?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
HAMMER VFS - Fix deadlock in rmdir
[dragonfly.git] / sys / vfs / hammer / hammer_vnops.c
blob84e36d1ae38306b6a1c1991ab2e3bb8d9289f784
1 /*
2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved.
3 *
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 *
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * SUCH DAMAGE.
33 *
34 * $DragonFly: src/sys/vfs/hammer/hammer_vnops.c,v 1.102 2008/10/16 17:24:16 dillon Exp $
35 */
36
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
40 #include <sys/fcntl.h>
41 #include <sys/namecache.h>
42 #include <sys/vnode.h>
43 #include <sys/lockf.h>
44 #include <sys/event.h>
45 #include <sys/stat.h>
46 #include <sys/dirent.h>
47 #include <sys/file.h>
48 #include <vm/vm_extern.h>
49 #include <vfs/fifofs/fifo.h>
50 #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 static int hammer_vop_specclose (struct vop_close_args *);
92 static int hammer_vop_specread (struct vop_read_args *);
93 static int hammer_vop_specwrite (struct vop_write_args *);
94
95 struct vop_ops hammer_vnode_vops = {
96 .vop_default = vop_defaultop,
97 .vop_fsync = hammer_vop_fsync,
98 .vop_getpages = vop_stdgetpages,
99 .vop_putpages = vop_stdputpages,
100 .vop_read = hammer_vop_read,
101 .vop_write = hammer_vop_write,
102 .vop_access = hammer_vop_access,
103 .vop_advlock = hammer_vop_advlock,
104 .vop_close = hammer_vop_close,
105 .vop_ncreate = hammer_vop_ncreate,
106 .vop_getattr = hammer_vop_getattr,
107 .vop_inactive = hammer_vop_inactive,
108 .vop_reclaim = hammer_vop_reclaim,
109 .vop_nresolve = hammer_vop_nresolve,
110 .vop_nlookupdotdot = hammer_vop_nlookupdotdot,
111 .vop_nlink = hammer_vop_nlink,
112 .vop_nmkdir = hammer_vop_nmkdir,
113 .vop_nmknod = hammer_vop_nmknod,
114 .vop_open = hammer_vop_open,
115 .vop_pathconf = vop_stdpathconf,
116 .vop_print = hammer_vop_print,
117 .vop_readdir = hammer_vop_readdir,
118 .vop_readlink = hammer_vop_readlink,
119 .vop_nremove = hammer_vop_nremove,
120 .vop_nrename = hammer_vop_nrename,
121 .vop_nrmdir = hammer_vop_nrmdir,
122 .vop_markatime = hammer_vop_markatime,
123 .vop_setattr = hammer_vop_setattr,
124 .vop_bmap = hammer_vop_bmap,
125 .vop_strategy = hammer_vop_strategy,
126 .vop_nsymlink = hammer_vop_nsymlink,
127 .vop_nwhiteout = hammer_vop_nwhiteout,
128 .vop_ioctl = hammer_vop_ioctl,
129 .vop_mountctl = hammer_vop_mountctl,
130 .vop_kqfilter = hammer_vop_kqfilter
131 };
132
133 struct vop_ops hammer_spec_vops = {
134 .vop_default = spec_vnoperate,
135 .vop_fsync = hammer_vop_fsync,
136 .vop_read = hammer_vop_specread,
137 .vop_write = hammer_vop_specwrite,
138 .vop_access = hammer_vop_access,
139 .vop_close = hammer_vop_specclose,
140 .vop_markatime = hammer_vop_markatime,
141 .vop_getattr = hammer_vop_getattr,
142 .vop_inactive = hammer_vop_inactive,
143 .vop_reclaim = hammer_vop_reclaim,
144 .vop_setattr = hammer_vop_setattr
145 };
146
147 struct vop_ops hammer_fifo_vops = {
148 .vop_default = fifo_vnoperate,
149 .vop_fsync = hammer_vop_fsync,
150 .vop_read = hammer_vop_fiforead,
151 .vop_write = hammer_vop_fifowrite,
152 .vop_access = hammer_vop_access,
153 .vop_close = hammer_vop_fifoclose,
154 .vop_markatime = hammer_vop_markatime,
155 .vop_getattr = hammer_vop_getattr,
156 .vop_inactive = hammer_vop_inactive,
157 .vop_reclaim = hammer_vop_reclaim,
158 .vop_setattr = hammer_vop_setattr,
159 .vop_kqfilter = hammer_vop_fifokqfilter
160 };
161
162 static __inline
163 void
164 hammer_knote(struct vnode *vp, int flags)
165 {
166 if (flags)
167 KNOTE(&vp->v_pollinfo.vpi_selinfo.si_note, flags);
168 }
169
170 #ifdef DEBUG_TRUNCATE
171 struct hammer_inode *HammerTruncIp;
172 #endif
173
174 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
175 struct vnode *dvp, struct ucred *cred,
176 int flags, int isdir);
177 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
178 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
179
180 #if 0
181 static
182 int
183 hammer_vop_vnoperate(struct vop_generic_args *)
184 {
185 return (VOCALL(&hammer_vnode_vops, ap));
186 }
187 #endif
188
189 /*
190 * hammer_vop_fsync { vp, waitfor }
191 *
192 * fsync() an inode to disk and wait for it to be completely committed
193 * such that the information would not be undone if a crash occured after
194 * return.
195 */
196 static
197 int
198 hammer_vop_fsync(struct vop_fsync_args *ap)
199 {
200 hammer_inode_t ip = VTOI(ap->a_vp);
201
202 ++hammer_count_fsyncs;
203 vfsync(ap->a_vp, ap->a_waitfor, 1, NULL, NULL);
204 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
205 if (ap->a_waitfor == MNT_WAIT) {
206 vn_unlock(ap->a_vp);
207 hammer_wait_inode(ip);
208 vn_lock(ap->a_vp, LK_EXCLUSIVE | LK_RETRY);
209 }
210 return (ip->error);
211 }
212
213 /*
214 * hammer_vop_read { vp, uio, ioflag, cred }
215 */
216 static
217 int
218 hammer_vop_read(struct vop_read_args *ap)
219 {
220 struct hammer_transaction trans;
221 hammer_inode_t ip;
222 off_t offset;
223 struct buf *bp;
224 struct uio *uio;
225 int error;
226 int n;
227 int seqcount;
228 int ioseqcount;
229 int blksize;
230
231 if (ap->a_vp->v_type != VREG)
232 return (EINVAL);
233 ip = VTOI(ap->a_vp);
234 error = 0;
235 uio = ap->a_uio;
236
237 /*
238 * Allow the UIO's size to override the sequential heuristic.
239 */
240 blksize = hammer_blocksize(uio->uio_offset);
241 seqcount = (uio->uio_resid + (blksize - 1)) / blksize;
242 ioseqcount = ap->a_ioflag >> 16;
243 if (seqcount < ioseqcount)
244 seqcount = ioseqcount;
245
246 hammer_start_transaction(&trans, ip->hmp);
247
248 /*
249 * Access the data typically in HAMMER_BUFSIZE blocks via the
250 * buffer cache, but HAMMER may use a variable block size based
251 * on the offset.
252 */
253 while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
254 int64_t base_offset;
255 int64_t file_limit;
256
257 blksize = hammer_blocksize(uio->uio_offset);
258 offset = (int)uio->uio_offset & (blksize - 1);
259 base_offset = uio->uio_offset - offset;
260
261 if (hammer_cluster_enable) {
262 /*
263 * Use file_limit to prevent cluster_read() from
264 * creating buffers of the wrong block size past
265 * the demarc.
266 */
267 file_limit = ip->ino_data.size;
268 if (base_offset < HAMMER_XDEMARC &&
269 file_limit > HAMMER_XDEMARC) {
270 file_limit = HAMMER_XDEMARC;
271 }
272 error = cluster_read(ap->a_vp,
273 file_limit, base_offset,
274 blksize, MAXPHYS,
275 seqcount, &bp);
276 } else {
277 error = bread(ap->a_vp, base_offset, blksize, &bp);
278 }
279 if (error) {
280 kprintf("error %d\n", error);
281 brelse(bp);
282 break;
283 }
284
285 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
286 n = blksize - offset;
287 if (n > uio->uio_resid)
288 n = uio->uio_resid;
289 if (n > ip->ino_data.size - uio->uio_offset)
290 n = (int)(ip->ino_data.size - uio->uio_offset);
291 error = uiomove((char *)bp->b_data + offset, n, uio);
292
293 /* data has a lower priority then meta-data */
294 bp->b_flags |= B_AGE;
295 bqrelse(bp);
296 if (error)
297 break;
298 hammer_stats_file_read += n;
299 }
300 if ((ip->flags & HAMMER_INODE_RO) == 0 &&
301 (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
302 ip->ino_data.atime = trans.time;
303 hammer_modify_inode(ip, HAMMER_INODE_ATIME);
304 }
305 hammer_done_transaction(&trans);
306 return (error);
307 }
308
309 /*
310 * hammer_vop_write { vp, uio, ioflag, cred }
311 */
312 static
313 int
314 hammer_vop_write(struct vop_write_args *ap)
315 {
316 struct hammer_transaction trans;
317 struct hammer_inode *ip;
318 hammer_mount_t hmp;
319 struct uio *uio;
320 int offset;
321 off_t base_offset;
322 struct buf *bp;
323 int kflags;
324 int error;
325 int n;
326 int flags;
327 int delta;
328 int seqcount;
329
330 if (ap->a_vp->v_type != VREG)
331 return (EINVAL);
332 ip = VTOI(ap->a_vp);
333 hmp = ip->hmp;
334 error = 0;
335 kflags = 0;
336 seqcount = ap->a_ioflag >> 16;
337
338 if (ip->flags & HAMMER_INODE_RO)
339 return (EROFS);
340
341 /*
342 * Create a transaction to cover the operations we perform.
343 */
344 hammer_start_transaction(&trans, hmp);
345 uio = ap->a_uio;
346
347 /*
348 * Check append mode
349 */
350 if (ap->a_ioflag & IO_APPEND)
351 uio->uio_offset = ip->ino_data.size;
352
353 /*
354 * Check for illegal write offsets. Valid range is 0...2^63-1.
355 *
356 * NOTE: the base_off assignment is required to work around what
357 * I consider to be a GCC-4 optimization bug.
358 */
359 if (uio->uio_offset < 0) {
360 hammer_done_transaction(&trans);
361 return (EFBIG);
362 }
363 base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
364 if (uio->uio_resid > 0 && base_offset <= 0) {
365 hammer_done_transaction(&trans);
366 return (EFBIG);
367 }
368
369 /*
370 * Access the data typically in HAMMER_BUFSIZE blocks via the
371 * buffer cache, but HAMMER may use a variable block size based
372 * on the offset.
373 */
374 while (uio->uio_resid > 0) {
375 int fixsize = 0;
376 int blksize;
377 int blkmask;
378
379 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
380 break;
381
382 blksize = hammer_blocksize(uio->uio_offset);
383
384 /*
385 * Do not allow HAMMER to blow out the buffer cache. Very
386 * large UIOs can lockout other processes due to bwillwrite()
387 * mechanics.
388 *
389 * The hammer inode is not locked during these operations.
390 * The vnode is locked which can interfere with the pageout
391 * daemon for non-UIO_NOCOPY writes but should not interfere
392 * with the buffer cache. Even so, we cannot afford to
393 * allow the pageout daemon to build up too many dirty buffer
394 * cache buffers.
395 *
396 * Only call this if we aren't being recursively called from
397 * a virtual disk device (vn), else we may deadlock.
398 */
399 if ((ap->a_ioflag & IO_RECURSE) == 0)
400 bwillwrite(blksize);
401
402 /*
403 * Do not allow HAMMER to blow out system memory by
404 * accumulating too many records. Records are so well
405 * decoupled from the buffer cache that it is possible
406 * for userland to push data out to the media via
407 * direct-write, but build up the records queued to the
408 * backend faster then the backend can flush them out.
409 * HAMMER has hit its write limit but the frontend has
410 * no pushback to slow it down.
411 */
412 if (hmp->rsv_recs > hammer_limit_recs / 2) {
413 /*
414 * Get the inode on the flush list
415 */
416 if (ip->rsv_recs >= 64)
417 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
418 else if (ip->rsv_recs >= 16)
419 hammer_flush_inode(ip, 0);
420
421 /*
422 * Keep the flusher going if the system keeps
423 * queueing records.
424 */
425 delta = hmp->count_newrecords -
426 hmp->last_newrecords;
427 if (delta < 0 || delta > hammer_limit_recs / 2) {
428 hmp->last_newrecords = hmp->count_newrecords;
429 hammer_sync_hmp(hmp, MNT_NOWAIT);
430 }
431
432 /*
433 * If we have gotten behind start slowing
434 * down the writers.
435 */
436 delta = (hmp->rsv_recs - hammer_limit_recs) *
437 hz / hammer_limit_recs;
438 if (delta > 0)
439 tsleep(&trans, 0, "hmrslo", delta);
440 }
441
442 /*
443 * Calculate the blocksize at the current offset and figure
444 * out how much we can actually write.
445 */
446 blkmask = blksize - 1;
447 offset = (int)uio->uio_offset & blkmask;
448 base_offset = uio->uio_offset & ~(int64_t)blkmask;
449 n = blksize - offset;
450 if (n > uio->uio_resid)
451 n = uio->uio_resid;
452 if (uio->uio_offset + n > ip->ino_data.size) {
453 vnode_pager_setsize(ap->a_vp, uio->uio_offset + n);
454 fixsize = 1;
455 kflags |= NOTE_EXTEND;
456 }
457
458 if (uio->uio_segflg == UIO_NOCOPY) {
459 /*
460 * Issuing a write with the same data backing the
461 * buffer. Instantiate the buffer to collect the
462 * backing vm pages, then read-in any missing bits.
463 *
464 * This case is used by vop_stdputpages().
465 */
466 bp = getblk(ap->a_vp, base_offset,
467 blksize, GETBLK_BHEAVY, 0);
468 if ((bp->b_flags & B_CACHE) == 0) {
469 bqrelse(bp);
470 error = bread(ap->a_vp, base_offset,
471 blksize, &bp);
472 }
473 } else if (offset == 0 && uio->uio_resid >= blksize) {
474 /*
475 * Even though we are entirely overwriting the buffer
476 * we may still have to zero it out to avoid a
477 * mmap/write visibility issue.
478 */
479 bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
480 if ((bp->b_flags & B_CACHE) == 0)
481 vfs_bio_clrbuf(bp);
482 } else if (base_offset >= ip->ino_data.size) {
483 /*
484 * If the base offset of the buffer is beyond the
485 * file EOF, we don't have to issue a read.
486 */
487 bp = getblk(ap->a_vp, base_offset,
488 blksize, GETBLK_BHEAVY, 0);
489 vfs_bio_clrbuf(bp);
490 } else {
491 /*
492 * Partial overwrite, read in any missing bits then
493 * replace the portion being written.
494 */
495 error = bread(ap->a_vp, base_offset, blksize, &bp);
496 if (error == 0)
497 bheavy(bp);
498 }
499 if (error == 0) {
500 error = uiomove((char *)bp->b_data + offset,
501 n, uio);
502 }
503
504 /*
505 * If we screwed up we have to undo any VM size changes we
506 * made.
507 */
508 if (error) {
509 brelse(bp);
510 if (fixsize) {
511 vtruncbuf(ap->a_vp, ip->ino_data.size,
512 hammer_blocksize(ip->ino_data.size));
513 }
514 break;
515 }
516 kflags |= NOTE_WRITE;
517 hammer_stats_file_write += n;
518 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
519 if (ip->ino_data.size < uio->uio_offset) {
520 ip->ino_data.size = uio->uio_offset;
521 flags = HAMMER_INODE_DDIRTY;
522 vnode_pager_setsize(ap->a_vp, ip->ino_data.size);
523 } else {
524 flags = 0;
525 }
526 ip->ino_data.mtime = trans.time;
527 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
528 hammer_modify_inode(ip, flags);
529
530 /*
531 * Once we dirty the buffer any cached zone-X offset
532 * becomes invalid. HAMMER NOTE: no-history mode cannot
533 * allow overwriting over the same data sector unless
534 * we provide UNDOs for the old data, which we don't.
535 */
536 bp->b_bio2.bio_offset = NOOFFSET;
537
538 /*
539 * Final buffer disposition.
540 */
541 bp->b_flags |= B_AGE;
542 if (ap->a_ioflag & IO_SYNC) {
543 bwrite(bp);
544 } else if (ap->a_ioflag & IO_DIRECT) {
545 bawrite(bp);
546 } else {
547 bdwrite(bp);
548 }
549 }
550 hammer_done_transaction(&trans);
551 hammer_knote(ap->a_vp, kflags);
552 return (error);
553 }
554
555 /*
556 * hammer_vop_access { vp, mode, cred }
557 */
558 static
559 int
560 hammer_vop_access(struct vop_access_args *ap)
561 {
562 struct hammer_inode *ip = VTOI(ap->a_vp);
563 uid_t uid;
564 gid_t gid;
565 int error;
566
567 ++hammer_stats_file_iopsr;
568 uid = hammer_to_unix_xid(&ip->ino_data.uid);
569 gid = hammer_to_unix_xid(&ip->ino_data.gid);
570
571 error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
572 ip->ino_data.uflags);
573 return (error);
574 }
575
576 /*
577 * hammer_vop_advlock { vp, id, op, fl, flags }
578 */
579 static
580 int
581 hammer_vop_advlock(struct vop_advlock_args *ap)
582 {
583 hammer_inode_t ip = VTOI(ap->a_vp);
584
585 return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
586 }
587
588 /*
589 * hammer_vop_close { vp, fflag }
590 */
591 static
592 int
593 hammer_vop_close(struct vop_close_args *ap)
594 {
595 /*hammer_inode_t ip = VTOI(ap->a_vp);*/
596 return (vop_stdclose(ap));
597 }
598
599 /*
600 * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
601 *
602 * The operating system has already ensured that the directory entry
603 * does not exist and done all appropriate namespace locking.
604 */
605 static
606 int
607 hammer_vop_ncreate(struct vop_ncreate_args *ap)
608 {
609 struct hammer_transaction trans;
610 struct hammer_inode *dip;
611 struct hammer_inode *nip;
612 struct nchandle *nch;
613 int error;
614
615 nch = ap->a_nch;
616 dip = VTOI(ap->a_dvp);
617
618 if (dip->flags & HAMMER_INODE_RO)
619 return (EROFS);
620 if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
621 return (error);
622
623 /*
624 * Create a transaction to cover the operations we perform.
625 */
626 hammer_start_transaction(&trans, dip->hmp);
627 ++hammer_stats_file_iopsw;
628
629 /*
630 * Create a new filesystem object of the requested type. The
631 * returned inode will be referenced and shared-locked to prevent
632 * it from being moved to the flusher.
633 */
634
635 error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
636 dip, NULL, &nip);
637 if (error) {
638 hkprintf("hammer_create_inode error %d\n", error);
639 hammer_done_transaction(&trans);
640 *ap->a_vpp = NULL;
641 return (error);
642 }
643
644 /*
645 * Add the new filesystem object to the directory. This will also
646 * bump the inode's link count.
647 */
648 error = hammer_ip_add_directory(&trans, dip,
649 nch->ncp->nc_name, nch->ncp->nc_nlen,
650 nip);
651 if (error)
652 hkprintf("hammer_ip_add_directory error %d\n", error);
653
654 /*
655 * Finish up.
656 */
657 if (error) {
658 hammer_rel_inode(nip, 0);
659 hammer_done_transaction(&trans);
660 *ap->a_vpp = NULL;
661 } else {
662 error = hammer_get_vnode(nip, ap->a_vpp);
663 hammer_done_transaction(&trans);
664 hammer_rel_inode(nip, 0);
665 if (error == 0) {
666 cache_setunresolved(ap->a_nch);
667 cache_setvp(ap->a_nch, *ap->a_vpp);
668 }
669 hammer_knote(ap->a_dvp, NOTE_WRITE);
670 }
671 return (error);
672 }
673
674 /*
675 * hammer_vop_getattr { vp, vap }
676 *
677 * Retrieve an inode's attribute information. When accessing inodes
678 * historically we fake the atime field to ensure consistent results.
679 * The atime field is stored in the B-Tree element and allowed to be
680 * updated without cycling the element.
681 */
682 static
683 int
684 hammer_vop_getattr(struct vop_getattr_args *ap)
685 {
686 struct hammer_inode *ip = VTOI(ap->a_vp);
687 struct vattr *vap = ap->a_vap;
688
689 /*
690 * We want the fsid to be different when accessing a filesystem
691 * with different as-of's so programs like diff don't think
692 * the files are the same.
693 *
694 * We also want the fsid to be the same when comparing snapshots,
695 * or when comparing mirrors (which might be backed by different
696 * physical devices). HAMMER fsids are based on the PFS's
697 * shared_uuid field.
698 *
699 * XXX there is a chance of collision here. The va_fsid reported
700 * by stat is different from the more involved fsid used in the
701 * mount structure.
702 */
703 ++hammer_stats_file_iopsr;
704 vap->va_fsid = ip->pfsm->fsid_udev ^ (u_int32_t)ip->obj_asof ^
705 (u_int32_t)(ip->obj_asof >> 32);
706
707 vap->va_fileid = ip->ino_leaf.base.obj_id;
708 vap->va_mode = ip->ino_data.mode;
709 vap->va_nlink = ip->ino_data.nlinks;
710 vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
711 vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
712 vap->va_rmajor = 0;
713 vap->va_rminor = 0;
714 vap->va_size = ip->ino_data.size;
715
716 /*
717 * Special case for @@PFS softlinks. The actual size of the
718 * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
719 */
720 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
721 ip->ino_data.size == 10 &&
722 ip->obj_asof == HAMMER_MAX_TID &&
723 ip->obj_localization == 0 &&
724 strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
725 vap->va_size = 26;
726 }
727
728 /*
729 * We must provide a consistent atime and mtime for snapshots
730 * so people can do a 'tar cf - ... | md5' on them and get
731 * consistent results.
732 */
733 if (ip->flags & HAMMER_INODE_RO) {
734 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
735 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
736 } else {
737 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
738 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
739 }
740 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
741 vap->va_flags = ip->ino_data.uflags;
742 vap->va_gen = 1; /* hammer inums are unique for all time */
743 vap->va_blocksize = HAMMER_BUFSIZE;
744 if (ip->ino_data.size >= HAMMER_XDEMARC) {
745 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
746 ~HAMMER_XBUFMASK64;
747 } else if (ip->ino_data.size > HAMMER_BUFSIZE / 2) {
748 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
749 ~HAMMER_BUFMASK64;
750 } else {
751 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
752 }
753
754 vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
755 vap->va_filerev = 0; /* XXX */
756 /* mtime uniquely identifies any adjustments made to the file XXX */
757 vap->va_fsmid = ip->ino_data.mtime;
758 vap->va_uid_uuid = ip->ino_data.uid;
759 vap->va_gid_uuid = ip->ino_data.gid;
760 vap->va_fsid_uuid = ip->hmp->fsid;
761 vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
762 VA_FSID_UUID_VALID;
763
764 switch (ip->ino_data.obj_type) {
765 case HAMMER_OBJTYPE_CDEV:
766 case HAMMER_OBJTYPE_BDEV:
767 vap->va_rmajor = ip->ino_data.rmajor;
768 vap->va_rminor = ip->ino_data.rminor;
769 break;
770 default:
771 break;
772 }
773 return(0);
774 }
775
776 /*
777 * hammer_vop_nresolve { nch, dvp, cred }
778 *
779 * Locate the requested directory entry.
780 */
781 static
782 int
783 hammer_vop_nresolve(struct vop_nresolve_args *ap)
784 {
785 struct hammer_transaction trans;
786 struct namecache *ncp;
787 hammer_inode_t dip;
788 hammer_inode_t ip;
789 hammer_tid_t asof;
790 struct hammer_cursor cursor;
791 struct vnode *vp;
792 int64_t namekey;
793 int error;
794 int i;
795 int nlen;
796 int flags;
797 int ispfs;
798 int64_t obj_id;
799 u_int32_t localization;
800 u_int32_t max_iterations;
801
802 /*
803 * Misc initialization, plus handle as-of name extensions. Look for
804 * the '@@' extension. Note that as-of files and directories cannot
805 * be modified.
806 */
807 dip = VTOI(ap->a_dvp);
808 ncp = ap->a_nch->ncp;
809 asof = dip->obj_asof;
810 localization = dip->obj_localization; /* for code consistency */
811 nlen = ncp->nc_nlen;
812 flags = dip->flags & HAMMER_INODE_RO;
813 ispfs = 0;
814
815 hammer_simple_transaction(&trans, dip->hmp);
816 ++hammer_stats_file_iopsr;
817
818 for (i = 0; i < nlen; ++i) {
819 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
820 error = hammer_str_to_tid(ncp->nc_name + i + 2,
821 &ispfs, &asof, &localization);
822 if (error != 0) {
823 i = nlen;
824 break;
825 }
826 if (asof != HAMMER_MAX_TID)
827 flags |= HAMMER_INODE_RO;
828 break;
829 }
830 }
831 nlen = i;
832
833 /*
834 * If this is a PFS softlink we dive into the PFS
835 */
836 if (ispfs && nlen == 0) {
837 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
838 asof, localization,
839 flags, &error);
840 if (error == 0) {
841 error = hammer_get_vnode(ip, &vp);
842 hammer_rel_inode(ip, 0);
843 } else {
844 vp = NULL;
845 }
846 if (error == 0) {
847 vn_unlock(vp);
848 cache_setvp(ap->a_nch, vp);
849 vrele(vp);
850 }
851 goto done;
852 }
853
854 /*
855 * If there is no path component the time extension is relative to
856 * dip.
857 */
858 if (nlen == 0) {
859 ip = hammer_get_inode(&trans, dip, dip->obj_id,
860 asof, dip->obj_localization,
861 flags, &error);
862 if (error == 0) {
863 error = hammer_get_vnode(ip, &vp);
864 hammer_rel_inode(ip, 0);
865 } else {
866 vp = NULL;
867 }
868 if (error == 0) {
869 vn_unlock(vp);
870 cache_setvp(ap->a_nch, vp);
871 vrele(vp);
872 }
873 goto done;
874 }
875
876 /*
877 * Calculate the namekey and setup the key range for the scan. This
878 * works kinda like a chained hash table where the lower 32 bits
879 * of the namekey synthesize the chain.
880 *
881 * The key range is inclusive of both key_beg and key_end.
882 */
883 namekey = hammer_directory_namekey(dip, ncp->nc_name, nlen,
884 &max_iterations);
885
886 error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
887 cursor.key_beg.localization = dip->obj_localization +
888 HAMMER_LOCALIZE_MISC;
889 cursor.key_beg.obj_id = dip->obj_id;
890 cursor.key_beg.key = namekey;
891 cursor.key_beg.create_tid = 0;
892 cursor.key_beg.delete_tid = 0;
893 cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
894 cursor.key_beg.obj_type = 0;
895
896 cursor.key_end = cursor.key_beg;
897 cursor.key_end.key += max_iterations;
898 cursor.asof = asof;
899 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
900
901 /*
902 * Scan all matching records (the chain), locate the one matching
903 * the requested path component.
904 *
905 * The hammer_ip_*() functions merge in-memory records with on-disk
906 * records for the purposes of the search.
907 */
908 obj_id = 0;
909 localization = HAMMER_DEF_LOCALIZATION;
910
911 if (error == 0) {
912 error = hammer_ip_first(&cursor);
913 while (error == 0) {
914 error = hammer_ip_resolve_data(&cursor);
915 if (error)
916 break;
917 if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
918 bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
919 obj_id = cursor.data->entry.obj_id;
920 localization = cursor.data->entry.localization;
921 break;
922 }
923 error = hammer_ip_next(&cursor);
924 }
925 }
926 hammer_done_cursor(&cursor);
927 if (error == 0) {
928 ip = hammer_get_inode(&trans, dip, obj_id,
929 asof, localization,
930 flags, &error);
931 if (error == 0) {
932 error = hammer_get_vnode(ip, &vp);
933 hammer_rel_inode(ip, 0);
934 } else {
935 vp = NULL;
936 }
937 if (error == 0) {
938 vn_unlock(vp);
939 cache_setvp(ap->a_nch, vp);
940 vrele(vp);
941 }
942 } else if (error == ENOENT) {
943 cache_setvp(ap->a_nch, NULL);
944 }
945 done:
946 hammer_done_transaction(&trans);
947 return (error);
948 }
949
950 /*
951 * hammer_vop_nlookupdotdot { dvp, vpp, cred }
952 *
953 * Locate the parent directory of a directory vnode.
954 *
955 * dvp is referenced but not locked. *vpp must be returned referenced and
956 * locked. A parent_obj_id of 0 does not necessarily indicate that we are
957 * at the root, instead it could indicate that the directory we were in was
958 * removed.
959 *
960 * NOTE: as-of sequences are not linked into the directory structure. If
961 * we are at the root with a different asof then the mount point, reload
962 * the same directory with the mount point's asof. I'm not sure what this
963 * will do to NFS. We encode ASOF stamps in NFS file handles so it might not
964 * get confused, but it hasn't been tested.
965 */
966 static
967 int
968 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
969 {
970 struct hammer_transaction trans;
971 struct hammer_inode *dip;
972 struct hammer_inode *ip;
973 int64_t parent_obj_id;
974 u_int32_t parent_obj_localization;
975 hammer_tid_t asof;
976 int error;
977
978 dip = VTOI(ap->a_dvp);
979 asof = dip->obj_asof;
980
981 /*
982 * Whos are parent? This could be the root of a pseudo-filesystem
983 * whos parent is in another localization domain.
984 */
985 parent_obj_id = dip->ino_data.parent_obj_id;
986 if (dip->obj_id == HAMMER_OBJID_ROOT)
987 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
988 else
989 parent_obj_localization = dip->obj_localization;
990
991 if (parent_obj_id == 0) {
992 if (dip->obj_id == HAMMER_OBJID_ROOT &&
993 asof != dip->hmp->asof) {
994 parent_obj_id = dip->obj_id;
995 asof = dip->hmp->asof;
996 *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
997 ksnprintf(*ap->a_fakename, 19, "0x%016llx",
998 dip->obj_asof);
999 } else {
1000 *ap->a_vpp = NULL;
1001 return ENOENT;
1002 }
1003 }
1004
1005 hammer_simple_transaction(&trans, dip->hmp);
1006 ++hammer_stats_file_iopsr;
1007
1008 ip = hammer_get_inode(&trans, dip, parent_obj_id,
1009 asof, parent_obj_localization,
1010 dip->flags, &error);
1011 if (ip) {
1012 error = hammer_get_vnode(ip, ap->a_vpp);
1013 hammer_rel_inode(ip, 0);
1014 } else {
1015 *ap->a_vpp = NULL;
1016 }
1017 hammer_done_transaction(&trans);
1018 return (error);
1019 }
1020
1021 /*
1022 * hammer_vop_nlink { nch, dvp, vp, cred }
1023 */
1024 static
1025 int
1026 hammer_vop_nlink(struct vop_nlink_args *ap)
1027 {
1028 struct hammer_transaction trans;
1029 struct hammer_inode *dip;
1030 struct hammer_inode *ip;
1031 struct nchandle *nch;
1032 int error;
1033
1034 if (ap->a_dvp->v_mount != ap->a_vp->v_mount)
1035 return(EXDEV);
1036
1037 nch = ap->a_nch;
1038 dip = VTOI(ap->a_dvp);
1039 ip = VTOI(ap->a_vp);
1040
1041 if (dip->obj_localization != ip->obj_localization)
1042 return(EXDEV);
1043
1044 if (dip->flags & HAMMER_INODE_RO)
1045 return (EROFS);
1046 if (ip->flags & HAMMER_INODE_RO)
1047 return (EROFS);
1048 if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1049 return (error);
1050
1051 /*
1052 * Create a transaction to cover the operations we perform.
1053 */
1054 hammer_start_transaction(&trans, dip->hmp);
1055 ++hammer_stats_file_iopsw;
1056
1057 /*
1058 * Add the filesystem object to the directory. Note that neither
1059 * dip nor ip are referenced or locked, but their vnodes are
1060 * referenced. This function will bump the inode's link count.
1061 */
1062 error = hammer_ip_add_directory(&trans, dip,
1063 nch->ncp->nc_name, nch->ncp->nc_nlen,
1064 ip);
1065
1066 /*
1067 * Finish up.
1068 */
1069 if (error == 0) {
1070 cache_setunresolved(nch);
1071 cache_setvp(nch, ap->a_vp);
1072 }
1073 hammer_done_transaction(&trans);
1074 hammer_knote(ap->a_vp, NOTE_LINK);
1075 hammer_knote(ap->a_dvp, NOTE_WRITE);
1076 return (error);
1077 }
1078
1079 /*
1080 * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1081 *
1082 * The operating system has already ensured that the directory entry
1083 * does not exist and done all appropriate namespace locking.
1084 */
1085 static
1086 int
1087 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1088 {
1089 struct hammer_transaction trans;
1090 struct hammer_inode *dip;
1091 struct hammer_inode *nip;
1092 struct nchandle *nch;
1093 int error;
1094
1095 nch = ap->a_nch;
1096 dip = VTOI(ap->a_dvp);
1097
1098 if (dip->flags & HAMMER_INODE_RO)
1099 return (EROFS);
1100 if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1101 return (error);
1102
1103 /*
1104 * Create a transaction to cover the operations we perform.
1105 */
1106 hammer_start_transaction(&trans, dip->hmp);
1107 ++hammer_stats_file_iopsw;
1108
1109 /*
1110 * Create a new filesystem object of the requested type. The
1111 * returned inode will be referenced but not locked.
1112 */
1113 error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1114 dip, NULL, &nip);
1115 if (error) {
1116 hkprintf("hammer_mkdir error %d\n", error);
1117 hammer_done_transaction(&trans);
1118 *ap->a_vpp = NULL;
1119 return (error);
1120 }
1121 /*
1122 * Add the new filesystem object to the directory. This will also
1123 * bump the inode's link count.
1124 */
1125 error = hammer_ip_add_directory(&trans, dip,
1126 nch->ncp->nc_name, nch->ncp->nc_nlen,
1127 nip);
1128 if (error)
1129 hkprintf("hammer_mkdir (add) error %d\n", error);
1130
1131 /*
1132 * Finish up.
1133 */
1134 if (error) {
1135 hammer_rel_inode(nip, 0);
1136 *ap->a_vpp = NULL;
1137 } else {
1138 error = hammer_get_vnode(nip, ap->a_vpp);
1139 hammer_rel_inode(nip, 0);
1140 if (error == 0) {
1141 cache_setunresolved(ap->a_nch);
1142 cache_setvp(ap->a_nch, *ap->a_vpp);
1143 }
1144 }
1145 hammer_done_transaction(&trans);
1146 if (error == 0)
1147 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1148 return (error);
1149 }
1150
1151 /*
1152 * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1153 *
1154 * The operating system has already ensured that the directory entry
1155 * does not exist and done all appropriate namespace locking.
1156 */
1157 static
1158 int
1159 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1160 {
1161 struct hammer_transaction trans;
1162 struct hammer_inode *dip;
1163 struct hammer_inode *nip;
1164 struct nchandle *nch;
1165 int error;
1166
1167 nch = ap->a_nch;
1168 dip = VTOI(ap->a_dvp);
1169
1170 if (dip->flags & HAMMER_INODE_RO)
1171 return (EROFS);
1172 if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1173 return (error);
1174
1175 /*
1176 * Create a transaction to cover the operations we perform.
1177 */
1178 hammer_start_transaction(&trans, dip->hmp);
1179 ++hammer_stats_file_iopsw;
1180
1181 /*
1182 * Create a new filesystem object of the requested type. The
1183 * returned inode will be referenced but not locked.
1184 *
1185 * If mknod specifies a directory a pseudo-fs is created.
1186 */
1187 error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1188 dip, NULL, &nip);
1189 if (error) {
1190 hammer_done_transaction(&trans);
1191 *ap->a_vpp = NULL;
1192 return (error);
1193 }
1194
1195 /*
1196 * Add the new filesystem object to the directory. This will also
1197 * bump the inode's link count.
1198 */
1199 error = hammer_ip_add_directory(&trans, dip,
1200 nch->ncp->nc_name, nch->ncp->nc_nlen,
1201 nip);
1202
1203 /*
1204 * Finish up.
1205 */
1206 if (error) {
1207 hammer_rel_inode(nip, 0);
1208 *ap->a_vpp = NULL;
1209 } else {
1210 error = hammer_get_vnode(nip, ap->a_vpp);
1211 hammer_rel_inode(nip, 0);
1212 if (error == 0) {
1213 cache_setunresolved(ap->a_nch);
1214 cache_setvp(ap->a_nch, *ap->a_vpp);
1215 }
1216 }
1217 hammer_done_transaction(&trans);
1218 if (error == 0)
1219 hammer_knote(ap->a_dvp, NOTE_WRITE);
1220 return (error);
1221 }
1222
1223 /*
1224 * hammer_vop_open { vp, mode, cred, fp }
1225 */
1226 static
1227 int
1228 hammer_vop_open(struct vop_open_args *ap)
1229 {
1230 hammer_inode_t ip;
1231
1232 ++hammer_stats_file_iopsr;
1233 ip = VTOI(ap->a_vp);
1234
1235 if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1236 return (EROFS);
1237 return(vop_stdopen(ap));
1238 }
1239
1240 /*
1241 * hammer_vop_print { vp }
1242 */
1243 static
1244 int
1245 hammer_vop_print(struct vop_print_args *ap)
1246 {
1247 return EOPNOTSUPP;
1248 }
1249
1250 /*
1251 * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1252 */
1253 static
1254 int
1255 hammer_vop_readdir(struct vop_readdir_args *ap)
1256 {
1257 struct hammer_transaction trans;
1258 struct hammer_cursor cursor;
1259 struct hammer_inode *ip;
1260 struct uio *uio;
1261 hammer_base_elm_t base;
1262 int error;
1263 int cookie_index;
1264 int ncookies;
1265 off_t *cookies;
1266 off_t saveoff;
1267 int r;
1268 int dtype;
1269
1270 ++hammer_stats_file_iopsr;
1271 ip = VTOI(ap->a_vp);
1272 uio = ap->a_uio;
1273 saveoff = uio->uio_offset;
1274
1275 if (ap->a_ncookies) {
1276 ncookies = uio->uio_resid / 16 + 1;
1277 if (ncookies > 1024)
1278 ncookies = 1024;
1279 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1280 cookie_index = 0;
1281 } else {
1282 ncookies = -1;
1283 cookies = NULL;
1284 cookie_index = 0;
1285 }
1286
1287 hammer_simple_transaction(&trans, ip->hmp);
1288
1289 /*
1290 * Handle artificial entries
1291 */
1292 error = 0;
1293 if (saveoff == 0) {
1294 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1295 if (r)
1296 goto done;
1297 if (cookies)
1298 cookies[cookie_index] = saveoff;
1299 ++saveoff;
1300 ++cookie_index;
1301 if (cookie_index == ncookies)
1302 goto done;
1303 }
1304 if (saveoff == 1) {
1305 if (ip->ino_data.parent_obj_id) {
1306 r = vop_write_dirent(&error, uio,
1307 ip->ino_data.parent_obj_id,
1308 DT_DIR, 2, "..");
1309 } else {
1310 r = vop_write_dirent(&error, uio,
1311 ip->obj_id, DT_DIR, 2, "..");
1312 }
1313 if (r)
1314 goto done;
1315 if (cookies)
1316 cookies[cookie_index] = saveoff;
1317 ++saveoff;
1318 ++cookie_index;
1319 if (cookie_index == ncookies)
1320 goto done;
1321 }
1322
1323 /*
1324 * Key range (begin and end inclusive) to scan. Directory keys
1325 * directly translate to a 64 bit 'seek' position.
1326 */
1327 hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1328 cursor.key_beg.localization = ip->obj_localization +
1329 HAMMER_LOCALIZE_MISC;
1330 cursor.key_beg.obj_id = ip->obj_id;
1331 cursor.key_beg.create_tid = 0;
1332 cursor.key_beg.delete_tid = 0;
1333 cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1334 cursor.key_beg.obj_type = 0;
1335 cursor.key_beg.key = saveoff;
1336
1337 cursor.key_end = cursor.key_beg;
1338 cursor.key_end.key = HAMMER_MAX_KEY;
1339 cursor.asof = ip->obj_asof;
1340 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1341
1342 error = hammer_ip_first(&cursor);
1343
1344 while (error == 0) {
1345 error = hammer_ip_resolve_data(&cursor);
1346 if (error)
1347 break;
1348 base = &cursor.leaf->base;
1349 saveoff = base->key;
1350 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1351
1352 if (base->obj_id != ip->obj_id)
1353 panic("readdir: bad record at %p", cursor.node);
1354
1355 /*
1356 * Convert pseudo-filesystems into softlinks
1357 */
1358 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1359 r = vop_write_dirent(
1360 &error, uio, cursor.data->entry.obj_id,
1361 dtype,
1362 cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1363 (void *)cursor.data->entry.name);
1364 if (r)
1365 break;
1366 ++saveoff;
1367 if (cookies)
1368 cookies[cookie_index] = base->key;
1369 ++cookie_index;
1370 if (cookie_index == ncookies)
1371 break;
1372 error = hammer_ip_next(&cursor);
1373 }
1374 hammer_done_cursor(&cursor);
1375
1376 done:
1377 hammer_done_transaction(&trans);
1378
1379 if (ap->a_eofflag)
1380 *ap->a_eofflag = (error == ENOENT);
1381 uio->uio_offset = saveoff;
1382 if (error && cookie_index == 0) {
1383 if (error == ENOENT)
1384 error = 0;
1385 if (cookies) {
1386 kfree(cookies, M_TEMP);
1387 *ap->a_ncookies = 0;
1388 *ap->a_cookies = NULL;
1389 }
1390 } else {
1391 if (error == ENOENT)
1392 error = 0;
1393 if (cookies) {
1394 *ap->a_ncookies = cookie_index;
1395 *ap->a_cookies = cookies;
1396 }
1397 }
1398 return(error);
1399 }
1400
1401 /*
1402 * hammer_vop_readlink { vp, uio, cred }
1403 */
1404 static
1405 int
1406 hammer_vop_readlink(struct vop_readlink_args *ap)
1407 {
1408 struct hammer_transaction trans;
1409 struct hammer_cursor cursor;
1410 struct hammer_inode *ip;
1411 char buf[32];
1412 u_int32_t localization;
1413 hammer_pseudofs_inmem_t pfsm;
1414 int error;
1415
1416 ip = VTOI(ap->a_vp);
1417
1418 /*
1419 * Shortcut if the symlink data was stuffed into ino_data.
1420 *
1421 * Also expand special "@@PFS%05d" softlinks (expansion only
1422 * occurs for non-historical (current) accesses made from the
1423 * primary filesystem).
1424 */
1425 if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1426 char *ptr;
1427 int bytes;
1428
1429 ptr = ip->ino_data.ext.symlink;
1430 bytes = (int)ip->ino_data.size;
1431 if (bytes == 10 &&
1432 ip->obj_asof == HAMMER_MAX_TID &&
1433 ip->obj_localization == 0 &&
1434 strncmp(ptr, "@@PFS", 5) == 0) {
1435 hammer_simple_transaction(&trans, ip->hmp);
1436 bcopy(ptr + 5, buf, 5);
1437 buf[5] = 0;
1438 localization = strtoul(buf, NULL, 10) << 16;
1439 pfsm = hammer_load_pseudofs(&trans, localization,
1440 &error);
1441 if (error == 0) {
1442 if (pfsm->pfsd.mirror_flags &
1443 HAMMER_PFSD_SLAVE) {
1444 ksnprintf(buf, sizeof(buf),
1445 "@@0x%016llx:%05d",
1446 pfsm->pfsd.sync_end_tid,
1447 localization >> 16);
1448 } else {
1449 ksnprintf(buf, sizeof(buf),
1450 "@@0x%016llx:%05d",
1451 HAMMER_MAX_TID,
1452 localization >> 16);
1453 }
1454 ptr = buf;
1455 bytes = strlen(buf);
1456 }
1457 if (pfsm)
1458 hammer_rel_pseudofs(trans.hmp, pfsm);
1459 hammer_done_transaction(&trans);
1460 }
1461 error = uiomove(ptr, bytes, ap->a_uio);
1462 return(error);
1463 }
1464
1465 /*
1466 * Long version
1467 */
1468 hammer_simple_transaction(&trans, ip->hmp);
1469 ++hammer_stats_file_iopsr;
1470 hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1471
1472 /*
1473 * Key range (begin and end inclusive) to scan. Directory keys
1474 * directly translate to a 64 bit 'seek' position.
1475 */
1476 cursor.key_beg.localization = ip->obj_localization +
1477 HAMMER_LOCALIZE_MISC;
1478 cursor.key_beg.obj_id = ip->obj_id;
1479 cursor.key_beg.create_tid = 0;
1480 cursor.key_beg.delete_tid = 0;
1481 cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1482 cursor.key_beg.obj_type = 0;
1483 cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1484 cursor.asof = ip->obj_asof;
1485 cursor.flags |= HAMMER_CURSOR_ASOF;
1486
1487 error = hammer_ip_lookup(&cursor);
1488 if (error == 0) {
1489 error = hammer_ip_resolve_data(&cursor);
1490 if (error == 0) {
1491 KKASSERT(cursor.leaf->data_len >=
1492 HAMMER_SYMLINK_NAME_OFF);
1493 error = uiomove(cursor.data->symlink.name,
1494 cursor.leaf->data_len -
1495 HAMMER_SYMLINK_NAME_OFF,
1496 ap->a_uio);
1497 }
1498 }
1499 hammer_done_cursor(&cursor);
1500 hammer_done_transaction(&trans);
1501 return(error);
1502 }
1503
1504 /*
1505 * hammer_vop_nremove { nch, dvp, cred }
1506 */
1507 static
1508 int
1509 hammer_vop_nremove(struct vop_nremove_args *ap)
1510 {
1511 struct hammer_transaction trans;
1512 struct hammer_inode *dip;
1513 int error;
1514
1515 dip = VTOI(ap->a_dvp);
1516
1517 if (hammer_nohistory(dip) == 0 &&
1518 (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1519 return (error);
1520 }
1521
1522 hammer_start_transaction(&trans, dip->hmp);
1523 ++hammer_stats_file_iopsw;
1524 error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1525 hammer_done_transaction(&trans);
1526 if (error == 0)
1527 hammer_knote(ap->a_dvp, NOTE_WRITE);
1528 return (error);
1529 }
1530
1531 /*
1532 * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1533 */
1534 static
1535 int
1536 hammer_vop_nrename(struct vop_nrename_args *ap)
1537 {
1538 struct hammer_transaction trans;
1539 struct namecache *fncp;
1540 struct namecache *tncp;
1541 struct hammer_inode *fdip;
1542 struct hammer_inode *tdip;
1543 struct hammer_inode *ip;
1544 struct hammer_cursor cursor;
1545 int64_t namekey;
1546 u_int32_t max_iterations;
1547 int nlen, error;
1548
1549 if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount)
1550 return(EXDEV);
1551 if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1552 return(EXDEV);
1553
1554 fdip = VTOI(ap->a_fdvp);
1555 tdip = VTOI(ap->a_tdvp);
1556 fncp = ap->a_fnch->ncp;
1557 tncp = ap->a_tnch->ncp;
1558 ip = VTOI(fncp->nc_vp);
1559 KKASSERT(ip != NULL);
1560
1561 if (fdip->obj_localization != tdip->obj_localization)
1562 return(EXDEV);
1563 if (fdip->obj_localization != ip->obj_localization)
1564 return(EXDEV);
1565
1566 if (fdip->flags & HAMMER_INODE_RO)
1567 return (EROFS);
1568 if (tdip->flags & HAMMER_INODE_RO)
1569 return (EROFS);
1570 if (ip->flags & HAMMER_INODE_RO)
1571 return (EROFS);
1572 if ((error = hammer_checkspace(fdip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1573 return (error);
1574
1575 hammer_start_transaction(&trans, fdip->hmp);
1576 ++hammer_stats_file_iopsw;
1577
1578 /*
1579 * Remove tncp from the target directory and then link ip as
1580 * tncp. XXX pass trans to dounlink
1581 *
1582 * Force the inode sync-time to match the transaction so it is
1583 * in-sync with the creation of the target directory entry.
1584 */
1585 error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
1586 ap->a_cred, 0, -1);
1587 if (error == 0 || error == ENOENT) {
1588 error = hammer_ip_add_directory(&trans, tdip,
1589 tncp->nc_name, tncp->nc_nlen,
1590 ip);
1591 if (error == 0) {
1592 ip->ino_data.parent_obj_id = tdip->obj_id;
1593 hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
1594 }
1595 }
1596 if (error)
1597 goto failed; /* XXX */
1598
1599 /*
1600 * Locate the record in the originating directory and remove it.
1601 *
1602 * Calculate the namekey and setup the key range for the scan. This
1603 * works kinda like a chained hash table where the lower 32 bits
1604 * of the namekey synthesize the chain.
1605 *
1606 * The key range is inclusive of both key_beg and key_end.
1607 */
1608 namekey = hammer_directory_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
1609 &max_iterations);
1610 retry:
1611 hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
1612 cursor.key_beg.localization = fdip->obj_localization +
1613 HAMMER_LOCALIZE_MISC;
1614 cursor.key_beg.obj_id = fdip->obj_id;
1615 cursor.key_beg.key = namekey;
1616 cursor.key_beg.create_tid = 0;
1617 cursor.key_beg.delete_tid = 0;
1618 cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1619 cursor.key_beg.obj_type = 0;
1620
1621 cursor.key_end = cursor.key_beg;
1622 cursor.key_end.key += max_iterations;
1623 cursor.asof = fdip->obj_asof;
1624 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1625
1626 /*
1627 * Scan all matching records (the chain), locate the one matching
1628 * the requested path component.
1629 *
1630 * The hammer_ip_*() functions merge in-memory records with on-disk
1631 * records for the purposes of the search.
1632 */
1633 error = hammer_ip_first(&cursor);
1634 while (error == 0) {
1635 if (hammer_ip_resolve_data(&cursor) != 0)
1636 break;
1637 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
1638 KKASSERT(nlen > 0);
1639 if (fncp->nc_nlen == nlen &&
1640 bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1641 break;
1642 }
1643 error = hammer_ip_next(&cursor);
1644 }
1645
1646 /*
1647 * If all is ok we have to get the inode so we can adjust nlinks.
1648 *
1649 * WARNING: hammer_ip_del_directory() may have to terminate the
1650 * cursor to avoid a recursion. It's ok to call hammer_done_cursor()
1651 * twice.
1652 */
1653 if (error == 0)
1654 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
1655
1656 /*
1657 * XXX A deadlock here will break rename's atomicy for the purposes
1658 * of crash recovery.
1659 */
1660 if (error == EDEADLK) {
1661 hammer_done_cursor(&cursor);
1662 goto retry;
1663 }
1664
1665 /*
1666 * Cleanup and tell the kernel that the rename succeeded.
1667 */
1668 hammer_done_cursor(&cursor);
1669 if (error == 0) {
1670 cache_rename(ap->a_fnch, ap->a_tnch);
1671 hammer_knote(ap->a_fdvp, NOTE_WRITE);
1672 hammer_knote(ap->a_tdvp, NOTE_WRITE);
1673 if (ip->vp)
1674 hammer_knote(ip->vp, NOTE_RENAME);
1675 }
1676
1677 failed:
1678 hammer_done_transaction(&trans);
1679 return (error);
1680 }
1681
1682 /*
1683 * hammer_vop_nrmdir { nch, dvp, cred }
1684 */
1685 static
1686 int
1687 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
1688 {
1689 struct hammer_transaction trans;
1690 struct hammer_inode *dip;
1691 int error;
1692
1693 dip = VTOI(ap->a_dvp);
1694
1695 if (hammer_nohistory(dip) == 0 &&
1696 (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1697 return (error);
1698 }
1699
1700 hammer_start_transaction(&trans, dip->hmp);
1701 ++hammer_stats_file_iopsw;
1702 error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
1703 hammer_done_transaction(&trans);
1704 if (error == 0)
1705 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1706 return (error);
1707 }
1708
1709 /*
1710 * hammer_vop_markatime { vp, cred }
1711 */
1712 static
1713 int
1714 hammer_vop_markatime(struct vop_markatime_args *ap)
1715 {
1716 struct hammer_transaction trans;
1717 struct hammer_inode *ip;
1718
1719 ip = VTOI(ap->a_vp);
1720 if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
1721 return (EROFS);
1722 if (ip->flags & HAMMER_INODE_RO)
1723 return (EROFS);
1724 if (ip->hmp->mp->mnt_flag & MNT_NOATIME)
1725 return (0);
1726 hammer_start_transaction(&trans, ip->hmp);
1727 ++hammer_stats_file_iopsw;
1728
1729 ip->ino_data.atime = trans.time;
1730 hammer_modify_inode(ip, HAMMER_INODE_ATIME);
1731 hammer_done_transaction(&trans);
1732 hammer_knote(ap->a_vp, NOTE_ATTRIB);
1733 return (0);
1734 }
1735
1736 /*
1737 * hammer_vop_setattr { vp, vap, cred }
1738 */
1739 static
1740 int
1741 hammer_vop_setattr(struct vop_setattr_args *ap)
1742 {
1743 struct hammer_transaction trans;
1744 struct vattr *vap;
1745 struct hammer_inode *ip;
1746 int modflags;
1747 int error;
1748 int truncating;
1749 int blksize;
1750 int kflags;
1751 int64_t aligned_size;
1752 u_int32_t flags;
1753
1754 vap = ap->a_vap;
1755 ip = ap->a_vp->v_data;
1756 modflags = 0;
1757 kflags = 0;
1758
1759 if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
1760 return(EROFS);
1761 if (ip->flags & HAMMER_INODE_RO)
1762 return (EROFS);
1763 if (hammer_nohistory(ip) == 0 &&
1764 (error = hammer_checkspace(ip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1765 return (error);
1766 }
1767
1768 hammer_start_transaction(&trans, ip->hmp);
1769 ++hammer_stats_file_iopsw;
1770 error = 0;
1771
1772 if (vap->va_flags != VNOVAL) {
1773 flags = ip->ino_data.uflags;
1774 error = vop_helper_setattr_flags(&flags, vap->va_flags,
1775 hammer_to_unix_xid(&ip->ino_data.uid),
1776 ap->a_cred);
1777 if (error == 0) {
1778 if (ip->ino_data.uflags != flags) {
1779 ip->ino_data.uflags = flags;
1780 modflags |= HAMMER_INODE_DDIRTY;
1781 kflags |= NOTE_ATTRIB;
1782 }
1783 if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
1784 error = 0;
1785 goto done;
1786 }
1787 }
1788 goto done;
1789 }
1790 if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
1791 error = EPERM;
1792 goto done;
1793 }
1794 if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
1795 mode_t cur_mode = ip->ino_data.mode;
1796 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1797 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1798 uuid_t uuid_uid;
1799 uuid_t uuid_gid;
1800
1801 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
1802 ap->a_cred,
1803 &cur_uid, &cur_gid, &cur_mode);
1804 if (error == 0) {
1805 hammer_guid_to_uuid(&uuid_uid, cur_uid);
1806 hammer_guid_to_uuid(&uuid_gid, cur_gid);
1807 if (bcmp(&uuid_uid, &ip->ino_data.uid,
1808 sizeof(uuid_uid)) ||
1809 bcmp(&uuid_gid, &ip->ino_data.gid,
1810 sizeof(uuid_gid)) ||
1811 ip->ino_data.mode != cur_mode
1812 ) {
1813 ip->ino_data.uid = uuid_uid;
1814 ip->ino_data.gid = uuid_gid;
1815 ip->ino_data.mode = cur_mode;
1816 }
1817 modflags |= HAMMER_INODE_DDIRTY;
1818 kflags |= NOTE_ATTRIB;
1819 }
1820 }
1821 while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
1822 switch(ap->a_vp->v_type) {
1823 case VREG:
1824 if (vap->va_size == ip->ino_data.size)
1825 break;
1826 /*
1827 * XXX break atomicy, we can deadlock the backend
1828 * if we do not release the lock. Probably not a
1829 * big deal here.
1830 */
1831 blksize = hammer_blocksize(vap->va_size);
1832 if (vap->va_size < ip->ino_data.size) {
1833 vtruncbuf(ap->a_vp, vap->va_size, blksize);
1834 truncating = 1;
1835 kflags |= NOTE_WRITE;
1836 } else {
1837 vnode_pager_setsize(ap->a_vp, vap->va_size);
1838 truncating = 0;
1839 kflags |= NOTE_WRITE | NOTE_EXTEND;
1840 }
1841 ip->ino_data.size = vap->va_size;
1842 modflags |= HAMMER_INODE_DDIRTY;
1843
1844 /*
1845 * on-media truncation is cached in the inode until
1846 * the inode is synchronized.
1847 */
1848 if (truncating) {
1849 hammer_ip_frontend_trunc(ip, vap->va_size);
1850 #ifdef DEBUG_TRUNCATE
1851 if (HammerTruncIp == NULL)
1852 HammerTruncIp = ip;
1853 #endif
1854 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
1855 ip->flags |= HAMMER_INODE_TRUNCATED;
1856 ip->trunc_off = vap->va_size;
1857 #ifdef DEBUG_TRUNCATE
1858 if (ip == HammerTruncIp)
1859 kprintf("truncate1 %016llx\n", ip->trunc_off);
1860 #endif
1861 } else if (ip->trunc_off > vap->va_size) {
1862 ip->trunc_off = vap->va_size;
1863 #ifdef DEBUG_TRUNCATE
1864 if (ip == HammerTruncIp)
1865 kprintf("truncate2 %016llx\n", ip->trunc_off);
1866 #endif
1867 } else {
1868 #ifdef DEBUG_TRUNCATE
1869 if (ip == HammerTruncIp)
1870 kprintf("truncate3 %016llx (ignored)\n", vap->va_size);
1871 #endif
1872 }
1873 }
1874
1875 /*
1876 * If truncating we have to clean out a portion of
1877 * the last block on-disk. We do this in the
1878 * front-end buffer cache.
1879 */
1880 aligned_size = (vap->va_size + (blksize - 1)) &
1881 ~(int64_t)(blksize - 1);
1882 if (truncating && vap->va_size < aligned_size) {
1883 struct buf *bp;
1884 int offset;
1885
1886 aligned_size -= blksize;
1887
1888 offset = (int)vap->va_size & (blksize - 1);
1889 error = bread(ap->a_vp, aligned_size,
1890 blksize, &bp);
1891 hammer_ip_frontend_trunc(ip, aligned_size);
1892 if (error == 0) {
1893 bzero(bp->b_data + offset,
1894 blksize - offset);
1895 /* must de-cache direct-io offset */
1896 bp->b_bio2.bio_offset = NOOFFSET;
1897 bdwrite(bp);
1898 } else {
1899 kprintf("ERROR %d\n", error);
1900 brelse(bp);
1901 }
1902 }
1903 break;
1904 case VDATABASE:
1905 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
1906 ip->flags |= HAMMER_INODE_TRUNCATED;
1907 ip->trunc_off = vap->va_size;
1908 } else if (ip->trunc_off > vap->va_size) {
1909 ip->trunc_off = vap->va_size;
1910 }
1911 hammer_ip_frontend_trunc(ip, vap->va_size);
1912 ip->ino_data.size = vap->va_size;
1913 modflags |= HAMMER_INODE_DDIRTY;
1914 kflags |= NOTE_ATTRIB;
1915 break;
1916 default:
1917 error = EINVAL;
1918 goto done;
1919 }
1920 break;
1921 }
1922 if (vap->va_atime.tv_sec != VNOVAL) {
1923 ip->ino_data.atime =
1924 hammer_timespec_to_time(&vap->va_atime);
1925 modflags |= HAMMER_INODE_ATIME;
1926 kflags |= NOTE_ATTRIB;
1927 }
1928 if (vap->va_mtime.tv_sec != VNOVAL) {
1929 ip->ino_data.mtime =
1930 hammer_timespec_to_time(&vap->va_mtime);
1931 modflags |= HAMMER_INODE_MTIME;
1932 kflags |= NOTE_ATTRIB;
1933 }
1934 if (vap->va_mode != (mode_t)VNOVAL) {
1935 mode_t cur_mode = ip->ino_data.mode;
1936 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1937 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1938
1939 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
1940 cur_uid, cur_gid, &cur_mode);
1941 if (error == 0 && ip->ino_data.mode != cur_mode) {
1942 ip->ino_data.mode = cur_mode;
1943 modflags |= HAMMER_INODE_DDIRTY;
1944 kflags |= NOTE_ATTRIB;
1945 }
1946 }
1947 done:
1948 if (error == 0)
1949 hammer_modify_inode(ip, modflags);
1950 hammer_done_transaction(&trans);
1951 hammer_knote(ap->a_vp, kflags);
1952 return (error);
1953 }
1954
1955 /*
1956 * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
1957 */
1958 static
1959 int
1960 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
1961 {
1962 struct hammer_transaction trans;
1963 struct hammer_inode *dip;
1964 struct hammer_inode *nip;
1965 struct nchandle *nch;
1966 hammer_record_t record;
1967 int error;
1968 int bytes;
1969
1970 ap->a_vap->va_type = VLNK;
1971
1972 nch = ap->a_nch;
1973 dip = VTOI(ap->a_dvp);
1974
1975 if (dip->flags & HAMMER_INODE_RO)
1976 return (EROFS);
1977 if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1978 return (error);
1979
1980 /*
1981 * Create a transaction to cover the operations we perform.
1982 */
1983 hammer_start_transaction(&trans, dip->hmp);
1984 ++hammer_stats_file_iopsw;
1985
1986 /*
1987 * Create a new filesystem object of the requested type. The
1988 * returned inode will be referenced but not locked.
1989 */
1990
1991 error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1992 dip, NULL, &nip);
1993 if (error) {
1994 hammer_done_transaction(&trans);
1995 *ap->a_vpp = NULL;
1996 return (error);
1997 }
1998
1999 /*
2000 * Add a record representing the symlink. symlink stores the link
2001 * as pure data, not a string, and is no \0 terminated.
2002 */
2003 if (error == 0) {
2004 bytes = strlen(ap->a_target);
2005
2006 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2007 bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2008 } else {
2009 record = hammer_alloc_mem_record(nip, bytes);
2010 record->type = HAMMER_MEM_RECORD_GENERAL;
2011
2012 record->leaf.base.localization = nip->obj_localization +
2013 HAMMER_LOCALIZE_MISC;
2014 record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2015 record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2016 record->leaf.data_len = bytes;
2017 KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2018 bcopy(ap->a_target, record->data->symlink.name, bytes);
2019 error = hammer_ip_add_record(&trans, record);
2020 }
2021
2022 /*
2023 * Set the file size to the length of the link.
2024 */
2025 if (error == 0) {
2026 nip->ino_data.size = bytes;
2027 hammer_modify_inode(nip, HAMMER_INODE_DDIRTY);
2028 }
2029 }
2030 if (error == 0)
2031 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
2032 nch->ncp->nc_nlen, nip);
2033
2034 /*
2035 * Finish up.
2036 */
2037 if (error) {
2038 hammer_rel_inode(nip, 0);
2039 *ap->a_vpp = NULL;
2040 } else {
2041 error = hammer_get_vnode(nip, ap->a_vpp);
2042 hammer_rel_inode(nip, 0);
2043 if (error == 0) {
2044 cache_setunresolved(ap->a_nch);
2045 cache_setvp(ap->a_nch, *ap->a_vpp);
2046 hammer_knote(ap->a_dvp, NOTE_WRITE);
2047 }
2048 }
2049 hammer_done_transaction(&trans);
2050 return (error);
2051 }
2052
2053 /*
2054 * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2055 */
2056 static
2057 int
2058 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2059 {
2060 struct hammer_transaction trans;
2061 struct hammer_inode *dip;
2062 int error;
2063
2064 dip = VTOI(ap->a_dvp);
2065
2066 if (hammer_nohistory(dip) == 0 &&
2067 (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2068 return (error);
2069 }
2070
2071 hammer_start_transaction(&trans, dip->hmp);
2072 ++hammer_stats_file_iopsw;
2073 error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2074 ap->a_cred, ap->a_flags, -1);
2075 hammer_done_transaction(&trans);
2076
2077 return (error);
2078 }
2079
2080 /*
2081 * hammer_vop_ioctl { vp, command, data, fflag, cred }
2082 */
2083 static
2084 int
2085 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2086 {
2087 struct hammer_inode *ip = ap->a_vp->v_data;
2088
2089 ++hammer_stats_file_iopsr;
2090 return(hammer_ioctl(ip, ap->a_command, ap->a_data,
2091 ap->a_fflag, ap->a_cred));
2092 }
2093
2094 static
2095 int
2096 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2097 {
2098 struct mount *mp;
2099 int error;
2100
2101 mp = ap->a_head.a_ops->head.vv_mount;
2102
2103 switch(ap->a_op) {
2104 case MOUNTCTL_SET_EXPORT:
2105 if (ap->a_ctllen != sizeof(struct export_args))
2106 error = EINVAL;
2107 else
2108 error = hammer_vfs_export(mp, ap->a_op,
2109 (const struct export_args *)ap->a_ctl);
2110 break;
2111 default:
2112 error = journal_mountctl(ap);
2113 break;
2114 }
2115 return(error);
2116 }
2117
2118 /*
2119 * hammer_vop_strategy { vp, bio }
2120 *
2121 * Strategy call, used for regular file read & write only. Note that the
2122 * bp may represent a cluster.
2123 *
2124 * To simplify operation and allow better optimizations in the future,
2125 * this code does not make any assumptions with regards to buffer alignment
2126 * or size.
2127 */
2128 static
2129 int
2130 hammer_vop_strategy(struct vop_strategy_args *ap)
2131 {
2132 struct buf *bp;
2133 int error;
2134
2135 bp = ap->a_bio->bio_buf;
2136
2137 switch(bp->b_cmd) {
2138 case BUF_CMD_READ:
2139 error = hammer_vop_strategy_read(ap);
2140 break;
2141 case BUF_CMD_WRITE:
2142 error = hammer_vop_strategy_write(ap);
2143 break;
2144 default:
2145 bp->b_error = error = EINVAL;
2146 bp->b_flags |= B_ERROR;
2147 biodone(ap->a_bio);
2148 break;
2149 }
2150 return (error);
2151 }
2152
2153 /*
2154 * Read from a regular file. Iterate the related records and fill in the
2155 * BIO/BUF. Gaps are zero-filled.
2156 *
2157 * The support code in hammer_object.c should be used to deal with mixed
2158 * in-memory and on-disk records.
2159 *
2160 * NOTE: Can be called from the cluster code with an oversized buf.
2161 *
2162 * XXX atime update
2163 */
2164 static
2165 int
2166 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2167 {
2168 struct hammer_transaction trans;
2169 struct hammer_inode *ip;
2170 struct hammer_cursor cursor;
2171 hammer_base_elm_t base;
2172 hammer_off_t disk_offset;
2173 struct bio *bio;
2174 struct bio *nbio;
2175 struct buf *bp;
2176 int64_t rec_offset;
2177 int64_t ran_end;
2178 int64_t tmp64;
2179 int error;
2180 int boff;
2181 int roff;
2182 int n;
2183
2184 bio = ap->a_bio;
2185 bp = bio->bio_buf;
2186 ip = ap->a_vp->v_data;
2187
2188 /*
2189 * The zone-2 disk offset may have been set by the cluster code via
2190 * a BMAP operation, or else should be NOOFFSET.
2191 *
2192 * Checking the high bits for a match against zone-2 should suffice.
2193 */
2194 nbio = push_bio(bio);
2195 if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2196 HAMMER_ZONE_LARGE_DATA) {
2197 error = hammer_io_direct_read(ip->hmp, nbio, NULL);
2198 return (error);
2199 }
2200
2201 /*
2202 * Well, that sucked. Do it the hard way. If all the stars are
2203 * aligned we may still be able to issue a direct-read.
2204 */
2205 hammer_simple_transaction(&trans, ip->hmp);
2206 hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2207
2208 /*
2209 * Key range (begin and end inclusive) to scan. Note that the key's
2210 * stored in the actual records represent BASE+LEN, not BASE. The
2211 * first record containing bio_offset will have a key > bio_offset.
2212 */
2213 cursor.key_beg.localization = ip->obj_localization +
2214 HAMMER_LOCALIZE_MISC;
2215 cursor.key_beg.obj_id = ip->obj_id;
2216 cursor.key_beg.create_tid = 0;
2217 cursor.key_beg.delete_tid = 0;
2218 cursor.key_beg.obj_type = 0;
2219 cursor.key_beg.key = bio->bio_offset + 1;
2220 cursor.asof = ip->obj_asof;
2221 cursor.flags |= HAMMER_CURSOR_ASOF;
2222
2223 cursor.key_end = cursor.key_beg;
2224 KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2225 #if 0
2226 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2227 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2228 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2229 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2230 } else
2231 #endif
2232 {
2233 ran_end = bio->bio_offset + bp->b_bufsize;
2234 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2235 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2236 tmp64 = ran_end + MAXPHYS + 1; /* work-around GCC-4 bug */
2237 if (tmp64 < ran_end)
2238 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2239 else
2240 cursor.key_end.key = ran_end + MAXPHYS + 1;
2241 }
2242 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2243
2244 error = hammer_ip_first(&cursor);
2245 boff = 0;
2246
2247 while (error == 0) {
2248 /*
2249 * Get the base file offset of the record. The key for
2250 * data records is (base + bytes) rather then (base).
2251 */
2252 base = &cursor.leaf->base;
2253 rec_offset = base->key - cursor.leaf->data_len;
2254
2255 /*
2256 * Calculate the gap, if any, and zero-fill it.
2257 *
2258 * n is the offset of the start of the record verses our
2259 * current seek offset in the bio.
2260 */
2261 n = (int)(rec_offset - (bio->bio_offset + boff));
2262 if (n > 0) {
2263 if (n > bp->b_bufsize - boff)
2264 n = bp->b_bufsize - boff;
2265 bzero((char *)bp->b_data + boff, n);
2266 boff += n;
2267 n = 0;
2268 }
2269
2270 /*
2271 * Calculate the data offset in the record and the number
2272 * of bytes we can copy.
2273 *
2274 * There are two degenerate cases. First, boff may already
2275 * be at bp->b_bufsize. Secondly, the data offset within
2276 * the record may exceed the record's size.
2277 */
2278 roff = -n;
2279 rec_offset += roff;
2280 n = cursor.leaf->data_len - roff;
2281 if (n <= 0) {
2282 kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2283 n = 0;
2284 } else if (n > bp->b_bufsize - boff) {
2285 n = bp->b_bufsize - boff;
2286 }
2287
2288 /*
2289 * Deal with cached truncations. This cool bit of code
2290 * allows truncate()/ftruncate() to avoid having to sync
2291 * the file.
2292 *
2293 * If the frontend is truncated then all backend records are
2294 * subject to the frontend's truncation.
2295 *
2296 * If the backend is truncated then backend records on-disk
2297 * (but not in-memory) are subject to the backend's
2298 * truncation. In-memory records owned by the backend
2299 * represent data written after the truncation point on the
2300 * backend and must not be truncated.
2301 *
2302 * Truncate operations deal with frontend buffer cache
2303 * buffers and frontend-owned in-memory records synchronously.
2304 */
2305 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2306 if (hammer_cursor_ondisk(&cursor) ||
2307 cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2308 if (ip->trunc_off <= rec_offset)
2309 n = 0;
2310 else if (ip->trunc_off < rec_offset + n)
2311 n = (int)(ip->trunc_off - rec_offset);
2312 }
2313 }
2314 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2315 if (hammer_cursor_ondisk(&cursor)) {
2316 if (ip->sync_trunc_off <= rec_offset)
2317 n = 0;
2318 else if (ip->sync_trunc_off < rec_offset + n)
2319 n = (int)(ip->sync_trunc_off - rec_offset);
2320 }
2321 }
2322
2323 /*
2324 * Try to issue a direct read into our bio if possible,
2325 * otherwise resolve the element data into a hammer_buffer
2326 * and copy.
2327 *
2328 * The buffer on-disk should be zerod past any real
2329 * truncation point, but may not be for any synthesized
2330 * truncation point from above.
2331 */
2332 disk_offset = cursor.leaf->data_offset + roff;
2333 if (boff == 0 && n == bp->b_bufsize &&
2334 hammer_cursor_ondisk(&cursor) &&
2335 (disk_offset & HAMMER_BUFMASK) == 0) {
2336 KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2337 HAMMER_ZONE_LARGE_DATA);
2338 nbio->bio_offset = disk_offset;
2339 error = hammer_io_direct_read(trans.hmp, nbio,
2340 cursor.leaf);
2341 goto done;
2342 } else if (n) {
2343 error = hammer_ip_resolve_data(&cursor);
2344 if (error == 0) {
2345 bcopy((char *)cursor.data + roff,
2346 (char *)bp->b_data + boff, n);
2347 }
2348 }
2349 if (error)
2350 break;
2351
2352 /*
2353 * Iterate until we have filled the request.
2354 */
2355 boff += n;
2356 if (boff == bp->b_bufsize)
2357 break;
2358 error = hammer_ip_next(&cursor);
2359 }
2360
2361 /*
2362 * There may have been a gap after the last record
2363 */
2364 if (error == ENOENT)
2365 error = 0;
2366 if (error == 0 && boff != bp->b_bufsize) {
2367 KKASSERT(boff < bp->b_bufsize);
2368 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2369 /* boff = bp->b_bufsize; */
2370 }
2371 bp->b_resid = 0;
2372 bp->b_error = error;
2373 if (error)
2374 bp->b_flags |= B_ERROR;
2375 biodone(ap->a_bio);
2376
2377 done:
2378 if (cursor.node)
2379 hammer_cache_node(&ip->cache[1], cursor.node);
2380 hammer_done_cursor(&cursor);
2381 hammer_done_transaction(&trans);
2382 return(error);
2383 }
2384
2385 /*
2386 * BMAP operation - used to support cluster_read() only.
2387 *
2388 * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2389 *
2390 * This routine may return EOPNOTSUPP if the opration is not supported for
2391 * the specified offset. The contents of the pointer arguments do not
2392 * need to be initialized in that case.
2393 *
2394 * If a disk address is available and properly aligned return 0 with
2395 * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2396 * to the run-length relative to that offset. Callers may assume that
2397 * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2398 * large, so return EOPNOTSUPP if it is not sufficiently large.
2399 */
2400 static
2401 int
2402 hammer_vop_bmap(struct vop_bmap_args *ap)
2403 {
2404 struct hammer_transaction trans;
2405 struct hammer_inode *ip;
2406 struct hammer_cursor cursor;
2407 hammer_base_elm_t base;
2408 int64_t rec_offset;
2409 int64_t ran_end;
2410 int64_t tmp64;
2411 int64_t base_offset;
2412 int64_t base_disk_offset;
2413 int64_t last_offset;
2414 hammer_off_t last_disk_offset;
2415 hammer_off_t disk_offset;
2416 int rec_len;
2417 int error;
2418 int blksize;
2419
2420 ++hammer_stats_file_iopsr;
2421 ip = ap->a_vp->v_data;
2422
2423 /*
2424 * We can only BMAP regular files. We can't BMAP database files,
2425 * directories, etc.
2426 */
2427 if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
2428 return(EOPNOTSUPP);
2429
2430 /*
2431 * bmap is typically called with runp/runb both NULL when used
2432 * for writing. We do not support BMAP for writing atm.
2433 */
2434 if (ap->a_cmd != BUF_CMD_READ)
2435 return(EOPNOTSUPP);
2436
2437 /*
2438 * Scan the B-Tree to acquire blockmap addresses, then translate
2439 * to raw addresses.
2440 */
2441 hammer_simple_transaction(&trans, ip->hmp);
2442 #if 0
2443 kprintf("bmap_beg %016llx ip->cache %p\n", ap->a_loffset, ip->cache[1]);
2444 #endif
2445 hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2446
2447 /*
2448 * Key range (begin and end inclusive) to scan. Note that the key's
2449 * stored in the actual records represent BASE+LEN, not BASE. The
2450 * first record containing bio_offset will have a key > bio_offset.
2451 */
2452 cursor.key_beg.localization = ip->obj_localization +
2453 HAMMER_LOCALIZE_MISC;
2454 cursor.key_beg.obj_id = ip->obj_id;
2455 cursor.key_beg.create_tid = 0;
2456 cursor.key_beg.delete_tid = 0;
2457 cursor.key_beg.obj_type = 0;
2458 if (ap->a_runb)
2459 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
2460 else
2461 cursor.key_beg.key = ap->a_loffset + 1;
2462 if (cursor.key_beg.key < 0)
2463 cursor.key_beg.key = 0;
2464 cursor.asof = ip->obj_asof;
2465 cursor.flags |= HAMMER_CURSOR_ASOF;
2466
2467 cursor.key_end = cursor.key_beg;
2468 KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2469
2470 ran_end = ap->a_loffset + MAXPHYS;
2471 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2472 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2473 tmp64 = ran_end + MAXPHYS + 1; /* work-around GCC-4 bug */
2474 if (tmp64 < ran_end)
2475 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2476 else
2477 cursor.key_end.key = ran_end + MAXPHYS + 1;
2478
2479 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2480
2481 error = hammer_ip_first(&cursor);
2482 base_offset = last_offset = 0;
2483 base_disk_offset = last_disk_offset = 0;
2484
2485 while (error == 0) {
2486 /*
2487 * Get the base file offset of the record. The key for
2488 * data records is (base + bytes) rather then (base).
2489 *
2490 * NOTE: rec_offset + rec_len may exceed the end-of-file.
2491 * The extra bytes should be zero on-disk and the BMAP op
2492 * should still be ok.
2493 */
2494 base = &cursor.leaf->base;
2495 rec_offset = base->key - cursor.leaf->data_len;
2496 rec_len = cursor.leaf->data_len;
2497
2498 /*
2499 * Incorporate any cached truncation.
2500 *
2501 * NOTE: Modifications to rec_len based on synthesized
2502 * truncation points remove the guarantee that any extended
2503 * data on disk is zero (since the truncations may not have
2504 * taken place on-media yet).
2505 */
2506 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2507 if (hammer_cursor_ondisk(&cursor) ||
2508 cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2509 if (ip->trunc_off <= rec_offset)
2510 rec_len = 0;
2511 else if (ip->trunc_off < rec_offset + rec_len)
2512 rec_len = (int)(ip->trunc_off - rec_offset);
2513 }
2514 }
2515 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2516 if (hammer_cursor_ondisk(&cursor)) {
2517 if (ip->sync_trunc_off <= rec_offset)
2518 rec_len = 0;
2519 else if (ip->sync_trunc_off < rec_offset + rec_len)
2520 rec_len = (int)(ip->sync_trunc_off - rec_offset);
2521 }
2522 }
2523
2524 /*
2525 * Accumulate information. If we have hit a discontiguous
2526 * block reset base_offset unless we are already beyond the
2527 * requested offset. If we are, that's it, we stop.
2528 */
2529 if (error)
2530 break;
2531 if (hammer_cursor_ondisk(&cursor)) {
2532 disk_offset = cursor.leaf->data_offset;
2533 if (rec_offset != last_offset ||
2534 disk_offset != last_disk_offset) {
2535 if (rec_offset > ap->a_loffset)
2536 break;
2537 base_offset = rec_offset;
2538 base_disk_offset = disk_offset;
2539 }
2540 last_offset = rec_offset + rec_len;
2541 last_disk_offset = disk_offset + rec_len;
2542 }
2543 error = hammer_ip_next(&cursor);
2544 }
2545
2546 #if 0
2547 kprintf("BMAP %016llx: %016llx - %016llx\n",
2548 ap->a_loffset, base_offset, last_offset);
2549 kprintf("BMAP %16s: %016llx - %016llx\n",
2550 "", base_disk_offset, last_disk_offset);
2551 #endif
2552
2553 if (cursor.node) {
2554 hammer_cache_node(&ip->cache[1], cursor.node);
2555 #if 0
2556 kprintf("bmap_end2 %016llx ip->cache %p\n", ap->a_loffset, ip->cache[1]);
2557 #endif
2558 }
2559 hammer_done_cursor(&cursor);
2560 hammer_done_transaction(&trans);
2561
2562 /*
2563 * If we couldn't find any records or the records we did find were
2564 * all behind the requested offset, return failure. A forward
2565 * truncation can leave a hole w/ no on-disk records.
2566 */
2567 if (last_offset == 0 || last_offset < ap->a_loffset)
2568 return (EOPNOTSUPP);
2569
2570 /*
2571 * Figure out the block size at the requested offset and adjust
2572 * our limits so the cluster_read() does not create inappropriately
2573 * sized buffer cache buffers.
2574 */
2575 blksize = hammer_blocksize(ap->a_loffset);
2576 if (hammer_blocksize(base_offset) != blksize) {
2577 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
2578 }
2579 if (last_offset != ap->a_loffset &&
2580 hammer_blocksize(last_offset - 1) != blksize) {
2581 last_offset = hammer_blockdemarc(ap->a_loffset,
2582 last_offset - 1);
2583 }
2584
2585 /*
2586 * Returning EOPNOTSUPP simply prevents the direct-IO optimization
2587 * from occuring.
2588 */
2589 disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
2590
2591 if ((disk_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_LARGE_DATA) {
2592 /*
2593 * Only large-data zones can be direct-IOd
2594 */
2595 error = EOPNOTSUPP;
2596 } else if ((disk_offset & HAMMER_BUFMASK) ||
2597 (last_offset - ap->a_loffset) < blksize) {
2598 /*
2599 * doffsetp is not aligned or the forward run size does
2600 * not cover a whole buffer, disallow the direct I/O.
2601 */
2602 error = EOPNOTSUPP;
2603 } else {
2604 /*
2605 * We're good.
2606 */
2607 *ap->a_doffsetp = disk_offset;
2608 if (ap->a_runb) {
2609 *ap->a_runb = ap->a_loffset - base_offset;
2610 KKASSERT(*ap->a_runb >= 0);
2611 }
2612 if (ap->a_runp) {
2613 *ap->a_runp = last_offset - ap->a_loffset;
2614 KKASSERT(*ap->a_runp >= 0);
2615 }
2616 error = 0;
2617 }
2618 return(error);
2619 }
2620
2621 /*
2622 * Write to a regular file. Because this is a strategy call the OS is
2623 * trying to actually get data onto the media.
2624 */
2625 static
2626 int
2627 hammer_vop_strategy_write(struct vop_strategy_args *ap)
2628 {
2629 hammer_record_t record;
2630 hammer_mount_t hmp;
2631 hammer_inode_t ip;
2632 struct bio *bio;
2633 struct buf *bp;
2634 int blksize;
2635 int bytes;
2636 int error;
2637
2638 bio = ap->a_bio;
2639 bp = bio->bio_buf;
2640 ip = ap->a_vp->v_data;
2641 hmp = ip->hmp;
2642
2643 blksize = hammer_blocksize(bio->bio_offset);
2644 KKASSERT(bp->b_bufsize == blksize);
2645
2646 if (ip->flags & HAMMER_INODE_RO) {
2647 bp->b_error = EROFS;
2648 bp->b_flags |= B_ERROR;
2649 biodone(ap->a_bio);
2650 return(EROFS);
2651 }
2652
2653 /*
2654 * Interlock with inode destruction (no in-kernel or directory
2655 * topology visibility). If we queue new IO while trying to
2656 * destroy the inode we can deadlock the vtrunc call in
2657 * hammer_inode_unloadable_check().
2658 *
2659 * Besides, there's no point flushing a bp associated with an
2660 * inode that is being destroyed on-media and has no kernel
2661 * references.
2662 */
2663 if ((ip->flags | ip->sync_flags) &
2664 (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
2665 bp->b_resid = 0;
2666 biodone(ap->a_bio);
2667 return(0);
2668 }
2669
2670 /*
2671 * Reserve space and issue a direct-write from the front-end.
2672 * NOTE: The direct_io code will hammer_bread/bcopy smaller
2673 * allocations.
2674 *
2675 * An in-memory record will be installed to reference the storage
2676 * until the flusher can get to it.
2677 *
2678 * Since we own the high level bio the front-end will not try to
2679 * do a direct-read until the write completes.
2680 *
2681 * NOTE: The only time we do not reserve a full-sized buffers
2682 * worth of data is if the file is small. We do not try to
2683 * allocate a fragment (from the small-data zone) at the end of
2684 * an otherwise large file as this can lead to wildly separated
2685 * data.
2686 */
2687 KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
2688 KKASSERT(bio->bio_offset < ip->ino_data.size);
2689 if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
2690 bytes = bp->b_bufsize;
2691 else
2692 bytes = ((int)ip->ino_data.size + 15) & ~15;
2693
2694 record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
2695 bytes, &error);
2696 if (record) {
2697 hammer_io_direct_write(hmp, record, bio);
2698 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
2699 hammer_flush_inode(ip, 0);
2700 } else {
2701 bp->b_bio2.bio_offset = NOOFFSET;
2702 bp->b_error = error;
2703 bp->b_flags |= B_ERROR;
2704 biodone(ap->a_bio);
2705 }
2706 return(error);
2707 }
2708
2709 /*
2710 * dounlink - disconnect a directory entry
2711 *
2712 * XXX whiteout support not really in yet
2713 */
2714 static int
2715 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
2716 struct vnode *dvp, struct ucred *cred,
2717 int flags, int isdir)
2718 {
2719 struct namecache *ncp;
2720 hammer_inode_t dip;
2721 hammer_inode_t ip;
2722 struct hammer_cursor cursor;
2723 int64_t namekey;
2724 u_int32_t max_iterations;
2725 int nlen, error;
2726
2727 /*
2728 * Calculate the namekey and setup the key range for the scan. This
2729 * works kinda like a chained hash table where the lower 32 bits
2730 * of the namekey synthesize the chain.
2731 *
2732 * The key range is inclusive of both key_beg and key_end.
2733 */
2734 dip = VTOI(dvp);
2735 ncp = nch->ncp;
2736
2737 if (dip->flags & HAMMER_INODE_RO)
2738 return (EROFS);
2739
2740 namekey = hammer_directory_namekey(dip, ncp->nc_name, ncp->nc_nlen,
2741 &max_iterations);
2742 retry:
2743 hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
2744 cursor.key_beg.localization = dip->obj_localization +
2745 HAMMER_LOCALIZE_MISC;
2746 cursor.key_beg.obj_id = dip->obj_id;
2747 cursor.key_beg.key = namekey;
2748 cursor.key_beg.create_tid = 0;
2749 cursor.key_beg.delete_tid = 0;
2750 cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2751 cursor.key_beg.obj_type = 0;
2752
2753 cursor.key_end = cursor.key_beg;
2754 cursor.key_end.key += max_iterations;
2755 cursor.asof = dip->obj_asof;
2756 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2757
2758 /*
2759 * Scan all matching records (the chain), locate the one matching
2760 * the requested path component. info->last_error contains the
2761 * error code on search termination and could be 0, ENOENT, or
2762 * something else.
2763 *
2764 * The hammer_ip_*() functions merge in-memory records with on-disk
2765 * records for the purposes of the search.
2766 */
2767 error = hammer_ip_first(&cursor);
2768
2769 while (error == 0) {
2770 error = hammer_ip_resolve_data(&cursor);
2771 if (error)
2772 break;
2773 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2774 KKASSERT(nlen > 0);
2775 if (ncp->nc_nlen == nlen &&
2776 bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2777 break;
2778 }
2779 error = hammer_ip_next(&cursor);
2780 }
2781
2782 /*
2783 * If all is ok we have to get the inode so we can adjust nlinks.
2784 * To avoid a deadlock with the flusher we must release the inode
2785 * lock on the directory when acquiring the inode for the entry.
2786 *
2787 * If the target is a directory, it must be empty.
2788 */
2789 if (error == 0) {
2790 hammer_unlock(&cursor.ip->lock);
2791 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
2792 dip->hmp->asof,
2793 cursor.data->entry.localization,
2794 0, &error);
2795 hammer_lock_sh(&cursor.ip->lock);
2796 if (error == ENOENT) {
2797 kprintf("obj_id %016llx\n", cursor.data->entry.obj_id);
2798 Debugger("ENOENT unlinking object that should exist");
2799 }
2800
2801 /*
2802 * If isdir >= 0 we validate that the entry is or is not a
2803 * directory. If isdir < 0 we don't care.
2804 */
2805 if (error == 0 && isdir >= 0) {
2806 if (isdir &&
2807 ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
2808 error = ENOTDIR;
2809 } else if (isdir == 0 &&
2810 ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
2811 error = EISDIR;
2812 }
2813 }
2814
2815 /*
2816 * If we are trying to remove a directory the directory must
2817 * be empty.
2818 *
2819 * The check directory code can loop and deadlock/retry. Our
2820 * own cursor's node locks must be released to avoid a 3-way
2821 * deadlock with the flusher if the check directory code
2822 * blocks.
2823 *
2824 * If any changes whatsoever have been made to the cursor
2825 * set EDEADLK and retry.
2826 */
2827 if (error == 0 && ip->ino_data.obj_type ==
2828 HAMMER_OBJTYPE_DIRECTORY) {
2829 hammer_unlock_cursor(&cursor);
2830 error = hammer_ip_check_directory_empty(trans, ip);
2831 hammer_lock_cursor(&cursor);
2832 if (cursor.flags & HAMMER_CURSOR_RETEST) {
2833 kprintf("HAMMER: Warning: avoided deadlock "
2834 "on rmdir '%s'\n",
2835 ncp->nc_name);
2836 error = EDEADLK;
2837 }
2838 }
2839
2840 /*
2841 * Delete the directory entry.
2842 *
2843 * WARNING: hammer_ip_del_directory() may have to terminate
2844 * the cursor to avoid a deadlock. It is ok to call
2845 * hammer_done_cursor() twice.
2846 */
2847 if (error == 0) {
2848 error = hammer_ip_del_directory(trans, &cursor,
2849 dip, ip);
2850 }
2851 hammer_done_cursor(&cursor);
2852 if (error == 0) {
2853 cache_setunresolved(nch);
2854 cache_setvp(nch, NULL);
2855 /* XXX locking */
2856 if (ip->vp) {
2857 hammer_knote(ip->vp, NOTE_DELETE);
2858 cache_inval_vp(ip->vp, CINV_DESTROY);
2859 }
2860 }
2861 if (ip)
2862 hammer_rel_inode(ip, 0);
2863 } else {
2864 hammer_done_cursor(&cursor);
2865 }
2866 if (error == EDEADLK)
2867 goto retry;
2868
2869 return (error);
2870 }
2871
2872 /************************************************************************
2873 * FIFO AND SPECFS OPS *
2874 ************************************************************************
2875 *
2876 */
2877
2878 static int
2879 hammer_vop_fifoclose (struct vop_close_args *ap)
2880 {
2881 /* XXX update itimes */
2882 return (VOCALL(&fifo_vnode_vops, &ap->a_head));
2883 }
2884
2885 static int
2886 hammer_vop_fiforead (struct vop_read_args *ap)
2887 {
2888 int error;
2889
2890 error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2891 /* XXX update access time */
2892 return (error);
2893 }
2894
2895 static int
2896 hammer_vop_fifowrite (struct vop_write_args *ap)
2897 {
2898 int error;
2899
2900 error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2901 /* XXX update access time */
2902 return (error);
2903 }
2904
2905 static
2906 int
2907 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
2908 {
2909 int error;
2910
2911 error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2912 if (error)
2913 error = hammer_vop_kqfilter(ap);
2914 return(error);
2915 }
2916
2917 static int
2918 hammer_vop_specclose (struct vop_close_args *ap)
2919 {
2920 /* XXX update itimes */
2921 return (VOCALL(&spec_vnode_vops, &ap->a_head));
2922 }
2923
2924 static int
2925 hammer_vop_specread (struct vop_read_args *ap)
2926 {
2927 /* XXX update access time */
2928 return (VOCALL(&spec_vnode_vops, &ap->a_head));
2929 }
2930
2931 static int
2932 hammer_vop_specwrite (struct vop_write_args *ap)
2933 {
2934 /* XXX update last change time */
2935 return (VOCALL(&spec_vnode_vops, &ap->a_head));
2936 }
2937
2938 /************************************************************************
2939 * KQFILTER OPS *
2940 ************************************************************************
2941 *
2942 */
2943 static void filt_hammerdetach(struct knote *kn);
2944 static int filt_hammerread(struct knote *kn, long hint);
2945 static int filt_hammerwrite(struct knote *kn, long hint);
2946 static int filt_hammervnode(struct knote *kn, long hint);
2947
2948 static struct filterops hammerread_filtops =
2949 { 1, NULL, filt_hammerdetach, filt_hammerread };
2950 static struct filterops hammerwrite_filtops =
2951 { 1, NULL, filt_hammerdetach, filt_hammerwrite };
2952 static struct filterops hammervnode_filtops =
2953 { 1, NULL, filt_hammerdetach, filt_hammervnode };
2954
2955 static
2956 int
2957 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
2958 {
2959 struct vnode *vp = ap->a_vp;
2960 struct knote *kn = ap->a_kn;
2961 lwkt_tokref ilock;
2962
2963 switch (kn->kn_filter) {
2964 case EVFILT_READ:
2965 kn->kn_fop = &hammerread_filtops;
2966 break;
2967 case EVFILT_WRITE:
2968 kn->kn_fop = &hammerwrite_filtops;
2969 break;
2970 case EVFILT_VNODE:
2971 kn->kn_fop = &hammervnode_filtops;
2972 break;
2973 default:
2974 return (1);
2975 }
2976
2977 kn->kn_hook = (caddr_t)vp;
2978
2979 lwkt_gettoken(&ilock, &vp->v_pollinfo.vpi_token);
2980 SLIST_INSERT_HEAD(&vp->v_pollinfo.vpi_selinfo.si_note, kn, kn_selnext);
2981 lwkt_reltoken(&ilock);
2982
2983 return(0);
2984 }
2985
2986 static void
2987 filt_hammerdetach(struct knote *kn)
2988 {
2989 struct vnode *vp = (void *)kn->kn_hook;
2990 lwkt_tokref ilock;
2991
2992 lwkt_gettoken(&ilock, &vp->v_pollinfo.vpi_token);
2993 SLIST_REMOVE(&vp->v_pollinfo.vpi_selinfo.si_note,
2994 kn, knote, kn_selnext);
2995 lwkt_reltoken(&ilock);
2996 }
2997
2998 static int
2999 filt_hammerread(struct knote *kn, long hint)
3000 {
3001 struct vnode *vp = (void *)kn->kn_hook;
3002 hammer_inode_t ip = VTOI(vp);
3003
3004 if (hint == NOTE_REVOKE) {
3005 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3006 return(1);
3007 }
3008 kn->kn_data = ip->ino_data.size - kn->kn_fp->f_offset;
3009 return (kn->kn_data != 0);
3010 }
3011
3012 static int
3013 filt_hammerwrite(struct knote *kn, long hint)
3014 {
3015 if (hint == NOTE_REVOKE)
3016 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3017 kn->kn_data = 0;
3018 return (1);
3019 }
3020
3021 static int
3022 filt_hammervnode(struct knote *kn, long hint)
3023 {
3024 if (kn->kn_sfflags & hint)
3025 kn->kn_fflags |= hint;
3026 if (hint == NOTE_REVOKE) {
3027 kn->kn_flags |= EV_EOF;
3028 return (1);
3029 }
3030 return (kn->kn_fflags != 0);
3031 }
3032
DragonFly BSD