2 * Copyright (c) 2004 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
34 * $DragonFly: src/sys/kern/vfs_journal.c,v 1.17 2005/07/06 06:02:22 dillon Exp $
37 * Each mount point may have zero or more independantly configured journals
38 * attached to it. Each journal is represented by a memory FIFO and worker
39 * thread. Journal events are streamed through the FIFO to the thread,
40 * batched up (typically on one-second intervals), and written out by the
43 * Journal vnode ops are executed instead of mnt_vn_norm_ops when one or
44 * more journals have been installed on a mount point. It becomes the
45 * responsibility of the journal op to call the underlying normal op as
48 * The journaling protocol is intended to evolve into a two-way stream
49 * whereby transaction IDs can be acknowledged by the journaling target
50 * when the data has been committed to hard storage. Both implicit and
51 * explicit acknowledgement schemes will be supported, depending on the
52 * sophistication of the journaling stream, plus resynchronization and
53 * restart when a journaling stream is interrupted. This information will
54 * also be made available to journaling-aware filesystems to allow better
55 * management of their own physical storage synchronization mechanisms as
56 * well as to allow such filesystems to take direct advantage of the kernel's
57 * journaling layer so they don't have to roll their own.
59 * In addition, the worker thread will have access to much larger
60 * spooling areas then the memory buffer is able to provide by e.g.
61 * reserving swap space, in order to absorb potentially long interruptions
62 * of off-site journaling streams, and to prevent 'slow' off-site linkages
63 * from radically slowing down local filesystem operations.
65 * Because of the non-trivial algorithms the journaling system will be
66 * required to support, use of a worker thread is mandatory. Efficiencies
67 * are maintained by utilitizing the memory FIFO to batch transactions when
68 * possible, reducing the number of gratuitous thread switches and taking
69 * advantage of cpu caches through the use of shorter batched code paths
70 * rather then trying to do everything in the context of the process
71 * originating the filesystem op. In the future the memory FIFO can be
72 * made per-cpu to remove BGL or other locking requirements.
74 #include <sys/param.h>
75 #include <sys/systm.h>
78 #include <sys/kernel.h>
79 #include <sys/queue.h>
81 #include <sys/malloc.h>
82 #include <sys/mount.h>
83 #include <sys/unistd.h>
84 #include <sys/vnode.h>
86 #include <sys/mountctl.h>
87 #include <sys/journal.h>
90 #include <sys/msfbuf.h>
92 #include <machine/limits.h>
95 #include <vm/vm_object.h>
96 #include <vm/vm_page.h>
97 #include <vm/vm_pager.h>
98 #include <vm/vnode_pager.h>
100 #include <sys/file2.h>
101 #include <sys/thread2.h>
103 static int journal_attach(struct mount *mp);
104 static void journal_detach(struct mount *mp);
105 static int journal_install_vfs_journal(struct mount *mp, struct file *fp,
106 const struct mountctl_install_journal *info);
107 static int journal_remove_vfs_journal(struct mount *mp,
108 const struct mountctl_remove_journal *info);
109 static int journal_destroy(struct mount *mp, struct journal *jo, int flags);
110 static int journal_resync_vfs_journal(struct mount *mp, const void *ctl);
111 static int journal_status_vfs_journal(struct mount *mp,
112 const struct mountctl_status_journal *info,
113 struct mountctl_journal_ret_status *rstat,
114 int buflen, int *res);
115 static void journal_wthread(void *info);
116 static void journal_rthread(void *info);
118 static void *journal_reserve(struct journal *jo,
119 struct journal_rawrecbeg **rawpp,
120 int16_t streamid, int bytes);
121 static void *journal_extend(struct journal *jo,
122 struct journal_rawrecbeg **rawpp,
123 int truncbytes, int bytes, int *newstreamrecp);
124 static void journal_abort(struct journal *jo,
125 struct journal_rawrecbeg **rawpp);
126 static void journal_commit(struct journal *jo,
127 struct journal_rawrecbeg **rawpp,
128 int bytes, int closeout);
130 static void jrecord_init(struct journal *jo,
131 struct jrecord *jrec, int16_t streamid);
132 static struct journal_subrecord *jrecord_push(
133 struct jrecord *jrec, int16_t rectype);
134 static void jrecord_pop(struct jrecord *jrec, struct journal_subrecord *parent);
135 static struct journal_subrecord *jrecord_write(struct jrecord *jrec,
136 int16_t rectype, int bytes);
137 static void jrecord_data(struct jrecord *jrec, const void *buf, int bytes);
138 static void jrecord_done(struct jrecord *jrec, int abortit);
140 static int journal_setattr(struct vop_setattr_args *ap);
141 static int journal_write(struct vop_write_args *ap);
142 static int journal_fsync(struct vop_fsync_args *ap);
143 static int journal_putpages(struct vop_putpages_args *ap);
144 static int journal_setacl(struct vop_setacl_args *ap);
145 static int journal_setextattr(struct vop_setextattr_args *ap);
146 static int journal_ncreate(struct vop_ncreate_args *ap);
147 static int journal_nmknod(struct vop_nmknod_args *ap);
148 static int journal_nlink(struct vop_nlink_args *ap);
149 static int journal_nsymlink(struct vop_nsymlink_args *ap);
150 static int journal_nwhiteout(struct vop_nwhiteout_args *ap);
151 static int journal_nremove(struct vop_nremove_args *ap);
152 static int journal_nmkdir(struct vop_nmkdir_args *ap);
153 static int journal_nrmdir(struct vop_nrmdir_args *ap);
154 static int journal_nrename(struct vop_nrename_args *ap);
156 static struct vnodeopv_entry_desc journal_vnodeop_entries[] = {
157 { &vop_default_desc, vop_journal_operate_ap },
158 { &vop_mountctl_desc, (void *)journal_mountctl },
159 { &vop_setattr_desc, (void *)journal_setattr },
160 { &vop_write_desc, (void *)journal_write },
161 { &vop_fsync_desc, (void *)journal_fsync },
162 { &vop_putpages_desc, (void *)journal_putpages },
163 { &vop_setacl_desc, (void *)journal_setacl },
164 { &vop_setextattr_desc, (void *)journal_setextattr },
165 { &vop_ncreate_desc, (void *)journal_ncreate },
166 { &vop_nmknod_desc, (void *)journal_nmknod },
167 { &vop_nlink_desc, (void *)journal_nlink },
168 { &vop_nsymlink_desc, (void *)journal_nsymlink },
169 { &vop_nwhiteout_desc, (void *)journal_nwhiteout },
170 { &vop_nremove_desc, (void *)journal_nremove },
171 { &vop_nmkdir_desc, (void *)journal_nmkdir },
172 { &vop_nrmdir_desc, (void *)journal_nrmdir },
173 { &vop_nrename_desc, (void *)journal_nrename },
177 static MALLOC_DEFINE(M_JOURNAL, "journal", "Journaling structures");
178 static MALLOC_DEFINE(M_JFIFO, "journal-fifo", "Journal FIFO");
181 journal_mountctl(struct vop_mountctl_args *ap)
186 mp = ap->a_head.a_ops->vv_mount;
189 if (mp->mnt_vn_journal_ops == NULL) {
191 case MOUNTCTL_INSTALL_VFS_JOURNAL:
192 error = journal_attach(mp);
193 if (error == 0 && ap->a_ctllen != sizeof(struct mountctl_install_journal))
195 if (error == 0 && ap->a_fp == NULL)
198 error = journal_install_vfs_journal(mp, ap->a_fp, ap->a_ctl);
199 if (TAILQ_EMPTY(&mp->mnt_jlist))
202 case MOUNTCTL_REMOVE_VFS_JOURNAL:
203 case MOUNTCTL_RESYNC_VFS_JOURNAL:
204 case MOUNTCTL_STATUS_VFS_JOURNAL:
213 case MOUNTCTL_INSTALL_VFS_JOURNAL:
214 if (ap->a_ctllen != sizeof(struct mountctl_install_journal))
216 if (error == 0 && ap->a_fp == NULL)
219 error = journal_install_vfs_journal(mp, ap->a_fp, ap->a_ctl);
221 case MOUNTCTL_REMOVE_VFS_JOURNAL:
222 if (ap->a_ctllen != sizeof(struct mountctl_remove_journal))
225 error = journal_remove_vfs_journal(mp, ap->a_ctl);
226 if (TAILQ_EMPTY(&mp->mnt_jlist))
229 case MOUNTCTL_RESYNC_VFS_JOURNAL:
230 if (ap->a_ctllen != 0)
232 error = journal_resync_vfs_journal(mp, ap->a_ctl);
234 case MOUNTCTL_STATUS_VFS_JOURNAL:
235 if (ap->a_ctllen != sizeof(struct mountctl_status_journal))
238 error = journal_status_vfs_journal(mp, ap->a_ctl,
239 ap->a_buf, ap->a_buflen, ap->a_res);
251 * High level mount point setup. When a
254 journal_attach(struct mount *mp)
256 vfs_add_vnodeops(mp, &mp->mnt_vn_journal_ops, journal_vnodeop_entries);
261 journal_detach(struct mount *mp)
263 if (mp->mnt_vn_journal_ops)
264 vfs_rm_vnodeops(&mp->mnt_vn_journal_ops);
268 * Install a journal on a mount point. Each journal has an associated worker
269 * thread which is responsible for buffering and spooling the data to the
270 * target. A mount point may have multiple journals attached to it. An
271 * initial start record is generated when the journal is associated.
274 journal_install_vfs_journal(struct mount *mp, struct file *fp,
275 const struct mountctl_install_journal *info)
282 jo = malloc(sizeof(struct journal), M_JOURNAL, M_WAITOK|M_ZERO);
283 bcopy(info->id, jo->id, sizeof(jo->id));
284 jo->flags = info->flags & ~(MC_JOURNAL_WACTIVE | MC_JOURNAL_RACTIVE |
285 MC_JOURNAL_STOP_REQ);
288 * Memory FIFO size, round to nearest power of 2
290 if (info->membufsize) {
291 if (info->membufsize < 65536)
293 else if (info->membufsize > 128 * 1024 * 1024)
294 size = 128 * 1024 * 1024;
296 size = (int)info->membufsize;
301 while (jo->fifo.size < size)
305 * Other parameters. If not specified the starting transaction id
306 * will be the current date.
309 jo->transid = info->transid;
313 jo->transid = ((int64_t)ts.tv_sec << 30) | ts.tv_nsec;
319 * Allocate the memory FIFO
321 jo->fifo.mask = jo->fifo.size - 1;
322 jo->fifo.membase = malloc(jo->fifo.size, M_JFIFO, M_WAITOK|M_ZERO|M_NULLOK);
323 if (jo->fifo.membase == NULL)
327 * Create the worker threads and generate the association record.
333 jo->flags |= MC_JOURNAL_WACTIVE;
334 lwkt_create(journal_wthread, jo, NULL, &jo->wthread,
335 TDF_STOPREQ, -1, "journal w:%.*s", JIDMAX, jo->id);
336 lwkt_setpri(&jo->wthread, TDPRI_KERN_DAEMON);
337 lwkt_schedule(&jo->wthread);
339 if (jo->flags & MC_JOURNAL_WANT_FULLDUPLEX) {
340 jo->flags |= MC_JOURNAL_RACTIVE;
341 lwkt_create(journal_rthread, jo, NULL, &jo->rthread,
342 TDF_STOPREQ, -1, "journal r:%.*s", JIDMAX, jo->id);
343 lwkt_setpri(&jo->rthread, TDPRI_KERN_DAEMON);
344 lwkt_schedule(&jo->rthread);
346 jrecord_init(jo, &jrec, JREC_STREAMID_DISCONT);
347 jrecord_write(&jrec, JTYPE_ASSOCIATE, 0);
348 jrecord_done(&jrec, 0);
349 TAILQ_INSERT_TAIL(&mp->mnt_jlist, jo, jentry);
355 * Disassociate a journal from a mount point and terminate its worker thread.
356 * A final termination record is written out before the file pointer is
360 journal_remove_vfs_journal(struct mount *mp,
361 const struct mountctl_remove_journal *info)
366 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) {
367 if (bcmp(jo->id, info->id, sizeof(jo->id)) == 0)
371 error = journal_destroy(mp, jo, info->flags);
378 * Remove all journals associated with a mount point. Usually called
379 * by the umount code.
382 journal_remove_all_journals(struct mount *mp, int flags)
386 while ((jo = TAILQ_FIRST(&mp->mnt_jlist)) != NULL) {
387 journal_destroy(mp, jo, flags);
392 journal_destroy(struct mount *mp, struct journal *jo, int flags)
397 TAILQ_REMOVE(&mp->mnt_jlist, jo, jentry);
399 jrecord_init(jo, &jrec, JREC_STREAMID_DISCONT);
400 jrecord_write(&jrec, JTYPE_DISASSOCIATE, 0);
401 jrecord_done(&jrec, 0);
403 jo->flags |= MC_JOURNAL_STOP_REQ | (flags & MC_JOURNAL_STOP_IMM);
406 while (jo->flags & (MC_JOURNAL_WACTIVE | MC_JOURNAL_RACTIVE)) {
407 tsleep(jo, 0, "jwait", hz);
408 if (++wcount % 10 == 0) {
409 printf("Warning: journal %s waiting for descriptors to close\n",
413 lwkt_free_thread(&jo->wthread); /* XXX SMP */
414 if (jo->flags & MC_JOURNAL_WANT_FULLDUPLEX)
415 lwkt_free_thread(&jo->rthread); /* XXX SMP */
417 fdrop(jo->fp, curthread);
418 if (jo->fifo.membase)
419 free(jo->fifo.membase, M_JFIFO);
425 journal_resync_vfs_journal(struct mount *mp, const void *ctl)
431 journal_status_vfs_journal(struct mount *mp,
432 const struct mountctl_status_journal *info,
433 struct mountctl_journal_ret_status *rstat,
434 int buflen, int *res)
442 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) {
443 if (info->index == MC_JOURNAL_INDEX_ID) {
444 if (bcmp(jo->id, info->id, sizeof(jo->id)) != 0)
446 } else if (info->index >= 0) {
447 if (info->index < index)
449 } else if (info->index != MC_JOURNAL_INDEX_ALL) {
452 if (buflen < sizeof(*rstat)) {
454 rstat[-1].flags |= MC_JOURNAL_STATUS_MORETOCOME;
459 bzero(rstat, sizeof(*rstat));
460 rstat->recsize = sizeof(*rstat);
461 bcopy(jo->id, rstat->id, sizeof(jo->id));
462 rstat->index = index;
463 rstat->membufsize = jo->fifo.size;
464 rstat->membufused = jo->fifo.windex - jo->fifo.xindex;
465 rstat->membufunacked = jo->fifo.rindex - jo->fifo.xindex;
466 rstat->bytessent = jo->total_acked;
467 rstat->fifostalls = jo->fifostalls;
470 *res += sizeof(*rstat);
471 buflen -= sizeof(*rstat);
477 * The per-journal worker thread is responsible for writing out the
478 * journal's FIFO to the target stream.
481 journal_wthread(void *info)
483 struct journal *jo = info;
484 struct journal_rawrecbeg *rawp;
492 * Calculate the number of bytes available to write. This buffer
493 * area may contain reserved records so we can't just write it out
494 * without further checks.
496 bytes = jo->fifo.windex - jo->fifo.rindex;
499 * sleep if no bytes are available or if an incomplete record is
500 * encountered (it needs to be filled in before we can write it
501 * out), and skip any pad records that we encounter.
504 if (jo->flags & MC_JOURNAL_STOP_REQ)
506 tsleep(&jo->fifo, 0, "jfifo", hz);
511 * Sleep if we can not go any further due to hitting an incomplete
512 * record. This case should occur rarely but may have to be better
515 rawp = (void *)(jo->fifo.membase + (jo->fifo.rindex & jo->fifo.mask));
516 if (rawp->begmagic == JREC_INCOMPLETEMAGIC) {
517 tsleep(&jo->fifo, 0, "jpad", hz);
522 * Skip any pad records. We do not write out pad records if we can
525 if (rawp->streamid == JREC_STREAMID_PAD) {
526 if ((jo->flags & MC_JOURNAL_WANT_FULLDUPLEX) == 0) {
527 if (jo->fifo.rindex == jo->fifo.xindex) {
528 jo->fifo.xindex += (rawp->recsize + 15) & ~15;
529 jo->total_acked += (rawp->recsize + 15) & ~15;
532 jo->fifo.rindex += (rawp->recsize + 15) & ~15;
533 jo->total_acked += bytes;
534 KKASSERT(jo->fifo.windex - jo->fifo.rindex >= 0);
539 * 'bytes' is the amount of data that can potentially be written out.
540 * Calculate 'res', the amount of data that can actually be written
541 * out. res is bounded either by hitting the end of the physical
542 * memory buffer or by hitting an incomplete record. Incomplete
543 * records often occur due to the way the space reservation model
547 avail = jo->fifo.size - (jo->fifo.rindex & jo->fifo.mask);
548 while (res < bytes && rawp->begmagic == JREC_BEGMAGIC) {
549 res += (rawp->recsize + 15) & ~15;
551 KKASSERT(res == avail);
554 rawp = (void *)((char *)rawp + ((rawp->recsize + 15) & ~15));
558 * Issue the write and deal with any errors or other conditions.
559 * For now assume blocking I/O. Since we are record-aware the
560 * code cannot yet handle partial writes.
562 * We bump rindex prior to issuing the write to avoid racing
563 * the acknowledgement coming back (which could prevent the ack
564 * from bumping xindex). Restarts are always based on xindex so
565 * we do not try to undo the rindex if an error occurs.
567 * XXX EWOULDBLOCK/NBIO
568 * XXX notification on failure
569 * XXX permanent verses temporary failures
570 * XXX two-way acknowledgement stream in the return direction / xindex
573 jo->fifo.rindex += bytes;
574 error = fp_write(jo->fp,
575 jo->fifo.membase + ((jo->fifo.rindex - bytes) & jo->fifo.mask),
578 printf("journal_thread(%s) write, error %d\n", jo->id, error);
581 KKASSERT(res == bytes);
585 * Advance rindex. If the journal stream is not full duplex we also
586 * advance xindex, otherwise the rjournal thread is responsible for
589 if ((jo->flags & MC_JOURNAL_WANT_FULLDUPLEX) == 0) {
590 jo->fifo.xindex += bytes;
591 jo->total_acked += bytes;
593 KKASSERT(jo->fifo.windex - jo->fifo.rindex >= 0);
594 if ((jo->flags & MC_JOURNAL_WANT_FULLDUPLEX) == 0) {
595 if (jo->flags & MC_JOURNAL_WWAIT) {
596 jo->flags &= ~MC_JOURNAL_WWAIT; /* XXX hysteresis */
597 wakeup(&jo->fifo.windex);
601 jo->flags &= ~MC_JOURNAL_WACTIVE;
603 wakeup(&jo->fifo.windex);
607 * A second per-journal worker thread is created for two-way journaling
608 * streams to deal with the return acknowledgement stream.
611 journal_rthread(void *info)
613 struct journal_rawrecbeg *rawp;
614 struct journal_ackrecord ack;
615 struct journal *jo = info;
626 * We have been asked to stop
628 if (jo->flags & MC_JOURNAL_STOP_REQ)
632 * If we have no active transaction id, get one from the return
636 error = fp_read(jo->fp, &ack, sizeof(ack), &count, 1);
638 printf("fp_read ack error %d count %d\n", error, count);
640 if (error || count != sizeof(ack))
643 printf("read error %d on receive stream\n", error);
646 if (ack.rbeg.begmagic != JREC_BEGMAGIC ||
647 ack.rend.endmagic != JREC_ENDMAGIC
649 printf("bad begmagic or endmagic on receive stream\n");
652 transid = ack.rbeg.transid;
656 * Calculate the number of unacknowledged bytes. If there are no
657 * unacknowledged bytes then unsent data was acknowledged, report,
658 * sleep a bit, and loop in that case. This should not happen
659 * normally. The ack record is thrown away.
661 bytes = jo->fifo.rindex - jo->fifo.xindex;
664 printf("warning: unsent data acknowledged transid %08llx\n", transid);
665 tsleep(&jo->fifo.xindex, 0, "jrseq", hz);
671 * Since rindex has advanced, the record pointed to by xindex
672 * must be a valid record.
674 rawp = (void *)(jo->fifo.membase + (jo->fifo.xindex & jo->fifo.mask));
675 KKASSERT(rawp->begmagic == JREC_BEGMAGIC);
676 KKASSERT(rawp->recsize <= bytes);
679 * The target can acknowledge several records at once.
681 if (rawp->transid < transid) {
683 printf("ackskip %08llx/%08llx\n", rawp->transid, transid);
685 jo->fifo.xindex += (rawp->recsize + 15) & ~15;
686 jo->total_acked += (rawp->recsize + 15) & ~15;
687 if (jo->flags & MC_JOURNAL_WWAIT) {
688 jo->flags &= ~MC_JOURNAL_WWAIT; /* XXX hysteresis */
689 wakeup(&jo->fifo.windex);
693 if (rawp->transid == transid) {
695 printf("ackskip %08llx/%08llx\n", rawp->transid, transid);
697 jo->fifo.xindex += (rawp->recsize + 15) & ~15;
698 jo->total_acked += (rawp->recsize + 15) & ~15;
699 if (jo->flags & MC_JOURNAL_WWAIT) {
700 jo->flags &= ~MC_JOURNAL_WWAIT; /* XXX hysteresis */
701 wakeup(&jo->fifo.windex);
706 printf("warning: unsent data(2) acknowledged transid %08llx\n", transid);
709 jo->flags &= ~MC_JOURNAL_RACTIVE;
711 wakeup(&jo->fifo.windex);
715 * This builds a pad record which the journaling thread will skip over. Pad
716 * records are required when we are unable to reserve sufficient stream space
717 * due to insufficient space at the end of the physical memory fifo.
719 * Even though the record is not transmitted, a normal transid must be
720 * assigned to it so link recovery operations after a failure work properly.
724 journal_build_pad(struct journal_rawrecbeg *rawp, int recsize, int64_t transid)
726 struct journal_rawrecend *rendp;
728 KKASSERT((recsize & 15) == 0 && recsize >= 16);
730 rawp->streamid = JREC_STREAMID_PAD;
731 rawp->recsize = recsize; /* must be 16-byte aligned */
732 rawp->transid = transid;
734 * WARNING, rendp may overlap rawp->seqno. This is necessary to
735 * allow PAD records to fit in 16 bytes. Use cpu_ccfence() to
736 * hopefully cause the compiler to not make any assumptions.
738 rendp = (void *)((char *)rawp + rawp->recsize - sizeof(*rendp));
739 rendp->endmagic = JREC_ENDMAGIC;
741 rendp->recsize = rawp->recsize;
744 * Set the begin magic last. This is what will allow the journal
745 * thread to write the record out. Use a store fence to prevent
746 * compiler and cpu reordering of the writes.
749 rawp->begmagic = JREC_BEGMAGIC;
753 * Wake up the worker thread if the FIFO is more then half full or if
754 * someone is waiting for space to be freed up. Otherwise let the
755 * heartbeat deal with it. Being able to avoid waking up the worker
756 * is the key to the journal's cpu performance.
760 journal_commit_wakeup(struct journal *jo)
764 avail = jo->fifo.size - (jo->fifo.windex - jo->fifo.xindex);
765 KKASSERT(avail >= 0);
766 if ((avail < (jo->fifo.size >> 1)) || (jo->flags & MC_JOURNAL_WWAIT))
771 * Create a new BEGIN stream record with the specified streamid and the
772 * specified amount of payload space. *rawpp will be set to point to the
773 * base of the new stream record and a pointer to the base of the payload
774 * space will be returned. *rawpp does not need to be pre-NULLd prior to
775 * making this call. The raw record header will be partially initialized.
777 * A stream can be extended, aborted, or committed by other API calls
778 * below. This may result in a sequence of potentially disconnected
779 * stream records to be output to the journaling target. The first record
780 * (the one created by this function) will be marked JREC_STREAMCTL_BEGIN,
781 * while the last record on commit or abort will be marked JREC_STREAMCTL_END
782 * (and possibly also JREC_STREAMCTL_ABORTED). The last record could wind
783 * up being the same as the first, in which case the bits are all set in
786 * The stream record is created in an incomplete state by setting the begin
787 * magic to JREC_INCOMPLETEMAGIC. This prevents the worker thread from
788 * flushing the fifo past our record until we have finished populating it.
789 * Other threads can reserve and operate on their own space without stalling
790 * but the stream output will stall until we have completed operations. The
791 * memory FIFO is intended to be large enough to absorb such situations
792 * without stalling out other threads.
796 journal_reserve(struct journal *jo, struct journal_rawrecbeg **rawpp,
797 int16_t streamid, int bytes)
799 struct journal_rawrecbeg *rawp;
805 * Add header and trailer overheads to the passed payload. Note that
806 * the passed payload size need not be aligned in any way.
808 bytes += sizeof(struct journal_rawrecbeg);
809 bytes += sizeof(struct journal_rawrecend);
813 * First, check boundary conditions. If the request would wrap around
814 * we have to skip past the ending block and return to the beginning
815 * of the FIFO's buffer. Calculate 'req' which is the actual number
816 * of bytes being reserved, including wrap-around dead space.
818 * Neither 'bytes' or 'req' are aligned.
820 * Note that availtoend is not truncated to avail and so cannot be
821 * used to determine whether the reservation is possible by itself.
822 * Also, since all fifo ops are 16-byte aligned, we can check
823 * the size before calculating the aligned size.
825 availtoend = jo->fifo.size - (jo->fifo.windex & jo->fifo.mask);
826 KKASSERT((availtoend & 15) == 0);
827 if (bytes > availtoend)
828 req = bytes + availtoend; /* add pad to end */
833 * Next calculate the total available space and see if it is
834 * sufficient. We cannot overwrite previously buffered data
835 * past xindex because otherwise we would not be able to restart
836 * a broken link at the target's last point of commit.
838 avail = jo->fifo.size - (jo->fifo.windex - jo->fifo.xindex);
839 KKASSERT(avail >= 0 && (avail & 15) == 0);
842 /* XXX MC_JOURNAL_STOP_IMM */
843 jo->flags |= MC_JOURNAL_WWAIT;
845 tsleep(&jo->fifo.windex, 0, "jwrite", 0);
850 * Create a pad record for any dead space and create an incomplete
851 * record for the live space, then return a pointer to the
852 * contiguous buffer space that was requested.
854 * NOTE: The worker thread will not flush past an incomplete
855 * record, so the reserved space can be filled in at-will. The
856 * journaling code must also be aware the reserved sections occuring
857 * after this one will also not be written out even if completed
858 * until this one is completed.
860 * The transaction id must accomodate real and potential pad creation.
862 rawp = (void *)(jo->fifo.membase + (jo->fifo.windex & jo->fifo.mask));
864 journal_build_pad(rawp, availtoend, jo->transid);
866 rawp = (void *)jo->fifo.membase;
868 rawp->begmagic = JREC_INCOMPLETEMAGIC; /* updated by abort/commit */
869 rawp->recsize = bytes; /* (unaligned size) */
870 rawp->streamid = streamid | JREC_STREAMCTL_BEGIN;
871 rawp->transid = jo->transid;
875 * Issue a memory barrier to guarentee that the record data has been
876 * properly initialized before we advance the write index and return
877 * a pointer to the reserved record. Otherwise the worker thread
878 * could accidently run past us.
880 * Note that stream records are always 16-byte aligned.
883 jo->fifo.windex += (req + 15) & ~15;
893 * Attempt to extend the stream record by <bytes> worth of payload space.
895 * If it is possible to extend the existing stream record no truncation
896 * occurs and the record is extended as specified. A pointer to the
897 * truncation offset within the payload space is returned.
899 * If it is not possible to do this the existing stream record is truncated
900 * and committed, and a new stream record of size <bytes> is created. A
901 * pointer to the base of the new stream record's payload space is returned.
903 * *rawpp is set to the new reservation in the case of a new record but
904 * the caller cannot depend on a comparison with the old rawp to determine if
905 * this case occurs because we could end up using the same memory FIFO
906 * offset for the new stream record. Use *newstreamrecp instead.
909 journal_extend(struct journal *jo, struct journal_rawrecbeg **rawpp,
910 int truncbytes, int bytes, int *newstreamrecp)
912 struct journal_rawrecbeg *rawp;
923 osize = (rawp->recsize + 15) & ~15;
924 nsize = (rawp->recsize + bytes + 15) & ~15;
925 wbase = (char *)rawp - jo->fifo.membase;
928 * If the aligned record size does not change we can trivially adjust
931 if (nsize == osize) {
932 rawp->recsize += bytes;
933 return((char *)(rawp + 1) + truncbytes);
937 * If the fifo's write index hasn't been modified since we made the
938 * reservation and we do not hit any boundary conditions, we can
939 * trivially make the record smaller or larger.
941 if ((jo->fifo.windex & jo->fifo.mask) == wbase + osize) {
942 availtoend = jo->fifo.size - wbase;
943 avail = jo->fifo.size - (jo->fifo.windex - jo->fifo.xindex) + osize;
944 KKASSERT((availtoend & 15) == 0);
945 KKASSERT((avail & 15) == 0);
946 if (nsize <= avail && nsize <= availtoend) {
947 jo->fifo.windex += nsize - osize;
948 rawp->recsize += bytes;
949 return((char *)(rawp + 1) + truncbytes);
954 * It was not possible to extend the buffer. Commit the current
955 * buffer and create a new one. We manually clear the BEGIN mark that
956 * journal_reserve() creates (because this is a continuing record, not
957 * the start of a new stream).
959 streamid = rawp->streamid & JREC_STREAMID_MASK;
960 journal_commit(jo, rawpp, truncbytes, 0);
961 rptr = journal_reserve(jo, rawpp, streamid, bytes);
963 rawp->streamid &= ~JREC_STREAMCTL_BEGIN;
969 * Abort a journal record. If the transaction record represents a stream
970 * BEGIN and we can reverse the fifo's write index we can simply reverse
971 * index the entire record, as if it were never reserved in the first place.
973 * Otherwise we set the JREC_STREAMCTL_ABORTED bit and commit the record
974 * with the payload truncated to 0 bytes.
977 journal_abort(struct journal *jo, struct journal_rawrecbeg **rawpp)
979 struct journal_rawrecbeg *rawp;
983 osize = (rawp->recsize + 15) & ~15;
985 if ((rawp->streamid & JREC_STREAMCTL_BEGIN) &&
986 (jo->fifo.windex & jo->fifo.mask) ==
987 (char *)rawp - jo->fifo.membase + osize)
989 jo->fifo.windex -= osize;
992 rawp->streamid |= JREC_STREAMCTL_ABORTED;
993 journal_commit(jo, rawpp, 0, 1);
998 * Commit a journal record and potentially truncate it to the specified
999 * number of payload bytes. If you do not want to truncate the record,
1000 * simply pass -1 for the bytes parameter. Do not pass rawp->recsize, that
1001 * field includes header and trailer and will not be correct. Note that
1002 * passing 0 will truncate the entire data payload of the record.
1004 * The logical stream is terminated by this function.
1006 * If truncation occurs, and it is not possible to physically optimize the
1007 * memory FIFO due to other threads having reserved space after ours,
1008 * the remaining reserved space will be covered by a pad record.
1011 journal_commit(struct journal *jo, struct journal_rawrecbeg **rawpp,
1012 int bytes, int closeout)
1014 struct journal_rawrecbeg *rawp;
1015 struct journal_rawrecend *rendp;
1022 KKASSERT((char *)rawp >= jo->fifo.membase &&
1023 (char *)rawp + rawp->recsize <= jo->fifo.membase + jo->fifo.size);
1024 KKASSERT(((intptr_t)rawp & 15) == 0);
1027 * Truncate the record if necessary. If the FIFO write index as still
1028 * at the end of our record we can optimally backindex it. Otherwise
1029 * we have to insert a pad record to cover the dead space.
1031 * We calculate osize which is the 16-byte-aligned original recsize.
1032 * We calculate nsize which is the 16-byte-aligned new recsize.
1034 * Due to alignment issues or in case the passed truncation bytes is
1035 * the same as the original payload, nsize may be equal to osize even
1036 * if the committed bytes is less then the originally reserved bytes.
1039 KKASSERT(bytes >= 0 && bytes <= rawp->recsize - sizeof(struct journal_rawrecbeg) - sizeof(struct journal_rawrecend));
1040 osize = (rawp->recsize + 15) & ~15;
1041 rawp->recsize = bytes + sizeof(struct journal_rawrecbeg) +
1042 sizeof(struct journal_rawrecend);
1043 nsize = (rawp->recsize + 15) & ~15;
1044 KKASSERT(nsize <= osize);
1045 if (osize == nsize) {
1047 } else if ((jo->fifo.windex & jo->fifo.mask) == (char *)rawp - jo->fifo.membase + osize) {
1048 /* we are able to backindex the fifo */
1049 jo->fifo.windex -= osize - nsize;
1051 /* we cannot backindex the fifo, emplace a pad in the dead space */
1052 journal_build_pad((void *)((char *)rawp + nsize), osize - nsize,
1058 * Fill in the trailer. Note that unlike pad records, the trailer will
1059 * never overlap the header.
1061 rendp = (void *)((char *)rawp +
1062 ((rawp->recsize + 15) & ~15) - sizeof(*rendp));
1063 rendp->endmagic = JREC_ENDMAGIC;
1064 rendp->recsize = rawp->recsize;
1065 rendp->check = 0; /* XXX check word, disabled for now */
1068 * Fill in begmagic last. This will allow the worker thread to proceed.
1069 * Use a memory barrier to guarentee write ordering. Mark the stream
1070 * as terminated if closeout is set. This is the typical case.
1073 rawp->streamid |= JREC_STREAMCTL_END;
1074 cpu_sfence(); /* memory and compiler barrier */
1075 rawp->begmagic = JREC_BEGMAGIC;
1077 journal_commit_wakeup(jo);
1080 /************************************************************************
1081 * TRANSACTION SUPPORT ROUTINES *
1082 ************************************************************************
1084 * JRECORD_*() - routines to create subrecord transactions and embed them
1085 * in the logical streams managed by the journal_*() routines.
1088 static int16_t sid = JREC_STREAMID_JMIN;
1091 * Initialize the passed jrecord structure and start a new stream transaction
1092 * by reserving an initial build space in the journal's memory FIFO.
1095 jrecord_init(struct journal *jo, struct jrecord *jrec, int16_t streamid)
1097 bzero(jrec, sizeof(*jrec));
1100 streamid = sid++; /* XXX need to track stream ids! */
1101 if (sid == JREC_STREAMID_JMAX)
1102 sid = JREC_STREAMID_JMIN;
1104 jrec->streamid = streamid;
1105 jrec->stream_residual = JREC_DEFAULTSIZE;
1106 jrec->stream_reserved = jrec->stream_residual;
1108 journal_reserve(jo, &jrec->rawp, streamid, jrec->stream_reserved);
1112 * Push a recursive record type. All pushes should have matching pops.
1113 * The old parent is returned and the newly pushed record becomes the
1114 * new parent. Note that the old parent's pointer may already be invalid
1115 * or may become invalid if jrecord_write() had to build a new stream
1116 * record, so the caller should not mess with the returned pointer in
1117 * any way other then to save it.
1120 struct journal_subrecord *
1121 jrecord_push(struct jrecord *jrec, int16_t rectype)
1123 struct journal_subrecord *save;
1125 save = jrec->parent;
1126 jrec->parent = jrecord_write(jrec, rectype|JMASK_NESTED, 0);
1128 KKASSERT(jrec->parent != NULL);
1130 ++jrec->pushptrgood; /* cleared on flush */
1135 * Pop a previously pushed sub-transaction. We must set JMASK_LAST
1136 * on the last record written within the subtransaction. If the last
1137 * record written is not accessible or if the subtransaction is empty,
1138 * we must write out a pad record with JMASK_LAST set before popping.
1140 * When popping a subtransaction the parent record's recsize field
1141 * will be properly set. If the parent pointer is no longer valid
1142 * (which can occur if the data has already been flushed out to the
1143 * stream), the protocol spec allows us to leave it 0.
1145 * The saved parent pointer which we restore may or may not be valid,
1146 * and if not valid may or may not be NULL, depending on the value
1150 jrecord_pop(struct jrecord *jrec, struct journal_subrecord *save)
1152 struct journal_subrecord *last;
1154 KKASSERT(jrec->pushcount > 0);
1155 KKASSERT(jrec->residual == 0);
1158 * Set JMASK_LAST on the last record we wrote at the current
1159 * level. If last is NULL we either no longer have access to the
1160 * record or the subtransaction was empty and we must write out a pad
1163 if ((last = jrec->last) == NULL) {
1164 jrecord_write(jrec, JLEAF_PAD|JMASK_LAST, 0);
1165 last = jrec->last; /* reload after possible flush */
1167 last->rectype |= JMASK_LAST;
1171 * pushptrgood tells us how many levels of parent record pointers
1172 * are valid. The jrec only stores the current parent record pointer
1173 * (and it is only valid if pushptrgood != 0). The higher level parent
1174 * record pointers are saved by the routines calling jrecord_push() and
1175 * jrecord_pop(). These pointers may become stale and we determine
1176 * that fact by tracking the count of valid parent pointers with
1177 * pushptrgood. Pointers become invalid when their related stream
1178 * record gets pushed out.
1180 * If no pointer is available (the data has already been pushed out),
1181 * then no fixup of e.g. the length field is possible for non-leaf
1182 * nodes. The protocol allows for this situation by placing a larger
1183 * burden on the program scanning the stream on the other end.
1190 * (pop B) see NOTE B
1191 * (pop A) see NOTE A
1193 * NOTE B: This pop sets LAST in node Z if the node is still accessible,
1194 * else a PAD record is appended and LAST is set in that.
1196 * This pop sets the record size in parentB if parentB is still
1197 * accessible, else the record size is left 0 (the scanner must
1200 * This pop sets the new 'last' record to parentB, the pointer
1201 * to which may or may not still be accessible.
1203 * NOTE A: This pop sets LAST in parentB if the node is still accessible,
1204 * else a PAD record is appended and LAST is set in that.
1206 * This pop sets the record size in parentA if parentA is still
1207 * accessible, else the record size is left 0 (the scanner must
1210 * This pop sets the new 'last' record to parentA, the pointer
1211 * to which may or may not still be accessible.
1213 * Also note that the last record in the stream transaction, which in
1214 * the above example is parentA, does not currently have the LAST bit
1217 * The current parent becomes the last record relative to the
1218 * saved parent passed into us. It's validity is based on
1219 * whether pushptrgood is non-zero prior to decrementing. The saved
1220 * parent becomes the new parent, and its validity is based on whether
1221 * pushptrgood is non-zero after decrementing.
1223 * The old jrec->parent may be NULL if it is no longer accessible.
1224 * If pushptrgood is non-zero, however, it is guarenteed to not
1225 * be NULL (since no flush occured).
1227 jrec->last = jrec->parent;
1229 if (jrec->pushptrgood) {
1230 KKASSERT(jrec->last != NULL && last != NULL);
1231 if (--jrec->pushptrgood == 0) {
1232 jrec->parent = NULL; /* 'save' contains garbage or NULL */
1234 KKASSERT(save != NULL);
1235 jrec->parent = save; /* 'save' must not be NULL */
1239 * Set the record size in the old parent. 'last' still points to
1240 * the original last record in the subtransaction being popped,
1241 * jrec->last points to the old parent (which became the last
1242 * record relative to the new parent being popped into).
1244 jrec->last->recsize = (char *)last + last->recsize - (char *)jrec->last;
1246 jrec->parent = NULL;
1247 KKASSERT(jrec->last == NULL);
1252 * Write out a leaf record, including associated data.
1256 jrecord_leaf(struct jrecord *jrec, int16_t rectype, void *ptr, int bytes)
1258 jrecord_write(jrec, rectype, bytes);
1259 jrecord_data(jrec, ptr, bytes);
1263 * Write a leaf record out and return a pointer to its base. The leaf
1264 * record may contain potentially megabytes of data which is supplied
1265 * in jrecord_data() calls. The exact amount must be specified in this
1268 * THE RETURNED SUBRECORD POINTER IS ONLY VALID IMMEDIATELY AFTER THE
1269 * CALL AND MAY BECOME INVALID AT ANY TIME. ONLY THE PUSH/POP CODE SHOULD
1270 * USE THE RETURN VALUE.
1273 struct journal_subrecord *
1274 jrecord_write(struct jrecord *jrec, int16_t rectype, int bytes)
1276 struct journal_subrecord *last;
1280 * Try to catch some obvious errors. Nesting records must specify a
1281 * size of 0, and there should be no left-overs from previous operations
1282 * (such as incomplete data writeouts).
1284 KKASSERT(bytes == 0 || (rectype & JMASK_NESTED) == 0);
1285 KKASSERT(jrec->residual == 0);
1288 * Check to see if the current stream record has enough room for
1289 * the new subrecord header. If it doesn't we extend the current
1292 * This may have the side effect of pushing out the current stream record
1293 * and creating a new one. We must adjust our stream tracking fields
1296 if (jrec->stream_residual < sizeof(struct journal_subrecord)) {
1297 jrec->stream_ptr = journal_extend(jrec->jo, &jrec->rawp,
1298 jrec->stream_reserved - jrec->stream_residual,
1299 JREC_DEFAULTSIZE, &pusheditout);
1302 * If a pushout occured, the pushed out stream record was
1303 * truncated as specified and the new record is exactly the
1304 * extension size specified.
1306 jrec->stream_reserved = JREC_DEFAULTSIZE;
1307 jrec->stream_residual = JREC_DEFAULTSIZE;
1308 jrec->parent = NULL; /* no longer accessible */
1309 jrec->pushptrgood = 0; /* restored parents in pops no good */
1312 * If no pushout occured the stream record is NOT truncated and
1315 jrec->stream_reserved += JREC_DEFAULTSIZE;
1316 jrec->stream_residual += JREC_DEFAULTSIZE;
1319 last = (void *)jrec->stream_ptr;
1320 last->rectype = rectype;
1324 * We may not know the record size for recursive records and the
1325 * header may become unavailable due to limited FIFO space. Write
1326 * -1 to indicate this special case.
1328 if ((rectype & JMASK_NESTED) && bytes == 0)
1331 last->recsize = sizeof(struct journal_subrecord) + bytes;
1333 jrec->residual = bytes; /* remaining data to be posted */
1334 jrec->residual_align = -bytes & 7; /* post-data alignment required */
1335 jrec->stream_ptr += sizeof(*last); /* current write pointer */
1336 jrec->stream_residual -= sizeof(*last); /* space remaining in stream */
1341 * Write out the data associated with a leaf record. Any number of calls
1342 * to this routine may be made as long as the byte count adds up to the
1343 * amount originally specified in jrecord_write().
1345 * The act of writing out the leaf data may result in numerous stream records
1346 * being pushed out. Callers should be aware that even the associated
1347 * subrecord header may become inaccessible due to stream record pushouts.
1350 jrecord_data(struct jrecord *jrec, const void *buf, int bytes)
1355 KKASSERT(bytes >= 0 && bytes <= jrec->residual);
1358 * Push out stream records as long as there is insufficient room to hold
1359 * the remaining data.
1361 while (jrec->stream_residual < bytes) {
1363 * Fill in any remaining space in the current stream record.
1365 bcopy(buf, jrec->stream_ptr, jrec->stream_residual);
1366 buf = (const char *)buf + jrec->stream_residual;
1367 bytes -= jrec->stream_residual;
1368 /*jrec->stream_ptr += jrec->stream_residual;*/
1369 jrec->residual -= jrec->stream_residual;
1370 jrec->stream_residual = 0;
1373 * Try to extend the current stream record, but no more then 1/4
1374 * the size of the FIFO.
1376 extsize = jrec->jo->fifo.size >> 2;
1377 if (extsize > bytes)
1378 extsize = (bytes + 15) & ~15;
1380 jrec->stream_ptr = journal_extend(jrec->jo, &jrec->rawp,
1381 jrec->stream_reserved - jrec->stream_residual,
1382 extsize, &pusheditout);
1384 jrec->stream_reserved = extsize;
1385 jrec->stream_residual = extsize;
1386 jrec->parent = NULL; /* no longer accessible */
1387 jrec->last = NULL; /* no longer accessible */
1388 jrec->pushptrgood = 0; /* restored parents in pops no good */
1390 jrec->stream_reserved += extsize;
1391 jrec->stream_residual += extsize;
1396 * Push out any remaining bytes into the current stream record.
1399 bcopy(buf, jrec->stream_ptr, bytes);
1400 jrec->stream_ptr += bytes;
1401 jrec->stream_residual -= bytes;
1402 jrec->residual -= bytes;
1406 * Handle data alignment requirements for the subrecord. Because the
1407 * stream record's data space is more strictly aligned, it must already
1408 * have sufficient space to hold any subrecord alignment slop.
1410 if (jrec->residual == 0 && jrec->residual_align) {
1411 KKASSERT(jrec->residual_align <= jrec->stream_residual);
1412 bzero(jrec->stream_ptr, jrec->residual_align);
1413 jrec->stream_ptr += jrec->residual_align;
1414 jrec->stream_residual -= jrec->residual_align;
1415 jrec->residual_align = 0;
1420 * We are finished with the transaction. This closes the transaction created
1421 * by jrecord_init().
1423 * NOTE: If abortit is not set then we must be at the top level with no
1424 * residual subrecord data left to output.
1426 * If abortit is set then we can be in any state, all pushes will be
1427 * popped and it is ok for there to be residual data. This works
1428 * because the virtual stream itself is truncated. Scanners must deal
1429 * with this situation.
1431 * The stream record will be committed or aborted as specified and jrecord
1432 * resources will be cleaned up.
1435 jrecord_done(struct jrecord *jrec, int abortit)
1437 KKASSERT(jrec->rawp != NULL);
1440 journal_abort(jrec->jo, &jrec->rawp);
1442 KKASSERT(jrec->pushcount == 0 && jrec->residual == 0);
1443 journal_commit(jrec->jo, &jrec->rawp,
1444 jrec->stream_reserved - jrec->stream_residual, 1);
1448 * jrec should not be used beyond this point without another init,
1449 * but clean up some fields to ensure that we panic if it is.
1451 * Note that jrec->rawp is NULLd out by journal_abort/journal_commit.
1454 jrec->stream_ptr = NULL;
1457 /************************************************************************
1458 * LOW LEVEL RECORD SUPPORT ROUTINES *
1459 ************************************************************************
1461 * These routine create low level recursive and leaf subrecords representing
1462 * common filesystem structures.
1466 * Write out a filename path relative to the base of the mount point.
1467 * rectype is typically JLEAF_PATH{1,2,3,4}.
1470 jrecord_write_path(struct jrecord *jrec, int16_t rectype, struct namecache *ncp)
1472 char buf[64]; /* local buffer if it fits, else malloced */
1476 struct namecache *scan;
1479 * Pass 1 - figure out the number of bytes required. Include terminating
1480 * \0 on last element and '/' separator on other elements.
1485 scan && (scan->nc_flag & NCF_MOUNTPT) == 0;
1486 scan = scan->nc_parent
1488 pathlen += scan->nc_nlen + 1;
1491 if (pathlen <= sizeof(buf))
1494 base = malloc(pathlen, M_TEMP, M_INTWAIT);
1497 * Pass 2 - generate the path buffer
1501 scan && (scan->nc_flag & NCF_MOUNTPT) == 0;
1502 scan = scan->nc_parent
1504 if (scan->nc_nlen >= index) {
1509 if (index == pathlen)
1512 base[--index] = '/';
1513 index -= scan->nc_nlen;
1514 bcopy(scan->nc_name, base + index, scan->nc_nlen);
1516 jrecord_leaf(jrec, rectype, base + index, pathlen - index);
1522 * Write out a file attribute structure. While somewhat inefficient, using
1523 * a recursive data structure is the most portable and extensible way.
1526 jrecord_write_vattr(struct jrecord *jrec, struct vattr *vat)
1530 save = jrecord_push(jrec, JTYPE_VATTR);
1531 if (vat->va_type != VNON)
1532 jrecord_leaf(jrec, JLEAF_VTYPE, &vat->va_type, sizeof(vat->va_type));
1533 if (vat->va_mode != (mode_t)VNOVAL)
1534 jrecord_leaf(jrec, JLEAF_MODES, &vat->va_mode, sizeof(vat->va_mode));
1535 if (vat->va_nlink != VNOVAL)
1536 jrecord_leaf(jrec, JLEAF_NLINK, &vat->va_nlink, sizeof(vat->va_nlink));
1537 if (vat->va_uid != VNOVAL)
1538 jrecord_leaf(jrec, JLEAF_UID, &vat->va_uid, sizeof(vat->va_uid));
1539 if (vat->va_gid != VNOVAL)
1540 jrecord_leaf(jrec, JLEAF_GID, &vat->va_gid, sizeof(vat->va_gid));
1541 if (vat->va_fsid != VNOVAL)
1542 jrecord_leaf(jrec, JLEAF_FSID, &vat->va_fsid, sizeof(vat->va_fsid));
1543 if (vat->va_fileid != VNOVAL)
1544 jrecord_leaf(jrec, JLEAF_INUM, &vat->va_fileid, sizeof(vat->va_fileid));
1545 if (vat->va_size != VNOVAL)
1546 jrecord_leaf(jrec, JLEAF_SIZE, &vat->va_size, sizeof(vat->va_size));
1547 if (vat->va_atime.tv_sec != VNOVAL)
1548 jrecord_leaf(jrec, JLEAF_ATIME, &vat->va_atime, sizeof(vat->va_atime));
1549 if (vat->va_mtime.tv_sec != VNOVAL)
1550 jrecord_leaf(jrec, JLEAF_MTIME, &vat->va_mtime, sizeof(vat->va_mtime));
1551 if (vat->va_ctime.tv_sec != VNOVAL)
1552 jrecord_leaf(jrec, JLEAF_CTIME, &vat->va_ctime, sizeof(vat->va_ctime));
1553 if (vat->va_gen != VNOVAL)
1554 jrecord_leaf(jrec, JLEAF_GEN, &vat->va_gen, sizeof(vat->va_gen));
1555 if (vat->va_flags != VNOVAL)
1556 jrecord_leaf(jrec, JLEAF_FLAGS, &vat->va_flags, sizeof(vat->va_flags));
1557 if (vat->va_rdev != VNOVAL)
1558 jrecord_leaf(jrec, JLEAF_UDEV, &vat->va_rdev, sizeof(vat->va_rdev));
1560 if (vat->va_filerev != VNOVAL)
1561 jrecord_leaf(jrec, JLEAF_FILEREV, &vat->va_filerev, sizeof(vat->va_filerev));
1563 jrecord_pop(jrec, save);
1567 * Write out the creds used to issue a file operation. If a process is
1568 * available write out additional tracking information related to the
1571 * XXX additional tracking info
1575 jrecord_write_cred(struct jrecord *jrec, struct thread *td, struct ucred *cred)
1580 save = jrecord_push(jrec, JTYPE_CRED);
1581 jrecord_leaf(jrec, JLEAF_UID, &cred->cr_uid, sizeof(cred->cr_uid));
1582 jrecord_leaf(jrec, JLEAF_GID, &cred->cr_gid, sizeof(cred->cr_gid));
1583 if (td && (p = td->td_proc) != NULL) {
1584 jrecord_leaf(jrec, JLEAF_PID, &p->p_pid, sizeof(p->p_pid));
1585 jrecord_leaf(jrec, JLEAF_COMM, p->p_comm, sizeof(p->p_comm));
1587 jrecord_pop(jrec, save);
1591 * Write out information required to identify a vnode
1593 * XXX this needs work. We should write out the inode number as well,
1594 * and in fact avoid writing out the file path for seqential writes
1595 * occuring within e.g. a certain period of time.
1598 jrecord_write_vnode_ref(struct jrecord *jrec, struct vnode *vp)
1600 struct namecache *ncp;
1602 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
1603 if ((ncp->nc_flag & (NCF_UNRESOLVED|NCF_DESTROYED)) == 0)
1607 jrecord_write_path(jrec, JLEAF_PATH_REF, ncp);
1611 jrecord_write_vnode_link(struct jrecord *jrec, struct vnode *vp,
1612 struct namecache *notncp)
1614 struct namecache *ncp;
1616 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
1619 if ((ncp->nc_flag & (NCF_UNRESOLVED|NCF_DESTROYED)) == 0)
1623 jrecord_write_path(jrec, JLEAF_PATH_REF, ncp);
1628 * Write out the current contents of the file within the specified
1629 * range. This is typically called from within an UNDO section. A
1630 * locked vnode must be passed.
1633 jrecord_write_filearea(struct jrecord *jrec, struct vnode *vp,
1634 off_t begoff, off_t endoff)
1640 * Write out the data represented by a pagelist
1643 jrecord_write_pagelist(struct jrecord *jrec, int16_t rectype,
1644 struct vm_page **pglist, int *rtvals, int pgcount,
1647 struct msf_buf *msf;
1653 while (i < pgcount) {
1655 * Find the next valid section. Skip any invalid elements
1657 if (rtvals[i] != VM_PAGER_OK) {
1659 offset += PAGE_SIZE;
1664 * Figure out how big the valid section is, capping I/O at what the
1665 * MSFBUF can represent.
1668 while (i < pgcount && i - b != XIO_INTERNAL_PAGES &&
1669 rtvals[i] == VM_PAGER_OK
1678 error = msf_map_pagelist(&msf, pglist + b, i - b, 0);
1680 printf("RECORD PUTPAGES %d\n", msf_buf_bytes(msf));
1681 jrecord_leaf(jrec, JLEAF_SEEKPOS, &offset, sizeof(offset));
1682 jrecord_leaf(jrec, rectype,
1683 msf_buf_kva(msf), msf_buf_bytes(msf));
1686 printf("jrecord_write_pagelist: mapping failure\n");
1688 offset += (off_t)(i - b) << PAGE_SHIFT;
1694 * Write out the data represented by a UIO.
1697 struct jrecord *jrec;
1701 static int jrecord_write_uio_callback(void *info, char *buf, int bytes);
1704 jrecord_write_uio(struct jrecord *jrec, int16_t rectype, struct uio *uio)
1706 struct jwuio_info info = { jrec, rectype };
1709 if (uio->uio_segflg != UIO_NOCOPY) {
1710 jrecord_leaf(jrec, JLEAF_SEEKPOS, &uio->uio_offset,
1711 sizeof(uio->uio_offset));
1712 error = msf_uio_iterate(uio, jrecord_write_uio_callback, &info);
1714 printf("XXX warning uio iterate failed %d\n", error);
1719 jrecord_write_uio_callback(void *info_arg, char *buf, int bytes)
1721 struct jwuio_info *info = info_arg;
1723 jrecord_leaf(info->jrec, info->rectype, buf, bytes);
1727 /************************************************************************
1729 ************************************************************************
1731 * These are function shims replacing the normal filesystem ops. We become
1732 * responsible for calling the underlying filesystem ops. We have the choice
1733 * of executing the underlying op first and then generating the journal entry,
1734 * or starting the journal entry, executing the underlying op, and then
1735 * either completing or aborting it.
1737 * The journal is supposed to be a high-level entity, which generally means
1738 * identifying files by name rather then by inode. Supplying both allows
1739 * the journal to be used both for inode-number-compatible 'mirrors' and
1740 * for simple filesystem replication.
1742 * Writes are particularly difficult to deal with because a single write may
1743 * represent a hundred megabyte buffer or more, and both writes and truncations
1744 * require the 'old' data to be written out as well as the new data if the
1745 * log is reversable. Other issues:
1747 * - How to deal with operations on unlinked files (no path available),
1748 * but which may still be filesystem visible due to hard links.
1750 * - How to deal with modifications made via a memory map.
1752 * - Future cache coherency support will require cache coherency API calls
1753 * both prior to and after the call to the underlying VFS.
1755 * ALSO NOTE: We do not have to shim compatibility VOPs like MKDIR which have
1756 * new VFS equivalents (NMKDIR).
1760 * Journal vop_settattr { a_vp, a_vap, a_cred, a_td }
1764 journal_setattr(struct vop_setattr_args *ap)
1768 struct jrecord jrec;
1769 void *save; /* warning, save pointers do not always remain valid */
1772 error = vop_journal_operate_ap(&ap->a_head);
1773 mp = ap->a_head.a_ops->vv_mount;
1775 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) {
1776 jrecord_init(jo, &jrec, -1);
1777 save = jrecord_push(&jrec, JTYPE_SETATTR);
1778 jrecord_write_cred(&jrec, ap->a_td, ap->a_cred);
1779 jrecord_write_vnode_ref(&jrec, ap->a_vp);
1780 jrecord_write_vattr(&jrec, ap->a_vap);
1781 jrecord_pop(&jrec, save);
1782 jrecord_done(&jrec, 0);
1789 * Journal vop_write { a_vp, a_uio, a_ioflag, a_cred }
1793 journal_write(struct vop_write_args *ap)
1797 struct jrecord jrec;
1798 struct uio uio_copy;
1799 struct iovec uio_one_iovec;
1800 void *save; /* warning, save pointers do not always remain valid */
1804 * This is really nasty. UIO's don't retain sufficient information to
1805 * be reusable once they've gone through the VOP chain. The iovecs get
1806 * cleared, so we have to copy the UIO.
1808 * XXX fix the UIO code to not destroy iov's during a scan so we can
1809 * reuse the uio over and over again.
1811 * XXX UNDO code needs to journal the old data prior to the write.
1813 uio_copy = *ap->a_uio;
1814 if (uio_copy.uio_iovcnt == 1) {
1815 uio_one_iovec = ap->a_uio->uio_iov[0];
1816 uio_copy.uio_iov = &uio_one_iovec;
1818 uio_copy.uio_iov = malloc(uio_copy.uio_iovcnt * sizeof(struct iovec),
1819 M_JOURNAL, M_WAITOK);
1820 bcopy(ap->a_uio->uio_iov, uio_copy.uio_iov,
1821 uio_copy.uio_iovcnt * sizeof(struct iovec));
1824 error = vop_journal_operate_ap(&ap->a_head);
1827 * XXX bad hack to figure out the offset for O_APPEND writes (note:
1828 * uio field state after the VFS operation).
1830 uio_copy.uio_offset = ap->a_uio->uio_offset -
1831 (uio_copy.uio_resid - ap->a_uio->uio_resid);
1833 mp = ap->a_head.a_ops->vv_mount;
1835 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) {
1836 jrecord_init(jo, &jrec, -1);
1837 save = jrecord_push(&jrec, JTYPE_WRITE);
1838 jrecord_write_cred(&jrec, NULL, ap->a_cred);
1839 jrecord_write_vnode_ref(&jrec, ap->a_vp);
1840 jrecord_write_uio(&jrec, JLEAF_FILEDATA, &uio_copy);
1841 jrecord_pop(&jrec, save);
1842 jrecord_done(&jrec, 0);
1846 if (uio_copy.uio_iov != &uio_one_iovec)
1847 free(uio_copy.uio_iov, M_JOURNAL);
1854 * Journal vop_fsync { a_vp, a_waitfor, a_td }
1858 journal_fsync(struct vop_fsync_args *ap)
1864 error = vop_journal_operate_ap(&ap->a_head);
1865 mp = ap->a_head.a_ops->vv_mount;
1867 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) {
1868 /* XXX synchronize pending journal records */
1875 * Journal vop_putpages { a_vp, a_m, a_count, a_sync, a_rtvals, a_offset }
1877 * note: a_count is in bytes.
1881 journal_putpages(struct vop_putpages_args *ap)
1885 struct jrecord jrec;
1886 void *save; /* warning, save pointers do not always remain valid */
1889 error = vop_journal_operate_ap(&ap->a_head);
1890 mp = ap->a_head.a_ops->vv_mount;
1891 if (error == 0 && ap->a_count > 0) {
1892 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) {
1893 jrecord_init(jo, &jrec, -1);
1894 save = jrecord_push(&jrec, JTYPE_PUTPAGES);
1895 jrecord_write_vnode_ref(&jrec, ap->a_vp);
1896 jrecord_write_pagelist(&jrec, JLEAF_FILEDATA,
1897 ap->a_m, ap->a_rtvals, btoc(ap->a_count), ap->a_offset);
1898 jrecord_pop(&jrec, save);
1899 jrecord_done(&jrec, 0);
1906 * Journal vop_setacl { a_vp, a_type, a_aclp, a_cred, a_td }
1910 journal_setacl(struct vop_setacl_args *ap)
1914 struct jrecord jrec;
1915 void *save; /* warning, save pointers do not always remain valid */
1918 error = vop_journal_operate_ap(&ap->a_head);
1919 mp = ap->a_head.a_ops->vv_mount;
1921 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) {
1922 jrecord_init(jo, &jrec, -1);
1923 save = jrecord_push(&jrec, JTYPE_SETACL);
1924 jrecord_write_cred(&jrec, ap->a_td, ap->a_cred);
1925 jrecord_write_vnode_ref(&jrec, ap->a_vp);
1926 /* XXX type, aclp */
1927 jrecord_pop(&jrec, save);
1928 jrecord_done(&jrec, 0);
1935 * Journal vop_setextattr { a_vp, a_name, a_uio, a_cred, a_td }
1939 journal_setextattr(struct vop_setextattr_args *ap)
1943 struct jrecord jrec;
1944 void *save; /* warning, save pointers do not always remain valid */
1947 error = vop_journal_operate_ap(&ap->a_head);
1948 mp = ap->a_head.a_ops->vv_mount;
1950 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) {
1951 jrecord_init(jo, &jrec, -1);
1952 save = jrecord_push(&jrec, JTYPE_SETEXTATTR);
1953 jrecord_write_cred(&jrec, ap->a_td, ap->a_cred);
1954 jrecord_write_vnode_ref(&jrec, ap->a_vp);
1955 jrecord_leaf(&jrec, JLEAF_ATTRNAME, ap->a_name, strlen(ap->a_name));
1956 jrecord_write_uio(&jrec, JLEAF_FILEDATA, ap->a_uio);
1957 jrecord_pop(&jrec, save);
1958 jrecord_done(&jrec, 0);
1965 * Journal vop_ncreate { a_ncp, a_vpp, a_cred, a_vap }
1969 journal_ncreate(struct vop_ncreate_args *ap)
1973 struct jrecord jrec;
1974 void *save; /* warning, save pointers do not always remain valid */
1977 error = vop_journal_operate_ap(&ap->a_head);
1978 mp = ap->a_head.a_ops->vv_mount;
1980 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) {
1981 jrecord_init(jo, &jrec, -1);
1982 save = jrecord_push(&jrec, JTYPE_CREATE);
1983 jrecord_write_cred(&jrec, NULL, ap->a_cred);
1984 jrecord_write_path(&jrec, JLEAF_PATH1, ap->a_ncp);
1986 jrecord_write_vnode_ref(&jrec, *ap->a_vpp);
1987 jrecord_write_vattr(&jrec, ap->a_vap);
1988 jrecord_pop(&jrec, save);
1989 jrecord_done(&jrec, 0);
1996 * Journal vop_nmknod { a_ncp, a_vpp, a_cred, a_vap }
2000 journal_nmknod(struct vop_nmknod_args *ap)
2004 struct jrecord jrec;
2005 void *save; /* warning, save pointers do not always remain valid */
2008 error = vop_journal_operate_ap(&ap->a_head);
2009 mp = ap->a_head.a_ops->vv_mount;
2011 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) {
2012 jrecord_init(jo, &jrec, -1);
2013 save = jrecord_push(&jrec, JTYPE_MKNOD);
2014 jrecord_write_cred(&jrec, NULL, ap->a_cred);
2015 jrecord_write_path(&jrec, JLEAF_PATH1, ap->a_ncp);
2016 jrecord_write_vattr(&jrec, ap->a_vap);
2018 jrecord_write_vnode_ref(&jrec, *ap->a_vpp);
2019 jrecord_pop(&jrec, save);
2020 jrecord_done(&jrec, 0);
2027 * Journal vop_nlink { a_ncp, a_vp, a_cred }
2031 journal_nlink(struct vop_nlink_args *ap)
2035 struct jrecord jrec;
2036 void *save; /* warning, save pointers do not always remain valid */
2039 error = vop_journal_operate_ap(&ap->a_head);
2040 mp = ap->a_head.a_ops->vv_mount;
2042 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) {
2043 jrecord_init(jo, &jrec, -1);
2044 save = jrecord_push(&jrec, JTYPE_LINK);
2045 jrecord_write_cred(&jrec, NULL, ap->a_cred);
2046 jrecord_write_path(&jrec, JLEAF_PATH1, ap->a_ncp);
2047 /* XXX PATH to VP and inode number */
2048 /* XXX this call may not record the correct path when
2049 * multiple paths are available */
2050 jrecord_write_vnode_link(&jrec, ap->a_vp, ap->a_ncp);
2051 jrecord_pop(&jrec, save);
2052 jrecord_done(&jrec, 0);
2059 * Journal vop_symlink { a_ncp, a_vpp, a_cred, a_vap, a_target }
2063 journal_nsymlink(struct vop_nsymlink_args *ap)
2067 struct jrecord jrec;
2068 void *save; /* warning, save pointers do not always remain valid */
2071 error = vop_journal_operate_ap(&ap->a_head);
2072 mp = ap->a_head.a_ops->vv_mount;
2074 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) {
2075 jrecord_init(jo, &jrec, -1);
2076 save = jrecord_push(&jrec, JTYPE_SYMLINK);
2077 jrecord_write_cred(&jrec, NULL, ap->a_cred);
2078 jrecord_write_path(&jrec, JLEAF_PATH1, ap->a_ncp);
2079 jrecord_leaf(&jrec, JLEAF_SYMLINKDATA,
2080 ap->a_target, strlen(ap->a_target));
2082 jrecord_write_vnode_ref(&jrec, *ap->a_vpp);
2083 jrecord_pop(&jrec, save);
2084 jrecord_done(&jrec, 0);
2091 * Journal vop_nwhiteout { a_ncp, a_cred, a_flags }
2095 journal_nwhiteout(struct vop_nwhiteout_args *ap)
2099 struct jrecord jrec;
2100 void *save; /* warning, save pointers do not always remain valid */
2103 error = vop_journal_operate_ap(&ap->a_head);
2104 mp = ap->a_head.a_ops->vv_mount;
2106 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) {
2107 jrecord_init(jo, &jrec, -1);
2108 save = jrecord_push(&jrec, JTYPE_WHITEOUT);
2109 jrecord_write_cred(&jrec, NULL, ap->a_cred);
2110 jrecord_write_path(&jrec, JLEAF_PATH1, ap->a_ncp);
2111 jrecord_pop(&jrec, save);
2112 jrecord_done(&jrec, 0);
2119 * Journal vop_nremove { a_ncp, a_cred }
2123 journal_nremove(struct vop_nremove_args *ap)
2127 struct jrecord jrec;
2128 void *save; /* warning, save pointers do not always remain valid */
2131 error = vop_journal_operate_ap(&ap->a_head);
2132 mp = ap->a_head.a_ops->vv_mount;
2134 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) {
2135 jrecord_init(jo, &jrec, -1);
2136 save = jrecord_push(&jrec, JTYPE_REMOVE);
2137 jrecord_write_cred(&jrec, NULL, ap->a_cred);
2138 jrecord_write_path(&jrec, JLEAF_PATH1, ap->a_ncp);
2139 jrecord_pop(&jrec, save);
2140 jrecord_done(&jrec, 0);
2147 * Journal vop_nmkdir { a_ncp, a_vpp, a_cred, a_vap }
2151 journal_nmkdir(struct vop_nmkdir_args *ap)
2155 struct jrecord jrec;
2156 void *save; /* warning, save pointers do not always remain valid */
2159 error = vop_journal_operate_ap(&ap->a_head);
2160 mp = ap->a_head.a_ops->vv_mount;
2162 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) {
2163 jrecord_init(jo, &jrec, -1);
2164 if (jo->flags & MC_JOURNAL_WANT_REVERSABLE) {
2165 save = jrecord_push(&jrec, JTYPE_UNDO);
2166 /* XXX undo operations */
2167 jrecord_pop(&jrec, save);
2170 if (jo->flags & MC_JOURNAL_WANT_AUDIT) {
2171 jrecord_write_audit(&jrec);
2174 save = jrecord_push(&jrec, JTYPE_MKDIR);
2175 jrecord_write_path(&jrec, JLEAF_PATH1, ap->a_ncp);
2176 jrecord_write_cred(&jrec, NULL, ap->a_cred);
2177 jrecord_write_vattr(&jrec, ap->a_vap);
2178 jrecord_write_path(&jrec, JLEAF_PATH1, ap->a_ncp);
2180 jrecord_write_vnode_ref(&jrec, *ap->a_vpp);
2181 jrecord_pop(&jrec, save);
2182 jrecord_done(&jrec, 0);
2189 * Journal vop_nrmdir { a_ncp, a_cred }
2193 journal_nrmdir(struct vop_nrmdir_args *ap)
2197 struct jrecord jrec;
2198 void *save; /* warning, save pointers do not always remain valid */
2201 error = vop_journal_operate_ap(&ap->a_head);
2202 mp = ap->a_head.a_ops->vv_mount;
2204 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) {
2205 jrecord_init(jo, &jrec, -1);
2206 save = jrecord_push(&jrec, JTYPE_RMDIR);
2207 jrecord_write_cred(&jrec, NULL, ap->a_cred);
2208 jrecord_write_path(&jrec, JLEAF_PATH1, ap->a_ncp);
2209 jrecord_pop(&jrec, save);
2210 jrecord_done(&jrec, 0);
2217 * Journal vop_nrename { a_fncp, a_tncp, a_cred }
2221 journal_nrename(struct vop_nrename_args *ap)
2225 struct jrecord jrec;
2226 void *save; /* warning, save pointers do not always remain valid */
2229 error = vop_journal_operate_ap(&ap->a_head);
2230 mp = ap->a_head.a_ops->vv_mount;
2232 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) {
2233 jrecord_init(jo, &jrec, -1);
2234 save = jrecord_push(&jrec, JTYPE_RENAME);
2235 jrecord_write_cred(&jrec, NULL, ap->a_cred);
2236 jrecord_write_path(&jrec, JLEAF_PATH1, ap->a_fncp);
2237 jrecord_write_path(&jrec, JLEAF_PATH2, ap->a_tncp);
2238 jrecord_pop(&jrec, save);
2239 jrecord_done(&jrec, 0);