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 * Note that userspace hammer command does not allow users to
1052 * create a @@PFS softlink under an existing other PFS (id!=0)
1053 * so the ip localization here for @@PFS softlink is always 0.
1055 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
1056 ip->ino_data.size == 10 &&
1057 ip->obj_asof == HAMMER_MAX_TID &&
1058 ip->obj_localization == 0 &&
1059 strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
1060 if (ip->pfsm->pfsd.mirror_flags & HAMMER_PFSD_SLAVE)
1067 * We must provide a consistent atime and mtime for snapshots
1068 * so people can do a 'tar cf - ... | md5' on them and get
1069 * consistent results.
1071 if (ip->flags & HAMMER_INODE_RO) {
1072 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
1073 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
1075 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
1076 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
1078 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
1079 vap->va_flags = ip->ino_data.uflags;
1080 vap->va_gen = 1; /* hammer inums are unique for all time */
1081 vap->va_blocksize = HAMMER_BUFSIZE;
1082 if (ip->ino_data.size >= HAMMER_XDEMARC) {
1083 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
1085 } else if (ip->ino_data.size > HAMMER_BUFSIZE / 2) {
1086 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
1089 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
1092 vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
1093 vap->va_filerev = 0; /* XXX */
1094 vap->va_uid_uuid = ip->ino_data.uid;
1095 vap->va_gid_uuid = ip->ino_data.gid;
1096 vap->va_fsid_uuid = ip->hmp->fsid;
1097 vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
1100 switch (ip->ino_data.obj_type) {
1101 case HAMMER_OBJTYPE_CDEV:
1102 case HAMMER_OBJTYPE_BDEV:
1103 vap->va_rmajor = ip->ino_data.rmajor;
1104 vap->va_rminor = ip->ino_data.rminor;
1109 hammer_unlock(&ip->lock);
1114 * hammer_vop_nresolve { nch, dvp, cred }
1116 * Locate the requested directory entry.
1120 hammer_vop_nresolve(struct vop_nresolve_args *ap)
1122 struct hammer_transaction trans;
1123 struct namecache *ncp;
1128 struct hammer_cursor cursor;
1137 u_int32_t localization;
1138 u_int32_t max_iterations;
1141 * Misc initialization, plus handle as-of name extensions. Look for
1142 * the '@@' extension. Note that as-of files and directories cannot
1145 dip = VTOI(ap->a_dvp);
1146 ncp = ap->a_nch->ncp;
1147 asof = dip->obj_asof;
1148 localization = dip->obj_localization; /* for code consistency */
1149 nlen = ncp->nc_nlen;
1150 flags = dip->flags & HAMMER_INODE_RO;
1154 lwkt_gettoken(&hmp->fs_token);
1155 hammer_simple_transaction(&trans, hmp);
1156 ++hammer_stats_file_iopsr;
1158 for (i = 0; i < nlen; ++i) {
1159 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
1160 error = hammer_str_to_tid(ncp->nc_name + i + 2,
1161 &ispfs, &asof, &localization);
1166 if (asof != HAMMER_MAX_TID)
1167 flags |= HAMMER_INODE_RO;
1174 * If this is a PFS softlink we dive into the PFS
1176 if (ispfs && nlen == 0) {
1177 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1181 error = hammer_get_vnode(ip, &vp);
1182 hammer_rel_inode(ip, 0);
1188 cache_setvp(ap->a_nch, vp);
1195 * If there is no path component the time extension is relative to dip.
1196 * e.g. "fubar/@@<snapshot>"
1198 * "." is handled by the kernel, but ".@@<snapshot>" is not.
1199 * e.g. "fubar/.@@<snapshot>"
1201 * ".." is handled by the kernel. We do not currently handle
1204 if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1205 ip = hammer_get_inode(&trans, dip, dip->obj_id,
1206 asof, dip->obj_localization,
1209 error = hammer_get_vnode(ip, &vp);
1210 hammer_rel_inode(ip, 0);
1216 cache_setvp(ap->a_nch, vp);
1223 * Calculate the namekey and setup the key range for the scan. This
1224 * works kinda like a chained hash table where the lower 32 bits
1225 * of the namekey synthesize the chain.
1227 * The key range is inclusive of both key_beg and key_end.
1229 namekey = hammer_directory_namekey(dip, ncp->nc_name, nlen,
1232 error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1233 cursor.key_beg.localization = dip->obj_localization +
1234 hammer_dir_localization(dip);
1235 cursor.key_beg.obj_id = dip->obj_id;
1236 cursor.key_beg.key = namekey;
1237 cursor.key_beg.create_tid = 0;
1238 cursor.key_beg.delete_tid = 0;
1239 cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1240 cursor.key_beg.obj_type = 0;
1242 cursor.key_end = cursor.key_beg;
1243 cursor.key_end.key += max_iterations;
1245 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1248 * Scan all matching records (the chain), locate the one matching
1249 * the requested path component.
1251 * The hammer_ip_*() functions merge in-memory records with on-disk
1252 * records for the purposes of the search.
1255 localization = HAMMER_DEF_LOCALIZATION;
1258 error = hammer_ip_first(&cursor);
1259 while (error == 0) {
1260 error = hammer_ip_resolve_data(&cursor);
1263 if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1264 bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1265 obj_id = cursor.data->entry.obj_id;
1266 localization = cursor.data->entry.localization;
1269 error = hammer_ip_next(&cursor);
1272 hammer_done_cursor(&cursor);
1275 * Lookup the obj_id. This should always succeed. If it does not
1276 * the filesystem may be damaged and we return a dummy inode.
1279 ip = hammer_get_inode(&trans, dip, obj_id,
1282 if (error == ENOENT) {
1283 kprintf("HAMMER: WARNING: Missing "
1284 "inode for dirent \"%s\"\n"
1285 "\tobj_id = %016llx, asof=%016llx, lo=%08x\n",
1287 (long long)obj_id, (long long)asof,
1290 ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1295 error = hammer_get_vnode(ip, &vp);
1296 hammer_rel_inode(ip, 0);
1302 cache_setvp(ap->a_nch, vp);
1305 } else if (error == ENOENT) {
1306 cache_setvp(ap->a_nch, NULL);
1309 hammer_done_transaction(&trans);
1310 lwkt_reltoken(&hmp->fs_token);
1315 * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1317 * Locate the parent directory of a directory vnode.
1319 * dvp is referenced but not locked. *vpp must be returned referenced and
1320 * locked. A parent_obj_id of 0 does not necessarily indicate that we are
1321 * at the root, instead it could indicate that the directory we were in was
1324 * NOTE: as-of sequences are not linked into the directory structure. If
1325 * we are at the root with a different asof then the mount point, reload
1326 * the same directory with the mount point's asof. I'm not sure what this
1327 * will do to NFS. We encode ASOF stamps in NFS file handles so it might not
1328 * get confused, but it hasn't been tested.
1332 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1334 struct hammer_transaction trans;
1335 struct hammer_inode *dip;
1336 struct hammer_inode *ip;
1338 int64_t parent_obj_id;
1339 u_int32_t parent_obj_localization;
1343 dip = VTOI(ap->a_dvp);
1344 asof = dip->obj_asof;
1348 * Whos are parent? This could be the root of a pseudo-filesystem
1349 * whos parent is in another localization domain.
1351 lwkt_gettoken(&hmp->fs_token);
1352 parent_obj_id = dip->ino_data.parent_obj_id;
1353 if (dip->obj_id == HAMMER_OBJID_ROOT)
1354 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
1356 parent_obj_localization = dip->obj_localization;
1359 * It's probably a PFS root when dip->ino_data.parent_obj_id is 0.
1361 if (parent_obj_id == 0) {
1362 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1363 asof != hmp->asof) {
1364 parent_obj_id = dip->obj_id;
1366 *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1367 ksnprintf(*ap->a_fakename, 19, "0x%016llx",
1368 (long long)dip->obj_asof);
1371 lwkt_reltoken(&hmp->fs_token);
1376 hammer_simple_transaction(&trans, hmp);
1377 ++hammer_stats_file_iopsr;
1379 ip = hammer_get_inode(&trans, dip, parent_obj_id,
1380 asof, parent_obj_localization,
1381 dip->flags, &error);
1383 error = hammer_get_vnode(ip, ap->a_vpp);
1384 hammer_rel_inode(ip, 0);
1388 hammer_done_transaction(&trans);
1389 lwkt_reltoken(&hmp->fs_token);
1394 * hammer_vop_nlink { nch, dvp, vp, cred }
1398 hammer_vop_nlink(struct vop_nlink_args *ap)
1400 struct hammer_transaction trans;
1401 struct hammer_inode *dip;
1402 struct hammer_inode *ip;
1403 struct nchandle *nch;
1407 if (ap->a_dvp->v_mount != ap->a_vp->v_mount)
1411 dip = VTOI(ap->a_dvp);
1412 ip = VTOI(ap->a_vp);
1415 if (dip->obj_localization != ip->obj_localization)
1418 if (dip->flags & HAMMER_INODE_RO)
1420 if (ip->flags & HAMMER_INODE_RO)
1422 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1426 * Create a transaction to cover the operations we perform.
1428 lwkt_gettoken(&hmp->fs_token);
1429 hammer_start_transaction(&trans, hmp);
1430 ++hammer_stats_file_iopsw;
1433 * Add the filesystem object to the directory. Note that neither
1434 * dip nor ip are referenced or locked, but their vnodes are
1435 * referenced. This function will bump the inode's link count.
1437 error = hammer_ip_add_directory(&trans, dip,
1438 nch->ncp->nc_name, nch->ncp->nc_nlen,
1445 cache_setunresolved(nch);
1446 cache_setvp(nch, ap->a_vp);
1448 hammer_done_transaction(&trans);
1449 hammer_knote(ap->a_vp, NOTE_LINK);
1450 hammer_knote(ap->a_dvp, NOTE_WRITE);
1451 lwkt_reltoken(&hmp->fs_token);
1456 * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1458 * The operating system has already ensured that the directory entry
1459 * does not exist and done all appropriate namespace locking.
1463 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1465 struct hammer_transaction trans;
1466 struct hammer_inode *dip;
1467 struct hammer_inode *nip;
1468 struct nchandle *nch;
1473 dip = VTOI(ap->a_dvp);
1476 if (dip->flags & HAMMER_INODE_RO)
1478 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1482 * Create a transaction to cover the operations we perform.
1484 lwkt_gettoken(&hmp->fs_token);
1485 hammer_start_transaction(&trans, hmp);
1486 ++hammer_stats_file_iopsw;
1489 * Create a new filesystem object of the requested type. The
1490 * returned inode will be referenced but not locked.
1492 error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1493 dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1496 hkprintf("hammer_mkdir error %d\n", error);
1497 hammer_done_transaction(&trans);
1499 lwkt_reltoken(&hmp->fs_token);
1503 * Add the new filesystem object to the directory. This will also
1504 * bump the inode's link count.
1506 error = hammer_ip_add_directory(&trans, dip,
1507 nch->ncp->nc_name, nch->ncp->nc_nlen,
1510 hkprintf("hammer_mkdir (add) error %d\n", error);
1516 hammer_rel_inode(nip, 0);
1519 error = hammer_get_vnode(nip, ap->a_vpp);
1520 hammer_rel_inode(nip, 0);
1522 cache_setunresolved(ap->a_nch);
1523 cache_setvp(ap->a_nch, *ap->a_vpp);
1526 hammer_done_transaction(&trans);
1528 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1529 lwkt_reltoken(&hmp->fs_token);
1534 * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1536 * The operating system has already ensured that the directory entry
1537 * does not exist and done all appropriate namespace locking.
1541 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1543 struct hammer_transaction trans;
1544 struct hammer_inode *dip;
1545 struct hammer_inode *nip;
1546 struct nchandle *nch;
1551 dip = VTOI(ap->a_dvp);
1554 if (dip->flags & HAMMER_INODE_RO)
1556 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1560 * Create a transaction to cover the operations we perform.
1562 lwkt_gettoken(&hmp->fs_token);
1563 hammer_start_transaction(&trans, hmp);
1564 ++hammer_stats_file_iopsw;
1567 * Create a new filesystem object of the requested type. The
1568 * returned inode will be referenced but not locked.
1570 * If mknod specifies a directory a pseudo-fs is created.
1572 error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1573 dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1576 hammer_done_transaction(&trans);
1578 lwkt_reltoken(&hmp->fs_token);
1583 * Add the new filesystem object to the directory. This will also
1584 * bump the inode's link count.
1586 error = hammer_ip_add_directory(&trans, dip,
1587 nch->ncp->nc_name, nch->ncp->nc_nlen,
1594 hammer_rel_inode(nip, 0);
1597 error = hammer_get_vnode(nip, ap->a_vpp);
1598 hammer_rel_inode(nip, 0);
1600 cache_setunresolved(ap->a_nch);
1601 cache_setvp(ap->a_nch, *ap->a_vpp);
1604 hammer_done_transaction(&trans);
1606 hammer_knote(ap->a_dvp, NOTE_WRITE);
1607 lwkt_reltoken(&hmp->fs_token);
1612 * hammer_vop_open { vp, mode, cred, fp }
1614 * MPSAFE (does not require fs_token)
1618 hammer_vop_open(struct vop_open_args *ap)
1622 ++hammer_stats_file_iopsr;
1623 ip = VTOI(ap->a_vp);
1625 if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1627 return(vop_stdopen(ap));
1631 * hammer_vop_print { vp }
1635 hammer_vop_print(struct vop_print_args *ap)
1641 * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1645 hammer_vop_readdir(struct vop_readdir_args *ap)
1647 struct hammer_transaction trans;
1648 struct hammer_cursor cursor;
1649 struct hammer_inode *ip;
1652 hammer_base_elm_t base;
1661 ++hammer_stats_file_iopsr;
1662 ip = VTOI(ap->a_vp);
1664 saveoff = uio->uio_offset;
1667 if (ap->a_ncookies) {
1668 ncookies = uio->uio_resid / 16 + 1;
1669 if (ncookies > 1024)
1671 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1679 lwkt_gettoken(&hmp->fs_token);
1680 hammer_simple_transaction(&trans, hmp);
1683 * Handle artificial entries
1685 * It should be noted that the minimum value for a directory
1686 * hash key on-media is 0x0000000100000000, so we can use anything
1687 * less then that to represent our 'special' key space.
1691 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1695 cookies[cookie_index] = saveoff;
1698 if (cookie_index == ncookies)
1702 if (ip->ino_data.parent_obj_id) {
1703 r = vop_write_dirent(&error, uio,
1704 ip->ino_data.parent_obj_id,
1707 r = vop_write_dirent(&error, uio,
1708 ip->obj_id, DT_DIR, 2, "..");
1713 cookies[cookie_index] = saveoff;
1716 if (cookie_index == ncookies)
1721 * Key range (begin and end inclusive) to scan. Directory keys
1722 * directly translate to a 64 bit 'seek' position.
1724 hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1725 cursor.key_beg.localization = ip->obj_localization +
1726 hammer_dir_localization(ip);
1727 cursor.key_beg.obj_id = ip->obj_id;
1728 cursor.key_beg.create_tid = 0;
1729 cursor.key_beg.delete_tid = 0;
1730 cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1731 cursor.key_beg.obj_type = 0;
1732 cursor.key_beg.key = saveoff;
1734 cursor.key_end = cursor.key_beg;
1735 cursor.key_end.key = HAMMER_MAX_KEY;
1736 cursor.asof = ip->obj_asof;
1737 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1739 error = hammer_ip_first(&cursor);
1741 while (error == 0) {
1742 error = hammer_ip_resolve_data(&cursor);
1745 base = &cursor.leaf->base;
1746 saveoff = base->key;
1747 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1749 if (base->obj_id != ip->obj_id)
1750 panic("readdir: bad record at %p", cursor.node);
1753 * Convert pseudo-filesystems into softlinks
1755 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1756 r = vop_write_dirent(
1757 &error, uio, cursor.data->entry.obj_id,
1759 cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1760 (void *)cursor.data->entry.name);
1765 cookies[cookie_index] = base->key;
1767 if (cookie_index == ncookies)
1769 error = hammer_ip_next(&cursor);
1771 hammer_done_cursor(&cursor);
1774 hammer_done_transaction(&trans);
1777 *ap->a_eofflag = (error == ENOENT);
1778 uio->uio_offset = saveoff;
1779 if (error && cookie_index == 0) {
1780 if (error == ENOENT)
1783 kfree(cookies, M_TEMP);
1784 *ap->a_ncookies = 0;
1785 *ap->a_cookies = NULL;
1788 if (error == ENOENT)
1791 *ap->a_ncookies = cookie_index;
1792 *ap->a_cookies = cookies;
1795 lwkt_reltoken(&hmp->fs_token);
1800 * hammer_vop_readlink { vp, uio, cred }
1804 hammer_vop_readlink(struct vop_readlink_args *ap)
1806 struct hammer_transaction trans;
1807 struct hammer_cursor cursor;
1808 struct hammer_inode *ip;
1811 u_int32_t localization;
1812 hammer_pseudofs_inmem_t pfsm;
1815 ip = VTOI(ap->a_vp);
1818 lwkt_gettoken(&hmp->fs_token);
1821 * Shortcut if the symlink data was stuffed into ino_data.
1823 * Also expand special "@@PFS%05d" softlinks (expansion only
1824 * occurs for non-historical (current) accesses made from the
1825 * primary filesystem).
1827 * Note that userspace hammer command does not allow users to
1828 * create a @@PFS softlink under an existing other PFS (id!=0)
1829 * so the ip localization here for @@PFS softlink is always 0.
1831 if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1835 ptr = ip->ino_data.ext.symlink;
1836 bytes = (int)ip->ino_data.size;
1838 ip->obj_asof == HAMMER_MAX_TID &&
1839 ip->obj_localization == 0 &&
1840 strncmp(ptr, "@@PFS", 5) == 0) {
1841 hammer_simple_transaction(&trans, hmp);
1842 bcopy(ptr + 5, buf, 5);
1844 localization = strtoul(buf, NULL, 10) << 16;
1845 pfsm = hammer_load_pseudofs(&trans, localization,
1848 if (pfsm->pfsd.mirror_flags &
1849 HAMMER_PFSD_SLAVE) {
1850 /* vap->va_size == 26 */
1851 ksnprintf(buf, sizeof(buf),
1853 (long long)pfsm->pfsd.sync_end_tid,
1854 localization >> 16);
1856 /* vap->va_size == 10 */
1857 ksnprintf(buf, sizeof(buf),
1859 localization >> 16);
1861 ksnprintf(buf, sizeof(buf),
1863 (long long)HAMMER_MAX_TID,
1864 localization >> 16);
1868 bytes = strlen(buf);
1871 hammer_rel_pseudofs(hmp, pfsm);
1872 hammer_done_transaction(&trans);
1874 error = uiomove(ptr, bytes, ap->a_uio);
1875 lwkt_reltoken(&hmp->fs_token);
1882 hammer_simple_transaction(&trans, hmp);
1883 ++hammer_stats_file_iopsr;
1884 hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1887 * Key range (begin and end inclusive) to scan. Directory keys
1888 * directly translate to a 64 bit 'seek' position.
1890 cursor.key_beg.localization = ip->obj_localization +
1891 HAMMER_LOCALIZE_MISC;
1892 cursor.key_beg.obj_id = ip->obj_id;
1893 cursor.key_beg.create_tid = 0;
1894 cursor.key_beg.delete_tid = 0;
1895 cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1896 cursor.key_beg.obj_type = 0;
1897 cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1898 cursor.asof = ip->obj_asof;
1899 cursor.flags |= HAMMER_CURSOR_ASOF;
1901 error = hammer_ip_lookup(&cursor);
1903 error = hammer_ip_resolve_data(&cursor);
1905 KKASSERT(cursor.leaf->data_len >=
1906 HAMMER_SYMLINK_NAME_OFF);
1907 error = uiomove(cursor.data->symlink.name,
1908 cursor.leaf->data_len -
1909 HAMMER_SYMLINK_NAME_OFF,
1913 hammer_done_cursor(&cursor);
1914 hammer_done_transaction(&trans);
1915 lwkt_reltoken(&hmp->fs_token);
1920 * hammer_vop_nremove { nch, dvp, cred }
1924 hammer_vop_nremove(struct vop_nremove_args *ap)
1926 struct hammer_transaction trans;
1927 struct hammer_inode *dip;
1931 dip = VTOI(ap->a_dvp);
1934 if (hammer_nohistory(dip) == 0 &&
1935 (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1939 lwkt_gettoken(&hmp->fs_token);
1940 hammer_start_transaction(&trans, hmp);
1941 ++hammer_stats_file_iopsw;
1942 error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1943 hammer_done_transaction(&trans);
1945 hammer_knote(ap->a_dvp, NOTE_WRITE);
1946 lwkt_reltoken(&hmp->fs_token);
1951 * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1955 hammer_vop_nrename(struct vop_nrename_args *ap)
1957 struct hammer_transaction trans;
1958 struct namecache *fncp;
1959 struct namecache *tncp;
1960 struct hammer_inode *fdip;
1961 struct hammer_inode *tdip;
1962 struct hammer_inode *ip;
1964 struct hammer_cursor cursor;
1966 u_int32_t max_iterations;
1969 if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount)
1971 if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1974 fdip = VTOI(ap->a_fdvp);
1975 tdip = VTOI(ap->a_tdvp);
1976 fncp = ap->a_fnch->ncp;
1977 tncp = ap->a_tnch->ncp;
1978 ip = VTOI(fncp->nc_vp);
1979 KKASSERT(ip != NULL);
1983 if (fdip->obj_localization != tdip->obj_localization)
1985 if (fdip->obj_localization != ip->obj_localization)
1988 if (fdip->flags & HAMMER_INODE_RO)
1990 if (tdip->flags & HAMMER_INODE_RO)
1992 if (ip->flags & HAMMER_INODE_RO)
1994 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1997 lwkt_gettoken(&hmp->fs_token);
1998 hammer_start_transaction(&trans, hmp);
1999 ++hammer_stats_file_iopsw;
2002 * Remove tncp from the target directory and then link ip as
2003 * tncp. XXX pass trans to dounlink
2005 * Force the inode sync-time to match the transaction so it is
2006 * in-sync with the creation of the target directory entry.
2008 error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
2010 if (error == 0 || error == ENOENT) {
2011 error = hammer_ip_add_directory(&trans, tdip,
2012 tncp->nc_name, tncp->nc_nlen,
2015 ip->ino_data.parent_obj_id = tdip->obj_id;
2016 ip->ino_data.ctime = trans.time;
2017 hammer_modify_inode(&trans, ip, HAMMER_INODE_DDIRTY);
2021 goto failed; /* XXX */
2024 * Locate the record in the originating directory and remove it.
2026 * Calculate the namekey and setup the key range for the scan. This
2027 * works kinda like a chained hash table where the lower 32 bits
2028 * of the namekey synthesize the chain.
2030 * The key range is inclusive of both key_beg and key_end.
2032 namekey = hammer_directory_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
2035 hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
2036 cursor.key_beg.localization = fdip->obj_localization +
2037 hammer_dir_localization(fdip);
2038 cursor.key_beg.obj_id = fdip->obj_id;
2039 cursor.key_beg.key = namekey;
2040 cursor.key_beg.create_tid = 0;
2041 cursor.key_beg.delete_tid = 0;
2042 cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2043 cursor.key_beg.obj_type = 0;
2045 cursor.key_end = cursor.key_beg;
2046 cursor.key_end.key += max_iterations;
2047 cursor.asof = fdip->obj_asof;
2048 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2051 * Scan all matching records (the chain), locate the one matching
2052 * the requested path component.
2054 * The hammer_ip_*() functions merge in-memory records with on-disk
2055 * records for the purposes of the search.
2057 error = hammer_ip_first(&cursor);
2058 while (error == 0) {
2059 if (hammer_ip_resolve_data(&cursor) != 0)
2061 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2063 if (fncp->nc_nlen == nlen &&
2064 bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2067 error = hammer_ip_next(&cursor);
2071 * If all is ok we have to get the inode so we can adjust nlinks.
2073 * WARNING: hammer_ip_del_directory() may have to terminate the
2074 * cursor to avoid a recursion. It's ok to call hammer_done_cursor()
2078 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
2081 * XXX A deadlock here will break rename's atomicy for the purposes
2082 * of crash recovery.
2084 if (error == EDEADLK) {
2085 hammer_done_cursor(&cursor);
2090 * Cleanup and tell the kernel that the rename succeeded.
2092 * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2093 * without formally acquiring the vp since the vp might
2094 * have zero refs on it, or in the middle of a reclaim,
2097 hammer_done_cursor(&cursor);
2099 cache_rename(ap->a_fnch, ap->a_tnch);
2100 hammer_knote(ap->a_fdvp, NOTE_WRITE);
2101 hammer_knote(ap->a_tdvp, NOTE_WRITE);
2105 error = hammer_get_vnode(ip, &vp);
2106 if (error == 0 && vp) {
2108 hammer_knote(ip->vp, NOTE_RENAME);
2112 kprintf("Debug: HAMMER ip/vp race2 avoided\n");
2117 hammer_done_transaction(&trans);
2118 lwkt_reltoken(&hmp->fs_token);
2123 * hammer_vop_nrmdir { nch, dvp, cred }
2127 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
2129 struct hammer_transaction trans;
2130 struct hammer_inode *dip;
2134 dip = VTOI(ap->a_dvp);
2137 if (hammer_nohistory(dip) == 0 &&
2138 (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2142 lwkt_gettoken(&hmp->fs_token);
2143 hammer_start_transaction(&trans, hmp);
2144 ++hammer_stats_file_iopsw;
2145 error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
2146 hammer_done_transaction(&trans);
2148 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
2149 lwkt_reltoken(&hmp->fs_token);
2154 * hammer_vop_markatime { vp, cred }
2158 hammer_vop_markatime(struct vop_markatime_args *ap)
2160 struct hammer_transaction trans;
2161 struct hammer_inode *ip;
2164 ip = VTOI(ap->a_vp);
2165 if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2167 if (ip->flags & HAMMER_INODE_RO)
2170 if (hmp->mp->mnt_flag & MNT_NOATIME)
2172 lwkt_gettoken(&hmp->fs_token);
2173 hammer_start_transaction(&trans, hmp);
2174 ++hammer_stats_file_iopsw;
2176 ip->ino_data.atime = trans.time;
2177 hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
2178 hammer_done_transaction(&trans);
2179 hammer_knote(ap->a_vp, NOTE_ATTRIB);
2180 lwkt_reltoken(&hmp->fs_token);
2185 * hammer_vop_setattr { vp, vap, cred }
2189 hammer_vop_setattr(struct vop_setattr_args *ap)
2191 struct hammer_transaction trans;
2192 struct hammer_inode *ip;
2201 int64_t aligned_size;
2206 ip = ap->a_vp->v_data;
2211 if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2213 if (ip->flags & HAMMER_INODE_RO)
2215 if (hammer_nohistory(ip) == 0 &&
2216 (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2220 lwkt_gettoken(&hmp->fs_token);
2221 hammer_start_transaction(&trans, hmp);
2222 ++hammer_stats_file_iopsw;
2225 if (vap->va_flags != VNOVAL) {
2226 flags = ip->ino_data.uflags;
2227 error = vop_helper_setattr_flags(&flags, vap->va_flags,
2228 hammer_to_unix_xid(&ip->ino_data.uid),
2231 if (ip->ino_data.uflags != flags) {
2232 ip->ino_data.uflags = flags;
2233 ip->ino_data.ctime = trans.time;
2234 modflags |= HAMMER_INODE_DDIRTY;
2235 kflags |= NOTE_ATTRIB;
2237 if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2244 if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2248 if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2249 mode_t cur_mode = ip->ino_data.mode;
2250 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2251 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2255 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2257 &cur_uid, &cur_gid, &cur_mode);
2259 hammer_guid_to_uuid(&uuid_uid, cur_uid);
2260 hammer_guid_to_uuid(&uuid_gid, cur_gid);
2261 if (bcmp(&uuid_uid, &ip->ino_data.uid,
2262 sizeof(uuid_uid)) ||
2263 bcmp(&uuid_gid, &ip->ino_data.gid,
2264 sizeof(uuid_gid)) ||
2265 ip->ino_data.mode != cur_mode
2267 ip->ino_data.uid = uuid_uid;
2268 ip->ino_data.gid = uuid_gid;
2269 ip->ino_data.mode = cur_mode;
2270 ip->ino_data.ctime = trans.time;
2271 modflags |= HAMMER_INODE_DDIRTY;
2273 kflags |= NOTE_ATTRIB;
2276 while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2277 switch(ap->a_vp->v_type) {
2279 if (vap->va_size == ip->ino_data.size)
2283 * Log the operation if in fast-fsync mode or if
2284 * there are unterminated redo write records present.
2286 * The second check is needed so the recovery code
2287 * properly truncates write redos even if nominal
2288 * REDO operations is turned off due to excessive
2289 * writes, because the related records might be
2290 * destroyed and never lay down a TERM_WRITE.
2292 if ((ip->flags & HAMMER_INODE_REDO) ||
2293 (ip->flags & HAMMER_INODE_RDIRTY)) {
2294 error = hammer_generate_redo(&trans, ip,
2299 blksize = hammer_blocksize(vap->va_size);
2302 * XXX break atomicy, we can deadlock the backend
2303 * if we do not release the lock. Probably not a
2306 if (vap->va_size < ip->ino_data.size) {
2307 nvtruncbuf(ap->a_vp, vap->va_size,
2309 hammer_blockoff(vap->va_size),
2312 kflags |= NOTE_WRITE;
2314 nvextendbuf(ap->a_vp,
2317 hammer_blocksize(ip->ino_data.size),
2318 hammer_blocksize(vap->va_size),
2319 hammer_blockoff(ip->ino_data.size),
2320 hammer_blockoff(vap->va_size),
2323 kflags |= NOTE_WRITE | NOTE_EXTEND;
2325 ip->ino_data.size = vap->va_size;
2326 ip->ino_data.mtime = trans.time;
2327 /* XXX safe to use SDIRTY instead of DDIRTY here? */
2328 modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2331 * On-media truncation is cached in the inode until
2332 * the inode is synchronized. We must immediately
2333 * handle any frontend records.
2336 hammer_ip_frontend_trunc(ip, vap->va_size);
2337 #ifdef DEBUG_TRUNCATE
2338 if (HammerTruncIp == NULL)
2341 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2342 ip->flags |= HAMMER_INODE_TRUNCATED;
2343 ip->trunc_off = vap->va_size;
2344 hammer_inode_dirty(ip);
2345 #ifdef DEBUG_TRUNCATE
2346 if (ip == HammerTruncIp)
2347 kprintf("truncate1 %016llx\n",
2348 (long long)ip->trunc_off);
2350 } else if (ip->trunc_off > vap->va_size) {
2351 ip->trunc_off = vap->va_size;
2352 #ifdef DEBUG_TRUNCATE
2353 if (ip == HammerTruncIp)
2354 kprintf("truncate2 %016llx\n",
2355 (long long)ip->trunc_off);
2358 #ifdef DEBUG_TRUNCATE
2359 if (ip == HammerTruncIp)
2360 kprintf("truncate3 %016llx (ignored)\n",
2361 (long long)vap->va_size);
2368 * When truncating, nvtruncbuf() may have cleaned out
2369 * a portion of the last block on-disk in the buffer
2370 * cache. We must clean out any frontend records
2371 * for blocks beyond the new last block.
2373 aligned_size = (vap->va_size + (blksize - 1)) &
2374 ~(int64_t)(blksize - 1);
2375 if (truncating && vap->va_size < aligned_size) {
2376 aligned_size -= blksize;
2377 hammer_ip_frontend_trunc(ip, aligned_size);
2382 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2383 ip->flags |= HAMMER_INODE_TRUNCATED;
2384 ip->trunc_off = vap->va_size;
2385 hammer_inode_dirty(ip);
2386 } else if (ip->trunc_off > vap->va_size) {
2387 ip->trunc_off = vap->va_size;
2389 hammer_ip_frontend_trunc(ip, vap->va_size);
2390 ip->ino_data.size = vap->va_size;
2391 ip->ino_data.mtime = trans.time;
2392 modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2393 kflags |= NOTE_ATTRIB;
2401 if (vap->va_atime.tv_sec != VNOVAL) {
2402 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2403 modflags |= HAMMER_INODE_ATIME;
2404 kflags |= NOTE_ATTRIB;
2406 if (vap->va_mtime.tv_sec != VNOVAL) {
2407 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2408 modflags |= HAMMER_INODE_MTIME;
2409 kflags |= NOTE_ATTRIB;
2411 if (vap->va_mode != (mode_t)VNOVAL) {
2412 mode_t cur_mode = ip->ino_data.mode;
2413 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2414 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2416 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2417 cur_uid, cur_gid, &cur_mode);
2418 if (error == 0 && ip->ino_data.mode != cur_mode) {
2419 ip->ino_data.mode = cur_mode;
2420 ip->ino_data.ctime = trans.time;
2421 modflags |= HAMMER_INODE_DDIRTY;
2422 kflags |= NOTE_ATTRIB;
2427 hammer_modify_inode(&trans, ip, modflags);
2428 hammer_done_transaction(&trans);
2429 hammer_knote(ap->a_vp, kflags);
2430 lwkt_reltoken(&hmp->fs_token);
2435 * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2439 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2441 struct hammer_transaction trans;
2442 struct hammer_inode *dip;
2443 struct hammer_inode *nip;
2444 hammer_record_t record;
2445 struct nchandle *nch;
2450 ap->a_vap->va_type = VLNK;
2453 dip = VTOI(ap->a_dvp);
2456 if (dip->flags & HAMMER_INODE_RO)
2458 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
2462 * Create a transaction to cover the operations we perform.
2464 lwkt_gettoken(&hmp->fs_token);
2465 hammer_start_transaction(&trans, hmp);
2466 ++hammer_stats_file_iopsw;
2469 * Create a new filesystem object of the requested type. The
2470 * returned inode will be referenced but not locked.
2473 error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2474 dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2477 hammer_done_transaction(&trans);
2479 lwkt_reltoken(&hmp->fs_token);
2484 * Add a record representing the symlink. symlink stores the link
2485 * as pure data, not a string, and is no \0 terminated.
2488 bytes = strlen(ap->a_target);
2490 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2491 bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2493 record = hammer_alloc_mem_record(nip, bytes);
2494 record->type = HAMMER_MEM_RECORD_GENERAL;
2496 record->leaf.base.localization = nip->obj_localization +
2497 HAMMER_LOCALIZE_MISC;
2498 record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2499 record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2500 record->leaf.data_len = bytes;
2501 KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2502 bcopy(ap->a_target, record->data->symlink.name, bytes);
2503 error = hammer_ip_add_record(&trans, record);
2507 * Set the file size to the length of the link.
2510 nip->ino_data.size = bytes;
2511 hammer_modify_inode(&trans, nip, HAMMER_INODE_DDIRTY);
2515 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
2516 nch->ncp->nc_nlen, nip);
2522 hammer_rel_inode(nip, 0);
2525 error = hammer_get_vnode(nip, ap->a_vpp);
2526 hammer_rel_inode(nip, 0);
2528 cache_setunresolved(ap->a_nch);
2529 cache_setvp(ap->a_nch, *ap->a_vpp);
2530 hammer_knote(ap->a_dvp, NOTE_WRITE);
2533 hammer_done_transaction(&trans);
2534 lwkt_reltoken(&hmp->fs_token);
2539 * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2543 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2545 struct hammer_transaction trans;
2546 struct hammer_inode *dip;
2550 dip = VTOI(ap->a_dvp);
2553 if (hammer_nohistory(dip) == 0 &&
2554 (error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2558 lwkt_gettoken(&hmp->fs_token);
2559 hammer_start_transaction(&trans, hmp);
2560 ++hammer_stats_file_iopsw;
2561 error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2562 ap->a_cred, ap->a_flags, -1);
2563 hammer_done_transaction(&trans);
2564 lwkt_reltoken(&hmp->fs_token);
2570 * hammer_vop_ioctl { vp, command, data, fflag, cred }
2574 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2576 struct hammer_inode *ip = ap->a_vp->v_data;
2577 hammer_mount_t hmp = ip->hmp;
2580 ++hammer_stats_file_iopsr;
2581 lwkt_gettoken(&hmp->fs_token);
2582 error = hammer_ioctl(ip, ap->a_command, ap->a_data,
2583 ap->a_fflag, ap->a_cred);
2584 lwkt_reltoken(&hmp->fs_token);
2590 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2592 static const struct mountctl_opt extraopt[] = {
2593 { HMNT_NOHISTORY, "nohistory" },
2594 { HMNT_MASTERID, "master" },
2598 struct hammer_mount *hmp;
2605 mp = ap->a_head.a_ops->head.vv_mount;
2606 KKASSERT(mp->mnt_data != NULL);
2607 hmp = (struct hammer_mount *)mp->mnt_data;
2609 lwkt_gettoken(&hmp->fs_token);
2612 case MOUNTCTL_SET_EXPORT:
2613 if (ap->a_ctllen != sizeof(struct export_args))
2616 error = hammer_vfs_export(mp, ap->a_op,
2617 (const struct export_args *)ap->a_ctl);
2619 case MOUNTCTL_MOUNTFLAGS:
2622 * Call standard mountctl VOP function
2623 * so we get user mount flags.
2625 error = vop_stdmountctl(ap);
2629 usedbytes = *ap->a_res;
2631 if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2632 usedbytes += vfs_flagstostr(hmp->hflags, extraopt,
2634 ap->a_buflen - usedbytes,
2638 *ap->a_res += usedbytes;
2642 error = vop_stdmountctl(ap);
2645 lwkt_reltoken(&hmp->fs_token);
2650 * hammer_vop_strategy { vp, bio }
2652 * Strategy call, used for regular file read & write only. Note that the
2653 * bp may represent a cluster.
2655 * To simplify operation and allow better optimizations in the future,
2656 * this code does not make any assumptions with regards to buffer alignment
2661 hammer_vop_strategy(struct vop_strategy_args *ap)
2666 bp = ap->a_bio->bio_buf;
2670 error = hammer_vop_strategy_read(ap);
2673 error = hammer_vop_strategy_write(ap);
2676 bp->b_error = error = EINVAL;
2677 bp->b_flags |= B_ERROR;
2682 /* hammer_dump_dedup_cache(((hammer_inode_t)ap->a_vp->v_data)->hmp); */
2688 * Read from a regular file. Iterate the related records and fill in the
2689 * BIO/BUF. Gaps are zero-filled.
2691 * The support code in hammer_object.c should be used to deal with mixed
2692 * in-memory and on-disk records.
2694 * NOTE: Can be called from the cluster code with an oversized buf.
2700 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2702 struct hammer_transaction trans;
2703 struct hammer_inode *ip;
2704 struct hammer_inode *dip;
2706 struct hammer_cursor cursor;
2707 hammer_base_elm_t base;
2708 hammer_off_t disk_offset;
2723 ip = ap->a_vp->v_data;
2727 * The zone-2 disk offset may have been set by the cluster code via
2728 * a BMAP operation, or else should be NOOFFSET.
2730 * Checking the high bits for a match against zone-2 should suffice.
2732 * In cases where a lot of data duplication is present it may be
2733 * more beneficial to drop through and doubule-buffer through the
2736 nbio = push_bio(bio);
2737 if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2738 HAMMER_ZONE_LARGE_DATA) {
2739 if (hammer_double_buffer == 0) {
2740 lwkt_gettoken(&hmp->fs_token);
2741 error = hammer_io_direct_read(hmp, nbio, NULL);
2742 lwkt_reltoken(&hmp->fs_token);
2747 * Try to shortcut requests for double_buffer mode too.
2748 * Since this mode runs through the device buffer cache
2749 * only compatible buffer sizes (meaning those generated
2750 * by normal filesystem buffers) are legal.
2752 if (hammer_live_dedup == 0 && (bp->b_flags & B_PAGING) == 0) {
2753 lwkt_gettoken(&hmp->fs_token);
2754 error = hammer_io_indirect_read(hmp, nbio, NULL);
2755 lwkt_reltoken(&hmp->fs_token);
2761 * Well, that sucked. Do it the hard way. If all the stars are
2762 * aligned we may still be able to issue a direct-read.
2764 lwkt_gettoken(&hmp->fs_token);
2765 hammer_simple_transaction(&trans, hmp);
2766 hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2769 * Key range (begin and end inclusive) to scan. Note that the key's
2770 * stored in the actual records represent BASE+LEN, not BASE. The
2771 * first record containing bio_offset will have a key > bio_offset.
2773 cursor.key_beg.localization = ip->obj_localization +
2774 HAMMER_LOCALIZE_MISC;
2775 cursor.key_beg.obj_id = ip->obj_id;
2776 cursor.key_beg.create_tid = 0;
2777 cursor.key_beg.delete_tid = 0;
2778 cursor.key_beg.obj_type = 0;
2779 cursor.key_beg.key = bio->bio_offset + 1;
2780 cursor.asof = ip->obj_asof;
2781 cursor.flags |= HAMMER_CURSOR_ASOF;
2783 cursor.key_end = cursor.key_beg;
2784 KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2786 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2787 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2788 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2789 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2793 ran_end = bio->bio_offset + bp->b_bufsize;
2794 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2795 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2796 tmp64 = ran_end + MAXPHYS + 1; /* work-around GCC-4 bug */
2797 if (tmp64 < ran_end)
2798 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2800 cursor.key_end.key = ran_end + MAXPHYS + 1;
2802 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2805 * Set NOSWAPCACHE for cursor data extraction if double buffering
2806 * is disabled or (if the file is not marked cacheable via chflags
2807 * and vm.swapcache_use_chflags is enabled).
2809 if (hammer_double_buffer == 0 ||
2810 ((ap->a_vp->v_flag & VSWAPCACHE) == 0 &&
2811 vm_swapcache_use_chflags)) {
2812 cursor.flags |= HAMMER_CURSOR_NOSWAPCACHE;
2815 error = hammer_ip_first(&cursor);
2818 while (error == 0) {
2820 * Get the base file offset of the record. The key for
2821 * data records is (base + bytes) rather then (base).
2823 base = &cursor.leaf->base;
2824 rec_offset = base->key - cursor.leaf->data_len;
2827 * Calculate the gap, if any, and zero-fill it.
2829 * n is the offset of the start of the record verses our
2830 * current seek offset in the bio.
2832 n = (int)(rec_offset - (bio->bio_offset + boff));
2834 if (n > bp->b_bufsize - boff)
2835 n = bp->b_bufsize - boff;
2836 bzero((char *)bp->b_data + boff, n);
2842 * Calculate the data offset in the record and the number
2843 * of bytes we can copy.
2845 * There are two degenerate cases. First, boff may already
2846 * be at bp->b_bufsize. Secondly, the data offset within
2847 * the record may exceed the record's size.
2851 n = cursor.leaf->data_len - roff;
2853 kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2855 } else if (n > bp->b_bufsize - boff) {
2856 n = bp->b_bufsize - boff;
2860 * Deal with cached truncations. This cool bit of code
2861 * allows truncate()/ftruncate() to avoid having to sync
2864 * If the frontend is truncated then all backend records are
2865 * subject to the frontend's truncation.
2867 * If the backend is truncated then backend records on-disk
2868 * (but not in-memory) are subject to the backend's
2869 * truncation. In-memory records owned by the backend
2870 * represent data written after the truncation point on the
2871 * backend and must not be truncated.
2873 * Truncate operations deal with frontend buffer cache
2874 * buffers and frontend-owned in-memory records synchronously.
2876 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2877 if (hammer_cursor_ondisk(&cursor)/* ||
2878 cursor.iprec->flush_state == HAMMER_FST_FLUSH*/) {
2879 if (ip->trunc_off <= rec_offset)
2881 else if (ip->trunc_off < rec_offset + n)
2882 n = (int)(ip->trunc_off - rec_offset);
2885 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2886 if (hammer_cursor_ondisk(&cursor)) {
2887 if (ip->sync_trunc_off <= rec_offset)
2889 else if (ip->sync_trunc_off < rec_offset + n)
2890 n = (int)(ip->sync_trunc_off - rec_offset);
2895 * Try to issue a direct read into our bio if possible,
2896 * otherwise resolve the element data into a hammer_buffer
2899 * The buffer on-disk should be zerod past any real
2900 * truncation point, but may not be for any synthesized
2901 * truncation point from above.
2903 * NOTE: disk_offset is only valid if the cursor data is
2906 disk_offset = cursor.leaf->data_offset + roff;
2907 isdedupable = (boff == 0 && n == bp->b_bufsize &&
2908 hammer_cursor_ondisk(&cursor) &&
2909 ((int)disk_offset & HAMMER_BUFMASK) == 0);
2911 if (isdedupable && hammer_double_buffer == 0) {
2915 KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2916 HAMMER_ZONE_LARGE_DATA);
2917 nbio->bio_offset = disk_offset;
2918 error = hammer_io_direct_read(hmp, nbio, cursor.leaf);
2919 if (hammer_live_dedup && error == 0)
2920 hammer_dedup_cache_add(ip, cursor.leaf);
2922 } else if (isdedupable) {
2924 * Async I/O case for reading from backing store
2925 * and copying the data to the filesystem buffer.
2926 * live-dedup has to verify the data anyway if it
2927 * gets a hit later so we can just add the entry
2930 KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2931 HAMMER_ZONE_LARGE_DATA);
2932 nbio->bio_offset = disk_offset;
2933 if (hammer_live_dedup)
2934 hammer_dedup_cache_add(ip, cursor.leaf);
2935 error = hammer_io_indirect_read(hmp, nbio, cursor.leaf);
2938 error = hammer_ip_resolve_data(&cursor);
2940 if (hammer_live_dedup && isdedupable)
2941 hammer_dedup_cache_add(ip, cursor.leaf);
2942 bcopy((char *)cursor.data + roff,
2943 (char *)bp->b_data + boff, n);
2950 * We have to be sure that the only elements added to the
2951 * dedup cache are those which are already on-media.
2953 if (hammer_live_dedup && hammer_cursor_ondisk(&cursor))
2954 hammer_dedup_cache_add(ip, cursor.leaf);
2957 * Iterate until we have filled the request.
2960 if (boff == bp->b_bufsize)
2962 error = hammer_ip_next(&cursor);
2966 * There may have been a gap after the last record
2968 if (error == ENOENT)
2970 if (error == 0 && boff != bp->b_bufsize) {
2971 KKASSERT(boff < bp->b_bufsize);
2972 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2973 /* boff = bp->b_bufsize; */
2977 * Disallow swapcache operation on the vnode buffer if double
2978 * buffering is enabled, the swapcache will get the data via
2979 * the block device buffer.
2981 if (hammer_double_buffer)
2982 bp->b_flags |= B_NOTMETA;
2988 bp->b_error = error;
2990 bp->b_flags |= B_ERROR;
2995 * Cache the b-tree node for the last data read in cache[1].
2997 * If we hit the file EOF then also cache the node in the
2998 * governing director's cache[3], it will be used to initialize
2999 * the inode's cache[1] for any inodes looked up via the directory.
3001 * This doesn't reduce disk accesses since the B-Tree chain is
3002 * likely cached, but it does reduce cpu overhead when looking
3003 * up file offsets for cpdup/tar/cpio style iterations.
3006 hammer_cache_node(&ip->cache[1], cursor.node);
3007 if (ran_end >= ip->ino_data.size) {
3008 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
3009 ip->obj_asof, ip->obj_localization);
3011 hammer_cache_node(&dip->cache[3], cursor.node);
3012 hammer_rel_inode(dip, 0);
3015 hammer_done_cursor(&cursor);
3016 hammer_done_transaction(&trans);
3017 lwkt_reltoken(&hmp->fs_token);
3022 * BMAP operation - used to support cluster_read() only.
3024 * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
3026 * This routine may return EOPNOTSUPP if the opration is not supported for
3027 * the specified offset. The contents of the pointer arguments do not
3028 * need to be initialized in that case.
3030 * If a disk address is available and properly aligned return 0 with
3031 * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
3032 * to the run-length relative to that offset. Callers may assume that
3033 * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
3034 * large, so return EOPNOTSUPP if it is not sufficiently large.
3038 hammer_vop_bmap(struct vop_bmap_args *ap)
3040 struct hammer_transaction trans;
3041 struct hammer_inode *ip;
3043 struct hammer_cursor cursor;
3044 hammer_base_elm_t base;
3048 int64_t base_offset;
3049 int64_t base_disk_offset;
3050 int64_t last_offset;
3051 hammer_off_t last_disk_offset;
3052 hammer_off_t disk_offset;
3057 ++hammer_stats_file_iopsr;
3058 ip = ap->a_vp->v_data;
3062 * We can only BMAP regular files. We can't BMAP database files,
3065 if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
3069 * bmap is typically called with runp/runb both NULL when used
3070 * for writing. We do not support BMAP for writing atm.
3072 if (ap->a_cmd != BUF_CMD_READ)
3076 * Scan the B-Tree to acquire blockmap addresses, then translate
3079 lwkt_gettoken(&hmp->fs_token);
3080 hammer_simple_transaction(&trans, hmp);
3082 kprintf("bmap_beg %016llx ip->cache %p\n",
3083 (long long)ap->a_loffset, ip->cache[1]);
3085 hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
3088 * Key range (begin and end inclusive) to scan. Note that the key's
3089 * stored in the actual records represent BASE+LEN, not BASE. The
3090 * first record containing bio_offset will have a key > bio_offset.
3092 cursor.key_beg.localization = ip->obj_localization +
3093 HAMMER_LOCALIZE_MISC;
3094 cursor.key_beg.obj_id = ip->obj_id;
3095 cursor.key_beg.create_tid = 0;
3096 cursor.key_beg.delete_tid = 0;
3097 cursor.key_beg.obj_type = 0;
3099 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
3101 cursor.key_beg.key = ap->a_loffset + 1;
3102 if (cursor.key_beg.key < 0)
3103 cursor.key_beg.key = 0;
3104 cursor.asof = ip->obj_asof;
3105 cursor.flags |= HAMMER_CURSOR_ASOF;
3107 cursor.key_end = cursor.key_beg;
3108 KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
3110 ran_end = ap->a_loffset + MAXPHYS;
3111 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
3112 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
3113 tmp64 = ran_end + MAXPHYS + 1; /* work-around GCC-4 bug */
3114 if (tmp64 < ran_end)
3115 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
3117 cursor.key_end.key = ran_end + MAXPHYS + 1;
3119 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
3121 error = hammer_ip_first(&cursor);
3122 base_offset = last_offset = 0;
3123 base_disk_offset = last_disk_offset = 0;
3125 while (error == 0) {
3127 * Get the base file offset of the record. The key for
3128 * data records is (base + bytes) rather then (base).
3130 * NOTE: rec_offset + rec_len may exceed the end-of-file.
3131 * The extra bytes should be zero on-disk and the BMAP op
3132 * should still be ok.
3134 base = &cursor.leaf->base;
3135 rec_offset = base->key - cursor.leaf->data_len;
3136 rec_len = cursor.leaf->data_len;
3139 * Incorporate any cached truncation.
3141 * NOTE: Modifications to rec_len based on synthesized
3142 * truncation points remove the guarantee that any extended
3143 * data on disk is zero (since the truncations may not have
3144 * taken place on-media yet).
3146 if (ip->flags & HAMMER_INODE_TRUNCATED) {
3147 if (hammer_cursor_ondisk(&cursor) ||
3148 cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
3149 if (ip->trunc_off <= rec_offset)
3151 else if (ip->trunc_off < rec_offset + rec_len)
3152 rec_len = (int)(ip->trunc_off - rec_offset);
3155 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
3156 if (hammer_cursor_ondisk(&cursor)) {
3157 if (ip->sync_trunc_off <= rec_offset)
3159 else if (ip->sync_trunc_off < rec_offset + rec_len)
3160 rec_len = (int)(ip->sync_trunc_off - rec_offset);
3165 * Accumulate information. If we have hit a discontiguous
3166 * block reset base_offset unless we are already beyond the
3167 * requested offset. If we are, that's it, we stop.
3171 if (hammer_cursor_ondisk(&cursor)) {
3172 disk_offset = cursor.leaf->data_offset;
3173 if (rec_offset != last_offset ||
3174 disk_offset != last_disk_offset) {
3175 if (rec_offset > ap->a_loffset)
3177 base_offset = rec_offset;
3178 base_disk_offset = disk_offset;
3180 last_offset = rec_offset + rec_len;
3181 last_disk_offset = disk_offset + rec_len;
3183 if (hammer_live_dedup)
3184 hammer_dedup_cache_add(ip, cursor.leaf);
3187 error = hammer_ip_next(&cursor);
3191 kprintf("BMAP %016llx: %016llx - %016llx\n",
3192 (long long)ap->a_loffset,
3193 (long long)base_offset,
3194 (long long)last_offset);
3195 kprintf("BMAP %16s: %016llx - %016llx\n", "",
3196 (long long)base_disk_offset,
3197 (long long)last_disk_offset);
3201 hammer_cache_node(&ip->cache[1], cursor.node);
3203 kprintf("bmap_end2 %016llx ip->cache %p\n",
3204 (long long)ap->a_loffset, ip->cache[1]);
3207 hammer_done_cursor(&cursor);
3208 hammer_done_transaction(&trans);
3209 lwkt_reltoken(&hmp->fs_token);
3212 * If we couldn't find any records or the records we did find were
3213 * all behind the requested offset, return failure. A forward
3214 * truncation can leave a hole w/ no on-disk records.
3216 if (last_offset == 0 || last_offset < ap->a_loffset)
3217 return (EOPNOTSUPP);
3220 * Figure out the block size at the requested offset and adjust
3221 * our limits so the cluster_read() does not create inappropriately
3222 * sized buffer cache buffers.
3224 blksize = hammer_blocksize(ap->a_loffset);
3225 if (hammer_blocksize(base_offset) != blksize) {
3226 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
3228 if (last_offset != ap->a_loffset &&
3229 hammer_blocksize(last_offset - 1) != blksize) {
3230 last_offset = hammer_blockdemarc(ap->a_loffset,
3235 * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3238 disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
3240 if ((disk_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_LARGE_DATA) {
3242 * Only large-data zones can be direct-IOd
3245 } else if ((disk_offset & HAMMER_BUFMASK) ||
3246 (last_offset - ap->a_loffset) < blksize) {
3248 * doffsetp is not aligned or the forward run size does
3249 * not cover a whole buffer, disallow the direct I/O.
3256 *ap->a_doffsetp = disk_offset;
3258 *ap->a_runb = ap->a_loffset - base_offset;
3259 KKASSERT(*ap->a_runb >= 0);
3262 *ap->a_runp = last_offset - ap->a_loffset;
3263 KKASSERT(*ap->a_runp >= 0);
3271 * Write to a regular file. Because this is a strategy call the OS is
3272 * trying to actually get data onto the media.
3276 hammer_vop_strategy_write(struct vop_strategy_args *ap)
3278 hammer_record_t record;
3283 int blksize __debugvar;
3289 ip = ap->a_vp->v_data;
3292 blksize = hammer_blocksize(bio->bio_offset);
3293 KKASSERT(bp->b_bufsize == blksize);
3295 if (ip->flags & HAMMER_INODE_RO) {
3296 bp->b_error = EROFS;
3297 bp->b_flags |= B_ERROR;
3302 lwkt_gettoken(&hmp->fs_token);
3305 * Disallow swapcache operation on the vnode buffer if double
3306 * buffering is enabled, the swapcache will get the data via
3307 * the block device buffer.
3309 if (hammer_double_buffer)
3310 bp->b_flags |= B_NOTMETA;
3313 * Interlock with inode destruction (no in-kernel or directory
3314 * topology visibility). If we queue new IO while trying to
3315 * destroy the inode we can deadlock the vtrunc call in
3316 * hammer_inode_unloadable_check().
3318 * Besides, there's no point flushing a bp associated with an
3319 * inode that is being destroyed on-media and has no kernel
3322 if ((ip->flags | ip->sync_flags) &
3323 (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
3326 lwkt_reltoken(&hmp->fs_token);
3331 * Reserve space and issue a direct-write from the front-end.
3332 * NOTE: The direct_io code will hammer_bread/bcopy smaller
3335 * An in-memory record will be installed to reference the storage
3336 * until the flusher can get to it.
3338 * Since we own the high level bio the front-end will not try to
3339 * do a direct-read until the write completes.
3341 * NOTE: The only time we do not reserve a full-sized buffers
3342 * worth of data is if the file is small. We do not try to
3343 * allocate a fragment (from the small-data zone) at the end of
3344 * an otherwise large file as this can lead to wildly separated
3347 KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
3348 KKASSERT(bio->bio_offset < ip->ino_data.size);
3349 if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
3350 bytes = bp->b_bufsize;
3352 bytes = ((int)ip->ino_data.size + 15) & ~15;
3354 record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
3358 * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3359 * in hammer_vop_write(). We must flag the record so the proper
3360 * REDO_TERM_WRITE entry is generated during the flush.
3363 if (bp->b_flags & B_VFSFLAG1) {
3364 record->flags |= HAMMER_RECF_REDO;
3365 bp->b_flags &= ~B_VFSFLAG1;
3367 if (record->flags & HAMMER_RECF_DEDUPED) {
3369 hammer_ip_replace_bulk(hmp, record);
3372 hammer_io_direct_write(hmp, bio, record);
3374 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
3375 hammer_flush_inode(ip, 0);
3377 bp->b_bio2.bio_offset = NOOFFSET;
3378 bp->b_error = error;
3379 bp->b_flags |= B_ERROR;
3382 lwkt_reltoken(&hmp->fs_token);
3387 * dounlink - disconnect a directory entry
3389 * XXX whiteout support not really in yet
3392 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
3393 struct vnode *dvp, struct ucred *cred,
3394 int flags, int isdir)
3396 struct namecache *ncp;
3400 struct hammer_cursor cursor;
3402 u_int32_t max_iterations;
3406 * Calculate the namekey and setup the key range for the scan. This
3407 * works kinda like a chained hash table where the lower 32 bits
3408 * of the namekey synthesize the chain.
3410 * The key range is inclusive of both key_beg and key_end.
3416 if (dip->flags & HAMMER_INODE_RO)
3419 namekey = hammer_directory_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3422 hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3423 cursor.key_beg.localization = dip->obj_localization +
3424 hammer_dir_localization(dip);
3425 cursor.key_beg.obj_id = dip->obj_id;
3426 cursor.key_beg.key = namekey;
3427 cursor.key_beg.create_tid = 0;
3428 cursor.key_beg.delete_tid = 0;
3429 cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3430 cursor.key_beg.obj_type = 0;
3432 cursor.key_end = cursor.key_beg;
3433 cursor.key_end.key += max_iterations;
3434 cursor.asof = dip->obj_asof;
3435 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3438 * Scan all matching records (the chain), locate the one matching
3439 * the requested path component. info->last_error contains the
3440 * error code on search termination and could be 0, ENOENT, or
3443 * The hammer_ip_*() functions merge in-memory records with on-disk
3444 * records for the purposes of the search.
3446 error = hammer_ip_first(&cursor);
3448 while (error == 0) {
3449 error = hammer_ip_resolve_data(&cursor);
3452 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3454 if (ncp->nc_nlen == nlen &&
3455 bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3458 error = hammer_ip_next(&cursor);
3462 * If all is ok we have to get the inode so we can adjust nlinks.
3463 * To avoid a deadlock with the flusher we must release the inode
3464 * lock on the directory when acquiring the inode for the entry.
3466 * If the target is a directory, it must be empty.
3469 hammer_unlock(&cursor.ip->lock);
3470 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3472 cursor.data->entry.localization,
3474 hammer_lock_sh(&cursor.ip->lock);
3475 if (error == ENOENT) {
3476 kprintf("HAMMER: WARNING: Removing "
3477 "dirent w/missing inode \"%s\"\n"
3478 "\tobj_id = %016llx\n",
3480 (long long)cursor.data->entry.obj_id);
3485 * If isdir >= 0 we validate that the entry is or is not a
3486 * directory. If isdir < 0 we don't care.
3488 if (error == 0 && isdir >= 0 && ip) {
3490 ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3492 } else if (isdir == 0 &&
3493 ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3499 * If we are trying to remove a directory the directory must
3502 * The check directory code can loop and deadlock/retry. Our
3503 * own cursor's node locks must be released to avoid a 3-way
3504 * deadlock with the flusher if the check directory code
3507 * If any changes whatsoever have been made to the cursor
3508 * set EDEADLK and retry.
3510 * WARNING: See warnings in hammer_unlock_cursor()
3513 if (error == 0 && ip && ip->ino_data.obj_type ==
3514 HAMMER_OBJTYPE_DIRECTORY) {
3515 hammer_unlock_cursor(&cursor);
3516 error = hammer_ip_check_directory_empty(trans, ip);
3517 hammer_lock_cursor(&cursor);
3518 if (cursor.flags & HAMMER_CURSOR_RETEST) {
3519 kprintf("HAMMER: Warning: avoided deadlock "
3527 * Delete the directory entry.
3529 * WARNING: hammer_ip_del_directory() may have to terminate
3530 * the cursor to avoid a deadlock. It is ok to call
3531 * hammer_done_cursor() twice.
3534 error = hammer_ip_del_directory(trans, &cursor,
3537 hammer_done_cursor(&cursor);
3540 * Tell the namecache that we are now unlinked.
3545 * NOTE: ip->vp, if non-NULL, cannot be directly
3546 * referenced without formally acquiring the
3547 * vp since the vp might have zero refs on it,
3548 * or in the middle of a reclaim, etc.
3550 * NOTE: The cache_setunresolved() can rip the vp
3551 * out from under us since the vp may not have
3552 * any refs, in which case ip->vp will be NULL
3555 while (ip && ip->vp) {
3558 error = hammer_get_vnode(ip, &vp);
3559 if (error == 0 && vp) {
3561 hammer_knote(ip->vp, NOTE_DELETE);
3564 * Don't do this, it can deadlock
3565 * on concurrent rm's of hardlinks.
3566 * Shouldn't be needed any more.
3568 cache_inval_vp(ip->vp, CINV_DESTROY);
3573 kprintf("Debug: HAMMER ip/vp race1 avoided\n");
3577 hammer_rel_inode(ip, 0);
3579 hammer_done_cursor(&cursor);
3581 if (error == EDEADLK)
3587 /************************************************************************
3588 * FIFO AND SPECFS OPS *
3589 ************************************************************************
3593 hammer_vop_fifoclose (struct vop_close_args *ap)
3595 /* XXX update itimes */
3596 return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3600 hammer_vop_fiforead (struct vop_read_args *ap)
3604 error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3605 /* XXX update access time */
3610 hammer_vop_fifowrite (struct vop_write_args *ap)
3614 error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3615 /* XXX update access time */
3621 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3625 error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3627 error = hammer_vop_kqfilter(ap);
3631 /************************************************************************
3633 ************************************************************************
3636 static void filt_hammerdetach(struct knote *kn);
3637 static int filt_hammerread(struct knote *kn, long hint);
3638 static int filt_hammerwrite(struct knote *kn, long hint);
3639 static int filt_hammervnode(struct knote *kn, long hint);
3641 static struct filterops hammerread_filtops =
3642 { FILTEROP_ISFD | FILTEROP_MPSAFE,
3643 NULL, filt_hammerdetach, filt_hammerread };
3644 static struct filterops hammerwrite_filtops =
3645 { FILTEROP_ISFD | FILTEROP_MPSAFE,
3646 NULL, filt_hammerdetach, filt_hammerwrite };
3647 static struct filterops hammervnode_filtops =
3648 { FILTEROP_ISFD | FILTEROP_MPSAFE,
3649 NULL, filt_hammerdetach, filt_hammervnode };
3653 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3655 struct vnode *vp = ap->a_vp;
3656 struct knote *kn = ap->a_kn;
3658 switch (kn->kn_filter) {
3660 kn->kn_fop = &hammerread_filtops;
3663 kn->kn_fop = &hammerwrite_filtops;
3666 kn->kn_fop = &hammervnode_filtops;
3669 return (EOPNOTSUPP);
3672 kn->kn_hook = (caddr_t)vp;
3674 knote_insert(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3680 filt_hammerdetach(struct knote *kn)
3682 struct vnode *vp = (void *)kn->kn_hook;
3684 knote_remove(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3688 filt_hammerread(struct knote *kn, long hint)
3690 struct vnode *vp = (void *)kn->kn_hook;
3691 hammer_inode_t ip = VTOI(vp);
3692 hammer_mount_t hmp = ip->hmp;
3695 if (hint == NOTE_REVOKE) {
3696 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3699 lwkt_gettoken(&hmp->fs_token); /* XXX use per-ip-token */
3700 off = ip->ino_data.size - kn->kn_fp->f_offset;
3701 kn->kn_data = (off < INTPTR_MAX) ? off : INTPTR_MAX;
3702 lwkt_reltoken(&hmp->fs_token);
3703 if (kn->kn_sfflags & NOTE_OLDAPI)
3705 return (kn->kn_data != 0);
3709 filt_hammerwrite(struct knote *kn, long hint)
3711 if (hint == NOTE_REVOKE)
3712 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3718 filt_hammervnode(struct knote *kn, long hint)
3720 if (kn->kn_sfflags & hint)
3721 kn->kn_fflags |= hint;
3722 if (hint == NOTE_REVOKE) {
3723 kn->kn_flags |= (EV_EOF | EV_NODATA);
3726 return (kn->kn_fflags != 0);