| blob | 80c3f51d2eb21c8f1c2f866f996a372f8f3c1a54 |
1 /*
2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved.
3 *
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 *
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * SUCH DAMAGE.
33 *
34 * $DragonFly: src/sys/vfs/hammer/hammer_inode.c,v 1.90 2008/06/30 02:45:30 dillon Exp $
35 */
38 #include <vm/vm_extern.h>
39 #include <sys/buf.h>
40 #include <sys/buf2.h>
50 #ifdef DEBUG_TRUNCATE
52 #endif
54 /*
55 * Red-Black tree support for inode structures.
56 *
57 * Insertions
58 */
59 int
61 {
75 }
77 /*
78 * LOOKUP_INFO
79 */
80 static int
82 {
96 }
98 /*
99 * Used by hammer_scan_inode_snapshots() to locate all of an object's
100 * snapshots. Note that the asof field is not tested, which we can get
101 * away with because it is the lowest-priority field.
102 */
103 static int
105 {
117 }
123 /*
124 * The kernel is not actively referencing this vnode but is still holding
125 * it cached.
126 *
127 * This is called from the frontend.
128 */
129 int
131 {
134 /*
135 * Degenerate case
136 */
140 }
142 /*
143 * If the inode no longer has visibility in the filesystem try to
144 * recycle it immediately, even if the inode is dirty. Recycling
145 * it quickly allows the system to reclaim buffer cache and VM
146 * resources which can matter a lot in a heavily loaded system.
147 *
148 * This can deadlock in vfsync() if we aren't careful.
149 *
150 * Do not queue the inode to the flusher if we still have visibility,
151 * otherwise namespace calls such as chmod will unnecessarily generate
152 * multiple inode updates.
153 */
159 }
161 }
163 /*
164 * Release the vnode association. This is typically (but not always)
165 * the last reference on the inode.
166 *
167 * Once the association is lost we are on our own with regards to
168 * flushing the inode.
169 */
170 int
172 {
174 hammer_mount_t hmp;
191 }
192 }
194 }
196 }
198 /*
199 * Return a locked vnode for the specified inode. The inode must be
200 * referenced but NOT LOCKED on entry and will remain referenced on
201 * return.
202 *
203 * Called from the frontend.
204 */
205 int
207 {
208 hammer_mount_t hmp;
225 }
246 }
248 /*
249 * Only mark as the root vnode if the ip is not
250 * historical, otherwise the VFS cache will get
251 * confused. The other half of the special handling
252 * is in hammer_vop_nlookupdotdot().
253 *
254 * Pseudo-filesystem roots also do not count.
255 */
260 }
263 /* vnode locked by getnewvnode() */
264 /* make related vnode dirty if inode dirty? */
269 }
271 /*
272 * loop if the vget fails (aka races), or if the vp
273 * no longer matches ip->vp.
274 */
279 }
280 }
283 }
285 /*
286 * Locate all copies of the inode for obj_id compatible with the specified
287 * asof, reference, and issue the related call-back. This routine is used
288 * for direct-io invalidation and does not create any new inodes.
289 */
290 void
294 {
296 hammer_inode_info_cmp_all_history,
298 }
300 /*
301 * Acquire a HAMMER inode. The returned inode is not locked. These functions
302 * do not attach or detach the related vnode (use hammer_get_vnode() for
303 * that).
304 *
305 * The flags argument is only applied for newly created inodes, and only
306 * certain flags are inherited.
307 *
308 * Called from the frontend.
309 */
314 {
320 /*
321 * Determine if we already have an inode cached. If we do then
322 * we are golden.
323 */
327 loop:
333 }
335 /*
336 * Allocate a new inode structure and deal with races later.
337 */
355 /*
356 * Locate the on-disk inode.
357 */
358 retry:
369 HAMMER_CURSOR_ASOF;
375 }
377 /*
378 * On success the B-Tree lookup will hold the appropriate
379 * buffer cache buffers and provide a pointer to the requested
380 * information. Copy the information to the in-memory inode
381 * and cache the B-Tree node to improve future operations.
382 */
387 /*
388 * cache[0] tries to cache the location of the object inode.
389 * The assumption is that it is near the directory inode.
390 *
391 * cache[1] tries to cache the location of the object data.
392 * The assumption is that it is near the directory data.
393 */
398 /*
399 * The file should not contain any data past the file size
400 * stored in the inode. Setting save_trunc_off to the
401 * file size instead of max reduces B-Tree lookup overheads
402 * on append by allowing the flusher to avoid checking for
403 * record overwrites.
404 */
406 }
408 /*
409 * The inode is placed on the red-black tree and will be synced to
410 * the media when flushed or by the filesystem sync. If this races
411 * another instantiation/lookup the insertion will fail.
412 */
424 }
430 }
438 }
441 }
443 /*
444 * Create a new filesystem object, returning the inode in *ipp. The
445 * returned inode will be referenced.
446 *
447 * The inode is created in-memory.
448 */
449 int
453 {
454 hammer_mount_t hmp;
455 hammer_inode_t ip;
456 uid_t xuid;
457 u_int32_t localization;
462 /*
463 * Assign the localization domain. If if dip is NULL we are creating
464 * a pseudo-fs and must locate an unused localization domain.
465 */
477 }
480 }
483 }
489 /*
490 * Allocate a new object id. If creating a new pseudo-fs the
491 * obj_id is 1.
492 */
495 else
509 /* ip->save_trunc_off = 0; (already zero) */
518 /*
519 * A nohistory designator on the parent directory is inherited by
520 * the child. We will do this even for pseudo-fs creation... the
521 * sysad can turn it off.
522 */
528 HAMMER_LOCALIZE_INODE;
541 /*
542 * Setup the ".." pointer. This only needs to be done for directories
543 * but we do it for all objects as a recovery aid.
544 *
545 * The parent_obj_localization field only applies to pseudo-fs roots.
546 */
552 }
562 }
564 /*
565 * Calculate default uid/gid and overwrite with information from
566 * the vap.
567 */
577 else
584 else
591 }
594 }
596 /*
597 * Called by hammer_sync_inode().
598 */
599 static int
601 {
603 hammer_record_t record;
607 retry:
610 /*
611 * If the inode has a presence on-disk then locate it and mark
612 * it deleted, setting DELONDISK.
613 *
614 * The record may or may not be physically deleted, depending on
615 * the retention policy.
616 */
618 HAMMER_INODE_ONDISK) {
621 HAMMER_LOCALIZE_INODE;
639 }
648 }
651 }
654 }
663 }
664 }
666 /*
667 * Ok, write out the initial record or a new record (after deleting
668 * the old one), unless the DELETED flag is set. This routine will
669 * clear DELONDISK if it writes out a record.
670 *
671 * Update our inode statistics if this is the first application of
672 * the inode on-disk.
673 */
675 /*
676 * Generate a record and write it to the media
677 */
688 /*
689 * If this flag is set we cannot sync the new file size
690 * because we haven't finished related truncations. The
691 * inode will be flushed in another flush group to finish
692 * the job.
693 */
700 }
716 }
720 }
722 /*
723 * The record isn't managed by the inode's record tree,
724 * destroy it whether we succeed or fail.
725 */
731 /*
732 * Finish up.
733 */
738 HAMMER_INODE_ATIME |
739 HAMMER_INODE_MTIME);
744 /*
745 * Root volume count of inodes
746 */
750 vol0_stat_inodes);
756 }
757 }
758 }
760 /*
761 * If the inode has been destroyed, clean out any left-over flags
762 * that may have been set by the frontend.
763 */
766 HAMMER_INODE_ATIME |
767 HAMMER_INODE_MTIME);
768 }
770 }
772 /*
773 * Update only the itimes fields.
774 *
775 * ATIME can be updated without generating any UNDO. MTIME is updated
776 * with UNDO so it is guaranteed to be synchronized properly in case of
777 * a crash.
778 *
779 * Neither field is included in the B-Tree leaf element's CRC, which is how
780 * we can get away with updating ATIME the way we do.
781 */
782 static int
784 {
788 retry:
790 HAMMER_INODE_ONDISK) {
792 }
796 HAMMER_LOCALIZE_INODE;
814 }
818 /*
819 * Updating MTIME requires an UNDO. Just cover
820 * both atime and mtime.
821 */
824 HAMMER_ITIMES_BYTES);
829 /*
830 * Updating atime only can be done in-place with
831 * no UNDO.
832 */
837 }
839 }
846 }
848 }
850 /*
851 * Release a reference on an inode, flush as requested.
852 *
853 * On the last reference we queue the inode to the flusher for its final
854 * disposition.
855 */
856 void
858 {
861 /*
862 * Handle disposition when dropping the last ref.
863 */
866 /*
867 * Determine whether on-disk action is needed for
868 * the inode's final disposition.
869 */
875 HAMMER_FLUSH_SIGNAL);
878 }
882 }
887 /*
888 * The inode still has multiple refs, try to drop
889 * one ref.
890 */
895 }
896 }
897 }
898 }
900 /*
901 * Unload and destroy the specified inode. Must be called with one remaining
902 * reference. The reference is disposed of.
903 *
904 * This can only be called in the context of the flusher.
905 */
906 static int
908 {
935 }
937 /*
938 * Called on mount -u when switching from RW to RO or vise-versa. Adjust
939 * the read-only flag for cached inodes.
940 *
941 * This routine is called from a RB_SCAN().
942 */
943 int
945 {
950 else
953 }
955 /*
956 * A transaction has modified an inode, requiring updates as specified by
957 * the passed flags.
958 *
959 * HAMMER_INODE_DDIRTY: Inode data has been updated
960 * HAMMER_INODE_XDIRTY: Dirty in-memory records
961 * HAMMER_INODE_BUFS: Dirty buffer cache buffers
962 * HAMMER_INODE_DELETED: Inode record/data must be deleted
963 * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
964 */
965 void
967 {
975 }
978 }
980 /*
981 * Request that an inode be flushed. This whole mess cannot block and may
982 * recurse (if not synchronous). Once requested HAMMER will attempt to
983 * actively flush the inode until the flush can be done.
984 *
985 * The inode may already be flushing, or may be in a setup state. We can
986 * place the inode in a flushing state if it is currently idle and flag it
987 * to reflush if it is currently flushing.
988 *
989 * If the HAMMER_FLUSH_SYNCHRONOUS flag is specified we will attempt to
990 * flush the indoe synchronously using the caller's context.
991 */
992 void
994 {
997 /*
998 * Trivial 'nothing to flush' case. If the inode is ina SETUP
999 * state we have to put it back into an IDLE state so we can
1000 * drop the extra ref.
1001 */
1006 }
1008 }
1010 /*
1011 * Our flush action will depend on the current state.
1012 */
1015 /*
1016 * We have no dependancies and can flush immediately. Some
1017 * our children may not be flushable so we have to re-test
1018 * with that additional knowledge.
1019 */
1023 /*
1024 * Recurse upwards through dependancies via target_list
1025 * and start their flusher actions going if possible.
1026 *
1027 * 'good' is our connectivity. -1 means we have none and
1028 * can't flush, 0 means there weren't any dependancies, and
1029 * 1 means we have good connectivity.
1030 */
1033 /*
1034 * We can continue if good >= 0. Determine how many records
1035 * under our inode can be flushed (and mark them).
1036 */
1044 }
1045 }
1048 /*
1049 * We are already flushing, flag the inode to reflush
1050 * if needed after it completes its current flush.
1051 */
1057 }
1059 }
1060 }
1062 /*
1063 * Scan ip->target_list, which is a list of records owned by PARENTS to our
1064 * ip which reference our ip.
1065 *
1066 * XXX This is a huge mess of recursive code, but not one bit of it blocks
1067 * so for now do not ref/deref the structures. Note that if we use the
1068 * ref/rel code later, the rel CAN block.
1069 */
1070 static int
1072 {
1073 hammer_record_t depend;
1074 #if 0
1075 hammer_record_t next;
1076 hammer_inode_t pip;
1077 #endif
1089 }
1092 #if 0
1093 retry:
1099 }
1106 }
1112 }
1121 }
1123 #endif
1124 }
1126 /*
1127 * This helper function takes a record representing the dependancy between
1128 * the parent inode and child inode.
1129 *
1130 * record->ip = parent inode
1131 * record->target_ip = child inode
1132 *
1133 * We are asked to recurse upwards and convert the record from SETUP
1134 * to FLUSH if possible.
1135 *
1136 * Return 1 if the record gives us connectivity
1137 *
1138 * Return 0 if the record is not relevant
1139 *
1140 * Return -1 if we can't resolve the dependancy and there is no connectivity.
1141 */
1142 static int
1144 {
1145 hammer_mount_t hmp;
1146 hammer_inode_t pip;
1153 /*
1154 * If the record is already flushing, is it in our flush group?
1155 *
1156 * If it is in our flush group but it is a general record or a
1157 * delete-on-disk, it does not improve our connectivity (return 0),
1158 * and if the target inode is not trying to destroy itself we can't
1159 * allow the operation yet anyway (the second return -1).
1160 */
1165 }
1168 /* GENERAL or DEL */
1170 }
1172 /*
1173 * It must be a setup record. Try to resolve the setup dependancies
1174 * by recursing upwards so we can place ip on the flush list.
1175 */
1180 /*
1181 * We can't flush ip because it has no connectivity (XXX also check
1182 * nlinks for pre-existing connectivity!). Flag it so any resolution
1183 * recurses back down.
1184 */
1188 }
1190 /*
1191 * We are go, place the parent inode in a flushing state so we can
1192 * place its record in a flushing state. Note that the parent
1193 * may already be flushing. The record must be in the same flush
1194 * group as the parent.
1195 */
1201 #if 0
1204 /*
1205 * Regardless of flushing state we cannot sync this path if the
1206 * record represents a delete-on-disk but the target inode
1207 * is not ready to sync its own deletion.
1208 *
1209 * XXX need to count effective nlinks to determine whether
1210 * the flush is ok, otherwise removing a hardlink will
1211 * just leave the DEL record to rot.
1212 */
1216 #endif
1218 /*
1219 * This is the record we wanted to synchronize. If the
1220 * record went into a flush state while we blocked it
1221 * had better be in the correct flush group.
1222 */
1229 }
1233 /*
1234 * A general or delete-on-disk record does not contribute
1235 * to our visibility. We can still flush it, however.
1236 */
1239 /*
1240 * We couldn't resolve the dependancies, request that the
1241 * inode be flushed when the dependancies can be resolved.
1242 */
1245 }
1246 }
1248 /*
1249 * This is the core routine placing an inode into the FST_FLUSH state.
1250 */
1251 static void
1253 {
1256 /*
1257 * Set flush state and prevent the flusher from cycling into
1258 * the next flush group. Do not place the ip on the list yet.
1259 * Inodes not in the idle state get an extra reference.
1260 */
1270 /*
1271 * We need to be able to vfsync/truncate from the backend.
1272 */
1277 }
1279 /*
1280 * Figure out how many in-memory records we can actually flush
1281 * (not including inode meta-data, buffers, etc).
1282 *
1283 * Do not add new records to the flush if this is a recursion or
1284 * if we must still complete a flush from the previous flush cycle.
1285 */
1295 }
1297 /*
1298 * This is a more involved test that includes go_count. If we
1299 * can't flush, flag the inode and return. If go_count is 0 we
1300 * were are unable to flush any records in our rec_tree and
1301 * must ignore the XDIRTY flag.
1302 */
1314 }
1318 }
1322 }
1323 }
1325 /*
1326 * Snapshot the state of the inode for the backend flusher.
1327 *
1328 * We continue to retain save_trunc_off even when all truncations
1329 * have been resolved as an optimization to determine if we can
1330 * skip the B-Tree lookup for overwrite deletions.
1331 *
1332 * NOTE: The DELETING flag is a mod flag, but it is also sticky,
1333 * and stays in ip->flags. Once set, it stays set until the
1334 * inode is destroyed.
1335 *
1336 * NOTE: If a truncation from a previous flush cycle had to be
1337 * continued into this one, the TRUNCATED flag will still be
1338 * set in sync_flags as will WOULDBLOCK. When this occurs
1339 * we CANNOT safely integrate a new truncation from the front-end
1340 * because there may be data records in-memory assigned a flush
1341 * state from the previous cycle that are supposed to be flushed
1342 * before the next frontend truncation.
1343 */
1345 HAMMER_INODE_TRUNCATED) {
1352 /*
1353 * The save_trunc_off used to cache whether the B-Tree
1354 * holds any records past that point is not used until
1355 * after the truncation has succeeded, so we can safely
1356 * set it now.
1357 */
1360 }
1366 #ifdef DEBUG_TRUNCATE
1369 #endif
1371 /*
1372 * The flusher list inherits our inode and reference.
1373 */
1380 }
1381 }
1383 /*
1384 * Callback for scan of ip->rec_tree. Try to include each record in our
1385 * flush. ip->flush_group has been set but the inode has not yet been
1386 * moved into a flushing state.
1387 *
1388 * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1389 * both inodes.
1390 *
1391 * We return 1 for any record placed or found in FST_FLUSH, which prevents
1392 * the caller from shortcutting the flush.
1393 */
1394 static int
1396 {
1397 hammer_inode_t target_ip;
1398 hammer_inode_t ip;
1401 /*
1402 * Deleted records are ignored. Note that the flush detects deleted
1403 * front-end records at multiple points to deal with races. This is
1404 * just the first line of defense. The only time DELETED_FE cannot
1405 * be set is when HAMMER_RECF_INTERLOCK_BE is set.
1406 *
1407 * Don't get confused between record deletion and, say, directory
1408 * entry deletion. The deletion of a directory entry that is on
1409 * the media has nothing to do with the record deletion flags.
1410 *
1411 * The flush_group for a record already in a flush state must
1412 * be updated. This case can only occur if the inode deleting
1413 * too many records had to be moved to the next flush group.
1414 */
1422 }
1424 }
1426 /*
1427 * If the record is in an idle state it has no dependancies and
1428 * can be flushed.
1429 */
1435 /*
1436 * Record has no setup dependancy, we can flush it.
1437 */
1445 /*
1446 * Record has a setup dependancy. Try to include the
1447 * target ip in the flush.
1448 *
1449 * We have to be careful here, if we do not do the right
1450 * thing we can lose track of dirty inodes and the system
1451 * will lockup trying to allocate buffers.
1452 */
1457 /*
1458 * If the target IP is already flushing in our group
1459 * we are golden, otherwise make sure the target
1460 * reflushes.
1461 */
1469 }
1471 /*
1472 * If the target IP is not flushing we can force
1473 * it to flush, even if it is unable to write out
1474 * any of its own records we have at least one in
1475 * hand that we CAN deal with.
1476 */
1481 HAMMER_FLUSH_RECURSION);
1484 /*
1485 * General or delete-on-disk record.
1486 *
1487 * XXX this needs help. If a delete-on-disk we could
1488 * disconnect the target. If the target has its own
1489 * dependancies they really need to be flushed.
1490 *
1491 * XXX
1492 */
1497 HAMMER_FLUSH_RECURSION);
1499 }
1502 /*
1503 * If the WOULDBLOCK flag is set records may have been left
1504 * over from a previous flush attempt and should be moved
1505 * to the current flush group. If it is not set then all
1506 * such records had better have been flushed already or
1507 * already associated with the current flush group.
1508 */
1513 }
1516 }
1518 }
1520 /*
1521 * This version just moves records already in a flush state to the new
1522 * flush group and that is it.
1523 */
1524 static int
1526 {
1535 }
1539 }
1541 }
1543 /*
1544 * Wait for a previously queued flush to complete. Not only do we need to
1545 * wait for the inode to sync out, we also may have to run the flusher again
1546 * to get it past the UNDO position pertaining to the flush so a crash does
1547 * not 'undo' our flush.
1548 */
1549 void
1551 {
1561 }
1562 /* XXX can we make this != FST_IDLE ? check SETUP depends */
1567 }
1571 }
1572 }
1574 /*
1575 * Called by the backend code when a flush has been completed.
1576 * The inode has already been removed from the flush list.
1577 *
1578 * A pipelined flush can occur, in which case we must re-enter the
1579 * inode on the list and re-copy its fields.
1580 */
1581 void
1583 {
1584 hammer_mount_t hmp;
1591 /*
1592 * Merge left-over flags back into the frontend and fix the state.
1593 * Incomplete truncations are retained by the backend.
1594 */
1598 /*
1599 * The backend may have adjusted nlinks, so if the adjusted nlinks
1600 * does not match the fronttend set the frontend's RDIRTY flag again.
1601 */
1605 /*
1606 * Fix up the dirty buffer status. IO completions will also
1607 * try to clean up rsv_databufs.
1608 */
1614 }
1616 /*
1617 * Re-set the XDIRTY flag if some of the inode's in-memory records
1618 * could not be flushed.
1619 */
1625 /*
1626 * Do not lose track of inodes which no longer have vnode
1627 * assocations, otherwise they may never get flushed again.
1628 */
1632 /*
1633 * Clean up the vnode ref
1634 */
1638 }
1640 /*
1641 * Adjust flush_state. The target state (idle or setup) shouldn't
1642 * be terribly important since we will reflush if we really need
1643 * to do anything.
1644 *
1645 * If the WOULDBLOCK flag is set we must re-flush immediately
1646 * to continue a potentially large deletion. The flag also causes
1647 * the hammer_setup_child_callback() to move records in the old
1648 * flush group to the new one.
1649 */
1662 }
1667 /*
1668 * If the frontend made more changes and requested another flush,
1669 * then try to get it running.
1670 */
1678 }
1679 }
1681 /*
1682 * If the inode is now clean drop the space reservation.
1683 */
1688 }
1690 /*
1691 * Finally, if the frontend is waiting for a flush to complete,
1692 * wake it up.
1693 */
1698 }
1699 }
1702 }
1704 /*
1705 * Called from hammer_sync_inode() to synchronize in-memory records
1706 * to the media.
1707 */
1708 static int
1710 {
1715 /*
1716 * Skip records that do not belong to the current flush.
1717 */
1722 #if 1
1724 kprintf("sync_record %p ip %p bad flush group %d %d\n", record, record->ip, record->flush_group ,record->ip->flush_group);
1727 }
1728 #endif
1731 /*
1732 * Interlock the record using the BE flag. Once BE is set the
1733 * frontend cannot change the state of FE.
1734 *
1735 * NOTE: If FE is set prior to us setting BE we still sync the
1736 * record out, but the flush completion code converts it to
1737 * a delete-on-disk record instead of destroying it.
1738 */
1742 /*
1743 * The backend may have already disposed of the record.
1744 */
1748 }
1750 /*
1751 * If the whole inode is being deleting all on-disk records will
1752 * be deleted very soon, we can't sync any new records to disk
1753 * because they will be deleted in the same transaction they were
1754 * created in (delete_tid == create_tid), which will assert.
1755 *
1756 * XXX There may be a case with RECORD_ADD with DELETED_FE set
1757 * that we currently panic on.
1758 */
1762 /*
1763 * We don't have to do anything, if the record was
1764 * committed the space will have been accounted for
1765 * in the blockmap.
1766 */
1767 /* fall through */
1782 /*
1783 * Follow through and issue the on-disk deletion
1784 */
1786 }
1787 }
1789 /*
1790 * If DELETED_FE is set special handling is needed for directory
1791 * entries. Dependant pieces related to the directory entry may
1792 * have already been synced to disk. If this occurs we have to
1793 * sync the directory entry and then change the in-memory record
1794 * from an ADD to a DELETE to cover the fact that it's been
1795 * deleted by the frontend.
1796 *
1797 * A directory delete covering record (MEM_RECORD_DEL) can never
1798 * be deleted by the frontend.
1799 *
1800 * Any other record type (aka DATA) can be deleted by the frontend.
1801 * XXX At the moment the flusher must skip it because there may
1802 * be another data record in the flush group for the same block,
1803 * meaning that some frontend data changes can leak into the backend's
1804 * synchronization point.
1805 */
1814 }
1815 }
1817 /*
1818 * Assign the create_tid for new records. Deletions already
1819 * have the record's entire key properly set up.
1820 */
1833 }
1842 }
1843 }
1844 done:
1847 }
1849 /*
1850 * XXX error handling
1851 */
1852 int
1854 {
1857 hammer_node_t tmp_node;
1858 hammer_record_t depend;
1859 hammer_record_t next;
1861 u_int64_t nlinks;
1871 /*
1872 * Any directory records referencing this inode which are not in
1873 * our current flush group must adjust our nlink count for the
1874 * purposes of synchronization to disk.
1875 *
1876 * Records which are in our flush group can be unlinked from our
1877 * inode now, potentially allowing the inode to be physically
1878 * deleted.
1879 *
1880 * This cannot block.
1881 */
1888 /*
1889 * If this is an ADD that was deleted by the frontend
1890 * the frontend nlinks count will have already been
1891 * decremented, but the backend is going to sync its
1892 * directory entry and must account for it. The
1893 * record will be converted to a delete-on-disk when
1894 * it gets synced.
1895 *
1896 * If the ADD was not deleted by the frontend we
1897 * can remove the dependancy from our target_list.
1898 */
1903 target_entry);
1905 }
1907 /*
1908 * Not part of our flush group
1909 */
1920 }
1921 }
1922 }
1924 /*
1925 * Set dirty if we had to modify the link count.
1926 */
1931 }
1933 /*
1934 * If there is a trunction queued destroy any data past the (aligned)
1935 * truncation point. Userland will have dealt with the buffer
1936 * containing the truncation point for us.
1937 *
1938 * We don't flush pending frontend data buffers until after we've
1939 * dealt with the truncation.
1940 */
1942 /*
1943 * Interlock trunc_off. The VOP front-end may continue to
1944 * make adjustments to it while we are blocked.
1945 */
1946 off_t trunc_off;
1947 off_t aligned_trunc_off;
1954 /*
1955 * Delete any whole blocks on-media. The front-end has
1956 * already cleaned out any partial block and made it
1957 * pending. The front-end may have updated trunc_off
1958 * while we were blocked so we only use sync_trunc_off.
1959 *
1960 * This operation can blow out the buffer cache, EWOULDBLOCK
1961 * means we were unable to complete the deletion. The
1962 * deletion will update sync_trunc_off in that case.
1963 */
1965 aligned_trunc_off,
1971 }
1976 /*
1977 * Clear the truncation flag on the backend after we have
1978 * complete the deletions. Backend data is now good again
1979 * (including new records we are about to sync, below).
1980 *
1981 * Leave sync_trunc_off intact. As we write additional
1982 * records the backend will update sync_trunc_off. This
1983 * tells the backend whether it can skip the overwrite
1984 * test. This should work properly even when the backend
1985 * writes full blocks where the truncation point straddles
1986 * the block because the comparison is against the base
1987 * offset of the record.
1988 */
1990 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
1993 }
1995 /*
1996 * Now sync related records. These will typically be directory
1997 * entries or delete-on-disk records.
1998 *
1999 * Not all records will be flushed, but clear XDIRTY anyway. We
2000 * will set it again in the frontend hammer_flush_inode_done()
2001 * if records remain.
2002 */
2010 }
2013 /*
2014 * Re-seek for inode update, assuming our cache hasn't been ripped
2015 * out from under us.
2016 */
2023 }
2025 }
2027 /*
2028 * If we are deleting the inode the frontend had better not have
2029 * any active references on elements making up the inode.
2030 *
2031 * The call to hammer_ip_delete_clean() cleans up auxillary records
2032 * but not DB or DATA records. Those must have already been deleted
2033 * by the normal truncation mechanic.
2034 */
2048 /*
2049 * Set delete_tid in both the frontend and backend
2050 * copy of the inode record. The DELETED flag handles
2051 * this, do not set RDIRTY.
2052 */
2059 /*
2060 * Adjust the inode count in the volume header
2061 */
2065 vol0_stat_inodes);
2068 }
2071 }
2072 }
2079 defer_buffer_flush:
2080 /*
2081 * Now update the inode's on-disk inode-data and/or on-disk record.
2082 * DELETED and ONDISK are managed only in ip->flags.
2083 *
2084 * In the case of a defered buffer flush we still update the on-disk
2085 * inode to satisfy visibility requirements if there happen to be
2086 * directory dependancies.
2087 */
2090 /*
2091 * If deleted and on-disk, don't set any additional flags.
2092 * the delete flag takes care of things.
2093 *
2094 * Clear flags which may have been set by the frontend.
2095 */
2098 HAMMER_INODE_DELETING);
2101 /*
2102 * Take care of the case where a deleted inode was never
2103 * flushed to the disk in the first place.
2104 *
2105 * Clear flags which may have been set by the frontend.
2106 */
2109 HAMMER_INODE_DELETING);
2117 }
2120 /*
2121 * If already on-disk, do not set any additional flags.
2122 */
2125 /*
2126 * If not on-disk and not deleted, set DDIRTY to force
2127 * an initial record to be written.
2128 *
2129 * Also set the create_tid in both the frontend and backend
2130 * copy of the inode record.
2131 */
2138 }
2140 /*
2141 * If RDIRTY or DDIRTY is set, write out a new record. If the inode
2142 * is already on-disk the old record is marked as deleted.
2143 *
2144 * If DELETED is set hammer_update_inode() will delete the existing
2145 * record without writing out a new one.
2146 *
2147 * If *ONLY* the ITIMES flag is set we can update the record in-place.
2148 */
2158 }
2161 done:
2162 /*
2163 * Save the TID we used to sync the inode with to make sure we
2164 * do not improperly reuse it.
2165 */
2169 }
2171 /*
2172 * This routine is called when the OS is no longer actively referencing
2173 * the inode (but might still be keeping it cached), or when releasing
2174 * the last reference to an inode.
2175 *
2176 * At this point if the inode's nlinks count is zero we want to destroy
2177 * it, which may mean destroying it on-media too.
2178 */
2179 void
2181 {
2184 /*
2185 * Set the DELETING flag when the link count drops to 0 and the
2186 * OS no longer has any opens on the inode.
2187 *
2188 * The backend will clear DELETING (a mod flag) and set DELETED
2189 * (a state flag) when it is actually able to perform the
2190 * operation.
2191 */
2201 }
2203 /*
2204 * Final cleanup
2205 */
2209 }
2212 }
2213 }
2214 }
2216 /*
2217 * Re-test an inode when a dependancy had gone away to see if we
2218 * can chain flush it.
2219 */
2220 void
2222 {
2231 }
2233 }
2234 }
2236 /*
2237 * Clear the RECLAIM flag on an inode. This occurs when the inode is
2238 * reassociated with a vp or just before it gets freed.
2239 *
2240 * Wakeup one thread blocked waiting on reclaims to complete. Note that
2241 * the inode the thread is waiting on behalf of is a different inode then
2242 * the inode we are called with. This is to create a pipeline.
2243 */
2244 static void
2246 {
2261 }
2262 }
2264 /*
2265 * Setup our reclaim pipeline. We only let so many detached (and dirty)
2266 * inodes build up before we start blocking.
2267 *
2268 * When we block we don't care *which* inode has finished reclaiming,
2269 * as lone as one does. This is somewhat heuristical... we also put a
2270 * cap on how long we are willing to wait.
2271 */
2272 void
2274 {
2284 }
2288 HAMMER_RECLAIM_WAIT;
2293 }
2294 }