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/mountctl.h>
36 #include <sys/namecache.h>
38 #include <vfs/fifofs/fifo.h>
45 static int hammer_vop_fsync(struct vop_fsync_args *);
46 static int hammer_vop_read(struct vop_read_args *);
47 static int hammer_vop_write(struct vop_write_args *);
48 static int hammer_vop_access(struct vop_access_args *);
49 static int hammer_vop_advlock(struct vop_advlock_args *);
50 static int hammer_vop_close(struct vop_close_args *);
51 static int hammer_vop_ncreate(struct vop_ncreate_args *);
52 static int hammer_vop_getattr(struct vop_getattr_args *);
53 static int hammer_vop_nresolve(struct vop_nresolve_args *);
54 static int hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *);
55 static int hammer_vop_nlink(struct vop_nlink_args *);
56 static int hammer_vop_nmkdir(struct vop_nmkdir_args *);
57 static int hammer_vop_nmknod(struct vop_nmknod_args *);
58 static int hammer_vop_open(struct vop_open_args *);
59 static int hammer_vop_print(struct vop_print_args *);
60 static int hammer_vop_readdir(struct vop_readdir_args *);
61 static int hammer_vop_readlink(struct vop_readlink_args *);
62 static int hammer_vop_nremove(struct vop_nremove_args *);
63 static int hammer_vop_nrename(struct vop_nrename_args *);
64 static int hammer_vop_nrmdir(struct vop_nrmdir_args *);
65 static int hammer_vop_markatime(struct vop_markatime_args *);
66 static int hammer_vop_setattr(struct vop_setattr_args *);
67 static int hammer_vop_strategy(struct vop_strategy_args *);
68 static int hammer_vop_bmap(struct vop_bmap_args *ap);
69 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
70 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
71 static int hammer_vop_ioctl(struct vop_ioctl_args *);
72 static int hammer_vop_mountctl(struct vop_mountctl_args *);
73 static int hammer_vop_kqfilter (struct vop_kqfilter_args *);
75 static int hammer_vop_fifoclose (struct vop_close_args *);
76 static int hammer_vop_fiforead (struct vop_read_args *);
77 static int hammer_vop_fifowrite (struct vop_write_args *);
78 static int hammer_vop_fifokqfilter (struct vop_kqfilter_args *);
80 struct vop_ops hammer_vnode_vops = {
81 .vop_default = vop_defaultop,
82 .vop_fsync = hammer_vop_fsync,
83 .vop_getpages = vop_stdgetpages,
84 .vop_putpages = vop_stdputpages,
85 .vop_read = hammer_vop_read,
86 .vop_write = hammer_vop_write,
87 .vop_access = hammer_vop_access,
88 .vop_advlock = hammer_vop_advlock,
89 .vop_close = hammer_vop_close,
90 .vop_ncreate = hammer_vop_ncreate,
91 .vop_getattr = hammer_vop_getattr,
92 .vop_inactive = hammer_vop_inactive,
93 .vop_reclaim = hammer_vop_reclaim,
94 .vop_nresolve = hammer_vop_nresolve,
95 .vop_nlookupdotdot = hammer_vop_nlookupdotdot,
96 .vop_nlink = hammer_vop_nlink,
97 .vop_nmkdir = hammer_vop_nmkdir,
98 .vop_nmknod = hammer_vop_nmknod,
99 .vop_open = hammer_vop_open,
100 .vop_pathconf = vop_stdpathconf,
101 .vop_print = hammer_vop_print,
102 .vop_readdir = hammer_vop_readdir,
103 .vop_readlink = hammer_vop_readlink,
104 .vop_nremove = hammer_vop_nremove,
105 .vop_nrename = hammer_vop_nrename,
106 .vop_nrmdir = hammer_vop_nrmdir,
107 .vop_markatime = hammer_vop_markatime,
108 .vop_setattr = hammer_vop_setattr,
109 .vop_bmap = hammer_vop_bmap,
110 .vop_strategy = hammer_vop_strategy,
111 .vop_nsymlink = hammer_vop_nsymlink,
112 .vop_nwhiteout = hammer_vop_nwhiteout,
113 .vop_ioctl = hammer_vop_ioctl,
114 .vop_mountctl = hammer_vop_mountctl,
115 .vop_kqfilter = hammer_vop_kqfilter
118 struct vop_ops hammer_spec_vops = {
119 .vop_default = vop_defaultop,
120 .vop_fsync = hammer_vop_fsync,
121 .vop_read = vop_stdnoread,
122 .vop_write = vop_stdnowrite,
123 .vop_access = hammer_vop_access,
124 .vop_close = hammer_vop_close,
125 .vop_markatime = hammer_vop_markatime,
126 .vop_getattr = hammer_vop_getattr,
127 .vop_inactive = hammer_vop_inactive,
128 .vop_reclaim = hammer_vop_reclaim,
129 .vop_setattr = hammer_vop_setattr
132 struct vop_ops hammer_fifo_vops = {
133 .vop_default = fifo_vnoperate,
134 .vop_fsync = hammer_vop_fsync,
135 .vop_read = hammer_vop_fiforead,
136 .vop_write = hammer_vop_fifowrite,
137 .vop_access = hammer_vop_access,
138 .vop_close = hammer_vop_fifoclose,
139 .vop_markatime = hammer_vop_markatime,
140 .vop_getattr = hammer_vop_getattr,
141 .vop_inactive = hammer_vop_inactive,
142 .vop_reclaim = hammer_vop_reclaim,
143 .vop_setattr = hammer_vop_setattr,
144 .vop_kqfilter = hammer_vop_fifokqfilter
149 hammer_knote(struct vnode *vp, int flags)
152 KNOTE(&vp->v_pollinfo.vpi_kqinfo.ki_note, flags);
155 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
156 struct vnode *dvp, struct ucred *cred,
157 int flags, int isdir);
158 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
159 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
162 * hammer_vop_fsync { vp, waitfor }
164 * fsync() an inode to disk and wait for it to be completely committed
165 * such that the information would not be undone if a crash occured after
168 * NOTE: HAMMER's fsync()'s are going to remain expensive until we implement
169 * a REDO log. A sysctl is provided to relax HAMMER's fsync()
172 * Ultimately the combination of a REDO log and use of fast storage
173 * to front-end cluster caches will make fsync fast, but it aint
174 * here yet. And, in anycase, we need real transactional
175 * all-or-nothing features which are not restricted to a single file.
179 hammer_vop_fsync(struct vop_fsync_args *ap)
181 hammer_inode_t ip = VTOI(ap->a_vp);
182 hammer_mount_t hmp = ip->hmp;
183 int waitfor = ap->a_waitfor;
186 lwkt_gettoken(&hmp->fs_token);
189 * Fsync rule relaxation (default is either full synchronous flush
190 * or REDO semantics with synchronous flush).
192 if (ap->a_flags & VOP_FSYNC_SYSCALL) {
193 switch(hammer_fsync_mode) {
196 /* no REDO, full synchronous flush */
200 /* no REDO, full asynchronous flush */
201 if (waitfor == MNT_WAIT)
202 waitfor = MNT_NOWAIT;
205 /* REDO semantics, synchronous flush */
206 if (hmp->version < HAMMER_VOL_VERSION_FOUR)
208 mode = HAMMER_FLUSH_UNDOS_AUTO;
211 /* REDO semantics, relaxed asynchronous flush */
212 if (hmp->version < HAMMER_VOL_VERSION_FOUR)
214 mode = HAMMER_FLUSH_UNDOS_RELAXED;
215 if (waitfor == MNT_WAIT)
216 waitfor = MNT_NOWAIT;
219 /* ignore the fsync() system call */
220 lwkt_reltoken(&hmp->fs_token);
223 /* we have to do something */
224 mode = HAMMER_FLUSH_UNDOS_RELAXED;
225 if (waitfor == MNT_WAIT)
226 waitfor = MNT_NOWAIT;
231 * Fast fsync only needs to flush the UNDO/REDO fifo if
232 * HAMMER_INODE_REDO is non-zero and the only modifications
233 * made to the file are write or write-extends.
235 if ((ip->flags & HAMMER_INODE_REDO) &&
236 (ip->flags & HAMMER_INODE_MODMASK_NOREDO) == 0) {
237 ++hammer_count_fsyncs;
238 hammer_flusher_flush_undos(hmp, mode);
240 if (ip->vp && (ip->flags & HAMMER_INODE_MODMASK) == 0)
242 lwkt_reltoken(&hmp->fs_token);
247 * REDO is enabled by fsync(), the idea being we really only
248 * want to lay down REDO records when programs are using
249 * fsync() heavily. The first fsync() on the file starts
250 * the gravy train going and later fsync()s keep it hot by
251 * resetting the redo_count.
253 * We weren't running REDOs before now so we have to fall
254 * through and do a full fsync of what we have.
256 if (hmp->version >= HAMMER_VOL_VERSION_FOUR &&
257 (hmp->flags & HAMMER_MOUNT_REDO_RECOVERY_RUN) == 0) {
258 ip->flags |= HAMMER_INODE_REDO;
265 * Do a full flush sequence.
267 * Attempt to release the vnode while waiting for the inode to
268 * finish flushing. This can really mess up inactive->reclaim
269 * sequences so only do it if the vnode is active.
271 * WARNING! The VX lock functions must be used. vn_lock() will
272 * fail when this is part of a VOP_RECLAIM sequence.
274 ++hammer_count_fsyncs;
275 vfsync(ap->a_vp, waitfor, 1, NULL, NULL);
276 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
277 if (waitfor == MNT_WAIT) {
280 if ((ap->a_vp->v_flag & VRECLAIMED) == 0) {
286 hammer_wait_inode(ip);
290 if (ip->vp && (ip->flags & HAMMER_INODE_MODMASK) == 0)
292 lwkt_reltoken(&hmp->fs_token);
297 * hammer_vop_read { vp, uio, ioflag, cred }
299 * MPSAFE (for the cache safe does not require fs_token)
303 hammer_vop_read(struct vop_read_args *ap)
305 struct hammer_transaction trans;
320 if (ap->a_vp->v_type != VREG)
329 * Attempt to shortcut directly to the VM object using lwbufs.
330 * This is much faster than instantiating buffer cache buffers.
332 resid = uio->uio_resid;
333 error = vop_helper_read_shortcut(ap);
334 hammer_stats_file_read += resid - uio->uio_resid;
337 if (uio->uio_resid == 0)
341 * Allow the UIO's size to override the sequential heuristic.
343 blksize = hammer_blocksize(uio->uio_offset);
344 seqcount = (uio->uio_resid + (MAXBSIZE - 1)) / MAXBSIZE;
345 ioseqcount = (ap->a_ioflag >> 16);
346 if (seqcount < ioseqcount)
347 seqcount = ioseqcount;
350 * If reading or writing a huge amount of data we have to break
351 * atomicy and allow the operation to be interrupted by a signal
352 * or it can DOS the machine.
354 bigread = (uio->uio_resid > 100 * 1024 * 1024);
357 * Access the data typically in HAMMER_BUFSIZE blocks via the
358 * buffer cache, but HAMMER may use a variable block size based
361 * XXX Temporary hack, delay the start transaction while we remain
362 * MPSAFE. NOTE: ino_data.size cannot change while vnode is
365 while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
369 blksize = hammer_blocksize(uio->uio_offset);
370 offset = (int)uio->uio_offset & (blksize - 1);
371 base_offset = uio->uio_offset - offset;
373 if (bigread && (error = hammer_signal_check(ip->hmp)) != 0)
379 bp = getblk(ap->a_vp, base_offset, blksize, 0, 0);
380 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == B_CACHE) {
381 bp->b_flags &= ~B_AGE;
385 if (ap->a_ioflag & IO_NRDELAY) {
387 return (EWOULDBLOCK);
393 if (got_trans == 0) {
394 hammer_start_transaction(&trans, ip->hmp);
399 * NOTE: A valid bp has already been acquired, but was not
402 if (hammer_cluster_enable) {
404 * Use file_limit to prevent cluster_read() from
405 * creating buffers of the wrong block size past
408 file_limit = ip->ino_data.size;
409 if (base_offset < HAMMER_XDEMARC &&
410 file_limit > HAMMER_XDEMARC) {
411 file_limit = HAMMER_XDEMARC;
413 error = cluster_readx(ap->a_vp,
414 file_limit, base_offset,
420 error = breadnx(ap->a_vp, base_offset,
429 if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IOISSUED)) {
430 hdkprintf("zone2_offset %016jx read file %016jx@%016jx\n",
431 (intmax_t)bp->b_bio2.bio_offset,
432 (intmax_t)ip->obj_id,
433 (intmax_t)bp->b_loffset);
435 bp->b_flags &= ~B_IOISSUED;
436 if (blksize == HAMMER_XBUFSIZE)
437 bp->b_flags |= B_CLUSTEROK;
439 n = blksize - offset;
440 if (n > uio->uio_resid)
442 if (n > ip->ino_data.size - uio->uio_offset)
443 n = (int)(ip->ino_data.size - uio->uio_offset);
446 * Set B_AGE, data has a lower priority than meta-data.
448 * Use a hold/unlock/drop sequence to run the uiomove
449 * with the buffer unlocked, avoiding deadlocks against
450 * read()s on mmap()'d spaces.
452 bp->b_flags |= B_AGE;
453 error = uiomovebp(bp, (char *)bp->b_data + offset, n, uio);
458 hammer_stats_file_read += n;
464 * Try to update the atime with just the inode lock for maximum
465 * concurrency. If we can't shortcut it we have to get the full
468 if (got_trans == 0 && hammer_update_atime_quick(ip) < 0) {
469 hammer_start_transaction(&trans, ip->hmp);
474 if ((ip->flags & HAMMER_INODE_RO) == 0 &&
475 (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
476 lwkt_gettoken(&hmp->fs_token);
477 ip->ino_data.atime = trans.time;
478 hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
479 hammer_done_transaction(&trans);
480 lwkt_reltoken(&hmp->fs_token);
482 hammer_done_transaction(&trans);
489 * hammer_vop_write { vp, uio, ioflag, cred }
493 hammer_vop_write(struct vop_write_args *ap)
495 struct hammer_transaction trans;
513 if (vp->v_type != VREG)
519 seqcount = ap->a_ioflag >> 16;
521 if (ip->flags & HAMMER_INODE_RO)
525 * Create a transaction to cover the operations we perform.
527 hammer_start_transaction(&trans, hmp);
531 * Use v_lastwrite_ts if file not open for writing
532 * (i.e. a late msync)
534 if (uio->uio_segflg == UIO_NOCOPY) {
535 if (vp->v_flag & VLASTWRITETS) {
536 trans.time = vp->v_lastwrite_ts.tv_sec * 1000000 +
537 vp->v_lastwrite_ts.tv_nsec / 1000;
539 trans.time = ip->ino_data.mtime;
542 vclrflags(vp, VLASTWRITETS);
548 if (ap->a_ioflag & IO_APPEND)
549 uio->uio_offset = ip->ino_data.size;
552 * Check for illegal write offsets. Valid range is 0...2^63-1.
554 * NOTE: the base_off assignment is required to work around what
555 * I consider to be a GCC-4 optimization bug.
557 if (uio->uio_offset < 0) {
558 hammer_done_transaction(&trans);
561 base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
562 if (uio->uio_resid > 0 && base_offset <= uio->uio_offset) {
563 hammer_done_transaction(&trans);
567 if (uio->uio_resid > 0 && (td = uio->uio_td) != NULL && td->td_proc &&
568 base_offset > td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
569 hammer_done_transaction(&trans);
570 lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ);
575 * If reading or writing a huge amount of data we have to break
576 * atomicy and allow the operation to be interrupted by a signal
577 * or it can DOS the machine.
579 * Preset redo_count so we stop generating REDOs earlier if the
582 * redo_count is heuristical, SMP races are ok
584 bigwrite = (uio->uio_resid > 100 * 1024 * 1024);
585 if ((ip->flags & HAMMER_INODE_REDO) &&
586 ip->redo_count < hammer_limit_redo) {
587 ip->redo_count += uio->uio_resid;
591 * Access the data typically in HAMMER_BUFSIZE blocks via the
592 * buffer cache, but HAMMER may use a variable block size based
595 while (uio->uio_resid > 0) {
603 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
605 if (bigwrite && (error = hammer_signal_check(hmp)) != 0)
608 blksize = hammer_blocksize(uio->uio_offset);
611 * Control the number of pending records associated with
612 * this inode. If too many have accumulated start a
613 * flush. Try to maintain a pipeline with the flusher.
615 * NOTE: It is possible for other sources to grow the
616 * records but not necessarily issue another flush,
617 * so use a timeout and ensure that a re-flush occurs.
619 if (ip->rsv_recs >= hammer_limit_inode_recs) {
620 lwkt_gettoken(&hmp->fs_token);
621 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
622 while (ip->rsv_recs >= hammer_limit_inode_recs * 2) {
623 ip->flags |= HAMMER_INODE_RECSW;
624 tsleep(&ip->rsv_recs, 0, "hmrwww", hz);
625 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
627 lwkt_reltoken(&hmp->fs_token);
631 * Do not allow HAMMER to blow out the buffer cache. Very
632 * large UIOs can lockout other processes due to bwillwrite()
635 * The hammer inode is not locked during these operations.
636 * The vnode is locked which can interfere with the pageout
637 * daemon for non-UIO_NOCOPY writes but should not interfere
638 * with the buffer cache. Even so, we cannot afford to
639 * allow the pageout daemon to build up too many dirty buffer
642 * Only call this if we aren't being recursively called from
643 * a virtual disk device (vn), else we may deadlock.
645 if ((ap->a_ioflag & IO_RECURSE) == 0)
649 * Calculate the blocksize at the current offset and figure
650 * out how much we can actually write.
652 blkmask = blksize - 1;
653 offset = (int)uio->uio_offset & blkmask;
654 base_offset = uio->uio_offset & ~(int64_t)blkmask;
655 n = blksize - offset;
656 if (n > uio->uio_resid) {
662 nsize = uio->uio_offset + n;
663 if (nsize > ip->ino_data.size) {
664 if (uio->uio_offset > ip->ino_data.size)
668 nvextendbuf(ap->a_vp,
671 hammer_blocksize(ip->ino_data.size),
672 hammer_blocksize(nsize),
673 hammer_blockoff(ip->ino_data.size),
674 hammer_blockoff(nsize),
677 kflags |= NOTE_EXTEND;
680 if (uio->uio_segflg == UIO_NOCOPY) {
682 * Issuing a write with the same data backing the
683 * buffer. Instantiate the buffer to collect the
684 * backing vm pages, then read-in any missing bits.
686 * This case is used by vop_stdputpages().
688 bp = getblk(ap->a_vp, base_offset,
689 blksize, GETBLK_BHEAVY, 0);
690 if ((bp->b_flags & B_CACHE) == 0) {
692 error = bread(ap->a_vp, base_offset,
695 } else if (offset == 0 && uio->uio_resid >= blksize) {
697 * Even though we are entirely overwriting the buffer
698 * we may still have to zero it out to avoid a
699 * mmap/write visibility issue.
701 bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
702 if ((bp->b_flags & B_CACHE) == 0)
704 } else if (base_offset >= ip->ino_data.size) {
706 * If the base offset of the buffer is beyond the
707 * file EOF, we don't have to issue a read.
709 bp = getblk(ap->a_vp, base_offset,
710 blksize, GETBLK_BHEAVY, 0);
714 * Partial overwrite, read in any missing bits then
715 * replace the portion being written.
717 error = bread(ap->a_vp, base_offset, blksize, &bp);
722 error = uiomovebp(bp, bp->b_data + offset, n, uio);
724 lwkt_gettoken(&hmp->fs_token);
727 * Generate REDO records if enabled and redo_count will not
728 * exceeded the limit.
730 * If redo_count exceeds the limit we stop generating records
731 * and clear HAMMER_INODE_REDO. This will cause the next
732 * fsync() to do a full meta-data sync instead of just an
733 * UNDO/REDO fifo update.
735 * When clearing HAMMER_INODE_REDO any pre-existing REDOs
736 * will still be tracked. The tracks will be terminated
737 * when the related meta-data (including possible data
738 * modifications which are not tracked via REDO) is
741 if ((ip->flags & HAMMER_INODE_REDO) && error == 0) {
742 if (ip->redo_count < hammer_limit_redo) {
743 bp->b_flags |= B_VFSFLAG1;
744 error = hammer_generate_redo(&trans, ip,
745 base_offset + offset,
750 ip->flags &= ~HAMMER_INODE_REDO;
755 * If we screwed up we have to undo any VM size changes we
761 nvtruncbuf(ap->a_vp, ip->ino_data.size,
762 hammer_blocksize(ip->ino_data.size),
763 hammer_blockoff(ip->ino_data.size),
766 lwkt_reltoken(&hmp->fs_token);
769 kflags |= NOTE_WRITE;
770 hammer_stats_file_write += n;
771 if (blksize == HAMMER_XBUFSIZE)
772 bp->b_flags |= B_CLUSTEROK;
773 if (ip->ino_data.size < uio->uio_offset) {
774 ip->ino_data.size = uio->uio_offset;
775 flags = HAMMER_INODE_SDIRTY;
779 ip->ino_data.mtime = trans.time;
780 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
781 hammer_modify_inode(&trans, ip, flags);
784 * Once we dirty the buffer any cached zone-X offset
785 * becomes invalid. HAMMER NOTE: no-history mode cannot
786 * allow overwriting over the same data sector unless
787 * we provide UNDOs for the old data, which we don't.
789 bp->b_bio2.bio_offset = NOOFFSET;
791 lwkt_reltoken(&hmp->fs_token);
794 * Final buffer disposition.
796 * Because meta-data updates are deferred, HAMMER is
797 * especially sensitive to excessive bdwrite()s because
798 * the I/O stream is not broken up by disk reads. So the
799 * buffer cache simply cannot keep up.
801 * WARNING! blksize is variable. cluster_write() is
802 * expected to not blow up if it encounters
803 * buffers that do not match the passed blksize.
805 * NOTE! Hammer shouldn't need to bawrite()/cluster_write().
806 * The ip->rsv_recs check should burst-flush the data.
807 * If we queue it immediately the buf could be left
808 * locked on the device queue for a very long time.
810 * However, failing to flush a dirty buffer out when
811 * issued from the pageout daemon can result in a low
812 * memory deadlock against bio_page_alloc(), so we
813 * have to bawrite() on IO_ASYNC as well.
815 * NOTE! To avoid degenerate stalls due to mismatched block
816 * sizes we only honor IO_DIRECT on the write which
817 * abuts the end of the buffer. However, we must
818 * honor IO_SYNC in case someone is silly enough to
819 * configure a HAMMER file as swap, or when HAMMER
820 * is serving NFS (for commits). Ick ick.
822 bp->b_flags |= B_AGE;
823 if (blksize == HAMMER_XBUFSIZE)
824 bp->b_flags |= B_CLUSTEROK;
826 if (ap->a_ioflag & IO_SYNC) {
828 } else if ((ap->a_ioflag & IO_DIRECT) && endofblk) {
830 } else if (ap->a_ioflag & IO_ASYNC) {
832 } else if (hammer_cluster_enable &&
833 !(ap->a_vp->v_mount->mnt_flag & MNT_NOCLUSTERW)) {
834 if (base_offset < HAMMER_XDEMARC)
835 cluster_eof = hammer_blockdemarc(base_offset,
838 cluster_eof = ip->ino_data.size;
839 cluster_write(bp, cluster_eof, blksize, seqcount);
844 hammer_done_transaction(&trans);
845 hammer_knote(ap->a_vp, kflags);
851 * hammer_vop_access { vp, mode, cred }
853 * MPSAFE - does not require fs_token
857 hammer_vop_access(struct vop_access_args *ap)
859 hammer_inode_t ip = VTOI(ap->a_vp);
864 uid = hammer_to_unix_xid(&ip->ino_data.uid);
865 gid = hammer_to_unix_xid(&ip->ino_data.gid);
867 error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
868 ip->ino_data.uflags);
873 * hammer_vop_advlock { vp, id, op, fl, flags }
875 * MPSAFE - does not require fs_token
879 hammer_vop_advlock(struct vop_advlock_args *ap)
881 hammer_inode_t ip = VTOI(ap->a_vp);
883 return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
887 * hammer_vop_close { vp, fflag }
889 * We can only sync-on-close for normal closes. XXX disabled for now.
893 hammer_vop_close(struct vop_close_args *ap)
896 struct vnode *vp = ap->a_vp;
897 hammer_inode_t ip = VTOI(vp);
899 if (ip->flags & (HAMMER_INODE_CLOSESYNC|HAMMER_INODE_CLOSEASYNC)) {
900 if (vn_islocked(vp) == LK_EXCLUSIVE &&
901 (vp->v_flag & (VINACTIVE|VRECLAIMED)) == 0) {
902 if (ip->flags & HAMMER_INODE_CLOSESYNC)
905 waitfor = MNT_NOWAIT;
906 ip->flags &= ~(HAMMER_INODE_CLOSESYNC |
907 HAMMER_INODE_CLOSEASYNC);
908 VOP_FSYNC(vp, MNT_NOWAIT, waitfor);
912 return (vop_stdclose(ap));
916 * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
918 * The operating system has already ensured that the directory entry
919 * does not exist and done all appropriate namespace locking.
923 hammer_vop_ncreate(struct vop_ncreate_args *ap)
925 struct hammer_transaction trans;
928 struct nchandle *nch;
933 dip = VTOI(ap->a_dvp);
936 if (dip->flags & HAMMER_INODE_RO)
938 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
942 * Create a transaction to cover the operations we perform.
944 lwkt_gettoken(&hmp->fs_token);
945 hammer_start_transaction(&trans, hmp);
948 * Create a new filesystem object of the requested type. The
949 * returned inode will be referenced and shared-locked to prevent
950 * it from being moved to the flusher.
952 error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
953 dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
956 hkprintf("hammer_create_inode error %d\n", error);
957 hammer_done_transaction(&trans);
959 lwkt_reltoken(&hmp->fs_token);
964 * Add the new filesystem object to the directory. This will also
965 * bump the inode's link count.
967 error = hammer_ip_add_direntry(&trans, dip,
968 nch->ncp->nc_name, nch->ncp->nc_nlen,
971 hkprintf("hammer_ip_add_direntry error %d\n", error);
977 hammer_rel_inode(nip, 0);
978 hammer_done_transaction(&trans);
981 error = hammer_get_vnode(nip, ap->a_vpp);
982 hammer_done_transaction(&trans);
983 hammer_rel_inode(nip, 0);
985 cache_setunresolved(ap->a_nch);
986 cache_setvp(ap->a_nch, *ap->a_vpp);
988 hammer_knote(ap->a_dvp, NOTE_WRITE);
990 lwkt_reltoken(&hmp->fs_token);
995 * hammer_vop_getattr { vp, vap }
997 * Retrieve an inode's attribute information. When accessing inodes
998 * historically we fake the atime field to ensure consistent results.
999 * The atime field is stored in the B-Tree element and allowed to be
1000 * updated without cycling the element.
1002 * MPSAFE - does not require fs_token
1006 hammer_vop_getattr(struct vop_getattr_args *ap)
1008 hammer_inode_t ip = VTOI(ap->a_vp);
1009 struct vattr *vap = ap->a_vap;
1012 * We want the fsid to be different when accessing a filesystem
1013 * with different as-of's so programs like diff don't think
1014 * the files are the same.
1016 * We also want the fsid to be the same when comparing snapshots,
1017 * or when comparing mirrors (which might be backed by different
1018 * physical devices). HAMMER fsids are based on the PFS's
1019 * shared_uuid field.
1021 * XXX there is a chance of collision here. The va_fsid reported
1022 * by stat is different from the more involved fsid used in the
1025 hammer_lock_sh(&ip->lock);
1026 vap->va_fsid = ip->pfsm->fsid_udev ^ (uint32_t)ip->obj_asof ^
1027 (uint32_t)(ip->obj_asof >> 32);
1029 vap->va_fileid = ip->ino_leaf.base.obj_id;
1030 vap->va_mode = ip->ino_data.mode;
1031 vap->va_nlink = ip->ino_data.nlinks;
1032 vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1033 vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1036 vap->va_size = ip->ino_data.size;
1039 * Special case for @@PFS softlinks. The actual size of the
1040 * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
1041 * or for MAX_TID is "@@-1:%05d" == 10 bytes.
1043 * Note that userspace hammer command does not allow users to
1044 * create a @@PFS softlink under an existing other PFS (id!=0)
1045 * so the ip localization here for @@PFS softlink is always 0.
1047 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
1048 ip->ino_data.size == 10 &&
1049 ip->obj_asof == HAMMER_MAX_TID &&
1050 ip->obj_localization == HAMMER_DEF_LOCALIZATION &&
1051 strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
1052 if (hammer_is_pfs_slave(&ip->pfsm->pfsd))
1059 * We must provide a consistent atime and mtime for snapshots
1060 * so people can do a 'tar cf - ... | md5' on them and get
1061 * consistent results.
1063 if (ip->flags & HAMMER_INODE_RO) {
1064 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
1065 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
1067 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
1068 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
1070 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
1071 vap->va_flags = ip->ino_data.uflags;
1072 vap->va_gen = 1; /* hammer inums are unique for all time */
1073 vap->va_blocksize = HAMMER_BUFSIZE;
1074 if (ip->ino_data.size >= HAMMER_XDEMARC) {
1075 vap->va_bytes = HAMMER_XBUFSIZE64_DOALIGN(ip->ino_data.size);
1076 } else if (ip->ino_data.size > HAMMER_HBUFSIZE) {
1077 vap->va_bytes = HAMMER_BUFSIZE64_DOALIGN(ip->ino_data.size);
1079 vap->va_bytes = HAMMER_DATA_DOALIGN(ip->ino_data.size);
1082 vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
1083 vap->va_filerev = 0; /* XXX */
1084 vap->va_uid_uuid = ip->ino_data.uid;
1085 vap->va_gid_uuid = ip->ino_data.gid;
1086 vap->va_fsid_uuid = ip->hmp->fsid;
1087 vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
1090 switch (ip->ino_data.obj_type) {
1091 case HAMMER_OBJTYPE_CDEV:
1092 case HAMMER_OBJTYPE_BDEV:
1093 vap->va_rmajor = ip->ino_data.rmajor;
1094 vap->va_rminor = ip->ino_data.rminor;
1099 hammer_unlock(&ip->lock);
1104 * hammer_vop_nresolve { nch, dvp, cred }
1106 * Locate the requested directory entry.
1110 hammer_vop_nresolve(struct vop_nresolve_args *ap)
1112 struct hammer_transaction trans;
1113 struct namecache *ncp;
1118 struct hammer_cursor cursor;
1127 uint32_t localization;
1128 uint32_t max_iterations;
1131 * Misc initialization, plus handle as-of name extensions. Look for
1132 * the '@@' extension. Note that as-of files and directories cannot
1135 dip = VTOI(ap->a_dvp);
1136 ncp = ap->a_nch->ncp;
1137 asof = dip->obj_asof;
1138 localization = dip->obj_localization; /* for code consistency */
1139 nlen = ncp->nc_nlen;
1140 flags = dip->flags & HAMMER_INODE_RO;
1144 lwkt_gettoken(&hmp->fs_token);
1145 hammer_simple_transaction(&trans, hmp);
1147 for (i = 0; i < nlen; ++i) {
1148 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
1149 error = hammer_str_to_tid(ncp->nc_name + i + 2,
1150 &ispfs, &asof, &localization);
1155 if (asof != HAMMER_MAX_TID)
1156 flags |= HAMMER_INODE_RO;
1163 * If this is a PFS we dive into the PFS root inode
1165 if (ispfs && nlen == 0) {
1166 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1170 error = hammer_get_vnode(ip, &vp);
1171 hammer_rel_inode(ip, 0);
1177 cache_setvp(ap->a_nch, vp);
1184 * If there is no path component the time extension is relative to dip.
1185 * e.g. "fubar/@@<snapshot>"
1187 * "." is handled by the kernel, but ".@@<snapshot>" is not.
1188 * e.g. "fubar/.@@<snapshot>"
1190 * ".." is handled by the kernel. We do not currently handle
1193 if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1194 ip = hammer_get_inode(&trans, dip, dip->obj_id,
1195 asof, dip->obj_localization,
1198 error = hammer_get_vnode(ip, &vp);
1199 hammer_rel_inode(ip, 0);
1205 cache_setvp(ap->a_nch, vp);
1212 * Calculate the namekey and setup the key range for the scan. This
1213 * works kinda like a chained hash table where the lower 32 bits
1214 * of the namekey synthesize the chain.
1216 * The key range is inclusive of both key_beg and key_end.
1218 namekey = hammer_direntry_namekey(dip, ncp->nc_name, nlen,
1221 error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1222 cursor.key_beg.localization = dip->obj_localization |
1223 hammer_dir_localization(dip);
1224 cursor.key_beg.obj_id = dip->obj_id;
1225 cursor.key_beg.key = namekey;
1226 cursor.key_beg.create_tid = 0;
1227 cursor.key_beg.delete_tid = 0;
1228 cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1229 cursor.key_beg.obj_type = 0;
1231 cursor.key_end = cursor.key_beg;
1232 cursor.key_end.key += max_iterations;
1234 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1237 * Scan all matching records (the chain), locate the one matching
1238 * the requested path component.
1240 * The hammer_ip_*() functions merge in-memory records with on-disk
1241 * records for the purposes of the search.
1244 localization = HAMMER_DEF_LOCALIZATION;
1247 error = hammer_ip_first(&cursor);
1248 while (error == 0) {
1249 error = hammer_ip_resolve_data(&cursor);
1252 if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1253 bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1254 obj_id = cursor.data->entry.obj_id;
1255 localization = cursor.data->entry.localization;
1258 error = hammer_ip_next(&cursor);
1261 hammer_done_cursor(&cursor);
1264 * Lookup the obj_id. This should always succeed. If it does not
1265 * the filesystem may be damaged and we return a dummy inode.
1268 ip = hammer_get_inode(&trans, dip, obj_id,
1271 if (error == ENOENT) {
1272 hkprintf("WARNING: Missing inode for dirent \"%s\"\n"
1273 "\tobj_id = %016jx, asof=%016jx, lo=%08x\n",
1275 (intmax_t)obj_id, (intmax_t)asof,
1278 ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1283 error = hammer_get_vnode(ip, &vp);
1284 hammer_rel_inode(ip, 0);
1290 cache_setvp(ap->a_nch, vp);
1293 } else if (error == ENOENT) {
1294 cache_setvp(ap->a_nch, NULL);
1297 hammer_done_transaction(&trans);
1298 lwkt_reltoken(&hmp->fs_token);
1303 * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1305 * Locate the parent directory of a directory vnode.
1307 * dvp is referenced but not locked. *vpp must be returned referenced and
1308 * locked. A parent_obj_id of 0 indicates that we are at the root.
1310 * NOTE: as-of sequences are not linked into the directory structure. If
1311 * we are at the root with a different asof then the mount point, reload
1312 * the same directory with the mount point's asof. I'm not sure what this
1313 * will do to NFS. We encode ASOF stamps in NFS file handles so it might not
1314 * get confused, but it hasn't been tested.
1318 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1320 struct hammer_transaction trans;
1324 int64_t parent_obj_id;
1325 uint32_t parent_obj_localization;
1329 dip = VTOI(ap->a_dvp);
1330 asof = dip->obj_asof;
1334 * Whos are parent? This could be the root of a pseudo-filesystem
1335 * whos parent is in another localization domain.
1337 lwkt_gettoken(&hmp->fs_token);
1338 parent_obj_id = dip->ino_data.parent_obj_id;
1339 if (dip->obj_id == HAMMER_OBJID_ROOT)
1340 parent_obj_localization = HAMMER_DEF_LOCALIZATION;
1342 parent_obj_localization = dip->obj_localization;
1345 * It's probably a PFS root when dip->ino_data.parent_obj_id is 0.
1347 if (parent_obj_id == 0) {
1348 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1349 asof != hmp->asof) {
1350 parent_obj_id = dip->obj_id;
1352 *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1353 ksnprintf(*ap->a_fakename, 19, "0x%016jx",
1354 (intmax_t)dip->obj_asof);
1357 lwkt_reltoken(&hmp->fs_token);
1362 hammer_simple_transaction(&trans, hmp);
1364 ip = hammer_get_inode(&trans, dip, parent_obj_id,
1365 asof, parent_obj_localization,
1366 dip->flags, &error);
1368 error = hammer_get_vnode(ip, ap->a_vpp);
1369 hammer_rel_inode(ip, 0);
1373 hammer_done_transaction(&trans);
1374 lwkt_reltoken(&hmp->fs_token);
1379 * hammer_vop_nlink { nch, dvp, vp, cred }
1383 hammer_vop_nlink(struct vop_nlink_args *ap)
1385 struct hammer_transaction trans;
1388 struct nchandle *nch;
1392 if (ap->a_dvp->v_mount != ap->a_vp->v_mount)
1396 dip = VTOI(ap->a_dvp);
1397 ip = VTOI(ap->a_vp);
1400 if (dip->obj_localization != ip->obj_localization)
1403 if (dip->flags & HAMMER_INODE_RO)
1405 if (ip->flags & HAMMER_INODE_RO)
1407 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1411 * Create a transaction to cover the operations we perform.
1413 lwkt_gettoken(&hmp->fs_token);
1414 hammer_start_transaction(&trans, hmp);
1417 * Add the filesystem object to the directory. Note that neither
1418 * dip nor ip are referenced or locked, but their vnodes are
1419 * referenced. This function will bump the inode's link count.
1421 error = hammer_ip_add_direntry(&trans, dip,
1422 nch->ncp->nc_name, nch->ncp->nc_nlen,
1429 cache_setunresolved(nch);
1430 cache_setvp(nch, ap->a_vp);
1432 hammer_done_transaction(&trans);
1433 hammer_knote(ap->a_vp, NOTE_LINK);
1434 hammer_knote(ap->a_dvp, NOTE_WRITE);
1435 lwkt_reltoken(&hmp->fs_token);
1440 * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1442 * The operating system has already ensured that the directory entry
1443 * does not exist and done all appropriate namespace locking.
1447 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1449 struct hammer_transaction trans;
1452 struct nchandle *nch;
1457 dip = VTOI(ap->a_dvp);
1460 if (dip->flags & HAMMER_INODE_RO)
1462 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1466 * Create a transaction to cover the operations we perform.
1468 lwkt_gettoken(&hmp->fs_token);
1469 hammer_start_transaction(&trans, hmp);
1472 * Create a new filesystem object of the requested type. The
1473 * returned inode will be referenced but not locked.
1475 error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1476 dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1479 hammer_done_transaction(&trans);
1481 lwkt_reltoken(&hmp->fs_token);
1485 * Add the new filesystem object to the directory. This will also
1486 * bump the inode's link count.
1488 error = hammer_ip_add_direntry(&trans, dip,
1489 nch->ncp->nc_name, nch->ncp->nc_nlen,
1492 hkprintf("hammer_mkdir (add) error %d\n", error);
1498 hammer_rel_inode(nip, 0);
1501 error = hammer_get_vnode(nip, ap->a_vpp);
1502 hammer_rel_inode(nip, 0);
1504 cache_setunresolved(ap->a_nch);
1505 cache_setvp(ap->a_nch, *ap->a_vpp);
1508 hammer_done_transaction(&trans);
1510 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1511 lwkt_reltoken(&hmp->fs_token);
1516 * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1518 * The operating system has already ensured that the directory entry
1519 * does not exist and done all appropriate namespace locking.
1523 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1525 struct hammer_transaction trans;
1528 struct nchandle *nch;
1533 dip = VTOI(ap->a_dvp);
1536 if (dip->flags & HAMMER_INODE_RO)
1538 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1542 * Create a transaction to cover the operations we perform.
1544 lwkt_gettoken(&hmp->fs_token);
1545 hammer_start_transaction(&trans, hmp);
1548 * Create a new filesystem object of the requested type. The
1549 * returned inode will be referenced but not locked.
1551 * If mknod specifies a directory a pseudo-fs is created.
1553 error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1554 dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1557 hammer_done_transaction(&trans);
1559 lwkt_reltoken(&hmp->fs_token);
1564 * Add the new filesystem object to the directory. This will also
1565 * bump the inode's link count.
1567 error = hammer_ip_add_direntry(&trans, dip,
1568 nch->ncp->nc_name, nch->ncp->nc_nlen,
1575 hammer_rel_inode(nip, 0);
1578 error = hammer_get_vnode(nip, ap->a_vpp);
1579 hammer_rel_inode(nip, 0);
1581 cache_setunresolved(ap->a_nch);
1582 cache_setvp(ap->a_nch, *ap->a_vpp);
1585 hammer_done_transaction(&trans);
1587 hammer_knote(ap->a_dvp, NOTE_WRITE);
1588 lwkt_reltoken(&hmp->fs_token);
1593 * hammer_vop_open { vp, mode, cred, fp }
1595 * MPSAFE (does not require fs_token)
1599 hammer_vop_open(struct vop_open_args *ap)
1603 ip = VTOI(ap->a_vp);
1605 if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1607 return(vop_stdopen(ap));
1611 * hammer_vop_print { vp }
1615 hammer_vop_print(struct vop_print_args *ap)
1621 * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1625 hammer_vop_readdir(struct vop_readdir_args *ap)
1627 struct hammer_transaction trans;
1628 struct hammer_cursor cursor;
1632 hammer_base_elm_t base;
1641 ip = VTOI(ap->a_vp);
1643 saveoff = uio->uio_offset;
1646 if (ap->a_ncookies) {
1647 ncookies = uio->uio_resid / 16 + 1;
1648 if (ncookies > 1024)
1650 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1658 lwkt_gettoken(&hmp->fs_token);
1659 hammer_simple_transaction(&trans, hmp);
1662 * Handle artificial entries
1664 * It should be noted that the minimum value for a directory
1665 * hash key on-media is 0x0000000100000000, so we can use anything
1666 * less then that to represent our 'special' key space.
1670 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1674 cookies[cookie_index] = saveoff;
1677 if (cookie_index == ncookies)
1681 if (ip->ino_data.parent_obj_id) {
1682 r = vop_write_dirent(&error, uio,
1683 ip->ino_data.parent_obj_id,
1686 r = vop_write_dirent(&error, uio,
1687 ip->obj_id, DT_DIR, 2, "..");
1692 cookies[cookie_index] = saveoff;
1695 if (cookie_index == ncookies)
1700 * Key range (begin and end inclusive) to scan. Directory keys
1701 * directly translate to a 64 bit 'seek' position.
1703 hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1704 cursor.key_beg.localization = ip->obj_localization |
1705 hammer_dir_localization(ip);
1706 cursor.key_beg.obj_id = ip->obj_id;
1707 cursor.key_beg.create_tid = 0;
1708 cursor.key_beg.delete_tid = 0;
1709 cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1710 cursor.key_beg.obj_type = 0;
1711 cursor.key_beg.key = saveoff;
1713 cursor.key_end = cursor.key_beg;
1714 cursor.key_end.key = HAMMER_MAX_KEY;
1715 cursor.asof = ip->obj_asof;
1716 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1718 error = hammer_ip_first(&cursor);
1720 while (error == 0) {
1721 error = hammer_ip_resolve_data(&cursor);
1724 base = &cursor.leaf->base;
1725 saveoff = base->key;
1726 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1728 if (base->obj_id != ip->obj_id)
1729 hpanic("bad record at %p", cursor.node);
1731 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1732 r = vop_write_dirent(
1733 &error, uio, cursor.data->entry.obj_id,
1735 cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1736 (void *)cursor.data->entry.name);
1741 cookies[cookie_index] = base->key;
1743 if (cookie_index == ncookies)
1745 error = hammer_ip_next(&cursor);
1747 hammer_done_cursor(&cursor);
1750 hammer_done_transaction(&trans);
1753 *ap->a_eofflag = (error == ENOENT);
1754 uio->uio_offset = saveoff;
1755 if (error && cookie_index == 0) {
1756 if (error == ENOENT)
1759 kfree(cookies, M_TEMP);
1760 *ap->a_ncookies = 0;
1761 *ap->a_cookies = NULL;
1764 if (error == ENOENT)
1767 *ap->a_ncookies = cookie_index;
1768 *ap->a_cookies = cookies;
1771 lwkt_reltoken(&hmp->fs_token);
1776 * hammer_vop_readlink { vp, uio, cred }
1780 hammer_vop_readlink(struct vop_readlink_args *ap)
1782 struct hammer_transaction trans;
1783 struct hammer_cursor cursor;
1787 uint32_t localization;
1788 hammer_pseudofs_inmem_t pfsm;
1791 ip = VTOI(ap->a_vp);
1794 lwkt_gettoken(&hmp->fs_token);
1797 * Shortcut if the symlink data was stuffed into ino_data.
1799 * Also expand special "@@PFS%05d" softlinks (expansion only
1800 * occurs for non-historical (current) accesses made from the
1801 * primary filesystem).
1803 * Note that userspace hammer command does not allow users to
1804 * create a @@PFS softlink under an existing other PFS (id!=0)
1805 * so the ip localization here for @@PFS softlink is always 0.
1807 if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1811 ptr = ip->ino_data.ext.symlink;
1812 bytes = (int)ip->ino_data.size;
1814 ip->obj_asof == HAMMER_MAX_TID &&
1815 ip->obj_localization == HAMMER_DEF_LOCALIZATION &&
1816 strncmp(ptr, "@@PFS", 5) == 0) {
1817 hammer_simple_transaction(&trans, hmp);
1818 bcopy(ptr + 5, buf, 5);
1820 localization = pfs_to_lo(strtoul(buf, NULL, 10));
1821 pfsm = hammer_load_pseudofs(&trans, localization,
1824 if (hammer_is_pfs_slave(&pfsm->pfsd)) {
1825 /* vap->va_size == 26 */
1826 ksnprintf(buf, sizeof(buf),
1828 (intmax_t)pfsm->pfsd.sync_end_tid,
1829 lo_to_pfs(localization));
1831 /* vap->va_size == 10 */
1832 ksnprintf(buf, sizeof(buf),
1834 lo_to_pfs(localization));
1837 bytes = strlen(buf);
1840 hammer_rel_pseudofs(hmp, pfsm);
1841 hammer_done_transaction(&trans);
1843 error = uiomove(ptr, bytes, ap->a_uio);
1844 lwkt_reltoken(&hmp->fs_token);
1851 hammer_simple_transaction(&trans, hmp);
1852 hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1855 * Key range (begin and end inclusive) to scan. Directory keys
1856 * directly translate to a 64 bit 'seek' position.
1858 cursor.key_beg.localization = ip->obj_localization |
1859 HAMMER_LOCALIZE_MISC;
1860 cursor.key_beg.obj_id = ip->obj_id;
1861 cursor.key_beg.create_tid = 0;
1862 cursor.key_beg.delete_tid = 0;
1863 cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1864 cursor.key_beg.obj_type = 0;
1865 cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1866 cursor.asof = ip->obj_asof;
1867 cursor.flags |= HAMMER_CURSOR_ASOF;
1869 error = hammer_ip_lookup(&cursor);
1871 error = hammer_ip_resolve_data(&cursor);
1873 KKASSERT(cursor.leaf->data_len >=
1874 HAMMER_SYMLINK_NAME_OFF);
1875 error = uiomove(cursor.data->symlink.name,
1876 cursor.leaf->data_len -
1877 HAMMER_SYMLINK_NAME_OFF,
1881 hammer_done_cursor(&cursor);
1882 hammer_done_transaction(&trans);
1883 lwkt_reltoken(&hmp->fs_token);
1888 * hammer_vop_nremove { nch, dvp, cred }
1892 hammer_vop_nremove(struct vop_nremove_args *ap)
1894 struct hammer_transaction trans;
1899 dip = VTOI(ap->a_dvp);
1902 if (hammer_nohistory(dip) == 0 &&
1903 (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1907 lwkt_gettoken(&hmp->fs_token);
1908 hammer_start_transaction(&trans, hmp);
1909 error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1910 hammer_done_transaction(&trans);
1912 hammer_knote(ap->a_dvp, NOTE_WRITE);
1913 lwkt_reltoken(&hmp->fs_token);
1918 * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1922 hammer_vop_nrename(struct vop_nrename_args *ap)
1924 struct hammer_transaction trans;
1925 struct namecache *fncp;
1926 struct namecache *tncp;
1927 hammer_inode_t fdip;
1928 hammer_inode_t tdip;
1931 struct hammer_cursor cursor;
1933 uint32_t max_iterations;
1936 if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount)
1938 if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1941 fdip = VTOI(ap->a_fdvp);
1942 tdip = VTOI(ap->a_tdvp);
1943 fncp = ap->a_fnch->ncp;
1944 tncp = ap->a_tnch->ncp;
1945 ip = VTOI(fncp->nc_vp);
1946 KKASSERT(ip != NULL);
1950 if (fdip->obj_localization != tdip->obj_localization)
1952 if (fdip->obj_localization != ip->obj_localization)
1955 if (fdip->flags & HAMMER_INODE_RO)
1957 if (tdip->flags & HAMMER_INODE_RO)
1959 if (ip->flags & HAMMER_INODE_RO)
1961 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1964 lwkt_gettoken(&hmp->fs_token);
1965 hammer_start_transaction(&trans, hmp);
1968 * Remove tncp from the target directory and then link ip as
1969 * tncp. XXX pass trans to dounlink
1971 * Force the inode sync-time to match the transaction so it is
1972 * in-sync with the creation of the target directory entry.
1974 error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
1976 if (error == 0 || error == ENOENT) {
1977 error = hammer_ip_add_direntry(&trans, tdip,
1978 tncp->nc_name, tncp->nc_nlen,
1981 ip->ino_data.parent_obj_id = tdip->obj_id;
1982 ip->ino_data.ctime = trans.time;
1983 hammer_modify_inode(&trans, ip, HAMMER_INODE_DDIRTY);
1987 goto failed; /* XXX */
1990 * Locate the record in the originating directory and remove it.
1992 * Calculate the namekey and setup the key range for the scan. This
1993 * works kinda like a chained hash table where the lower 32 bits
1994 * of the namekey synthesize the chain.
1996 * The key range is inclusive of both key_beg and key_end.
1998 namekey = hammer_direntry_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
2001 hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
2002 cursor.key_beg.localization = fdip->obj_localization |
2003 hammer_dir_localization(fdip);
2004 cursor.key_beg.obj_id = fdip->obj_id;
2005 cursor.key_beg.key = namekey;
2006 cursor.key_beg.create_tid = 0;
2007 cursor.key_beg.delete_tid = 0;
2008 cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2009 cursor.key_beg.obj_type = 0;
2011 cursor.key_end = cursor.key_beg;
2012 cursor.key_end.key += max_iterations;
2013 cursor.asof = fdip->obj_asof;
2014 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2017 * Scan all matching records (the chain), locate the one matching
2018 * the requested path component.
2020 * The hammer_ip_*() functions merge in-memory records with on-disk
2021 * records for the purposes of the search.
2023 error = hammer_ip_first(&cursor);
2024 while (error == 0) {
2025 if (hammer_ip_resolve_data(&cursor) != 0)
2027 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2029 if (fncp->nc_nlen == nlen &&
2030 bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2033 error = hammer_ip_next(&cursor);
2037 * If all is ok we have to get the inode so we can adjust nlinks.
2039 * WARNING: hammer_ip_del_direntry() may have to terminate the
2040 * cursor to avoid a recursion. It's ok to call hammer_done_cursor()
2044 error = hammer_ip_del_direntry(&trans, &cursor, fdip, ip);
2047 * XXX A deadlock here will break rename's atomicy for the purposes
2048 * of crash recovery.
2050 if (error == EDEADLK) {
2051 hammer_done_cursor(&cursor);
2056 * Cleanup and tell the kernel that the rename succeeded.
2058 * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2059 * without formally acquiring the vp since the vp might
2060 * have zero refs on it, or in the middle of a reclaim,
2063 hammer_done_cursor(&cursor);
2065 cache_rename(ap->a_fnch, ap->a_tnch);
2066 hammer_knote(ap->a_fdvp, NOTE_WRITE);
2067 hammer_knote(ap->a_tdvp, NOTE_WRITE);
2071 error = hammer_get_vnode(ip, &vp);
2072 if (error == 0 && vp) {
2074 hammer_knote(ip->vp, NOTE_RENAME);
2078 hdkprintf("ip/vp race2 avoided\n");
2083 hammer_done_transaction(&trans);
2084 lwkt_reltoken(&hmp->fs_token);
2089 * hammer_vop_nrmdir { nch, dvp, cred }
2093 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
2095 struct hammer_transaction trans;
2100 dip = VTOI(ap->a_dvp);
2103 if (hammer_nohistory(dip) == 0 &&
2104 (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2108 lwkt_gettoken(&hmp->fs_token);
2109 hammer_start_transaction(&trans, hmp);
2110 error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
2111 hammer_done_transaction(&trans);
2113 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
2114 lwkt_reltoken(&hmp->fs_token);
2119 * hammer_vop_markatime { vp, cred }
2123 hammer_vop_markatime(struct vop_markatime_args *ap)
2125 struct hammer_transaction trans;
2129 ip = VTOI(ap->a_vp);
2130 if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2132 if (ip->flags & HAMMER_INODE_RO)
2135 if (hmp->mp->mnt_flag & MNT_NOATIME)
2137 lwkt_gettoken(&hmp->fs_token);
2138 hammer_start_transaction(&trans, hmp);
2140 ip->ino_data.atime = trans.time;
2141 hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
2142 hammer_done_transaction(&trans);
2143 hammer_knote(ap->a_vp, NOTE_ATTRIB);
2144 lwkt_reltoken(&hmp->fs_token);
2149 * hammer_vop_setattr { vp, vap, cred }
2153 hammer_vop_setattr(struct vop_setattr_args *ap)
2155 struct hammer_transaction trans;
2165 int64_t aligned_size;
2170 ip = ap->a_vp->v_data;
2175 if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2177 if (ip->flags & HAMMER_INODE_RO)
2179 if (hammer_nohistory(ip) == 0 &&
2180 (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2184 lwkt_gettoken(&hmp->fs_token);
2185 hammer_start_transaction(&trans, hmp);
2188 if (vap->va_flags != VNOVAL) {
2189 flags = ip->ino_data.uflags;
2190 error = vop_helper_setattr_flags(&flags, vap->va_flags,
2191 hammer_to_unix_xid(&ip->ino_data.uid),
2194 if (ip->ino_data.uflags != flags) {
2195 ip->ino_data.uflags = flags;
2196 ip->ino_data.ctime = trans.time;
2197 modflags |= HAMMER_INODE_DDIRTY;
2198 kflags |= NOTE_ATTRIB;
2200 if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2207 if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2211 if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2212 mode_t cur_mode = ip->ino_data.mode;
2213 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2214 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2215 hammer_uuid_t uuid_uid;
2216 hammer_uuid_t uuid_gid;
2218 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2220 &cur_uid, &cur_gid, &cur_mode);
2222 hammer_guid_to_uuid(&uuid_uid, cur_uid);
2223 hammer_guid_to_uuid(&uuid_gid, cur_gid);
2224 if (kuuid_compare(&uuid_uid, &ip->ino_data.uid) ||
2225 kuuid_compare(&uuid_gid, &ip->ino_data.gid) ||
2226 ip->ino_data.mode != cur_mode) {
2227 ip->ino_data.uid = uuid_uid;
2228 ip->ino_data.gid = uuid_gid;
2229 ip->ino_data.mode = cur_mode;
2230 ip->ino_data.ctime = trans.time;
2231 modflags |= HAMMER_INODE_DDIRTY;
2233 kflags |= NOTE_ATTRIB;
2236 while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2237 switch(ap->a_vp->v_type) {
2239 if (vap->va_size == ip->ino_data.size)
2243 * Log the operation if in fast-fsync mode or if
2244 * there are unterminated redo write records present.
2246 * The second check is needed so the recovery code
2247 * properly truncates write redos even if nominal
2248 * REDO operations is turned off due to excessive
2249 * writes, because the related records might be
2250 * destroyed and never lay down a TERM_WRITE.
2252 if ((ip->flags & HAMMER_INODE_REDO) ||
2253 (ip->flags & HAMMER_INODE_RDIRTY)) {
2254 error = hammer_generate_redo(&trans, ip,
2259 blksize = hammer_blocksize(vap->va_size);
2262 * XXX break atomicy, we can deadlock the backend
2263 * if we do not release the lock. Probably not a
2266 if (vap->va_size < ip->ino_data.size) {
2267 nvtruncbuf(ap->a_vp, vap->va_size,
2269 hammer_blockoff(vap->va_size),
2272 kflags |= NOTE_WRITE;
2274 nvextendbuf(ap->a_vp,
2277 hammer_blocksize(ip->ino_data.size),
2278 hammer_blocksize(vap->va_size),
2279 hammer_blockoff(ip->ino_data.size),
2280 hammer_blockoff(vap->va_size),
2283 kflags |= NOTE_WRITE | NOTE_EXTEND;
2285 ip->ino_data.size = vap->va_size;
2286 ip->ino_data.mtime = trans.time;
2287 /* XXX safe to use SDIRTY instead of DDIRTY here? */
2288 modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2289 vclrflags(ap->a_vp, VLASTWRITETS);
2292 * On-media truncation is cached in the inode until
2293 * the inode is synchronized. We must immediately
2294 * handle any frontend records.
2297 hammer_ip_frontend_trunc(ip, vap->va_size);
2298 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2299 ip->flags |= HAMMER_INODE_TRUNCATED;
2300 ip->trunc_off = vap->va_size;
2301 hammer_inode_dirty(ip);
2302 } else if (ip->trunc_off > vap->va_size) {
2303 ip->trunc_off = vap->va_size;
2309 * When truncating, nvtruncbuf() may have cleaned out
2310 * a portion of the last block on-disk in the buffer
2311 * cache. We must clean out any frontend records
2312 * for blocks beyond the new last block.
2314 aligned_size = (vap->va_size + (blksize - 1)) &
2315 ~(int64_t)(blksize - 1);
2316 if (truncating && vap->va_size < aligned_size) {
2317 aligned_size -= blksize;
2318 hammer_ip_frontend_trunc(ip, aligned_size);
2323 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2324 ip->flags |= HAMMER_INODE_TRUNCATED;
2325 ip->trunc_off = vap->va_size;
2326 hammer_inode_dirty(ip);
2327 } else if (ip->trunc_off > vap->va_size) {
2328 ip->trunc_off = vap->va_size;
2330 hammer_ip_frontend_trunc(ip, vap->va_size);
2331 ip->ino_data.size = vap->va_size;
2332 ip->ino_data.mtime = trans.time;
2333 modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2334 vclrflags(ap->a_vp, VLASTWRITETS);
2335 kflags |= NOTE_ATTRIB;
2343 if (vap->va_atime.tv_sec != VNOVAL) {
2344 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2345 modflags |= HAMMER_INODE_ATIME;
2346 kflags |= NOTE_ATTRIB;
2348 if (vap->va_mtime.tv_sec != VNOVAL) {
2349 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2350 modflags |= HAMMER_INODE_MTIME;
2351 kflags |= NOTE_ATTRIB;
2352 vclrflags(ap->a_vp, VLASTWRITETS);
2354 if (vap->va_mode != (mode_t)VNOVAL) {
2355 mode_t cur_mode = ip->ino_data.mode;
2356 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2357 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2359 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2360 cur_uid, cur_gid, &cur_mode);
2361 if (error == 0 && ip->ino_data.mode != cur_mode) {
2362 ip->ino_data.mode = cur_mode;
2363 ip->ino_data.ctime = trans.time;
2364 modflags |= HAMMER_INODE_DDIRTY;
2365 kflags |= NOTE_ATTRIB;
2370 hammer_modify_inode(&trans, ip, modflags);
2371 hammer_done_transaction(&trans);
2372 hammer_knote(ap->a_vp, kflags);
2373 lwkt_reltoken(&hmp->fs_token);
2378 * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2382 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2384 struct hammer_transaction trans;
2387 hammer_record_t record;
2388 struct nchandle *nch;
2393 ap->a_vap->va_type = VLNK;
2396 dip = VTOI(ap->a_dvp);
2399 if (dip->flags & HAMMER_INODE_RO)
2401 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
2405 * Create a transaction to cover the operations we perform.
2407 lwkt_gettoken(&hmp->fs_token);
2408 hammer_start_transaction(&trans, hmp);
2411 * Create a new filesystem object of the requested type. The
2412 * returned inode will be referenced but not locked.
2415 error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2416 dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2419 hammer_done_transaction(&trans);
2421 lwkt_reltoken(&hmp->fs_token);
2426 * Add a record representing the symlink. symlink stores the link
2427 * as pure data, not a string, and is no \0 terminated.
2430 bytes = strlen(ap->a_target);
2432 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2433 bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2435 record = hammer_alloc_mem_record(nip, bytes);
2436 record->type = HAMMER_MEM_RECORD_GENERAL;
2438 record->leaf.base.localization = nip->obj_localization |
2439 HAMMER_LOCALIZE_MISC;
2440 record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2441 record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2442 record->leaf.data_len = bytes;
2443 KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2444 bcopy(ap->a_target, record->data->symlink.name, bytes);
2445 error = hammer_ip_add_record(&trans, record);
2449 * Set the file size to the length of the link.
2452 nip->ino_data.size = bytes;
2453 hammer_modify_inode(&trans, nip, HAMMER_INODE_DDIRTY);
2457 error = hammer_ip_add_direntry(&trans, dip, nch->ncp->nc_name,
2458 nch->ncp->nc_nlen, nip);
2464 hammer_rel_inode(nip, 0);
2467 error = hammer_get_vnode(nip, ap->a_vpp);
2468 hammer_rel_inode(nip, 0);
2470 cache_setunresolved(ap->a_nch);
2471 cache_setvp(ap->a_nch, *ap->a_vpp);
2472 hammer_knote(ap->a_dvp, NOTE_WRITE);
2475 hammer_done_transaction(&trans);
2476 lwkt_reltoken(&hmp->fs_token);
2481 * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2485 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2487 struct hammer_transaction trans;
2492 dip = VTOI(ap->a_dvp);
2495 if (hammer_nohistory(dip) == 0 &&
2496 (error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2500 lwkt_gettoken(&hmp->fs_token);
2501 hammer_start_transaction(&trans, hmp);
2502 error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2503 ap->a_cred, ap->a_flags, -1);
2504 hammer_done_transaction(&trans);
2505 lwkt_reltoken(&hmp->fs_token);
2511 * hammer_vop_ioctl { vp, command, data, fflag, cred }
2515 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2517 hammer_inode_t ip = ap->a_vp->v_data;
2518 hammer_mount_t hmp = ip->hmp;
2521 lwkt_gettoken(&hmp->fs_token);
2522 error = hammer_ioctl(ip, ap->a_command, ap->a_data,
2523 ap->a_fflag, ap->a_cred);
2524 lwkt_reltoken(&hmp->fs_token);
2530 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2532 static const struct mountctl_opt extraopt[] = {
2533 { HMNT_NOHISTORY, "nohistory" },
2534 { HMNT_MASTERID, "master" },
2535 { HMNT_NOMIRROR, "nomirror" },
2546 mp = ap->a_head.a_ops->head.vv_mount;
2547 KKASSERT(mp->mnt_data != NULL);
2548 hmp = (hammer_mount_t)mp->mnt_data;
2550 lwkt_gettoken(&hmp->fs_token);
2553 case MOUNTCTL_SET_EXPORT:
2554 if (ap->a_ctllen != sizeof(struct export_args))
2557 error = hammer_vfs_export(mp, ap->a_op,
2558 (const struct export_args *)ap->a_ctl);
2560 case MOUNTCTL_MOUNTFLAGS:
2562 * Call standard mountctl VOP function
2563 * so we get user mount flags.
2565 error = vop_stdmountctl(ap);
2569 usedbytes = *ap->a_res;
2571 if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2572 usedbytes += vfs_flagstostr(hmp->hflags, extraopt,
2574 ap->a_buflen - usedbytes,
2578 *ap->a_res += usedbytes;
2581 error = vop_stdmountctl(ap);
2584 lwkt_reltoken(&hmp->fs_token);
2589 * hammer_vop_strategy { vp, bio }
2591 * Strategy call, used for regular file read & write only. Note that the
2592 * bp may represent a cluster.
2594 * To simplify operation and allow better optimizations in the future,
2595 * this code does not make any assumptions with regards to buffer alignment
2600 hammer_vop_strategy(struct vop_strategy_args *ap)
2605 bp = ap->a_bio->bio_buf;
2609 error = hammer_vop_strategy_read(ap);
2612 error = hammer_vop_strategy_write(ap);
2615 bp->b_error = error = EINVAL;
2616 bp->b_flags |= B_ERROR;
2624 * Read from a regular file. Iterate the related records and fill in the
2625 * BIO/BUF. Gaps are zero-filled.
2627 * The support code in hammer_object.c should be used to deal with mixed
2628 * in-memory and on-disk records.
2630 * NOTE: Can be called from the cluster code with an oversized buf.
2636 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2638 struct hammer_transaction trans;
2642 struct hammer_cursor cursor;
2643 hammer_base_elm_t base;
2644 hammer_off_t disk_offset;
2659 ip = ap->a_vp->v_data;
2663 * The zone-2 disk offset may have been set by the cluster code via
2664 * a BMAP operation, or else should be NOOFFSET.
2666 * Checking the high bits for a match against zone-2 should suffice.
2668 * In cases where a lot of data duplication is present it may be
2669 * more beneficial to drop through and doubule-buffer through the
2672 nbio = push_bio(bio);
2673 if (hammer_is_zone_large_data(nbio->bio_offset)) {
2674 if (hammer_double_buffer == 0) {
2675 lwkt_gettoken(&hmp->fs_token);
2676 error = hammer_io_direct_read(hmp, nbio, NULL);
2677 lwkt_reltoken(&hmp->fs_token);
2682 * Try to shortcut requests for double_buffer mode too.
2683 * Since this mode runs through the device buffer cache
2684 * only compatible buffer sizes (meaning those generated
2685 * by normal filesystem buffers) are legal.
2687 if ((bp->b_flags & B_PAGING) == 0) {
2688 lwkt_gettoken(&hmp->fs_token);
2689 error = hammer_io_indirect_read(hmp, nbio, NULL);
2690 lwkt_reltoken(&hmp->fs_token);
2696 * Well, that sucked. Do it the hard way. If all the stars are
2697 * aligned we may still be able to issue a direct-read.
2699 lwkt_gettoken(&hmp->fs_token);
2700 hammer_simple_transaction(&trans, hmp);
2701 hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2704 * Key range (begin and end inclusive) to scan. Note that the key's
2705 * stored in the actual records represent BASE+LEN, not BASE. The
2706 * first record containing bio_offset will have a key > bio_offset.
2708 cursor.key_beg.localization = ip->obj_localization |
2709 HAMMER_LOCALIZE_MISC;
2710 cursor.key_beg.obj_id = ip->obj_id;
2711 cursor.key_beg.create_tid = 0;
2712 cursor.key_beg.delete_tid = 0;
2713 cursor.key_beg.obj_type = 0;
2714 cursor.key_beg.key = bio->bio_offset + 1;
2715 cursor.asof = ip->obj_asof;
2716 cursor.flags |= HAMMER_CURSOR_ASOF;
2718 cursor.key_end = cursor.key_beg;
2719 KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2721 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2722 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2723 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2724 cursor.key_end.key = HAMMER_MAX_KEY;
2728 ran_end = bio->bio_offset + bp->b_bufsize;
2729 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2730 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2731 tmp64 = ran_end + MAXPHYS + 1; /* work-around GCC-4 bug */
2732 if (tmp64 < ran_end)
2733 cursor.key_end.key = HAMMER_MAX_KEY;
2735 cursor.key_end.key = ran_end + MAXPHYS + 1;
2737 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2740 * Set NOSWAPCACHE for cursor data extraction if double buffering
2741 * is disabled or (if the file is not marked cacheable via chflags
2742 * and vm.swapcache_use_chflags is enabled).
2744 if (hammer_double_buffer == 0 ||
2745 ((ap->a_vp->v_flag & VSWAPCACHE) == 0 &&
2746 vm_swapcache_use_chflags)) {
2747 cursor.flags |= HAMMER_CURSOR_NOSWAPCACHE;
2750 error = hammer_ip_first(&cursor);
2753 while (error == 0) {
2755 * Get the base file offset of the record. The key for
2756 * data records is (base + bytes) rather then (base).
2758 base = &cursor.leaf->base;
2759 rec_offset = base->key - cursor.leaf->data_len;
2762 * Calculate the gap, if any, and zero-fill it.
2764 * n is the offset of the start of the record verses our
2765 * current seek offset in the bio.
2767 n = (int)(rec_offset - (bio->bio_offset + boff));
2769 if (n > bp->b_bufsize - boff)
2770 n = bp->b_bufsize - boff;
2771 bzero((char *)bp->b_data + boff, n);
2777 * Calculate the data offset in the record and the number
2778 * of bytes we can copy.
2780 * There are two degenerate cases. First, boff may already
2781 * be at bp->b_bufsize. Secondly, the data offset within
2782 * the record may exceed the record's size.
2786 n = cursor.leaf->data_len - roff;
2788 hdkprintf("bad n=%d roff=%d\n", n, roff);
2790 } else if (n > bp->b_bufsize - boff) {
2791 n = bp->b_bufsize - boff;
2795 * Deal with cached truncations. This cool bit of code
2796 * allows truncate()/ftruncate() to avoid having to sync
2799 * If the frontend is truncated then all backend records are
2800 * subject to the frontend's truncation.
2802 * If the backend is truncated then backend records on-disk
2803 * (but not in-memory) are subject to the backend's
2804 * truncation. In-memory records owned by the backend
2805 * represent data written after the truncation point on the
2806 * backend and must not be truncated.
2808 * Truncate operations deal with frontend buffer cache
2809 * buffers and frontend-owned in-memory records synchronously.
2811 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2812 if (hammer_cursor_ondisk(&cursor)/* ||
2813 cursor.iprec->flush_state == HAMMER_FST_FLUSH*/) {
2814 if (ip->trunc_off <= rec_offset)
2816 else if (ip->trunc_off < rec_offset + n)
2817 n = (int)(ip->trunc_off - rec_offset);
2820 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2821 if (hammer_cursor_ondisk(&cursor)) {
2822 if (ip->sync_trunc_off <= rec_offset)
2824 else if (ip->sync_trunc_off < rec_offset + n)
2825 n = (int)(ip->sync_trunc_off - rec_offset);
2830 * Try to issue a direct read into our bio if possible,
2831 * otherwise resolve the element data into a hammer_buffer
2834 * The buffer on-disk should be zerod past any real
2835 * truncation point, but may not be for any synthesized
2836 * truncation point from above.
2838 * NOTE: disk_offset is only valid if the cursor data is
2841 disk_offset = cursor.leaf->data_offset + roff;
2842 isdedupable = (boff == 0 && n == bp->b_bufsize &&
2843 hammer_cursor_ondisk(&cursor) &&
2844 ((int)disk_offset & HAMMER_BUFMASK) == 0);
2846 if (isdedupable && hammer_double_buffer == 0) {
2850 KKASSERT(hammer_is_zone_large_data(disk_offset));
2851 nbio->bio_offset = disk_offset;
2852 error = hammer_io_direct_read(hmp, nbio, cursor.leaf);
2854 } else if (isdedupable) {
2856 * Async I/O case for reading from backing store
2857 * and copying the data to the filesystem buffer.
2859 KKASSERT(hammer_is_zone_large_data(disk_offset));
2860 nbio->bio_offset = disk_offset;
2861 error = hammer_io_indirect_read(hmp, nbio, cursor.leaf);
2864 error = hammer_ip_resolve_data(&cursor);
2866 bcopy((char *)cursor.data + roff,
2867 (char *)bp->b_data + boff, n);
2874 * Iterate until we have filled the request.
2877 if (boff == bp->b_bufsize)
2879 error = hammer_ip_next(&cursor);
2883 * There may have been a gap after the last record
2885 if (error == ENOENT)
2887 if (error == 0 && boff != bp->b_bufsize) {
2888 KKASSERT(boff < bp->b_bufsize);
2889 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2890 /* boff = bp->b_bufsize; */
2894 * Disallow swapcache operation on the vnode buffer if double
2895 * buffering is enabled, the swapcache will get the data via
2896 * the block device buffer.
2898 if (hammer_double_buffer)
2899 bp->b_flags |= B_NOTMETA;
2905 bp->b_error = error;
2907 bp->b_flags |= B_ERROR;
2912 * Cache the b-tree node for the last data read in cache[1].
2914 * If we hit the file EOF then also cache the node in the
2915 * governing directory's cache[3], it will be used to initialize
2916 * the new inode's cache[1] for any inodes looked up via the directory.
2918 * This doesn't reduce disk accesses since the B-Tree chain is
2919 * likely cached, but it does reduce cpu overhead when looking
2920 * up file offsets for cpdup/tar/cpio style iterations.
2923 hammer_cache_node(&ip->cache[1], cursor.node);
2924 if (ran_end >= ip->ino_data.size) {
2925 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
2926 ip->obj_asof, ip->obj_localization);
2928 hammer_cache_node(&dip->cache[3], cursor.node);
2929 hammer_rel_inode(dip, 0);
2932 hammer_done_cursor(&cursor);
2933 hammer_done_transaction(&trans);
2934 lwkt_reltoken(&hmp->fs_token);
2939 * BMAP operation - used to support cluster_read() only.
2941 * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2943 * This routine may return EOPNOTSUPP if the opration is not supported for
2944 * the specified offset. The contents of the pointer arguments do not
2945 * need to be initialized in that case.
2947 * If a disk address is available and properly aligned return 0 with
2948 * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2949 * to the run-length relative to that offset. Callers may assume that
2950 * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2951 * large, so return EOPNOTSUPP if it is not sufficiently large.
2955 hammer_vop_bmap(struct vop_bmap_args *ap)
2957 struct hammer_transaction trans;
2960 struct hammer_cursor cursor;
2961 hammer_base_elm_t base;
2965 int64_t base_offset;
2966 int64_t base_disk_offset;
2967 int64_t last_offset;
2968 hammer_off_t last_disk_offset;
2969 hammer_off_t disk_offset;
2974 ip = ap->a_vp->v_data;
2978 * We can only BMAP regular files. We can't BMAP database files,
2981 if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
2985 * bmap is typically called with runp/runb both NULL when used
2986 * for writing. We do not support BMAP for writing atm.
2988 if (ap->a_cmd != BUF_CMD_READ)
2992 * Scan the B-Tree to acquire blockmap addresses, then translate
2995 lwkt_gettoken(&hmp->fs_token);
2996 hammer_simple_transaction(&trans, hmp);
2998 hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
3001 * Key range (begin and end inclusive) to scan. Note that the key's
3002 * stored in the actual records represent BASE+LEN, not BASE. The
3003 * first record containing bio_offset will have a key > bio_offset.
3005 cursor.key_beg.localization = ip->obj_localization |
3006 HAMMER_LOCALIZE_MISC;
3007 cursor.key_beg.obj_id = ip->obj_id;
3008 cursor.key_beg.create_tid = 0;
3009 cursor.key_beg.delete_tid = 0;
3010 cursor.key_beg.obj_type = 0;
3012 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
3014 cursor.key_beg.key = ap->a_loffset + 1;
3015 if (cursor.key_beg.key < 0)
3016 cursor.key_beg.key = 0;
3017 cursor.asof = ip->obj_asof;
3018 cursor.flags |= HAMMER_CURSOR_ASOF;
3020 cursor.key_end = cursor.key_beg;
3021 KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
3023 ran_end = ap->a_loffset + MAXPHYS;
3024 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
3025 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
3026 tmp64 = ran_end + MAXPHYS + 1; /* work-around GCC-4 bug */
3027 if (tmp64 < ran_end)
3028 cursor.key_end.key = HAMMER_MAX_KEY;
3030 cursor.key_end.key = ran_end + MAXPHYS + 1;
3032 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
3034 error = hammer_ip_first(&cursor);
3035 base_offset = last_offset = 0;
3036 base_disk_offset = last_disk_offset = 0;
3038 while (error == 0) {
3040 * Get the base file offset of the record. The key for
3041 * data records is (base + bytes) rather then (base).
3043 * NOTE: rec_offset + rec_len may exceed the end-of-file.
3044 * The extra bytes should be zero on-disk and the BMAP op
3045 * should still be ok.
3047 base = &cursor.leaf->base;
3048 rec_offset = base->key - cursor.leaf->data_len;
3049 rec_len = cursor.leaf->data_len;
3052 * Incorporate any cached truncation.
3054 * NOTE: Modifications to rec_len based on synthesized
3055 * truncation points remove the guarantee that any extended
3056 * data on disk is zero (since the truncations may not have
3057 * taken place on-media yet).
3059 if (ip->flags & HAMMER_INODE_TRUNCATED) {
3060 if (hammer_cursor_ondisk(&cursor) ||
3061 cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
3062 if (ip->trunc_off <= rec_offset)
3064 else if (ip->trunc_off < rec_offset + rec_len)
3065 rec_len = (int)(ip->trunc_off - rec_offset);
3068 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
3069 if (hammer_cursor_ondisk(&cursor)) {
3070 if (ip->sync_trunc_off <= rec_offset)
3072 else if (ip->sync_trunc_off < rec_offset + rec_len)
3073 rec_len = (int)(ip->sync_trunc_off - rec_offset);
3078 * Accumulate information. If we have hit a discontiguous
3079 * block reset base_offset unless we are already beyond the
3080 * requested offset. If we are, that's it, we stop.
3084 if (hammer_cursor_ondisk(&cursor)) {
3085 disk_offset = cursor.leaf->data_offset;
3086 if (rec_offset != last_offset ||
3087 disk_offset != last_disk_offset) {
3088 if (rec_offset > ap->a_loffset)
3090 base_offset = rec_offset;
3091 base_disk_offset = disk_offset;
3093 last_offset = rec_offset + rec_len;
3094 last_disk_offset = disk_offset + rec_len;
3096 error = hammer_ip_next(&cursor);
3100 hammer_cache_node(&ip->cache[1], cursor.node);
3102 hammer_done_cursor(&cursor);
3103 hammer_done_transaction(&trans);
3104 lwkt_reltoken(&hmp->fs_token);
3107 * If we couldn't find any records or the records we did find were
3108 * all behind the requested offset, return failure. A forward
3109 * truncation can leave a hole w/ no on-disk records.
3111 if (last_offset == 0 || last_offset < ap->a_loffset)
3112 return (EOPNOTSUPP);
3115 * Figure out the block size at the requested offset and adjust
3116 * our limits so the cluster_read() does not create inappropriately
3117 * sized buffer cache buffers.
3119 blksize = hammer_blocksize(ap->a_loffset);
3120 if (hammer_blocksize(base_offset) != blksize) {
3121 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
3123 if (last_offset != ap->a_loffset &&
3124 hammer_blocksize(last_offset - 1) != blksize) {
3125 last_offset = hammer_blockdemarc(ap->a_loffset,
3130 * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3133 disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
3135 if (!hammer_is_zone_large_data(disk_offset)) {
3137 * Only large-data zones can be direct-IOd
3140 } else if ((disk_offset & HAMMER_BUFMASK) ||
3141 (last_offset - ap->a_loffset) < blksize) {
3143 * doffsetp is not aligned or the forward run size does
3144 * not cover a whole buffer, disallow the direct I/O.
3151 *ap->a_doffsetp = disk_offset;
3153 *ap->a_runb = ap->a_loffset - base_offset;
3154 KKASSERT(*ap->a_runb >= 0);
3157 *ap->a_runp = last_offset - ap->a_loffset;
3158 KKASSERT(*ap->a_runp >= 0);
3166 * Write to a regular file. Because this is a strategy call the OS is
3167 * trying to actually get data onto the media.
3171 hammer_vop_strategy_write(struct vop_strategy_args *ap)
3173 hammer_record_t record;
3178 int blksize __debugvar;
3184 ip = ap->a_vp->v_data;
3187 blksize = hammer_blocksize(bio->bio_offset);
3188 KKASSERT(bp->b_bufsize == blksize);
3190 if (ip->flags & HAMMER_INODE_RO) {
3191 bp->b_error = EROFS;
3192 bp->b_flags |= B_ERROR;
3197 lwkt_gettoken(&hmp->fs_token);
3200 * Disallow swapcache operation on the vnode buffer if double
3201 * buffering is enabled, the swapcache will get the data via
3202 * the block device buffer.
3204 if (hammer_double_buffer)
3205 bp->b_flags |= B_NOTMETA;
3208 * Interlock with inode destruction (no in-kernel or directory
3209 * topology visibility). If we queue new IO while trying to
3210 * destroy the inode we can deadlock the vtrunc call in
3211 * hammer_inode_unloadable_check().
3213 * Besides, there's no point flushing a bp associated with an
3214 * inode that is being destroyed on-media and has no kernel
3217 if ((ip->flags | ip->sync_flags) &
3218 (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
3221 lwkt_reltoken(&hmp->fs_token);
3226 * Reserve space and issue a direct-write from the front-end.
3227 * NOTE: The direct_io code will hammer_bread/bcopy smaller
3230 * An in-memory record will be installed to reference the storage
3231 * until the flusher can get to it.
3233 * Since we own the high level bio the front-end will not try to
3234 * do a direct-read until the write completes.
3236 * NOTE: The only time we do not reserve a full-sized buffers
3237 * worth of data is if the file is small. We do not try to
3238 * allocate a fragment (from the small-data zone) at the end of
3239 * an otherwise large file as this can lead to wildly separated
3242 KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
3243 KKASSERT(bio->bio_offset < ip->ino_data.size);
3244 if (bio->bio_offset || ip->ino_data.size > HAMMER_HBUFSIZE)
3245 bytes = bp->b_bufsize;
3247 bytes = HAMMER_DATA_DOALIGN_WITH(int, ip->ino_data.size);
3249 record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
3253 * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3254 * in hammer_vop_write(). We must flag the record so the proper
3255 * REDO_TERM_WRITE entry is generated during the flush.
3258 if (bp->b_flags & B_VFSFLAG1) {
3259 record->flags |= HAMMER_RECF_REDO;
3260 bp->b_flags &= ~B_VFSFLAG1;
3262 hammer_io_direct_write(hmp, bio, record);
3263 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
3264 hammer_flush_inode(ip, 0);
3266 bp->b_bio2.bio_offset = NOOFFSET;
3267 bp->b_error = error;
3268 bp->b_flags |= B_ERROR;
3271 lwkt_reltoken(&hmp->fs_token);
3276 * dounlink - disconnect a directory entry
3278 * XXX whiteout support not really in yet
3281 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
3282 struct vnode *dvp, struct ucred *cred,
3283 int flags, int isdir)
3285 struct namecache *ncp;
3289 struct hammer_cursor cursor;
3291 uint32_t max_iterations;
3295 * Calculate the namekey and setup the key range for the scan. This
3296 * works kinda like a chained hash table where the lower 32 bits
3297 * of the namekey synthesize the chain.
3299 * The key range is inclusive of both key_beg and key_end.
3305 if (dip->flags & HAMMER_INODE_RO)
3308 namekey = hammer_direntry_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3311 hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3312 cursor.key_beg.localization = dip->obj_localization |
3313 hammer_dir_localization(dip);
3314 cursor.key_beg.obj_id = dip->obj_id;
3315 cursor.key_beg.key = namekey;
3316 cursor.key_beg.create_tid = 0;
3317 cursor.key_beg.delete_tid = 0;
3318 cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3319 cursor.key_beg.obj_type = 0;
3321 cursor.key_end = cursor.key_beg;
3322 cursor.key_end.key += max_iterations;
3323 cursor.asof = dip->obj_asof;
3324 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3327 * Scan all matching records (the chain), locate the one matching
3328 * the requested path component. info->last_error contains the
3329 * error code on search termination and could be 0, ENOENT, or
3332 * The hammer_ip_*() functions merge in-memory records with on-disk
3333 * records for the purposes of the search.
3335 error = hammer_ip_first(&cursor);
3337 while (error == 0) {
3338 error = hammer_ip_resolve_data(&cursor);
3341 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3343 if (ncp->nc_nlen == nlen &&
3344 bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3347 error = hammer_ip_next(&cursor);
3351 * If all is ok we have to get the inode so we can adjust nlinks.
3352 * To avoid a deadlock with the flusher we must release the inode
3353 * lock on the directory when acquiring the inode for the entry.
3355 * If the target is a directory, it must be empty.
3358 hammer_unlock(&cursor.ip->lock);
3359 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3361 cursor.data->entry.localization,
3363 hammer_lock_sh(&cursor.ip->lock);
3364 if (error == ENOENT) {
3365 hkprintf("WARNING: Removing dirent w/missing inode "
3367 "\tobj_id = %016jx\n",
3369 (intmax_t)cursor.data->entry.obj_id);
3374 * If isdir >= 0 we validate that the entry is or is not a
3375 * directory. If isdir < 0 we don't care.
3377 if (error == 0 && isdir >= 0 && ip) {
3379 ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3381 } else if (isdir == 0 &&
3382 ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3388 * If we are trying to remove a directory the directory must
3391 * The check directory code can loop and deadlock/retry. Our
3392 * own cursor's node locks must be released to avoid a 3-way
3393 * deadlock with the flusher if the check directory code
3396 * If any changes whatsoever have been made to the cursor
3397 * set EDEADLK and retry.
3399 * WARNING: See warnings in hammer_unlock_cursor()
3402 if (error == 0 && ip && ip->ino_data.obj_type ==
3403 HAMMER_OBJTYPE_DIRECTORY) {
3404 hammer_unlock_cursor(&cursor);
3405 error = hammer_ip_check_directory_empty(trans, ip);
3406 hammer_lock_cursor(&cursor);
3407 if (cursor.flags & HAMMER_CURSOR_RETEST) {
3408 hkprintf("Warning: avoided deadlock "
3416 * Delete the directory entry.
3418 * WARNING: hammer_ip_del_direntry() may have to terminate
3419 * the cursor to avoid a deadlock. It is ok to call
3420 * hammer_done_cursor() twice.
3423 error = hammer_ip_del_direntry(trans, &cursor,
3426 hammer_done_cursor(&cursor);
3429 * Tell the namecache that we are now unlinked.
3434 * NOTE: ip->vp, if non-NULL, cannot be directly
3435 * referenced without formally acquiring the
3436 * vp since the vp might have zero refs on it,
3437 * or in the middle of a reclaim, etc.
3439 * NOTE: The cache_setunresolved() can rip the vp
3440 * out from under us since the vp may not have
3441 * any refs, in which case ip->vp will be NULL
3444 while (ip && ip->vp) {
3447 error = hammer_get_vnode(ip, &vp);
3448 if (error == 0 && vp) {
3450 hammer_knote(ip->vp, NOTE_DELETE);
3453 * Don't do this, it can deadlock
3454 * on concurrent rm's of hardlinks.
3455 * Shouldn't be needed any more.
3457 cache_inval_vp(ip->vp, CINV_DESTROY);
3462 hdkprintf("ip/vp race1 avoided\n");
3466 hammer_rel_inode(ip, 0);
3468 hammer_done_cursor(&cursor);
3470 if (error == EDEADLK)
3476 /************************************************************************
3477 * FIFO AND SPECFS OPS *
3478 ************************************************************************
3482 hammer_vop_fifoclose (struct vop_close_args *ap)
3484 /* XXX update itimes */
3485 return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3489 hammer_vop_fiforead (struct vop_read_args *ap)
3493 error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3494 /* XXX update access time */
3499 hammer_vop_fifowrite (struct vop_write_args *ap)
3503 error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3504 /* XXX update access time */
3510 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3514 error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3516 error = hammer_vop_kqfilter(ap);
3520 /************************************************************************
3522 ************************************************************************
3525 static void filt_hammerdetach(struct knote *kn);
3526 static int filt_hammerread(struct knote *kn, long hint);
3527 static int filt_hammerwrite(struct knote *kn, long hint);
3528 static int filt_hammervnode(struct knote *kn, long hint);
3530 static struct filterops hammerread_filtops =
3531 { FILTEROP_ISFD | FILTEROP_MPSAFE,
3532 NULL, filt_hammerdetach, filt_hammerread };
3533 static struct filterops hammerwrite_filtops =
3534 { FILTEROP_ISFD | FILTEROP_MPSAFE,
3535 NULL, filt_hammerdetach, filt_hammerwrite };
3536 static struct filterops hammervnode_filtops =
3537 { FILTEROP_ISFD | FILTEROP_MPSAFE,
3538 NULL, filt_hammerdetach, filt_hammervnode };
3542 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3544 struct vnode *vp = ap->a_vp;
3545 struct knote *kn = ap->a_kn;
3547 switch (kn->kn_filter) {
3549 kn->kn_fop = &hammerread_filtops;
3552 kn->kn_fop = &hammerwrite_filtops;
3555 kn->kn_fop = &hammervnode_filtops;
3558 return (EOPNOTSUPP);
3561 kn->kn_hook = (caddr_t)vp;
3563 knote_insert(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3569 filt_hammerdetach(struct knote *kn)
3571 struct vnode *vp = (void *)kn->kn_hook;
3573 knote_remove(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3577 filt_hammerread(struct knote *kn, long hint)
3579 struct vnode *vp = (void *)kn->kn_hook;
3580 hammer_inode_t ip = VTOI(vp);
3581 hammer_mount_t hmp = ip->hmp;
3584 if (hint == NOTE_REVOKE) {
3585 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3588 lwkt_gettoken(&hmp->fs_token); /* XXX use per-ip-token */
3589 off = ip->ino_data.size - kn->kn_fp->f_offset;
3590 kn->kn_data = (off < INTPTR_MAX) ? off : INTPTR_MAX;
3591 lwkt_reltoken(&hmp->fs_token);
3592 if (kn->kn_sfflags & NOTE_OLDAPI)
3594 return (kn->kn_data != 0);
3598 filt_hammerwrite(struct knote *kn, long hint)
3600 if (hint == NOTE_REVOKE)
3601 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3607 filt_hammervnode(struct knote *kn, long hint)
3609 if (kn->kn_sfflags & hint)
3610 kn->kn_fflags |= hint;
3611 if (hint == NOTE_REVOKE) {
3612 kn->kn_flags |= (EV_EOF | EV_NODATA);
3615 return (kn->kn_fflags != 0);