2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
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
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.
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
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>
44 #include <sys/dirent.h>
46 #include <vm/vm_extern.h>
47 #include <vm/swap_pager.h>
48 #include <vfs/fifofs/fifo.h>
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 *);
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 *);
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
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
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
160 hammer_knote(struct vnode *vp, int flags)
163 KNOTE(&vp->v_pollinfo.vpi_kqinfo.ki_note, flags);
166 #ifdef DEBUG_TRUNCATE
167 struct hammer_inode *HammerTruncIp;
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);
179 hammer_vop_vnoperate(struct vop_generic_args *)
181 return (VOCALL(&hammer_vnode_vops, ap));
186 * hammer_vop_fsync { vp, waitfor }
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
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()
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.
203 hammer_vop_fsync(struct vop_fsync_args *ap)
205 hammer_inode_t ip = VTOI(ap->a_vp);
206 hammer_mount_t hmp = ip->hmp;
207 int waitfor = ap->a_waitfor;
210 lwkt_gettoken(&hmp->fs_token);
213 * Fsync rule relaxation (default is either full synchronous flush
214 * or REDO semantics with synchronous flush).
216 if (ap->a_flags & VOP_FSYNC_SYSCALL) {
217 switch(hammer_fsync_mode) {
220 /* no REDO, full synchronous flush */
224 /* no REDO, full asynchronous flush */
225 if (waitfor == MNT_WAIT)
226 waitfor = MNT_NOWAIT;
229 /* REDO semantics, synchronous flush */
230 if (hmp->version < HAMMER_VOL_VERSION_FOUR)
232 mode = HAMMER_FLUSH_UNDOS_AUTO;
235 /* REDO semantics, relaxed asynchronous flush */
236 if (hmp->version < HAMMER_VOL_VERSION_FOUR)
238 mode = HAMMER_FLUSH_UNDOS_RELAXED;
239 if (waitfor == MNT_WAIT)
240 waitfor = MNT_NOWAIT;
243 /* ignore the fsync() system call */
244 lwkt_reltoken(&hmp->fs_token);
247 /* we have to do something */
248 mode = HAMMER_FLUSH_UNDOS_RELAXED;
249 if (waitfor == MNT_WAIT)
250 waitfor = MNT_NOWAIT;
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.
259 if ((ip->flags & HAMMER_INODE_REDO) &&
260 (ip->flags & HAMMER_INODE_MODMASK_NOREDO) == 0
262 ++hammer_count_fsyncs;
263 hammer_flusher_flush_undos(hmp, mode);
265 if (ip->vp && (ip->flags & HAMMER_INODE_MODMASK) == 0)
267 lwkt_reltoken(&hmp->fs_token);
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.
278 * We weren't running REDOs before now so we have to fall
279 * through and do a full fsync of what we have.
281 if (hmp->version >= HAMMER_VOL_VERSION_FOUR &&
282 (hmp->flags & HAMMER_MOUNT_REDO_RECOVERY_RUN) == 0) {
283 ip->flags |= HAMMER_INODE_REDO;
290 * Do a full flush sequence.
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.
296 * WARNING! The VX lock functions must be used. vn_lock() will
297 * fail when this is part of a VOP_RECLAIM sequence.
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) {
305 if ((ap->a_vp->v_flag & VRECLAIMED) == 0) {
311 hammer_wait_inode(ip);
315 if (ip->vp && (ip->flags & HAMMER_INODE_MODMASK) == 0)
317 lwkt_reltoken(&hmp->fs_token);
322 * hammer_vop_read { vp, uio, ioflag, cred }
324 * MPSAFE (for the cache safe does not require fs_token)
328 hammer_vop_read(struct vop_read_args *ap)
330 struct hammer_transaction trans;
345 if (ap->a_vp->v_type != VREG)
354 * Attempt to shortcut directly to the VM object using lwbufs.
355 * This is much faster than instantiating buffer cache buffers.
357 resid = uio->uio_resid;
358 error = vop_helper_read_shortcut(ap);
359 hammer_stats_file_read += resid - uio->uio_resid;
362 if (uio->uio_resid == 0)
366 * Allow the UIO's size to override the sequential heuristic.
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;
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.
379 bigread = (uio->uio_resid > 100 * 1024 * 1024);
382 * Access the data typically in HAMMER_BUFSIZE blocks via the
383 * buffer cache, but HAMMER may use a variable block size based
386 * XXX Temporary hack, delay the start transaction while we remain
387 * MPSAFE. NOTE: ino_data.size cannot change while vnode is
390 while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
394 blksize = hammer_blocksize(uio->uio_offset);
395 offset = (int)uio->uio_offset & (blksize - 1);
396 base_offset = uio->uio_offset - offset;
398 if (bigread && (error = hammer_signal_check(ip->hmp)) != 0)
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;
410 if (ap->a_ioflag & IO_NRDELAY) {
412 return (EWOULDBLOCK);
418 if (got_trans == 0) {
419 hammer_start_transaction(&trans, ip->hmp);
424 * NOTE: A valid bp has already been acquired, but was not
427 if (hammer_cluster_enable) {
429 * Use file_limit to prevent cluster_read() from
430 * creating buffers of the wrong block size past
433 file_limit = ip->ino_data.size;
434 if (base_offset < HAMMER_XDEMARC &&
435 file_limit > HAMMER_XDEMARC) {
436 file_limit = HAMMER_XDEMARC;
438 error = cluster_readx(ap->a_vp,
439 file_limit, base_offset,
440 blksize, uio->uio_resid,
441 seqcount * BKVASIZE, &bp);
443 error = breadnx(ap->a_vp, base_offset, blksize,
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);
457 bp->b_flags &= ~B_IODEBUG;
458 if (blksize == HAMMER_XBUFSIZE)
459 bp->b_flags |= B_CLUSTEROK;
461 n = blksize - offset;
462 if (n > uio->uio_resid)
464 if (n > ip->ino_data.size - uio->uio_offset)
465 n = (int)(ip->ino_data.size - uio->uio_offset);
468 * Set B_AGE, data has a lower priority than meta-data.
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.
474 bp->b_flags |= B_AGE;
475 error = uiomovebp(bp, (char *)bp->b_data + offset, n, uio);
480 hammer_stats_file_read += n;
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
490 if (got_trans == 0 && hammer_update_atime_quick(ip) < 0) {
491 hammer_start_transaction(&trans, ip->hmp);
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);
504 hammer_done_transaction(&trans);
511 * hammer_vop_write { vp, uio, ioflag, cred }
515 hammer_vop_write(struct vop_write_args *ap)
517 struct hammer_transaction trans;
518 struct hammer_inode *ip;
533 if (ap->a_vp->v_type != VREG)
539 seqcount = ap->a_ioflag >> 16;
541 if (ip->flags & HAMMER_INODE_RO)
545 * Create a transaction to cover the operations we perform.
547 hammer_start_transaction(&trans, hmp);
553 if (ap->a_ioflag & IO_APPEND)
554 uio->uio_offset = ip->ino_data.size;
557 * Check for illegal write offsets. Valid range is 0...2^63-1.
559 * NOTE: the base_off assignment is required to work around what
560 * I consider to be a GCC-4 optimization bug.
562 if (uio->uio_offset < 0) {
563 hammer_done_transaction(&trans);
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);
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);
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.
584 * Preset redo_count so we stop generating REDOs earlier if the
587 * redo_count is heuristical, SMP races are ok
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;
596 * Access the data typically in HAMMER_BUFSIZE blocks via the
597 * buffer cache, but HAMMER may use a variable block size based
600 while (uio->uio_resid > 0) {
608 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
610 if (bigwrite && (error = hammer_signal_check(hmp)) != 0)
613 blksize = hammer_blocksize(uio->uio_offset);
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.
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.
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);
632 lwkt_reltoken(&hmp->fs_token);
636 * Do not allow HAMMER to blow out the buffer cache. Very
637 * large UIOs can lockout other processes due to bwillwrite()
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
647 * Only call this if we aren't being recursively called from
648 * a virtual disk device (vn), else we may deadlock.
650 if ((ap->a_ioflag & IO_RECURSE) == 0)
654 * Calculate the blocksize at the current offset and figure
655 * out how much we can actually write.
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) {
667 nsize = uio->uio_offset + n;
668 if (nsize > ip->ino_data.size) {
669 if (uio->uio_offset > ip->ino_data.size)
673 nvextendbuf(ap->a_vp,
676 hammer_blocksize(ip->ino_data.size),
677 hammer_blocksize(nsize),
678 hammer_blockoff(ip->ino_data.size),
679 hammer_blockoff(nsize),
682 kflags |= NOTE_EXTEND;
685 if (uio->uio_segflg == UIO_NOCOPY) {
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.
691 * This case is used by vop_stdputpages().
693 bp = getblk(ap->a_vp, base_offset,
694 blksize, GETBLK_BHEAVY, 0);
695 if ((bp->b_flags & B_CACHE) == 0) {
697 error = bread(ap->a_vp, base_offset,
700 } else if (offset == 0 && uio->uio_resid >= blksize) {
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.
706 bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
707 if ((bp->b_flags & B_CACHE) == 0)
709 } else if (base_offset >= ip->ino_data.size) {
711 * If the base offset of the buffer is beyond the
712 * file EOF, we don't have to issue a read.
714 bp = getblk(ap->a_vp, base_offset,
715 blksize, GETBLK_BHEAVY, 0);
719 * Partial overwrite, read in any missing bits then
720 * replace the portion being written.
722 error = bread(ap->a_vp, base_offset, blksize, &bp);
727 error = uiomovebp(bp, bp->b_data + offset, n, uio);
729 lwkt_gettoken(&hmp->fs_token);
732 * Generate REDO records if enabled and redo_count will not
733 * exceeded the limit.
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.
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
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,
755 ip->flags &= ~HAMMER_INODE_REDO;
760 * If we screwed up we have to undo any VM size changes we
766 nvtruncbuf(ap->a_vp, ip->ino_data.size,
767 hammer_blocksize(ip->ino_data.size),
768 hammer_blockoff(ip->ino_data.size),
771 lwkt_reltoken(&hmp->fs_token);
774 kflags |= NOTE_WRITE;
775 hammer_stats_file_write += n;
776 if (blksize == HAMMER_XBUFSIZE)
777 bp->b_flags |= B_CLUSTEROK;
778 if (ip->ino_data.size < uio->uio_offset) {
779 ip->ino_data.size = uio->uio_offset;
780 flags = HAMMER_INODE_SDIRTY;
784 ip->ino_data.mtime = trans.time;
785 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
786 hammer_modify_inode(&trans, ip, flags);
789 * Once we dirty the buffer any cached zone-X offset
790 * becomes invalid. HAMMER NOTE: no-history mode cannot
791 * allow overwriting over the same data sector unless
792 * we provide UNDOs for the old data, which we don't.
794 bp->b_bio2.bio_offset = NOOFFSET;
796 lwkt_reltoken(&hmp->fs_token);
799 * Final buffer disposition.
801 * Because meta-data updates are deferred, HAMMER is
802 * especially sensitive to excessive bdwrite()s because
803 * the I/O stream is not broken up by disk reads. So the
804 * buffer cache simply cannot keep up.
806 * WARNING! blksize is variable. cluster_write() is
807 * expected to not blow up if it encounters
808 * buffers that do not match the passed blksize.
810 * NOTE! Hammer shouldn't need to bawrite()/cluster_write().
811 * The ip->rsv_recs check should burst-flush the data.
812 * If we queue it immediately the buf could be left
813 * locked on the device queue for a very long time.
815 * However, failing to flush a dirty buffer out when
816 * issued from the pageout daemon can result in a low
817 * memory deadlock against bio_page_alloc(), so we
818 * have to bawrite() on IO_ASYNC as well.
820 * NOTE! To avoid degenerate stalls due to mismatched block
821 * sizes we only honor IO_DIRECT on the write which
822 * abuts the end of the buffer. However, we must
823 * honor IO_SYNC in case someone is silly enough to
824 * configure a HAMMER file as swap, or when HAMMER
825 * is serving NFS (for commits). Ick ick.
827 bp->b_flags |= B_AGE;
828 if (blksize == HAMMER_XBUFSIZE)
829 bp->b_flags |= B_CLUSTEROK;
831 if (ap->a_ioflag & IO_SYNC) {
833 } else if ((ap->a_ioflag & IO_DIRECT) && endofblk) {
835 } else if (ap->a_ioflag & IO_ASYNC) {
837 } else if (hammer_cluster_enable &&
838 !(ap->a_vp->v_mount->mnt_flag & MNT_NOCLUSTERW)) {
839 if (base_offset < HAMMER_XDEMARC)
840 cluster_eof = hammer_blockdemarc(base_offset,
843 cluster_eof = ip->ino_data.size;
844 cluster_write(bp, cluster_eof, blksize, seqcount);
849 hammer_done_transaction(&trans);
850 hammer_knote(ap->a_vp, kflags);
856 * hammer_vop_access { vp, mode, cred }
858 * MPSAFE - does not require fs_token
862 hammer_vop_access(struct vop_access_args *ap)
864 struct hammer_inode *ip = VTOI(ap->a_vp);
869 ++hammer_stats_file_iopsr;
870 uid = hammer_to_unix_xid(&ip->ino_data.uid);
871 gid = hammer_to_unix_xid(&ip->ino_data.gid);
873 error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
874 ip->ino_data.uflags);
879 * hammer_vop_advlock { vp, id, op, fl, flags }
881 * MPSAFE - does not require fs_token
885 hammer_vop_advlock(struct vop_advlock_args *ap)
887 hammer_inode_t ip = VTOI(ap->a_vp);
889 return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
893 * hammer_vop_close { vp, fflag }
895 * We can only sync-on-close for normal closes. XXX disabled for now.
899 hammer_vop_close(struct vop_close_args *ap)
902 struct vnode *vp = ap->a_vp;
903 hammer_inode_t ip = VTOI(vp);
905 if (ip->flags & (HAMMER_INODE_CLOSESYNC|HAMMER_INODE_CLOSEASYNC)) {
906 if (vn_islocked(vp) == LK_EXCLUSIVE &&
907 (vp->v_flag & (VINACTIVE|VRECLAIMED)) == 0) {
908 if (ip->flags & HAMMER_INODE_CLOSESYNC)
911 waitfor = MNT_NOWAIT;
912 ip->flags &= ~(HAMMER_INODE_CLOSESYNC |
913 HAMMER_INODE_CLOSEASYNC);
914 VOP_FSYNC(vp, MNT_NOWAIT, waitfor);
918 return (vop_stdclose(ap));
922 * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
924 * The operating system has already ensured that the directory entry
925 * does not exist and done all appropriate namespace locking.
929 hammer_vop_ncreate(struct vop_ncreate_args *ap)
931 struct hammer_transaction trans;
932 struct hammer_inode *dip;
933 struct hammer_inode *nip;
934 struct nchandle *nch;
939 dip = VTOI(ap->a_dvp);
942 if (dip->flags & HAMMER_INODE_RO)
944 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
948 * Create a transaction to cover the operations we perform.
950 lwkt_gettoken(&hmp->fs_token);
951 hammer_start_transaction(&trans, hmp);
952 ++hammer_stats_file_iopsw;
955 * Create a new filesystem object of the requested type. The
956 * returned inode will be referenced and shared-locked to prevent
957 * it from being moved to the flusher.
959 error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
960 dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
963 hkprintf("hammer_create_inode error %d\n", error);
964 hammer_done_transaction(&trans);
966 lwkt_reltoken(&hmp->fs_token);
971 * Add the new filesystem object to the directory. This will also
972 * bump the inode's link count.
974 error = hammer_ip_add_directory(&trans, dip,
975 nch->ncp->nc_name, nch->ncp->nc_nlen,
978 hkprintf("hammer_ip_add_directory error %d\n", error);
984 hammer_rel_inode(nip, 0);
985 hammer_done_transaction(&trans);
988 error = hammer_get_vnode(nip, ap->a_vpp);
989 hammer_done_transaction(&trans);
990 hammer_rel_inode(nip, 0);
992 cache_setunresolved(ap->a_nch);
993 cache_setvp(ap->a_nch, *ap->a_vpp);
995 hammer_knote(ap->a_dvp, NOTE_WRITE);
997 lwkt_reltoken(&hmp->fs_token);
1002 * hammer_vop_getattr { vp, vap }
1004 * Retrieve an inode's attribute information. When accessing inodes
1005 * historically we fake the atime field to ensure consistent results.
1006 * The atime field is stored in the B-Tree element and allowed to be
1007 * updated without cycling the element.
1009 * MPSAFE - does not require fs_token
1013 hammer_vop_getattr(struct vop_getattr_args *ap)
1015 struct hammer_inode *ip = VTOI(ap->a_vp);
1016 struct vattr *vap = ap->a_vap;
1019 * We want the fsid to be different when accessing a filesystem
1020 * with different as-of's so programs like diff don't think
1021 * the files are the same.
1023 * We also want the fsid to be the same when comparing snapshots,
1024 * or when comparing mirrors (which might be backed by different
1025 * physical devices). HAMMER fsids are based on the PFS's
1026 * shared_uuid field.
1028 * XXX there is a chance of collision here. The va_fsid reported
1029 * by stat is different from the more involved fsid used in the
1032 ++hammer_stats_file_iopsr;
1033 hammer_lock_sh(&ip->lock);
1034 vap->va_fsid = ip->pfsm->fsid_udev ^ (u_int32_t)ip->obj_asof ^
1035 (u_int32_t)(ip->obj_asof >> 32);
1037 vap->va_fileid = ip->ino_leaf.base.obj_id;
1038 vap->va_mode = ip->ino_data.mode;
1039 vap->va_nlink = ip->ino_data.nlinks;
1040 vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1041 vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1044 vap->va_size = ip->ino_data.size;
1047 * Special case for @@PFS softlinks. The actual size of the
1048 * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
1049 * or for MAX_TID is "@@-1:%05d" == 10 bytes.
1051 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
1052 ip->ino_data.size == 10 &&
1053 ip->obj_asof == HAMMER_MAX_TID &&
1054 ip->obj_localization == 0 &&
1055 strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
1056 if (ip->pfsm->pfsd.mirror_flags & HAMMER_PFSD_SLAVE)
1063 * We must provide a consistent atime and mtime for snapshots
1064 * so people can do a 'tar cf - ... | md5' on them and get
1065 * consistent results.
1067 if (ip->flags & HAMMER_INODE_RO) {
1068 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
1069 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
1071 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
1072 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
1074 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
1075 vap->va_flags = ip->ino_data.uflags;
1076 vap->va_gen = 1; /* hammer inums are unique for all time */
1077 vap->va_blocksize = HAMMER_BUFSIZE;
1078 if (ip->ino_data.size >= HAMMER_XDEMARC) {
1079 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
1081 } else if (ip->ino_data.size > HAMMER_BUFSIZE / 2) {
1082 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
1085 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
1088 vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
1089 vap->va_filerev = 0; /* XXX */
1090 vap->va_uid_uuid = ip->ino_data.uid;
1091 vap->va_gid_uuid = ip->ino_data.gid;
1092 vap->va_fsid_uuid = ip->hmp->fsid;
1093 vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
1096 switch (ip->ino_data.obj_type) {
1097 case HAMMER_OBJTYPE_CDEV:
1098 case HAMMER_OBJTYPE_BDEV:
1099 vap->va_rmajor = ip->ino_data.rmajor;
1100 vap->va_rminor = ip->ino_data.rminor;
1105 hammer_unlock(&ip->lock);
1110 * hammer_vop_nresolve { nch, dvp, cred }
1112 * Locate the requested directory entry.
1116 hammer_vop_nresolve(struct vop_nresolve_args *ap)
1118 struct hammer_transaction trans;
1119 struct namecache *ncp;
1124 struct hammer_cursor cursor;
1133 u_int32_t localization;
1134 u_int32_t max_iterations;
1137 * Misc initialization, plus handle as-of name extensions. Look for
1138 * the '@@' extension. Note that as-of files and directories cannot
1141 dip = VTOI(ap->a_dvp);
1142 ncp = ap->a_nch->ncp;
1143 asof = dip->obj_asof;
1144 localization = dip->obj_localization; /* for code consistency */
1145 nlen = ncp->nc_nlen;
1146 flags = dip->flags & HAMMER_INODE_RO;
1150 lwkt_gettoken(&hmp->fs_token);
1151 hammer_simple_transaction(&trans, hmp);
1152 ++hammer_stats_file_iopsr;
1154 for (i = 0; i < nlen; ++i) {
1155 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
1156 error = hammer_str_to_tid(ncp->nc_name + i + 2,
1157 &ispfs, &asof, &localization);
1162 if (asof != HAMMER_MAX_TID)
1163 flags |= HAMMER_INODE_RO;
1170 * If this is a PFS softlink we dive into the PFS
1172 if (ispfs && nlen == 0) {
1173 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1177 error = hammer_get_vnode(ip, &vp);
1178 hammer_rel_inode(ip, 0);
1184 cache_setvp(ap->a_nch, vp);
1191 * If there is no path component the time extension is relative to dip.
1192 * e.g. "fubar/@@<snapshot>"
1194 * "." is handled by the kernel, but ".@@<snapshot>" is not.
1195 * e.g. "fubar/.@@<snapshot>"
1197 * ".." is handled by the kernel. We do not currently handle
1200 if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1201 ip = hammer_get_inode(&trans, dip, dip->obj_id,
1202 asof, dip->obj_localization,
1205 error = hammer_get_vnode(ip, &vp);
1206 hammer_rel_inode(ip, 0);
1212 cache_setvp(ap->a_nch, vp);
1219 * Calculate the namekey and setup the key range for the scan. This
1220 * works kinda like a chained hash table where the lower 32 bits
1221 * of the namekey synthesize the chain.
1223 * The key range is inclusive of both key_beg and key_end.
1225 namekey = hammer_directory_namekey(dip, ncp->nc_name, nlen,
1228 error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1229 cursor.key_beg.localization = dip->obj_localization +
1230 hammer_dir_localization(dip);
1231 cursor.key_beg.obj_id = dip->obj_id;
1232 cursor.key_beg.key = namekey;
1233 cursor.key_beg.create_tid = 0;
1234 cursor.key_beg.delete_tid = 0;
1235 cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1236 cursor.key_beg.obj_type = 0;
1238 cursor.key_end = cursor.key_beg;
1239 cursor.key_end.key += max_iterations;
1241 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1244 * Scan all matching records (the chain), locate the one matching
1245 * the requested path component.
1247 * The hammer_ip_*() functions merge in-memory records with on-disk
1248 * records for the purposes of the search.
1251 localization = HAMMER_DEF_LOCALIZATION;
1254 error = hammer_ip_first(&cursor);
1255 while (error == 0) {
1256 error = hammer_ip_resolve_data(&cursor);
1259 if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1260 bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1261 obj_id = cursor.data->entry.obj_id;
1262 localization = cursor.data->entry.localization;
1265 error = hammer_ip_next(&cursor);
1268 hammer_done_cursor(&cursor);
1271 * Lookup the obj_id. This should always succeed. If it does not
1272 * the filesystem may be damaged and we return a dummy inode.
1275 ip = hammer_get_inode(&trans, dip, obj_id,
1278 if (error == ENOENT) {
1279 kprintf("HAMMER: WARNING: Missing "
1280 "inode for dirent \"%s\"\n"
1281 "\tobj_id = %016llx, asof=%016llx, lo=%08x\n",
1283 (long long)obj_id, (long long)asof,
1286 ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1291 error = hammer_get_vnode(ip, &vp);
1292 hammer_rel_inode(ip, 0);
1298 cache_setvp(ap->a_nch, vp);
1301 } else if (error == ENOENT) {
1302 cache_setvp(ap->a_nch, NULL);
1305 hammer_done_transaction(&trans);
1306 lwkt_reltoken(&hmp->fs_token);
1311 * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1313 * Locate the parent directory of a directory vnode.
1315 * dvp is referenced but not locked. *vpp must be returned referenced and
1316 * locked. A parent_obj_id of 0 does not necessarily indicate that we are
1317 * at the root, instead it could indicate that the directory we were in was
1320 * NOTE: as-of sequences are not linked into the directory structure. If
1321 * we are at the root with a different asof then the mount point, reload
1322 * the same directory with the mount point's asof. I'm not sure what this
1323 * will do to NFS. We encode ASOF stamps in NFS file handles so it might not
1324 * get confused, but it hasn't been tested.
1328 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1330 struct hammer_transaction trans;
1331 struct hammer_inode *dip;
1332 struct hammer_inode *ip;
1334 int64_t parent_obj_id;
1335 u_int32_t parent_obj_localization;
1339 dip = VTOI(ap->a_dvp);
1340 asof = dip->obj_asof;
1344 * Whos are parent? This could be the root of a pseudo-filesystem
1345 * whos parent is in another localization domain.
1347 lwkt_gettoken(&hmp->fs_token);
1348 parent_obj_id = dip->ino_data.parent_obj_id;
1349 if (dip->obj_id == HAMMER_OBJID_ROOT)
1350 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
1352 parent_obj_localization = dip->obj_localization;
1354 if (parent_obj_id == 0) {
1355 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1356 asof != hmp->asof) {
1357 parent_obj_id = dip->obj_id;
1359 *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1360 ksnprintf(*ap->a_fakename, 19, "0x%016llx",
1361 (long long)dip->obj_asof);
1364 lwkt_reltoken(&hmp->fs_token);
1369 hammer_simple_transaction(&trans, hmp);
1370 ++hammer_stats_file_iopsr;
1372 ip = hammer_get_inode(&trans, dip, parent_obj_id,
1373 asof, parent_obj_localization,
1374 dip->flags, &error);
1376 error = hammer_get_vnode(ip, ap->a_vpp);
1377 hammer_rel_inode(ip, 0);
1381 hammer_done_transaction(&trans);
1382 lwkt_reltoken(&hmp->fs_token);
1387 * hammer_vop_nlink { nch, dvp, vp, cred }
1391 hammer_vop_nlink(struct vop_nlink_args *ap)
1393 struct hammer_transaction trans;
1394 struct hammer_inode *dip;
1395 struct hammer_inode *ip;
1396 struct nchandle *nch;
1400 if (ap->a_dvp->v_mount != ap->a_vp->v_mount)
1404 dip = VTOI(ap->a_dvp);
1405 ip = VTOI(ap->a_vp);
1408 if (dip->obj_localization != ip->obj_localization)
1411 if (dip->flags & HAMMER_INODE_RO)
1413 if (ip->flags & HAMMER_INODE_RO)
1415 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1419 * Create a transaction to cover the operations we perform.
1421 lwkt_gettoken(&hmp->fs_token);
1422 hammer_start_transaction(&trans, hmp);
1423 ++hammer_stats_file_iopsw;
1426 * Add the filesystem object to the directory. Note that neither
1427 * dip nor ip are referenced or locked, but their vnodes are
1428 * referenced. This function will bump the inode's link count.
1430 error = hammer_ip_add_directory(&trans, dip,
1431 nch->ncp->nc_name, nch->ncp->nc_nlen,
1438 cache_setunresolved(nch);
1439 cache_setvp(nch, ap->a_vp);
1441 hammer_done_transaction(&trans);
1442 hammer_knote(ap->a_vp, NOTE_LINK);
1443 hammer_knote(ap->a_dvp, NOTE_WRITE);
1444 lwkt_reltoken(&hmp->fs_token);
1449 * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1451 * The operating system has already ensured that the directory entry
1452 * does not exist and done all appropriate namespace locking.
1456 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1458 struct hammer_transaction trans;
1459 struct hammer_inode *dip;
1460 struct hammer_inode *nip;
1461 struct nchandle *nch;
1466 dip = VTOI(ap->a_dvp);
1469 if (dip->flags & HAMMER_INODE_RO)
1471 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1475 * Create a transaction to cover the operations we perform.
1477 lwkt_gettoken(&hmp->fs_token);
1478 hammer_start_transaction(&trans, hmp);
1479 ++hammer_stats_file_iopsw;
1482 * Create a new filesystem object of the requested type. The
1483 * returned inode will be referenced but not locked.
1485 error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1486 dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1489 hkprintf("hammer_mkdir error %d\n", error);
1490 hammer_done_transaction(&trans);
1492 lwkt_reltoken(&hmp->fs_token);
1496 * Add the new filesystem object to the directory. This will also
1497 * bump the inode's link count.
1499 error = hammer_ip_add_directory(&trans, dip,
1500 nch->ncp->nc_name, nch->ncp->nc_nlen,
1503 hkprintf("hammer_mkdir (add) error %d\n", error);
1509 hammer_rel_inode(nip, 0);
1512 error = hammer_get_vnode(nip, ap->a_vpp);
1513 hammer_rel_inode(nip, 0);
1515 cache_setunresolved(ap->a_nch);
1516 cache_setvp(ap->a_nch, *ap->a_vpp);
1519 hammer_done_transaction(&trans);
1521 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1522 lwkt_reltoken(&hmp->fs_token);
1527 * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1529 * The operating system has already ensured that the directory entry
1530 * does not exist and done all appropriate namespace locking.
1534 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1536 struct hammer_transaction trans;
1537 struct hammer_inode *dip;
1538 struct hammer_inode *nip;
1539 struct nchandle *nch;
1544 dip = VTOI(ap->a_dvp);
1547 if (dip->flags & HAMMER_INODE_RO)
1549 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1553 * Create a transaction to cover the operations we perform.
1555 lwkt_gettoken(&hmp->fs_token);
1556 hammer_start_transaction(&trans, hmp);
1557 ++hammer_stats_file_iopsw;
1560 * Create a new filesystem object of the requested type. The
1561 * returned inode will be referenced but not locked.
1563 * If mknod specifies a directory a pseudo-fs is created.
1565 error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1566 dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1569 hammer_done_transaction(&trans);
1571 lwkt_reltoken(&hmp->fs_token);
1576 * Add the new filesystem object to the directory. This will also
1577 * bump the inode's link count.
1579 error = hammer_ip_add_directory(&trans, dip,
1580 nch->ncp->nc_name, nch->ncp->nc_nlen,
1587 hammer_rel_inode(nip, 0);
1590 error = hammer_get_vnode(nip, ap->a_vpp);
1591 hammer_rel_inode(nip, 0);
1593 cache_setunresolved(ap->a_nch);
1594 cache_setvp(ap->a_nch, *ap->a_vpp);
1597 hammer_done_transaction(&trans);
1599 hammer_knote(ap->a_dvp, NOTE_WRITE);
1600 lwkt_reltoken(&hmp->fs_token);
1605 * hammer_vop_open { vp, mode, cred, fp }
1607 * MPSAFE (does not require fs_token)
1611 hammer_vop_open(struct vop_open_args *ap)
1615 ++hammer_stats_file_iopsr;
1616 ip = VTOI(ap->a_vp);
1618 if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1620 return(vop_stdopen(ap));
1624 * hammer_vop_print { vp }
1628 hammer_vop_print(struct vop_print_args *ap)
1634 * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1638 hammer_vop_readdir(struct vop_readdir_args *ap)
1640 struct hammer_transaction trans;
1641 struct hammer_cursor cursor;
1642 struct hammer_inode *ip;
1645 hammer_base_elm_t base;
1654 ++hammer_stats_file_iopsr;
1655 ip = VTOI(ap->a_vp);
1657 saveoff = uio->uio_offset;
1660 if (ap->a_ncookies) {
1661 ncookies = uio->uio_resid / 16 + 1;
1662 if (ncookies > 1024)
1664 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1672 lwkt_gettoken(&hmp->fs_token);
1673 hammer_simple_transaction(&trans, hmp);
1676 * Handle artificial entries
1678 * It should be noted that the minimum value for a directory
1679 * hash key on-media is 0x0000000100000000, so we can use anything
1680 * less then that to represent our 'special' key space.
1684 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1688 cookies[cookie_index] = saveoff;
1691 if (cookie_index == ncookies)
1695 if (ip->ino_data.parent_obj_id) {
1696 r = vop_write_dirent(&error, uio,
1697 ip->ino_data.parent_obj_id,
1700 r = vop_write_dirent(&error, uio,
1701 ip->obj_id, DT_DIR, 2, "..");
1706 cookies[cookie_index] = saveoff;
1709 if (cookie_index == ncookies)
1714 * Key range (begin and end inclusive) to scan. Directory keys
1715 * directly translate to a 64 bit 'seek' position.
1717 hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1718 cursor.key_beg.localization = ip->obj_localization +
1719 hammer_dir_localization(ip);
1720 cursor.key_beg.obj_id = ip->obj_id;
1721 cursor.key_beg.create_tid = 0;
1722 cursor.key_beg.delete_tid = 0;
1723 cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1724 cursor.key_beg.obj_type = 0;
1725 cursor.key_beg.key = saveoff;
1727 cursor.key_end = cursor.key_beg;
1728 cursor.key_end.key = HAMMER_MAX_KEY;
1729 cursor.asof = ip->obj_asof;
1730 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1732 error = hammer_ip_first(&cursor);
1734 while (error == 0) {
1735 error = hammer_ip_resolve_data(&cursor);
1738 base = &cursor.leaf->base;
1739 saveoff = base->key;
1740 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1742 if (base->obj_id != ip->obj_id)
1743 panic("readdir: bad record at %p", cursor.node);
1746 * Convert pseudo-filesystems into softlinks
1748 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1749 r = vop_write_dirent(
1750 &error, uio, cursor.data->entry.obj_id,
1752 cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1753 (void *)cursor.data->entry.name);
1758 cookies[cookie_index] = base->key;
1760 if (cookie_index == ncookies)
1762 error = hammer_ip_next(&cursor);
1764 hammer_done_cursor(&cursor);
1767 hammer_done_transaction(&trans);
1770 *ap->a_eofflag = (error == ENOENT);
1771 uio->uio_offset = saveoff;
1772 if (error && cookie_index == 0) {
1773 if (error == ENOENT)
1776 kfree(cookies, M_TEMP);
1777 *ap->a_ncookies = 0;
1778 *ap->a_cookies = NULL;
1781 if (error == ENOENT)
1784 *ap->a_ncookies = cookie_index;
1785 *ap->a_cookies = cookies;
1788 lwkt_reltoken(&hmp->fs_token);
1793 * hammer_vop_readlink { vp, uio, cred }
1797 hammer_vop_readlink(struct vop_readlink_args *ap)
1799 struct hammer_transaction trans;
1800 struct hammer_cursor cursor;
1801 struct hammer_inode *ip;
1804 u_int32_t localization;
1805 hammer_pseudofs_inmem_t pfsm;
1808 ip = VTOI(ap->a_vp);
1811 lwkt_gettoken(&hmp->fs_token);
1814 * Shortcut if the symlink data was stuffed into ino_data.
1816 * Also expand special "@@PFS%05d" softlinks (expansion only
1817 * occurs for non-historical (current) accesses made from the
1818 * primary filesystem).
1820 if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1824 ptr = ip->ino_data.ext.symlink;
1825 bytes = (int)ip->ino_data.size;
1827 ip->obj_asof == HAMMER_MAX_TID &&
1828 ip->obj_localization == 0 &&
1829 strncmp(ptr, "@@PFS", 5) == 0) {
1830 hammer_simple_transaction(&trans, hmp);
1831 bcopy(ptr + 5, buf, 5);
1833 localization = strtoul(buf, NULL, 10) << 16;
1834 pfsm = hammer_load_pseudofs(&trans, localization,
1837 if (pfsm->pfsd.mirror_flags &
1838 HAMMER_PFSD_SLAVE) {
1839 /* vap->va_size == 26 */
1840 ksnprintf(buf, sizeof(buf),
1842 (long long)pfsm->pfsd.sync_end_tid,
1843 localization >> 16);
1845 /* vap->va_size == 10 */
1846 ksnprintf(buf, sizeof(buf),
1848 localization >> 16);
1850 ksnprintf(buf, sizeof(buf),
1852 (long long)HAMMER_MAX_TID,
1853 localization >> 16);
1857 bytes = strlen(buf);
1860 hammer_rel_pseudofs(hmp, pfsm);
1861 hammer_done_transaction(&trans);
1863 error = uiomove(ptr, bytes, ap->a_uio);
1864 lwkt_reltoken(&hmp->fs_token);
1871 hammer_simple_transaction(&trans, hmp);
1872 ++hammer_stats_file_iopsr;
1873 hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1876 * Key range (begin and end inclusive) to scan. Directory keys
1877 * directly translate to a 64 bit 'seek' position.
1879 cursor.key_beg.localization = ip->obj_localization +
1880 HAMMER_LOCALIZE_MISC;
1881 cursor.key_beg.obj_id = ip->obj_id;
1882 cursor.key_beg.create_tid = 0;
1883 cursor.key_beg.delete_tid = 0;
1884 cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1885 cursor.key_beg.obj_type = 0;
1886 cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1887 cursor.asof = ip->obj_asof;
1888 cursor.flags |= HAMMER_CURSOR_ASOF;
1890 error = hammer_ip_lookup(&cursor);
1892 error = hammer_ip_resolve_data(&cursor);
1894 KKASSERT(cursor.leaf->data_len >=
1895 HAMMER_SYMLINK_NAME_OFF);
1896 error = uiomove(cursor.data->symlink.name,
1897 cursor.leaf->data_len -
1898 HAMMER_SYMLINK_NAME_OFF,
1902 hammer_done_cursor(&cursor);
1903 hammer_done_transaction(&trans);
1904 lwkt_reltoken(&hmp->fs_token);
1909 * hammer_vop_nremove { nch, dvp, cred }
1913 hammer_vop_nremove(struct vop_nremove_args *ap)
1915 struct hammer_transaction trans;
1916 struct hammer_inode *dip;
1920 dip = VTOI(ap->a_dvp);
1923 if (hammer_nohistory(dip) == 0 &&
1924 (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1928 lwkt_gettoken(&hmp->fs_token);
1929 hammer_start_transaction(&trans, hmp);
1930 ++hammer_stats_file_iopsw;
1931 error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1932 hammer_done_transaction(&trans);
1934 hammer_knote(ap->a_dvp, NOTE_WRITE);
1935 lwkt_reltoken(&hmp->fs_token);
1940 * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1944 hammer_vop_nrename(struct vop_nrename_args *ap)
1946 struct hammer_transaction trans;
1947 struct namecache *fncp;
1948 struct namecache *tncp;
1949 struct hammer_inode *fdip;
1950 struct hammer_inode *tdip;
1951 struct hammer_inode *ip;
1953 struct hammer_cursor cursor;
1955 u_int32_t max_iterations;
1958 if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount)
1960 if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1963 fdip = VTOI(ap->a_fdvp);
1964 tdip = VTOI(ap->a_tdvp);
1965 fncp = ap->a_fnch->ncp;
1966 tncp = ap->a_tnch->ncp;
1967 ip = VTOI(fncp->nc_vp);
1968 KKASSERT(ip != NULL);
1972 if (fdip->obj_localization != tdip->obj_localization)
1974 if (fdip->obj_localization != ip->obj_localization)
1977 if (fdip->flags & HAMMER_INODE_RO)
1979 if (tdip->flags & HAMMER_INODE_RO)
1981 if (ip->flags & HAMMER_INODE_RO)
1983 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1986 lwkt_gettoken(&hmp->fs_token);
1987 hammer_start_transaction(&trans, hmp);
1988 ++hammer_stats_file_iopsw;
1991 * Remove tncp from the target directory and then link ip as
1992 * tncp. XXX pass trans to dounlink
1994 * Force the inode sync-time to match the transaction so it is
1995 * in-sync with the creation of the target directory entry.
1997 error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
1999 if (error == 0 || error == ENOENT) {
2000 error = hammer_ip_add_directory(&trans, tdip,
2001 tncp->nc_name, tncp->nc_nlen,
2004 ip->ino_data.parent_obj_id = tdip->obj_id;
2005 ip->ino_data.ctime = trans.time;
2006 hammer_modify_inode(&trans, ip, HAMMER_INODE_DDIRTY);
2010 goto failed; /* XXX */
2013 * Locate the record in the originating directory and remove it.
2015 * Calculate the namekey and setup the key range for the scan. This
2016 * works kinda like a chained hash table where the lower 32 bits
2017 * of the namekey synthesize the chain.
2019 * The key range is inclusive of both key_beg and key_end.
2021 namekey = hammer_directory_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
2024 hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
2025 cursor.key_beg.localization = fdip->obj_localization +
2026 hammer_dir_localization(fdip);
2027 cursor.key_beg.obj_id = fdip->obj_id;
2028 cursor.key_beg.key = namekey;
2029 cursor.key_beg.create_tid = 0;
2030 cursor.key_beg.delete_tid = 0;
2031 cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2032 cursor.key_beg.obj_type = 0;
2034 cursor.key_end = cursor.key_beg;
2035 cursor.key_end.key += max_iterations;
2036 cursor.asof = fdip->obj_asof;
2037 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2040 * Scan all matching records (the chain), locate the one matching
2041 * the requested path component.
2043 * The hammer_ip_*() functions merge in-memory records with on-disk
2044 * records for the purposes of the search.
2046 error = hammer_ip_first(&cursor);
2047 while (error == 0) {
2048 if (hammer_ip_resolve_data(&cursor) != 0)
2050 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2052 if (fncp->nc_nlen == nlen &&
2053 bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2056 error = hammer_ip_next(&cursor);
2060 * If all is ok we have to get the inode so we can adjust nlinks.
2062 * WARNING: hammer_ip_del_directory() may have to terminate the
2063 * cursor to avoid a recursion. It's ok to call hammer_done_cursor()
2067 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
2070 * XXX A deadlock here will break rename's atomicy for the purposes
2071 * of crash recovery.
2073 if (error == EDEADLK) {
2074 hammer_done_cursor(&cursor);
2079 * Cleanup and tell the kernel that the rename succeeded.
2081 * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2082 * without formally acquiring the vp since the vp might
2083 * have zero refs on it, or in the middle of a reclaim,
2086 hammer_done_cursor(&cursor);
2088 cache_rename(ap->a_fnch, ap->a_tnch);
2089 hammer_knote(ap->a_fdvp, NOTE_WRITE);
2090 hammer_knote(ap->a_tdvp, NOTE_WRITE);
2094 error = hammer_get_vnode(ip, &vp);
2095 if (error == 0 && vp) {
2097 hammer_knote(ip->vp, NOTE_RENAME);
2101 kprintf("Debug: HAMMER ip/vp race2 avoided\n");
2106 hammer_done_transaction(&trans);
2107 lwkt_reltoken(&hmp->fs_token);
2112 * hammer_vop_nrmdir { nch, dvp, cred }
2116 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
2118 struct hammer_transaction trans;
2119 struct hammer_inode *dip;
2123 dip = VTOI(ap->a_dvp);
2126 if (hammer_nohistory(dip) == 0 &&
2127 (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2131 lwkt_gettoken(&hmp->fs_token);
2132 hammer_start_transaction(&trans, hmp);
2133 ++hammer_stats_file_iopsw;
2134 error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
2135 hammer_done_transaction(&trans);
2137 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
2138 lwkt_reltoken(&hmp->fs_token);
2143 * hammer_vop_markatime { vp, cred }
2147 hammer_vop_markatime(struct vop_markatime_args *ap)
2149 struct hammer_transaction trans;
2150 struct hammer_inode *ip;
2153 ip = VTOI(ap->a_vp);
2154 if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2156 if (ip->flags & HAMMER_INODE_RO)
2159 if (hmp->mp->mnt_flag & MNT_NOATIME)
2161 lwkt_gettoken(&hmp->fs_token);
2162 hammer_start_transaction(&trans, hmp);
2163 ++hammer_stats_file_iopsw;
2165 ip->ino_data.atime = trans.time;
2166 hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
2167 hammer_done_transaction(&trans);
2168 hammer_knote(ap->a_vp, NOTE_ATTRIB);
2169 lwkt_reltoken(&hmp->fs_token);
2174 * hammer_vop_setattr { vp, vap, cred }
2178 hammer_vop_setattr(struct vop_setattr_args *ap)
2180 struct hammer_transaction trans;
2181 struct hammer_inode *ip;
2190 int64_t aligned_size;
2195 ip = ap->a_vp->v_data;
2200 if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2202 if (ip->flags & HAMMER_INODE_RO)
2204 if (hammer_nohistory(ip) == 0 &&
2205 (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2209 lwkt_gettoken(&hmp->fs_token);
2210 hammer_start_transaction(&trans, hmp);
2211 ++hammer_stats_file_iopsw;
2214 if (vap->va_flags != VNOVAL) {
2215 flags = ip->ino_data.uflags;
2216 error = vop_helper_setattr_flags(&flags, vap->va_flags,
2217 hammer_to_unix_xid(&ip->ino_data.uid),
2220 if (ip->ino_data.uflags != flags) {
2221 ip->ino_data.uflags = flags;
2222 ip->ino_data.ctime = trans.time;
2223 modflags |= HAMMER_INODE_DDIRTY;
2224 kflags |= NOTE_ATTRIB;
2226 if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2233 if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2237 if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2238 mode_t cur_mode = ip->ino_data.mode;
2239 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2240 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2244 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2246 &cur_uid, &cur_gid, &cur_mode);
2248 hammer_guid_to_uuid(&uuid_uid, cur_uid);
2249 hammer_guid_to_uuid(&uuid_gid, cur_gid);
2250 if (bcmp(&uuid_uid, &ip->ino_data.uid,
2251 sizeof(uuid_uid)) ||
2252 bcmp(&uuid_gid, &ip->ino_data.gid,
2253 sizeof(uuid_gid)) ||
2254 ip->ino_data.mode != cur_mode
2256 ip->ino_data.uid = uuid_uid;
2257 ip->ino_data.gid = uuid_gid;
2258 ip->ino_data.mode = cur_mode;
2259 ip->ino_data.ctime = trans.time;
2260 modflags |= HAMMER_INODE_DDIRTY;
2262 kflags |= NOTE_ATTRIB;
2265 while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2266 switch(ap->a_vp->v_type) {
2268 if (vap->va_size == ip->ino_data.size)
2272 * Log the operation if in fast-fsync mode or if
2273 * there are unterminated redo write records present.
2275 * The second check is needed so the recovery code
2276 * properly truncates write redos even if nominal
2277 * REDO operations is turned off due to excessive
2278 * writes, because the related records might be
2279 * destroyed and never lay down a TERM_WRITE.
2281 if ((ip->flags & HAMMER_INODE_REDO) ||
2282 (ip->flags & HAMMER_INODE_RDIRTY)) {
2283 error = hammer_generate_redo(&trans, ip,
2288 blksize = hammer_blocksize(vap->va_size);
2291 * XXX break atomicy, we can deadlock the backend
2292 * if we do not release the lock. Probably not a
2295 if (vap->va_size < ip->ino_data.size) {
2296 nvtruncbuf(ap->a_vp, vap->va_size,
2298 hammer_blockoff(vap->va_size),
2301 kflags |= NOTE_WRITE;
2303 nvextendbuf(ap->a_vp,
2306 hammer_blocksize(ip->ino_data.size),
2307 hammer_blocksize(vap->va_size),
2308 hammer_blockoff(ip->ino_data.size),
2309 hammer_blockoff(vap->va_size),
2312 kflags |= NOTE_WRITE | NOTE_EXTEND;
2314 ip->ino_data.size = vap->va_size;
2315 ip->ino_data.mtime = trans.time;
2316 /* XXX safe to use SDIRTY instead of DDIRTY here? */
2317 modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2320 * On-media truncation is cached in the inode until
2321 * the inode is synchronized. We must immediately
2322 * handle any frontend records.
2325 hammer_ip_frontend_trunc(ip, vap->va_size);
2326 #ifdef DEBUG_TRUNCATE
2327 if (HammerTruncIp == NULL)
2330 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2331 ip->flags |= HAMMER_INODE_TRUNCATED;
2332 ip->trunc_off = vap->va_size;
2333 hammer_inode_dirty(ip);
2334 #ifdef DEBUG_TRUNCATE
2335 if (ip == HammerTruncIp)
2336 kprintf("truncate1 %016llx\n",
2337 (long long)ip->trunc_off);
2339 } else if (ip->trunc_off > vap->va_size) {
2340 ip->trunc_off = vap->va_size;
2341 #ifdef DEBUG_TRUNCATE
2342 if (ip == HammerTruncIp)
2343 kprintf("truncate2 %016llx\n",
2344 (long long)ip->trunc_off);
2347 #ifdef DEBUG_TRUNCATE
2348 if (ip == HammerTruncIp)
2349 kprintf("truncate3 %016llx (ignored)\n",
2350 (long long)vap->va_size);
2357 * When truncating, nvtruncbuf() may have cleaned out
2358 * a portion of the last block on-disk in the buffer
2359 * cache. We must clean out any frontend records
2360 * for blocks beyond the new last block.
2362 aligned_size = (vap->va_size + (blksize - 1)) &
2363 ~(int64_t)(blksize - 1);
2364 if (truncating && vap->va_size < aligned_size) {
2365 aligned_size -= blksize;
2366 hammer_ip_frontend_trunc(ip, aligned_size);
2371 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2372 ip->flags |= HAMMER_INODE_TRUNCATED;
2373 ip->trunc_off = vap->va_size;
2374 hammer_inode_dirty(ip);
2375 } else if (ip->trunc_off > vap->va_size) {
2376 ip->trunc_off = vap->va_size;
2378 hammer_ip_frontend_trunc(ip, vap->va_size);
2379 ip->ino_data.size = vap->va_size;
2380 ip->ino_data.mtime = trans.time;
2381 modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2382 kflags |= NOTE_ATTRIB;
2390 if (vap->va_atime.tv_sec != VNOVAL) {
2391 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2392 modflags |= HAMMER_INODE_ATIME;
2393 kflags |= NOTE_ATTRIB;
2395 if (vap->va_mtime.tv_sec != VNOVAL) {
2396 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2397 modflags |= HAMMER_INODE_MTIME;
2398 kflags |= NOTE_ATTRIB;
2400 if (vap->va_mode != (mode_t)VNOVAL) {
2401 mode_t cur_mode = ip->ino_data.mode;
2402 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2403 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2405 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2406 cur_uid, cur_gid, &cur_mode);
2407 if (error == 0 && ip->ino_data.mode != cur_mode) {
2408 ip->ino_data.mode = cur_mode;
2409 ip->ino_data.ctime = trans.time;
2410 modflags |= HAMMER_INODE_DDIRTY;
2411 kflags |= NOTE_ATTRIB;
2416 hammer_modify_inode(&trans, ip, modflags);
2417 hammer_done_transaction(&trans);
2418 hammer_knote(ap->a_vp, kflags);
2419 lwkt_reltoken(&hmp->fs_token);
2424 * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2428 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2430 struct hammer_transaction trans;
2431 struct hammer_inode *dip;
2432 struct hammer_inode *nip;
2433 hammer_record_t record;
2434 struct nchandle *nch;
2439 ap->a_vap->va_type = VLNK;
2442 dip = VTOI(ap->a_dvp);
2445 if (dip->flags & HAMMER_INODE_RO)
2447 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
2451 * Create a transaction to cover the operations we perform.
2453 lwkt_gettoken(&hmp->fs_token);
2454 hammer_start_transaction(&trans, hmp);
2455 ++hammer_stats_file_iopsw;
2458 * Create a new filesystem object of the requested type. The
2459 * returned inode will be referenced but not locked.
2462 error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2463 dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2466 hammer_done_transaction(&trans);
2468 lwkt_reltoken(&hmp->fs_token);
2473 * Add a record representing the symlink. symlink stores the link
2474 * as pure data, not a string, and is no \0 terminated.
2477 bytes = strlen(ap->a_target);
2479 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2480 bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2482 record = hammer_alloc_mem_record(nip, bytes);
2483 record->type = HAMMER_MEM_RECORD_GENERAL;
2485 record->leaf.base.localization = nip->obj_localization +
2486 HAMMER_LOCALIZE_MISC;
2487 record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2488 record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2489 record->leaf.data_len = bytes;
2490 KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2491 bcopy(ap->a_target, record->data->symlink.name, bytes);
2492 error = hammer_ip_add_record(&trans, record);
2496 * Set the file size to the length of the link.
2499 nip->ino_data.size = bytes;
2500 hammer_modify_inode(&trans, nip, HAMMER_INODE_DDIRTY);
2504 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
2505 nch->ncp->nc_nlen, nip);
2511 hammer_rel_inode(nip, 0);
2514 error = hammer_get_vnode(nip, ap->a_vpp);
2515 hammer_rel_inode(nip, 0);
2517 cache_setunresolved(ap->a_nch);
2518 cache_setvp(ap->a_nch, *ap->a_vpp);
2519 hammer_knote(ap->a_dvp, NOTE_WRITE);
2522 hammer_done_transaction(&trans);
2523 lwkt_reltoken(&hmp->fs_token);
2528 * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2532 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2534 struct hammer_transaction trans;
2535 struct hammer_inode *dip;
2539 dip = VTOI(ap->a_dvp);
2542 if (hammer_nohistory(dip) == 0 &&
2543 (error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2547 lwkt_gettoken(&hmp->fs_token);
2548 hammer_start_transaction(&trans, hmp);
2549 ++hammer_stats_file_iopsw;
2550 error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2551 ap->a_cred, ap->a_flags, -1);
2552 hammer_done_transaction(&trans);
2553 lwkt_reltoken(&hmp->fs_token);
2559 * hammer_vop_ioctl { vp, command, data, fflag, cred }
2563 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2565 struct hammer_inode *ip = ap->a_vp->v_data;
2566 hammer_mount_t hmp = ip->hmp;
2569 ++hammer_stats_file_iopsr;
2570 lwkt_gettoken(&hmp->fs_token);
2571 error = hammer_ioctl(ip, ap->a_command, ap->a_data,
2572 ap->a_fflag, ap->a_cred);
2573 lwkt_reltoken(&hmp->fs_token);
2579 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2581 static const struct mountctl_opt extraopt[] = {
2582 { HMNT_NOHISTORY, "nohistory" },
2583 { HMNT_MASTERID, "master" },
2587 struct hammer_mount *hmp;
2594 mp = ap->a_head.a_ops->head.vv_mount;
2595 KKASSERT(mp->mnt_data != NULL);
2596 hmp = (struct hammer_mount *)mp->mnt_data;
2598 lwkt_gettoken(&hmp->fs_token);
2601 case MOUNTCTL_SET_EXPORT:
2602 if (ap->a_ctllen != sizeof(struct export_args))
2605 error = hammer_vfs_export(mp, ap->a_op,
2606 (const struct export_args *)ap->a_ctl);
2608 case MOUNTCTL_MOUNTFLAGS:
2611 * Call standard mountctl VOP function
2612 * so we get user mount flags.
2614 error = vop_stdmountctl(ap);
2618 usedbytes = *ap->a_res;
2620 if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2621 usedbytes += vfs_flagstostr(hmp->hflags, extraopt,
2623 ap->a_buflen - usedbytes,
2627 *ap->a_res += usedbytes;
2631 error = vop_stdmountctl(ap);
2634 lwkt_reltoken(&hmp->fs_token);
2639 * hammer_vop_strategy { vp, bio }
2641 * Strategy call, used for regular file read & write only. Note that the
2642 * bp may represent a cluster.
2644 * To simplify operation and allow better optimizations in the future,
2645 * this code does not make any assumptions with regards to buffer alignment
2650 hammer_vop_strategy(struct vop_strategy_args *ap)
2655 bp = ap->a_bio->bio_buf;
2659 error = hammer_vop_strategy_read(ap);
2662 error = hammer_vop_strategy_write(ap);
2665 bp->b_error = error = EINVAL;
2666 bp->b_flags |= B_ERROR;
2671 /* hammer_dump_dedup_cache(((hammer_inode_t)ap->a_vp->v_data)->hmp); */
2677 * Read from a regular file. Iterate the related records and fill in the
2678 * BIO/BUF. Gaps are zero-filled.
2680 * The support code in hammer_object.c should be used to deal with mixed
2681 * in-memory and on-disk records.
2683 * NOTE: Can be called from the cluster code with an oversized buf.
2689 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2691 struct hammer_transaction trans;
2692 struct hammer_inode *ip;
2693 struct hammer_inode *dip;
2695 struct hammer_cursor cursor;
2696 hammer_base_elm_t base;
2697 hammer_off_t disk_offset;
2712 ip = ap->a_vp->v_data;
2716 * The zone-2 disk offset may have been set by the cluster code via
2717 * a BMAP operation, or else should be NOOFFSET.
2719 * Checking the high bits for a match against zone-2 should suffice.
2721 * In cases where a lot of data duplication is present it may be
2722 * more beneficial to drop through and doubule-buffer through the
2725 nbio = push_bio(bio);
2726 if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2727 HAMMER_ZONE_LARGE_DATA) {
2728 if (hammer_double_buffer == 0) {
2729 lwkt_gettoken(&hmp->fs_token);
2730 error = hammer_io_direct_read(hmp, nbio, NULL);
2731 lwkt_reltoken(&hmp->fs_token);
2736 * Try to shortcut requests for double_buffer mode too.
2737 * Since this mode runs through the device buffer cache
2738 * only compatible buffer sizes (meaning those generated
2739 * by normal filesystem buffers) are legal.
2741 if (hammer_live_dedup == 0 && (bp->b_flags & B_PAGING) == 0) {
2742 lwkt_gettoken(&hmp->fs_token);
2743 error = hammer_io_indirect_read(hmp, nbio, NULL);
2744 lwkt_reltoken(&hmp->fs_token);
2750 * Well, that sucked. Do it the hard way. If all the stars are
2751 * aligned we may still be able to issue a direct-read.
2753 lwkt_gettoken(&hmp->fs_token);
2754 hammer_simple_transaction(&trans, hmp);
2755 hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2758 * Key range (begin and end inclusive) to scan. Note that the key's
2759 * stored in the actual records represent BASE+LEN, not BASE. The
2760 * first record containing bio_offset will have a key > bio_offset.
2762 cursor.key_beg.localization = ip->obj_localization +
2763 HAMMER_LOCALIZE_MISC;
2764 cursor.key_beg.obj_id = ip->obj_id;
2765 cursor.key_beg.create_tid = 0;
2766 cursor.key_beg.delete_tid = 0;
2767 cursor.key_beg.obj_type = 0;
2768 cursor.key_beg.key = bio->bio_offset + 1;
2769 cursor.asof = ip->obj_asof;
2770 cursor.flags |= HAMMER_CURSOR_ASOF;
2772 cursor.key_end = cursor.key_beg;
2773 KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2775 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2776 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2777 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2778 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2782 ran_end = bio->bio_offset + bp->b_bufsize;
2783 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2784 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2785 tmp64 = ran_end + MAXPHYS + 1; /* work-around GCC-4 bug */
2786 if (tmp64 < ran_end)
2787 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2789 cursor.key_end.key = ran_end + MAXPHYS + 1;
2791 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2794 * Set NOSWAPCACHE for cursor data extraction if double buffering
2795 * is disabled or (if the file is not marked cacheable via chflags
2796 * and vm.swapcache_use_chflags is enabled).
2798 if (hammer_double_buffer == 0 ||
2799 ((ap->a_vp->v_flag & VSWAPCACHE) == 0 &&
2800 vm_swapcache_use_chflags)) {
2801 cursor.flags |= HAMMER_CURSOR_NOSWAPCACHE;
2804 error = hammer_ip_first(&cursor);
2807 while (error == 0) {
2809 * Get the base file offset of the record. The key for
2810 * data records is (base + bytes) rather then (base).
2812 base = &cursor.leaf->base;
2813 rec_offset = base->key - cursor.leaf->data_len;
2816 * Calculate the gap, if any, and zero-fill it.
2818 * n is the offset of the start of the record verses our
2819 * current seek offset in the bio.
2821 n = (int)(rec_offset - (bio->bio_offset + boff));
2823 if (n > bp->b_bufsize - boff)
2824 n = bp->b_bufsize - boff;
2825 bzero((char *)bp->b_data + boff, n);
2831 * Calculate the data offset in the record and the number
2832 * of bytes we can copy.
2834 * There are two degenerate cases. First, boff may already
2835 * be at bp->b_bufsize. Secondly, the data offset within
2836 * the record may exceed the record's size.
2840 n = cursor.leaf->data_len - roff;
2842 kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2844 } else if (n > bp->b_bufsize - boff) {
2845 n = bp->b_bufsize - boff;
2849 * Deal with cached truncations. This cool bit of code
2850 * allows truncate()/ftruncate() to avoid having to sync
2853 * If the frontend is truncated then all backend records are
2854 * subject to the frontend's truncation.
2856 * If the backend is truncated then backend records on-disk
2857 * (but not in-memory) are subject to the backend's
2858 * truncation. In-memory records owned by the backend
2859 * represent data written after the truncation point on the
2860 * backend and must not be truncated.
2862 * Truncate operations deal with frontend buffer cache
2863 * buffers and frontend-owned in-memory records synchronously.
2865 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2866 if (hammer_cursor_ondisk(&cursor)/* ||
2867 cursor.iprec->flush_state == HAMMER_FST_FLUSH*/) {
2868 if (ip->trunc_off <= rec_offset)
2870 else if (ip->trunc_off < rec_offset + n)
2871 n = (int)(ip->trunc_off - rec_offset);
2874 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2875 if (hammer_cursor_ondisk(&cursor)) {
2876 if (ip->sync_trunc_off <= rec_offset)
2878 else if (ip->sync_trunc_off < rec_offset + n)
2879 n = (int)(ip->sync_trunc_off - rec_offset);
2884 * Try to issue a direct read into our bio if possible,
2885 * otherwise resolve the element data into a hammer_buffer
2888 * The buffer on-disk should be zerod past any real
2889 * truncation point, but may not be for any synthesized
2890 * truncation point from above.
2892 * NOTE: disk_offset is only valid if the cursor data is
2895 disk_offset = cursor.leaf->data_offset + roff;
2896 isdedupable = (boff == 0 && n == bp->b_bufsize &&
2897 hammer_cursor_ondisk(&cursor) &&
2898 ((int)disk_offset & HAMMER_BUFMASK) == 0);
2900 if (isdedupable && hammer_double_buffer == 0) {
2904 KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2905 HAMMER_ZONE_LARGE_DATA);
2906 nbio->bio_offset = disk_offset;
2907 error = hammer_io_direct_read(hmp, nbio, cursor.leaf);
2908 if (hammer_live_dedup && error == 0)
2909 hammer_dedup_cache_add(ip, cursor.leaf);
2911 } else if (isdedupable) {
2913 * Async I/O case for reading from backing store
2914 * and copying the data to the filesystem buffer.
2915 * live-dedup has to verify the data anyway if it
2916 * gets a hit later so we can just add the entry
2919 KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2920 HAMMER_ZONE_LARGE_DATA);
2921 nbio->bio_offset = disk_offset;
2922 if (hammer_live_dedup)
2923 hammer_dedup_cache_add(ip, cursor.leaf);
2924 error = hammer_io_indirect_read(hmp, nbio, cursor.leaf);
2927 error = hammer_ip_resolve_data(&cursor);
2929 if (hammer_live_dedup && isdedupable)
2930 hammer_dedup_cache_add(ip, cursor.leaf);
2931 bcopy((char *)cursor.data + roff,
2932 (char *)bp->b_data + boff, n);
2939 * We have to be sure that the only elements added to the
2940 * dedup cache are those which are already on-media.
2942 if (hammer_live_dedup && hammer_cursor_ondisk(&cursor))
2943 hammer_dedup_cache_add(ip, cursor.leaf);
2946 * Iterate until we have filled the request.
2949 if (boff == bp->b_bufsize)
2951 error = hammer_ip_next(&cursor);
2955 * There may have been a gap after the last record
2957 if (error == ENOENT)
2959 if (error == 0 && boff != bp->b_bufsize) {
2960 KKASSERT(boff < bp->b_bufsize);
2961 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2962 /* boff = bp->b_bufsize; */
2966 * Disallow swapcache operation on the vnode buffer if double
2967 * buffering is enabled, the swapcache will get the data via
2968 * the block device buffer.
2970 if (hammer_double_buffer)
2971 bp->b_flags |= B_NOTMETA;
2977 bp->b_error = error;
2979 bp->b_flags |= B_ERROR;
2984 * Cache the b-tree node for the last data read in cache[1].
2986 * If we hit the file EOF then also cache the node in the
2987 * governing director's cache[3], it will be used to initialize
2988 * the inode's cache[1] for any inodes looked up via the directory.
2990 * This doesn't reduce disk accesses since the B-Tree chain is
2991 * likely cached, but it does reduce cpu overhead when looking
2992 * up file offsets for cpdup/tar/cpio style iterations.
2995 hammer_cache_node(&ip->cache[1], cursor.node);
2996 if (ran_end >= ip->ino_data.size) {
2997 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
2998 ip->obj_asof, ip->obj_localization);
3000 hammer_cache_node(&dip->cache[3], cursor.node);
3001 hammer_rel_inode(dip, 0);
3004 hammer_done_cursor(&cursor);
3005 hammer_done_transaction(&trans);
3006 lwkt_reltoken(&hmp->fs_token);
3011 * BMAP operation - used to support cluster_read() only.
3013 * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
3015 * This routine may return EOPNOTSUPP if the opration is not supported for
3016 * the specified offset. The contents of the pointer arguments do not
3017 * need to be initialized in that case.
3019 * If a disk address is available and properly aligned return 0 with
3020 * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
3021 * to the run-length relative to that offset. Callers may assume that
3022 * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
3023 * large, so return EOPNOTSUPP if it is not sufficiently large.
3027 hammer_vop_bmap(struct vop_bmap_args *ap)
3029 struct hammer_transaction trans;
3030 struct hammer_inode *ip;
3032 struct hammer_cursor cursor;
3033 hammer_base_elm_t base;
3037 int64_t base_offset;
3038 int64_t base_disk_offset;
3039 int64_t last_offset;
3040 hammer_off_t last_disk_offset;
3041 hammer_off_t disk_offset;
3046 ++hammer_stats_file_iopsr;
3047 ip = ap->a_vp->v_data;
3051 * We can only BMAP regular files. We can't BMAP database files,
3054 if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
3058 * bmap is typically called with runp/runb both NULL when used
3059 * for writing. We do not support BMAP for writing atm.
3061 if (ap->a_cmd != BUF_CMD_READ)
3065 * Scan the B-Tree to acquire blockmap addresses, then translate
3068 lwkt_gettoken(&hmp->fs_token);
3069 hammer_simple_transaction(&trans, hmp);
3071 kprintf("bmap_beg %016llx ip->cache %p\n",
3072 (long long)ap->a_loffset, ip->cache[1]);
3074 hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
3077 * Key range (begin and end inclusive) to scan. Note that the key's
3078 * stored in the actual records represent BASE+LEN, not BASE. The
3079 * first record containing bio_offset will have a key > bio_offset.
3081 cursor.key_beg.localization = ip->obj_localization +
3082 HAMMER_LOCALIZE_MISC;
3083 cursor.key_beg.obj_id = ip->obj_id;
3084 cursor.key_beg.create_tid = 0;
3085 cursor.key_beg.delete_tid = 0;
3086 cursor.key_beg.obj_type = 0;
3088 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
3090 cursor.key_beg.key = ap->a_loffset + 1;
3091 if (cursor.key_beg.key < 0)
3092 cursor.key_beg.key = 0;
3093 cursor.asof = ip->obj_asof;
3094 cursor.flags |= HAMMER_CURSOR_ASOF;
3096 cursor.key_end = cursor.key_beg;
3097 KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
3099 ran_end = ap->a_loffset + MAXPHYS;
3100 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
3101 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
3102 tmp64 = ran_end + MAXPHYS + 1; /* work-around GCC-4 bug */
3103 if (tmp64 < ran_end)
3104 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
3106 cursor.key_end.key = ran_end + MAXPHYS + 1;
3108 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
3110 error = hammer_ip_first(&cursor);
3111 base_offset = last_offset = 0;
3112 base_disk_offset = last_disk_offset = 0;
3114 while (error == 0) {
3116 * Get the base file offset of the record. The key for
3117 * data records is (base + bytes) rather then (base).
3119 * NOTE: rec_offset + rec_len may exceed the end-of-file.
3120 * The extra bytes should be zero on-disk and the BMAP op
3121 * should still be ok.
3123 base = &cursor.leaf->base;
3124 rec_offset = base->key - cursor.leaf->data_len;
3125 rec_len = cursor.leaf->data_len;
3128 * Incorporate any cached truncation.
3130 * NOTE: Modifications to rec_len based on synthesized
3131 * truncation points remove the guarantee that any extended
3132 * data on disk is zero (since the truncations may not have
3133 * taken place on-media yet).
3135 if (ip->flags & HAMMER_INODE_TRUNCATED) {
3136 if (hammer_cursor_ondisk(&cursor) ||
3137 cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
3138 if (ip->trunc_off <= rec_offset)
3140 else if (ip->trunc_off < rec_offset + rec_len)
3141 rec_len = (int)(ip->trunc_off - rec_offset);
3144 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
3145 if (hammer_cursor_ondisk(&cursor)) {
3146 if (ip->sync_trunc_off <= rec_offset)
3148 else if (ip->sync_trunc_off < rec_offset + rec_len)
3149 rec_len = (int)(ip->sync_trunc_off - rec_offset);
3154 * Accumulate information. If we have hit a discontiguous
3155 * block reset base_offset unless we are already beyond the
3156 * requested offset. If we are, that's it, we stop.
3160 if (hammer_cursor_ondisk(&cursor)) {
3161 disk_offset = cursor.leaf->data_offset;
3162 if (rec_offset != last_offset ||
3163 disk_offset != last_disk_offset) {
3164 if (rec_offset > ap->a_loffset)
3166 base_offset = rec_offset;
3167 base_disk_offset = disk_offset;
3169 last_offset = rec_offset + rec_len;
3170 last_disk_offset = disk_offset + rec_len;
3172 if (hammer_live_dedup)
3173 hammer_dedup_cache_add(ip, cursor.leaf);
3176 error = hammer_ip_next(&cursor);
3180 kprintf("BMAP %016llx: %016llx - %016llx\n",
3181 (long long)ap->a_loffset,
3182 (long long)base_offset,
3183 (long long)last_offset);
3184 kprintf("BMAP %16s: %016llx - %016llx\n", "",
3185 (long long)base_disk_offset,
3186 (long long)last_disk_offset);
3190 hammer_cache_node(&ip->cache[1], cursor.node);
3192 kprintf("bmap_end2 %016llx ip->cache %p\n",
3193 (long long)ap->a_loffset, ip->cache[1]);
3196 hammer_done_cursor(&cursor);
3197 hammer_done_transaction(&trans);
3198 lwkt_reltoken(&hmp->fs_token);
3201 * If we couldn't find any records or the records we did find were
3202 * all behind the requested offset, return failure. A forward
3203 * truncation can leave a hole w/ no on-disk records.
3205 if (last_offset == 0 || last_offset < ap->a_loffset)
3206 return (EOPNOTSUPP);
3209 * Figure out the block size at the requested offset and adjust
3210 * our limits so the cluster_read() does not create inappropriately
3211 * sized buffer cache buffers.
3213 blksize = hammer_blocksize(ap->a_loffset);
3214 if (hammer_blocksize(base_offset) != blksize) {
3215 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
3217 if (last_offset != ap->a_loffset &&
3218 hammer_blocksize(last_offset - 1) != blksize) {
3219 last_offset = hammer_blockdemarc(ap->a_loffset,
3224 * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3227 disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
3229 if ((disk_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_LARGE_DATA) {
3231 * Only large-data zones can be direct-IOd
3234 } else if ((disk_offset & HAMMER_BUFMASK) ||
3235 (last_offset - ap->a_loffset) < blksize) {
3237 * doffsetp is not aligned or the forward run size does
3238 * not cover a whole buffer, disallow the direct I/O.
3245 *ap->a_doffsetp = disk_offset;
3247 *ap->a_runb = ap->a_loffset - base_offset;
3248 KKASSERT(*ap->a_runb >= 0);
3251 *ap->a_runp = last_offset - ap->a_loffset;
3252 KKASSERT(*ap->a_runp >= 0);
3260 * Write to a regular file. Because this is a strategy call the OS is
3261 * trying to actually get data onto the media.
3265 hammer_vop_strategy_write(struct vop_strategy_args *ap)
3267 hammer_record_t record;
3272 int blksize __debugvar;
3278 ip = ap->a_vp->v_data;
3281 blksize = hammer_blocksize(bio->bio_offset);
3282 KKASSERT(bp->b_bufsize == blksize);
3284 if (ip->flags & HAMMER_INODE_RO) {
3285 bp->b_error = EROFS;
3286 bp->b_flags |= B_ERROR;
3291 lwkt_gettoken(&hmp->fs_token);
3294 * Disallow swapcache operation on the vnode buffer if double
3295 * buffering is enabled, the swapcache will get the data via
3296 * the block device buffer.
3298 if (hammer_double_buffer)
3299 bp->b_flags |= B_NOTMETA;
3302 * Interlock with inode destruction (no in-kernel or directory
3303 * topology visibility). If we queue new IO while trying to
3304 * destroy the inode we can deadlock the vtrunc call in
3305 * hammer_inode_unloadable_check().
3307 * Besides, there's no point flushing a bp associated with an
3308 * inode that is being destroyed on-media and has no kernel
3311 if ((ip->flags | ip->sync_flags) &
3312 (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
3315 lwkt_reltoken(&hmp->fs_token);
3320 * Reserve space and issue a direct-write from the front-end.
3321 * NOTE: The direct_io code will hammer_bread/bcopy smaller
3324 * An in-memory record will be installed to reference the storage
3325 * until the flusher can get to it.
3327 * Since we own the high level bio the front-end will not try to
3328 * do a direct-read until the write completes.
3330 * NOTE: The only time we do not reserve a full-sized buffers
3331 * worth of data is if the file is small. We do not try to
3332 * allocate a fragment (from the small-data zone) at the end of
3333 * an otherwise large file as this can lead to wildly separated
3336 KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
3337 KKASSERT(bio->bio_offset < ip->ino_data.size);
3338 if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
3339 bytes = bp->b_bufsize;
3341 bytes = ((int)ip->ino_data.size + 15) & ~15;
3343 record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
3347 * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3348 * in hammer_vop_write(). We must flag the record so the proper
3349 * REDO_TERM_WRITE entry is generated during the flush.
3352 if (bp->b_flags & B_VFSFLAG1) {
3353 record->flags |= HAMMER_RECF_REDO;
3354 bp->b_flags &= ~B_VFSFLAG1;
3356 if (record->flags & HAMMER_RECF_DEDUPED) {
3358 hammer_ip_replace_bulk(hmp, record);
3361 hammer_io_direct_write(hmp, bio, record);
3363 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
3364 hammer_flush_inode(ip, 0);
3366 bp->b_bio2.bio_offset = NOOFFSET;
3367 bp->b_error = error;
3368 bp->b_flags |= B_ERROR;
3371 lwkt_reltoken(&hmp->fs_token);
3376 * dounlink - disconnect a directory entry
3378 * XXX whiteout support not really in yet
3381 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
3382 struct vnode *dvp, struct ucred *cred,
3383 int flags, int isdir)
3385 struct namecache *ncp;
3389 struct hammer_cursor cursor;
3391 u_int32_t max_iterations;
3395 * Calculate the namekey and setup the key range for the scan. This
3396 * works kinda like a chained hash table where the lower 32 bits
3397 * of the namekey synthesize the chain.
3399 * The key range is inclusive of both key_beg and key_end.
3405 if (dip->flags & HAMMER_INODE_RO)
3408 namekey = hammer_directory_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3411 hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3412 cursor.key_beg.localization = dip->obj_localization +
3413 hammer_dir_localization(dip);
3414 cursor.key_beg.obj_id = dip->obj_id;
3415 cursor.key_beg.key = namekey;
3416 cursor.key_beg.create_tid = 0;
3417 cursor.key_beg.delete_tid = 0;
3418 cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3419 cursor.key_beg.obj_type = 0;
3421 cursor.key_end = cursor.key_beg;
3422 cursor.key_end.key += max_iterations;
3423 cursor.asof = dip->obj_asof;
3424 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3427 * Scan all matching records (the chain), locate the one matching
3428 * the requested path component. info->last_error contains the
3429 * error code on search termination and could be 0, ENOENT, or
3432 * The hammer_ip_*() functions merge in-memory records with on-disk
3433 * records for the purposes of the search.
3435 error = hammer_ip_first(&cursor);
3437 while (error == 0) {
3438 error = hammer_ip_resolve_data(&cursor);
3441 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3443 if (ncp->nc_nlen == nlen &&
3444 bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3447 error = hammer_ip_next(&cursor);
3451 * If all is ok we have to get the inode so we can adjust nlinks.
3452 * To avoid a deadlock with the flusher we must release the inode
3453 * lock on the directory when acquiring the inode for the entry.
3455 * If the target is a directory, it must be empty.
3458 hammer_unlock(&cursor.ip->lock);
3459 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3461 cursor.data->entry.localization,
3463 hammer_lock_sh(&cursor.ip->lock);
3464 if (error == ENOENT) {
3465 kprintf("HAMMER: WARNING: Removing "
3466 "dirent w/missing inode \"%s\"\n"
3467 "\tobj_id = %016llx\n",
3469 (long long)cursor.data->entry.obj_id);
3474 * If isdir >= 0 we validate that the entry is or is not a
3475 * directory. If isdir < 0 we don't care.
3477 if (error == 0 && isdir >= 0 && ip) {
3479 ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3481 } else if (isdir == 0 &&
3482 ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3488 * If we are trying to remove a directory the directory must
3491 * The check directory code can loop and deadlock/retry. Our
3492 * own cursor's node locks must be released to avoid a 3-way
3493 * deadlock with the flusher if the check directory code
3496 * If any changes whatsoever have been made to the cursor
3497 * set EDEADLK and retry.
3499 * WARNING: See warnings in hammer_unlock_cursor()
3502 if (error == 0 && ip && ip->ino_data.obj_type ==
3503 HAMMER_OBJTYPE_DIRECTORY) {
3504 hammer_unlock_cursor(&cursor);
3505 error = hammer_ip_check_directory_empty(trans, ip);
3506 hammer_lock_cursor(&cursor);
3507 if (cursor.flags & HAMMER_CURSOR_RETEST) {
3508 kprintf("HAMMER: Warning: avoided deadlock "
3516 * Delete the directory entry.
3518 * WARNING: hammer_ip_del_directory() may have to terminate
3519 * the cursor to avoid a deadlock. It is ok to call
3520 * hammer_done_cursor() twice.
3523 error = hammer_ip_del_directory(trans, &cursor,
3526 hammer_done_cursor(&cursor);
3529 * Tell the namecache that we are now unlinked.
3534 * NOTE: ip->vp, if non-NULL, cannot be directly
3535 * referenced without formally acquiring the
3536 * vp since the vp might have zero refs on it,
3537 * or in the middle of a reclaim, etc.
3539 * NOTE: The cache_setunresolved() can rip the vp
3540 * out from under us since the vp may not have
3541 * any refs, in which case ip->vp will be NULL
3544 while (ip && ip->vp) {
3547 error = hammer_get_vnode(ip, &vp);
3548 if (error == 0 && vp) {
3550 hammer_knote(ip->vp, NOTE_DELETE);
3553 * Don't do this, it can deadlock
3554 * on concurrent rm's of hardlinks.
3555 * Shouldn't be needed any more.
3557 cache_inval_vp(ip->vp, CINV_DESTROY);
3562 kprintf("Debug: HAMMER ip/vp race1 avoided\n");
3566 hammer_rel_inode(ip, 0);
3568 hammer_done_cursor(&cursor);
3570 if (error == EDEADLK)
3576 /************************************************************************
3577 * FIFO AND SPECFS OPS *
3578 ************************************************************************
3582 hammer_vop_fifoclose (struct vop_close_args *ap)
3584 /* XXX update itimes */
3585 return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3589 hammer_vop_fiforead (struct vop_read_args *ap)
3593 error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3594 /* XXX update access time */
3599 hammer_vop_fifowrite (struct vop_write_args *ap)
3603 error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3604 /* XXX update access time */
3610 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3614 error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3616 error = hammer_vop_kqfilter(ap);
3620 /************************************************************************
3622 ************************************************************************
3625 static void filt_hammerdetach(struct knote *kn);
3626 static int filt_hammerread(struct knote *kn, long hint);
3627 static int filt_hammerwrite(struct knote *kn, long hint);
3628 static int filt_hammervnode(struct knote *kn, long hint);
3630 static struct filterops hammerread_filtops =
3631 { FILTEROP_ISFD | FILTEROP_MPSAFE,
3632 NULL, filt_hammerdetach, filt_hammerread };
3633 static struct filterops hammerwrite_filtops =
3634 { FILTEROP_ISFD | FILTEROP_MPSAFE,
3635 NULL, filt_hammerdetach, filt_hammerwrite };
3636 static struct filterops hammervnode_filtops =
3637 { FILTEROP_ISFD | FILTEROP_MPSAFE,
3638 NULL, filt_hammerdetach, filt_hammervnode };
3642 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3644 struct vnode *vp = ap->a_vp;
3645 struct knote *kn = ap->a_kn;
3647 switch (kn->kn_filter) {
3649 kn->kn_fop = &hammerread_filtops;
3652 kn->kn_fop = &hammerwrite_filtops;
3655 kn->kn_fop = &hammervnode_filtops;
3658 return (EOPNOTSUPP);
3661 kn->kn_hook = (caddr_t)vp;
3663 knote_insert(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3669 filt_hammerdetach(struct knote *kn)
3671 struct vnode *vp = (void *)kn->kn_hook;
3673 knote_remove(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3677 filt_hammerread(struct knote *kn, long hint)
3679 struct vnode *vp = (void *)kn->kn_hook;
3680 hammer_inode_t ip = VTOI(vp);
3681 hammer_mount_t hmp = ip->hmp;
3684 if (hint == NOTE_REVOKE) {
3685 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3688 lwkt_gettoken(&hmp->fs_token); /* XXX use per-ip-token */
3689 off = ip->ino_data.size - kn->kn_fp->f_offset;
3690 kn->kn_data = (off < INTPTR_MAX) ? off : INTPTR_MAX;
3691 lwkt_reltoken(&hmp->fs_token);
3692 if (kn->kn_sfflags & NOTE_OLDAPI)
3694 return (kn->kn_data != 0);
3698 filt_hammerwrite(struct knote *kn, long hint)
3700 if (hint == NOTE_REVOKE)
3701 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3707 filt_hammervnode(struct knote *kn, long hint)
3709 if (kn->kn_sfflags & hint)
3710 kn->kn_fflags |= hint;
3711 if (hint == NOTE_REVOKE) {
3712 kn->kn_flags |= (EV_EOF | EV_NODATA);
3715 return (kn->kn_fflags != 0);