415 cursor.key_beg.create_tid = 0;
3416 cursor.key_beg.delete_tid = 0;
3417 cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3418 cursor.key_beg.obj_type = 0;
3419
3420 cursor.key_end = cursor.key_beg;
3421 cursor.key_end.key += max_iterations;
3422 cursor.asof = dip->obj_asof;
3423 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3424
3425 /*
3426 * Scan all matching records (the chain), locate the one matching
3427 * the requested path component. info->last_error contains the
3428 * error code on search termination and could be 0, ENOENT, or
3429 * something else.
3430 *
3431 * The hammer_ip_*() functions merge in-memory records with on-disk
3432 * records for the purposes of the search.
3433 */
3434 error = hammer_ip_first(&cursor);
3435
3436 while (error == 0) {
3437 error = hammer_ip_resolve_data(&cursor);
3438 if (error)
3439 break;
3440 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3441 KKASSERT(nlen > 0);
3442 if (ncp->nc_nlen == nlen &&
3443 bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3444 break;
3445 }
3446 error = hammer_ip_next(&cursor);
3447 }
3448
3449 /*
3450 * If all is ok we have to get the inode so we can adjust nlinks.
3451 * To avoid a deadlock with the flusher we must release the inode
3452 * lock on the directory when acquiring the inode for the entry.
3453 *
3454 * If the target is a directory, it must be empty.
3455 */
3456 if (error == 0) {
3457 hammer_unlock(&cursor.ip->lock);
3458 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3459 hmp->asof,
3460 cursor.data->entry.localization,
3461 0, &error);
3462 hammer_lock_sh(&cursor.ip->lock);
3463 if (error == ENOENT) {
3464 kprintf("HAMMER: WARNING: Removing "
3465 "dirent w/missing inode \"%s\"\n"
3466 "\tobj_id = %016llx\n",
3467 ncp->nc_name,
3468 (long long)cursor.data->entry.obj_id);
3469 error = 0;
3470 }
3471
3472 /*
3473 * If isdir >= 0 we validate that the entry is or is not a
3474 * directory. If isdir < 0 we don't care.
3475 */
3476 if (error == 0 && isdir >= 0 && ip) {
3477 if (isdir &&
3478 ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3479 error = ENOTDIR;
3480 } else if (isdir == 0 &&
3481 ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3482 error = EISDIR;
3483 }
3484 }
3485
3486 /*
3487 * If we are trying to remove a directory the directory must
3488 * be empty.
3489 *
3490 * The check directory code can loop and deadlock/retry. Our
3491 * own cursor's node locks must be released to avoid a 3-way
3492 * deadlock with the flusher if the check directory code
3493 * blocks.
3494 *
3495 * If any changes whatsoever have been made to the cursor
3496 * set EDEADLK and retry.
3497 *
3498 * WARNING: See warnings in hammer_unlock_cursor()
3499 * function.
3500 */
3501 if (error == 0 && ip && ip->ino_data.obj_type ==
3502 HAMMER_OBJTYPE_DIRECTORY) {
3503 hammer_unlock_cursor(&cursor);
3504 error = hammer_ip_check_directory_empty(trans, ip);
3505 hammer_lock_cursor(&cursor);
3506 if (cursor.flags & HAMMER_CURSOR_RETEST) {
3507 kprintf("HAMMER: Warning: avoided deadlock "
3508 "on rmdir '%s'\n",
3509 ncp->nc_name);
3510 error = EDEADLK;
3511 }
3512 }
3513
3514 /*
3515 * Delete the directory entry.
3516 *
3517 * WARNING: hammer_ip_del_directory() may have to terminate
3518 * the cursor to avoid a deadlock. It is ok to call
3519 * hammer_done_cursor() twice.
3520 */
3521 if (error == 0) {
3522 error = hammer_ip_del_directory(trans, &cursor,
3523 dip, ip);
3524 }
3525 hammer_done_cursor(&cursor);
3526 if (error == 0) {
3527 /*
3528 * Tell the namecache that we are now unlinked.
3529 */
3530 cache_unlink(nch);
3531
3532 /*
3533 * NOTE: ip->vp, if non-NULL, cannot be directly
3534 * referenced without formally acquiring the
3535 * vp since the vp might have zero refs on it,
3536 * or in the middle of a reclaim, etc.
3537 *
3538 * NOTE: The cache_setunresolved() can rip the vp
3539 * out from under us since the vp may not have
3540 * any refs, in which case ip->vp will be NULL
3541 * from the outset.
3542 */
3543 while (ip && ip->vp) {
3544 struct vnode *vp;
3545
3546 error = hammer_get_vnode(ip, &vp);
3547 if (error == 0 && vp) {
3548 vn_unlock(vp);
3549 hammer_knote(ip->vp, NOTE_DELETE);
3550 #if 0
3551 /*
3552 * Don't do this, it can deadlock
3553 * on concurrent rm's of hardlinks.
3554 * Shouldn't be needed any more.
3555 */
3556 cache_inval_vp(ip->vp, CINV_DESTROY);
3557 #endif
3558 vrele(vp);
3559 break;
3560 }
3561 kprintf("Debug: HAMMER ip/vp race1 avoided\n");
3562 }
3563 }
3564 if (ip)
3565 hammer_rel_inode(ip, 0);
3566 } else {
3567 hammer_done_cursor(&cursor);
3568 }
3569 if (error == EDEADLK)
3570 goto retry;
3571
3572 return (error);
3573 }
3574
3575 /************************************************************************
3576 * FIFO AND SPECFS OPS *
3577 ************************************************************************
3578 *
3579 */
3580 static int
3581 hammer_vop_fifoclose (struct vop_close_args *ap)
3582 {
3583 /* XXX update itimes */
3584 return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3585 }
3586
3587 static int
3588 hammer_vop_fiforead (struct vop_read_args *ap)
3589 {
3590 int error;
3591
3592 error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3593 /* XXX update access time */
3594 return (error);
3595 }
3596
3597 static int
3598 hammer_vop_fifowrite (struct vop_write_args *ap)
3599 {
3600 int error;
3601
3602 error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3603 /* XXX update access time */
3604 return (error);
3605 }
3606
3607 static
3608 int
3609 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3610 {
3611 int error;
3612
3613 error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3614 if (error)
3615 error = hammer_vop_kqfilter(ap);
3616 return(error);
3617 }
3618
3619 /************************************************************************
3620 * KQFILTER OPS *
3621 ************************************************************************
3622 *
3623 */
3624 static void filt_hammerdetach(struct knote *kn);
3625 static int filt_hammerread(struct knote *kn, long hint);
3626 static int filt_hammerwrite(struct knote *kn, long hint);
3627 static int filt_hammervnode(struct knote *kn, long hint);
3628
3629 static struct filterops hammerread_filtops =
3630 { FILTEROP_ISFD | FILTEROP_MPSAFE,
3631 NULL, filt_hammerdetach, filt_hammerread };
3632 static struct filterops hammerwrite_filtops =
3633 { FILTEROP_ISFD | FILTEROP_MPSAFE,
3634 NULL, filt_hammerdetach, filt_hammerwrite };
3635 static struct filterops hammervnode_filtops =
3636 { FILTEROP_ISFD | FILTEROP_MPSAFE,
3637 NULL, filt_hammerdetach, filt_hammervnode };
3638
3639 static
3640 int
3641 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3642 {
3643 struct vnode *vp = ap->a_vp;
3644 struct knote *kn = ap->a_kn;
3645
3646 switch (kn->kn_filter) {
3647 case EVFILT_READ:
3648 kn->kn_fop = &hammerread_filtops;
3649 break;
3650 case EVFILT_WRITE:
3651 kn->kn_fop = &hammerwrite_filtops;
3652 break;
3653 case EVFILT_VNODE:
3654 kn->kn_fop = &hammervnode_filtops;
3655 break;
3656 default:
3657 return (EOPNOTSUPP);
3658 }
3659
3660 kn->kn_hook = (caddr_t)vp;
3661
3662 knote_insert(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3663
3664 return(0);
3665 }
3666
3667 static void
3668 filt_hammerdetach(struct knote *kn)
3669 {
3670 struct vnode *vp = (void *)kn->kn_hook;
3671
3672 knote_remove(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3673 }
3674
3675 static int
3676 filt_hammerread(struct knote *kn, long hint)
3677 {
3678 struct vnode *vp = (void *)kn->kn_hook;
3679 hammer_inode_t ip = VTOI(vp);
3680 hammer_mount_t hmp = ip->hmp;
3681 off_t off;
3682
3683 if (hint == NOTE_REVOKE) {
3684 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3685 return(1);
3686 }
3687 lwkt_gettoken(&hmp->fs_token); /* XXX use per-ip-token */
3688 off = ip->ino_data.size - kn->kn_fp->f_offset;
3689 kn->kn_data = (off < INTPTR_MAX) ? off : INTPTR_MAX;
3690 lwkt_reltoken(&hmp->fs_token);
3691 if (kn->kn_sfflags & NOTE_OLDAPI)
3692 return(1);
3693 return (kn->kn_data != 0);
3694 }
3695
3696 static int
3697 filt_hammerwrite(struct knote *kn, long hint)
3698 {
3699 if (hint == NOTE_REVOKE)
3700 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3701 kn->kn_data = 0;
3702 return (1);
3703 }
3704
3705 static int
3706 filt_hammervnode(struct knote *kn, long hint)
3707 {
3708 if (kn->kn_sfflags & hint)
3709 kn->kn_fflags |= hint;
3710 if (hint == NOTE_REVOKE) {
3711 kn->kn_flags |= (EV_EOF | EV_NODATA);
3712 return (1);
3713 }
3714 return (kn->kn_fflags != 0);
3715 }
3716
Cache object: 0b99dab3f554bcd276a95555e4d3072d
|