| blob | c0a53d897b0c07bab86746a39055c3f7e5995f12 |
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 1.95 2008/07/07 03:49:50 dillon Exp $
35 */
38 #include <vm/vm_extern.h>
39 #include <sys/buf.h>
40 #include <sys/buf2.h>
51 #ifdef DEBUG_TRUNCATE
53 #endif
55 /*
56 * RB-Tree support for inode structures
57 */
58 int
60 {
74 }
76 /*
77 * RB-Tree support for inode structures / special LOOKUP_INFO
78 */
79 static int
81 {
95 }
97 /*
98 * Used by hammer_scan_inode_snapshots() to locate all of an object's
99 * snapshots. Note that the asof field is not tested, which we can get
100 * away with because it is the lowest-priority field.
101 */
102 static int
104 {
116 }
118 /*
119 * RB-Tree support for pseudofs structures
120 */
121 static int
123 {
129 }
138 /*
139 * The kernel is not actively referencing this vnode but is still holding
140 * it cached.
141 *
142 * This is called from the frontend.
143 */
144 int
146 {
149 /*
150 * Degenerate case
151 */
155 }
157 /*
158 * If the inode no longer has visibility in the filesystem try to
159 * recycle it immediately, even if the inode is dirty. Recycling
160 * it quickly allows the system to reclaim buffer cache and VM
161 * resources which can matter a lot in a heavily loaded system.
162 *
163 * This can deadlock in vfsync() if we aren't careful.
164 *
165 * Do not queue the inode to the flusher if we still have visibility,
166 * otherwise namespace calls such as chmod will unnecessarily generate
167 * multiple inode updates.
168 */
174 }
176 }
178 /*
179 * Release the vnode association. This is typically (but not always)
180 * the last reference on the inode.
181 *
182 * Once the association is lost we are on our own with regards to
183 * flushing the inode.
184 */
185 int
187 {
189 hammer_mount_t hmp;
206 }
207 }
209 }
211 }
213 /*
214 * Return a locked vnode for the specified inode. The inode must be
215 * referenced but NOT LOCKED on entry and will remain referenced on
216 * return.
217 *
218 * Called from the frontend.
219 */
220 int
222 {
223 hammer_mount_t hmp;
240 }
261 }
263 /*
264 * Only mark as the root vnode if the ip is not
265 * historical, otherwise the VFS cache will get
266 * confused. The other half of the special handling
267 * is in hammer_vop_nlookupdotdot().
268 *
269 * Pseudo-filesystem roots also do not count.
270 */
275 }
278 /* vnode locked by getnewvnode() */
279 /* make related vnode dirty if inode dirty? */
284 }
286 /*
287 * loop if the vget fails (aka races), or if the vp
288 * no longer matches ip->vp.
289 */
294 }
295 }
298 }
300 /*
301 * Locate all copies of the inode for obj_id compatible with the specified
302 * asof, reference, and issue the related call-back. This routine is used
303 * for direct-io invalidation and does not create any new inodes.
304 */
305 void
309 {
311 hammer_inode_info_cmp_all_history,
313 }
315 /*
316 * Acquire a HAMMER inode. The returned inode is not locked. These functions
317 * do not attach or detach the related vnode (use hammer_get_vnode() for
318 * that).
319 *
320 * The flags argument is only applied for newly created inodes, and only
321 * certain flags are inherited.
322 *
323 * Called from the frontend.
324 */
329 {
336 /*
337 * Determine if we already have an inode cached. If we do then
338 * we are golden.
339 */
343 loop:
349 }
351 /*
352 * Allocate a new inode structure and deal with races later.
353 */
372 /*
373 * Locate the on-disk inode.
374 */
375 retry:
386 HAMMER_CURSOR_ASOF;
392 }
394 /*
395 * On success the B-Tree lookup will hold the appropriate
396 * buffer cache buffers and provide a pointer to the requested
397 * information. Copy the information to the in-memory inode
398 * and cache the B-Tree node to improve future operations.
399 */
404 /*
405 * cache[0] tries to cache the location of the object inode.
406 * The assumption is that it is near the directory inode.
407 *
408 * cache[1] tries to cache the location of the object data.
409 * The assumption is that it is near the directory data.
410 */
415 /*
416 * The file should not contain any data past the file size
417 * stored in the inode. Setting save_trunc_off to the
418 * file size instead of max reduces B-Tree lookup overheads
419 * on append by allowing the flusher to avoid checking for
420 * record overwrites.
421 */
424 /*
425 * Locate and assign the pseudofs management structure to
426 * the inode.
427 */
433 }
434 }
436 /*
437 * The inode is placed on the red-black tree and will be synced to
438 * the media when flushed or by the filesystem sync. If this races
439 * another instantiation/lookup the insertion will fail.
440 */
446 }
452 }
456 }
459 }
461 /*
462 * Create a new filesystem object, returning the inode in *ipp. The
463 * returned inode will be referenced.
464 *
465 * The inode is created in-memory.
466 */
467 int
471 {
472 hammer_mount_t hmp;
473 hammer_inode_t ip;
474 uid_t xuid;
475 u_int32_t localization;
480 /*
481 * Assign the localization domain. If if dip is NULL we are creating
482 * a pseudo-fs and must locate an unused localization domain.
483 */
495 }
498 }
501 }
507 /*
508 * Allocate a new object id. If creating a new pseudo-fs the
509 * obj_id is 1.
510 */
513 else
527 /* ip->save_trunc_off = 0; (already zero) */
536 /*
537 * A nohistory designator on the parent directory is inherited by
538 * the child. We will do this even for pseudo-fs creation... the
539 * sysad can turn it off.
540 */
546 HAMMER_LOCALIZE_INODE;
559 /*
560 * Setup the ".." pointer. This only needs to be done for directories
561 * but we do it for all objects as a recovery aid.
562 *
563 * The parent_obj_localization field only applies to pseudo-fs roots.
564 */
570 }
580 }
582 /*
583 * Calculate default uid/gid and overwrite with information from
584 * the vap.
585 */
595 else
602 else
613 }
620 /* not reached */
622 }
625 }
627 /*
628 * Final cleanup / freeing of an inode structure
629 */
630 static void
632 {
644 }
647 }
649 /*
650 * Retrieve pseudo-fs data.
651 */
652 int
654 {
656 hammer_pseudofs_inmem_t pfsm;
661 retry:
669 }
678 HAMMER_LOCALIZE_MISC;
696 }
699 }
709 }
712 /*
713 * Certain aspects of the pseudofs configuration are reflected
714 * in the inode.
715 */
721 }
725 }
727 }
729 /*
730 * Store pseudo-fs data. The backend will automatically delete any prior
731 * on-disk pseudo-fs data but we have to delete in-memory versions.
732 */
733 int
735 {
737 hammer_pseudofs_inmem_t pfsm;
738 hammer_record_t record;
741 retry:
746 HAMMER_LOCALIZE_MISC;
767 }
768 }
774 HAMMER_LOCALIZE_MISC;
780 }
785 /*
786 * Certain aspects of the pseudofs configuration are reflected
787 * in the inode. Note that we cannot mess with the as-of or
788 * clear the read-only state.
789 *
790 * If this inode represented a slave snapshot its asof will
791 * be set to a snapshot tid. When clearing slave mode any
792 * re-access of the inode via the parent directory will
793 * wind up using a different asof and thus will instantiate
794 * a new inode.
795 */
803 }
804 }
806 }
808 void
810 {
815 }
816 }
818 /*
819 * Called by hammer_sync_inode().
820 */
821 static int
823 {
825 hammer_record_t record;
829 retry:
832 /*
833 * If the inode has a presence on-disk then locate it and mark
834 * it deleted, setting DELONDISK.
835 *
836 * The record may or may not be physically deleted, depending on
837 * the retention policy.
838 */
840 HAMMER_INODE_ONDISK) {
843 HAMMER_LOCALIZE_INODE;
861 }
870 }
873 }
876 }
885 }
886 }
888 /*
889 * Ok, write out the initial record or a new record (after deleting
890 * the old one), unless the DELETED flag is set. This routine will
891 * clear DELONDISK if it writes out a record.
892 *
893 * Update our inode statistics if this is the first application of
894 * the inode on-disk.
895 */
897 /*
898 * Generate a record and write it to the media
899 */
910 /*
911 * If this flag is set we cannot sync the new file size
912 * because we haven't finished related truncations. The
913 * inode will be flushed in another flush group to finish
914 * the job.
915 */
922 }
938 }
942 }
944 /*
945 * The record isn't managed by the inode's record tree,
946 * destroy it whether we succeed or fail.
947 */
953 /*
954 * Finish up.
955 */
960 HAMMER_INODE_ATIME |
961 HAMMER_INODE_MTIME);
966 /*
967 * Root volume count of inodes
968 */
972 vol0_stat_inodes);
978 }
979 }
980 }
982 /*
983 * If the inode has been destroyed, clean out any left-over flags
984 * that may have been set by the frontend.
985 */
988 HAMMER_INODE_ATIME |
989 HAMMER_INODE_MTIME);
990 }
992 }
994 /*
995 * Update only the itimes fields.
996 *
997 * ATIME can be updated without generating any UNDO. MTIME is updated
998 * with UNDO so it is guaranteed to be synchronized properly in case of
999 * a crash.
1000 *
1001 * Neither field is included in the B-Tree leaf element's CRC, which is how
1002 * we can get away with updating ATIME the way we do.
1003 */
1004 static int
1006 {
1010 retry:
1012 HAMMER_INODE_ONDISK) {
1014 }
1018 HAMMER_LOCALIZE_INODE;
1036 }
1040 /*
1041 * Updating MTIME requires an UNDO. Just cover
1042 * both atime and mtime.
1043 */
1046 HAMMER_ITIMES_BYTES);
1051 /*
1052 * Updating atime only can be done in-place with
1053 * no UNDO.
1054 */
1059 }
1061 }
1068 }
1070 }
1072 /*
1073 * Release a reference on an inode, flush as requested.
1074 *
1075 * On the last reference we queue the inode to the flusher for its final
1076 * disposition.
1077 */
1078 void
1080 {
1083 /*
1084 * Handle disposition when dropping the last ref.
1085 */
1088 /*
1089 * Determine whether on-disk action is needed for
1090 * the inode's final disposition.
1091 */
1097 HAMMER_FLUSH_SIGNAL);
1100 }
1104 }
1109 /*
1110 * The inode still has multiple refs, try to drop
1111 * one ref.
1112 */
1117 }
1118 }
1119 }
1120 }
1122 /*
1123 * Unload and destroy the specified inode. Must be called with one remaining
1124 * reference. The reference is disposed of.
1125 *
1126 * This can only be called in the context of the flusher.
1127 */
1128 static int
1130 {
1148 }
1150 /*
1151 * Called on mount -u when switching from RW to RO or vise-versa. Adjust
1152 * the read-only flag for cached inodes.
1153 *
1154 * This routine is called from a RB_SCAN().
1155 */
1156 int
1158 {
1163 else
1166 }
1168 /*
1169 * A transaction has modified an inode, requiring updates as specified by
1170 * the passed flags.
1171 *
1172 * HAMMER_INODE_DDIRTY: Inode data has been updated
1173 * HAMMER_INODE_XDIRTY: Dirty in-memory records
1174 * HAMMER_INODE_BUFS: Dirty buffer cache buffers
1175 * HAMMER_INODE_DELETED: Inode record/data must be deleted
1176 * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1177 */
1178 void
1180 {
1188 }
1191 }
1193 /*
1194 * Request that an inode be flushed. This whole mess cannot block and may
1195 * recurse (if not synchronous). Once requested HAMMER will attempt to
1196 * actively flush the inode until the flush can be done.
1197 *
1198 * The inode may already be flushing, or may be in a setup state. We can
1199 * place the inode in a flushing state if it is currently idle and flag it
1200 * to reflush if it is currently flushing.
1201 *
1202 * If the HAMMER_FLUSH_SYNCHRONOUS flag is specified we will attempt to
1203 * flush the indoe synchronously using the caller's context.
1204 */
1205 void
1207 {
1210 /*
1211 * Trivial 'nothing to flush' case. If the inode is ina SETUP
1212 * state we have to put it back into an IDLE state so we can
1213 * drop the extra ref.
1214 */
1219 }
1221 }
1223 /*
1224 * Our flush action will depend on the current state.
1225 */
1228 /*
1229 * We have no dependancies and can flush immediately. Some
1230 * our children may not be flushable so we have to re-test
1231 * with that additional knowledge.
1232 */
1236 /*
1237 * Recurse upwards through dependancies via target_list
1238 * and start their flusher actions going if possible.
1239 *
1240 * 'good' is our connectivity. -1 means we have none and
1241 * can't flush, 0 means there weren't any dependancies, and
1242 * 1 means we have good connectivity.
1243 */
1246 /*
1247 * We can continue if good >= 0. Determine how many records
1248 * under our inode can be flushed (and mark them).
1249 */
1257 }
1258 }
1261 /*
1262 * We are already flushing, flag the inode to reflush
1263 * if needed after it completes its current flush.
1264 */
1270 }
1272 }
1273 }
1275 /*
1276 * Scan ip->target_list, which is a list of records owned by PARENTS to our
1277 * ip which reference our ip.
1278 *
1279 * XXX This is a huge mess of recursive code, but not one bit of it blocks
1280 * so for now do not ref/deref the structures. Note that if we use the
1281 * ref/rel code later, the rel CAN block.
1282 */
1283 static int
1285 {
1286 hammer_record_t depend;
1287 #if 0
1288 hammer_record_t next;
1289 hammer_inode_t pip;
1290 #endif
1302 }
1305 #if 0
1306 retry:
1312 }
1319 }
1325 }
1334 }
1336 #endif
1337 }
1339 /*
1340 * This helper function takes a record representing the dependancy between
1341 * the parent inode and child inode.
1342 *
1343 * record->ip = parent inode
1344 * record->target_ip = child inode
1345 *
1346 * We are asked to recurse upwards and convert the record from SETUP
1347 * to FLUSH if possible.
1348 *
1349 * Return 1 if the record gives us connectivity
1350 *
1351 * Return 0 if the record is not relevant
1352 *
1353 * Return -1 if we can't resolve the dependancy and there is no connectivity.
1354 */
1355 static int
1357 {
1358 hammer_mount_t hmp;
1359 hammer_inode_t pip;
1366 /*
1367 * If the record is already flushing, is it in our flush group?
1368 *
1369 * If it is in our flush group but it is a general record or a
1370 * delete-on-disk, it does not improve our connectivity (return 0),
1371 * and if the target inode is not trying to destroy itself we can't
1372 * allow the operation yet anyway (the second return -1).
1373 */
1378 }
1381 /* GENERAL or DEL */
1383 }
1385 /*
1386 * It must be a setup record. Try to resolve the setup dependancies
1387 * by recursing upwards so we can place ip on the flush list.
1388 */
1393 /*
1394 * We can't flush ip because it has no connectivity (XXX also check
1395 * nlinks for pre-existing connectivity!). Flag it so any resolution
1396 * recurses back down.
1397 */
1401 }
1403 /*
1404 * We are go, place the parent inode in a flushing state so we can
1405 * place its record in a flushing state. Note that the parent
1406 * may already be flushing. The record must be in the same flush
1407 * group as the parent.
1408 */
1414 #if 0
1417 /*
1418 * Regardless of flushing state we cannot sync this path if the
1419 * record represents a delete-on-disk but the target inode
1420 * is not ready to sync its own deletion.
1421 *
1422 * XXX need to count effective nlinks to determine whether
1423 * the flush is ok, otherwise removing a hardlink will
1424 * just leave the DEL record to rot.
1425 */
1429 #endif
1431 /*
1432 * This is the record we wanted to synchronize. If the
1433 * record went into a flush state while we blocked it
1434 * had better be in the correct flush group.
1435 */
1442 }
1446 /*
1447 * A general or delete-on-disk record does not contribute
1448 * to our visibility. We can still flush it, however.
1449 */
1452 /*
1453 * We couldn't resolve the dependancies, request that the
1454 * inode be flushed when the dependancies can be resolved.
1455 */
1458 }
1459 }
1461 /*
1462 * This is the core routine placing an inode into the FST_FLUSH state.
1463 */
1464 static void
1466 {
1469 /*
1470 * Set flush state and prevent the flusher from cycling into
1471 * the next flush group. Do not place the ip on the list yet.
1472 * Inodes not in the idle state get an extra reference.
1473 */
1483 /*
1484 * We need to be able to vfsync/truncate from the backend.
1485 */
1490 }
1492 /*
1493 * Figure out how many in-memory records we can actually flush
1494 * (not including inode meta-data, buffers, etc).
1495 *
1496 * Do not add new records to the flush if this is a recursion or
1497 * if we must still complete a flush from the previous flush cycle.
1498 */
1508 }
1510 /*
1511 * This is a more involved test that includes go_count. If we
1512 * can't flush, flag the inode and return. If go_count is 0 we
1513 * were are unable to flush any records in our rec_tree and
1514 * must ignore the XDIRTY flag.
1515 */
1527 }
1531 }
1535 }
1536 }
1538 /*
1539 * Snapshot the state of the inode for the backend flusher.
1540 *
1541 * We continue to retain save_trunc_off even when all truncations
1542 * have been resolved as an optimization to determine if we can
1543 * skip the B-Tree lookup for overwrite deletions.
1544 *
1545 * NOTE: The DELETING flag is a mod flag, but it is also sticky,
1546 * and stays in ip->flags. Once set, it stays set until the
1547 * inode is destroyed.
1548 *
1549 * NOTE: If a truncation from a previous flush cycle had to be
1550 * continued into this one, the TRUNCATED flag will still be
1551 * set in sync_flags as will WOULDBLOCK. When this occurs
1552 * we CANNOT safely integrate a new truncation from the front-end
1553 * because there may be data records in-memory assigned a flush
1554 * state from the previous cycle that are supposed to be flushed
1555 * before the next frontend truncation.
1556 */
1558 HAMMER_INODE_TRUNCATED) {
1565 /*
1566 * The save_trunc_off used to cache whether the B-Tree
1567 * holds any records past that point is not used until
1568 * after the truncation has succeeded, so we can safely
1569 * set it now.
1570 */
1573 }
1579 #ifdef DEBUG_TRUNCATE
1582 #endif
1584 /*
1585 * The flusher list inherits our inode and reference.
1586 */
1593 }
1594 }
1596 /*
1597 * Callback for scan of ip->rec_tree. Try to include each record in our
1598 * flush. ip->flush_group has been set but the inode has not yet been
1599 * moved into a flushing state.
1600 *
1601 * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1602 * both inodes.
1603 *
1604 * We return 1 for any record placed or found in FST_FLUSH, which prevents
1605 * the caller from shortcutting the flush.
1606 */
1607 static int
1609 {
1610 hammer_inode_t target_ip;
1611 hammer_inode_t ip;
1614 /*
1615 * Deleted records are ignored. Note that the flush detects deleted
1616 * front-end records at multiple points to deal with races. This is
1617 * just the first line of defense. The only time DELETED_FE cannot
1618 * be set is when HAMMER_RECF_INTERLOCK_BE is set.
1619 *
1620 * Don't get confused between record deletion and, say, directory
1621 * entry deletion. The deletion of a directory entry that is on
1622 * the media has nothing to do with the record deletion flags.
1623 *
1624 * The flush_group for a record already in a flush state must
1625 * be updated. This case can only occur if the inode deleting
1626 * too many records had to be moved to the next flush group.
1627 */
1635 }
1637 }
1639 /*
1640 * If the record is in an idle state it has no dependancies and
1641 * can be flushed.
1642 */
1648 /*
1649 * Record has no setup dependancy, we can flush it.
1650 */
1658 /*
1659 * Record has a setup dependancy. Try to include the
1660 * target ip in the flush.
1661 *
1662 * We have to be careful here, if we do not do the right
1663 * thing we can lose track of dirty inodes and the system
1664 * will lockup trying to allocate buffers.
1665 */
1670 /*
1671 * If the target IP is already flushing in our group
1672 * we are golden, otherwise make sure the target
1673 * reflushes.
1674 */
1682 }
1684 /*
1685 * If the target IP is not flushing we can force
1686 * it to flush, even if it is unable to write out
1687 * any of its own records we have at least one in
1688 * hand that we CAN deal with.
1689 */
1694 HAMMER_FLUSH_RECURSION);
1697 /*
1698 * General or delete-on-disk record.
1699 *
1700 * XXX this needs help. If a delete-on-disk we could
1701 * disconnect the target. If the target has its own
1702 * dependancies they really need to be flushed.
1703 *
1704 * XXX
1705 */
1710 HAMMER_FLUSH_RECURSION);
1712 }
1715 /*
1716 * If the WOULDBLOCK flag is set records may have been left
1717 * over from a previous flush attempt and should be moved
1718 * to the current flush group. If it is not set then all
1719 * such records had better have been flushed already or
1720 * already associated with the current flush group.
1721 */
1726 }
1729 }
1731 }
1733 /*
1734 * This version just moves records already in a flush state to the new
1735 * flush group and that is it.
1736 */
1737 static int
1739 {
1748 }
1752 }
1754 }
1756 /*
1757 * Wait for a previously queued flush to complete. Not only do we need to
1758 * wait for the inode to sync out, we also may have to run the flusher again
1759 * to get it past the UNDO position pertaining to the flush so a crash does
1760 * not 'undo' our flush.
1761 */
1762 void
1764 {
1774 }
1775 /* XXX can we make this != FST_IDLE ? check SETUP depends */
1780 }
1784 }
1785 }
1787 /*
1788 * Called by the backend code when a flush has been completed.
1789 * The inode has already been removed from the flush list.
1790 *
1791 * A pipelined flush can occur, in which case we must re-enter the
1792 * inode on the list and re-copy its fields.
1793 */
1794 void
1796 {
1797 hammer_mount_t hmp;
1804 /*
1805 * Merge left-over flags back into the frontend and fix the state.
1806 * Incomplete truncations are retained by the backend.
1807 */
1811 /*
1812 * The backend may have adjusted nlinks, so if the adjusted nlinks
1813 * does not match the fronttend set the frontend's RDIRTY flag again.
1814 */
1818 /*
1819 * Fix up the dirty buffer status.
1820 */
1823 }
1825 /*
1826 * Re-set the XDIRTY flag if some of the inode's in-memory records
1827 * could not be flushed.
1828 */
1834 /*
1835 * Do not lose track of inodes which no longer have vnode
1836 * assocations, otherwise they may never get flushed again.
1837 */
1841 /*
1842 * Clean up the vnode ref
1843 */
1847 }
1849 /*
1850 * Adjust flush_state. The target state (idle or setup) shouldn't
1851 * be terribly important since we will reflush if we really need
1852 * to do anything.
1853 *
1854 * If the WOULDBLOCK flag is set we must re-flush immediately
1855 * to continue a potentially large deletion. The flag also causes
1856 * the hammer_setup_child_callback() to move records in the old
1857 * flush group to the new one.
1858 */
1870 }
1875 /*
1876 * If the frontend made more changes and requested another flush,
1877 * then try to get it running.
1878 */
1886 }
1887 }
1889 /*
1890 * If the inode is now clean drop the space reservation.
1891 */
1896 }
1898 /*
1899 * Finally, if the frontend is waiting for a flush to complete,
1900 * wake it up.
1901 */
1906 }
1907 }
1910 }
1912 /*
1913 * Called from hammer_sync_inode() to synchronize in-memory records
1914 * to the media.
1915 */
1916 static int
1918 {
1923 /*
1924 * Skip records that do not belong to the current flush.
1925 */
1930 #if 1
1932 kprintf("sync_record %p ip %p bad flush group %d %d\n", record, record->ip, record->flush_group ,record->ip->flush_group);
1935 }
1936 #endif
1939 /*
1940 * Interlock the record using the BE flag. Once BE is set the
1941 * frontend cannot change the state of FE.
1942 *
1943 * NOTE: If FE is set prior to us setting BE we still sync the
1944 * record out, but the flush completion code converts it to
1945 * a delete-on-disk record instead of destroying it.
1946 */
1950 /*
1951 * The backend may have already disposed of the record.
1952 */
1956 }
1958 /*
1959 * If the whole inode is being deleting all on-disk records will
1960 * be deleted very soon, we can't sync any new records to disk
1961 * because they will be deleted in the same transaction they were
1962 * created in (delete_tid == create_tid), which will assert.
1963 *
1964 * XXX There may be a case with RECORD_ADD with DELETED_FE set
1965 * that we currently panic on.
1966 */
1970 /*
1971 * We don't have to do anything, if the record was
1972 * committed the space will have been accounted for
1973 * in the blockmap.
1974 */
1975 /* fall through */
1990 /*
1991 * Follow through and issue the on-disk deletion
1992 */
1994 }
1995 }
1997 /*
1998 * If DELETED_FE is set special handling is needed for directory
1999 * entries. Dependant pieces related to the directory entry may
2000 * have already been synced to disk. If this occurs we have to
2001 * sync the directory entry and then change the in-memory record
2002 * from an ADD to a DELETE to cover the fact that it's been
2003 * deleted by the frontend.
2004 *
2005 * A directory delete covering record (MEM_RECORD_DEL) can never
2006 * be deleted by the frontend.
2007 *
2008 * Any other record type (aka DATA) can be deleted by the frontend.
2009 * XXX At the moment the flusher must skip it because there may
2010 * be another data record in the flush group for the same block,
2011 * meaning that some frontend data changes can leak into the backend's
2012 * synchronization point.
2013 */
2022 }
2023 }
2025 /*
2026 * Assign the create_tid for new records. Deletions already
2027 * have the record's entire key properly set up.
2028 */
2041 }
2050 }
2051 }
2052 done:
2055 }
2057 /*
2058 * XXX error handling
2059 */
2060 int
2062 {
2065 hammer_node_t tmp_node;
2066 hammer_record_t depend;
2067 hammer_record_t next;
2069 u_int64_t nlinks;
2079 /*
2080 * Any directory records referencing this inode which are not in
2081 * our current flush group must adjust our nlink count for the
2082 * purposes of synchronization to disk.
2083 *
2084 * Records which are in our flush group can be unlinked from our
2085 * inode now, potentially allowing the inode to be physically
2086 * deleted.
2087 *
2088 * This cannot block.
2089 */
2096 /*
2097 * If this is an ADD that was deleted by the frontend
2098 * the frontend nlinks count will have already been
2099 * decremented, but the backend is going to sync its
2100 * directory entry and must account for it. The
2101 * record will be converted to a delete-on-disk when
2102 * it gets synced.
2103 *
2104 * If the ADD was not deleted by the frontend we
2105 * can remove the dependancy from our target_list.
2106 */
2111 target_entry);
2113 }
2115 /*
2116 * Not part of our flush group
2117 */
2128 }
2129 }
2130 }
2132 /*
2133 * Set dirty if we had to modify the link count.
2134 */
2139 }
2141 /*
2142 * If there is a trunction queued destroy any data past the (aligned)
2143 * truncation point. Userland will have dealt with the buffer
2144 * containing the truncation point for us.
2145 *
2146 * We don't flush pending frontend data buffers until after we've
2147 * dealt with the truncation.
2148 */
2150 /*
2151 * Interlock trunc_off. The VOP front-end may continue to
2152 * make adjustments to it while we are blocked.
2153 */
2154 off_t trunc_off;
2155 off_t aligned_trunc_off;
2162 /*
2163 * Delete any whole blocks on-media. The front-end has
2164 * already cleaned out any partial block and made it
2165 * pending. The front-end may have updated trunc_off
2166 * while we were blocked so we only use sync_trunc_off.
2167 *
2168 * This operation can blow out the buffer cache, EWOULDBLOCK
2169 * means we were unable to complete the deletion. The
2170 * deletion will update sync_trunc_off in that case.
2171 */
2173 aligned_trunc_off,
2179 }
2184 /*
2185 * Clear the truncation flag on the backend after we have
2186 * complete the deletions. Backend data is now good again
2187 * (including new records we are about to sync, below).
2188 *
2189 * Leave sync_trunc_off intact. As we write additional
2190 * records the backend will update sync_trunc_off. This
2191 * tells the backend whether it can skip the overwrite
2192 * test. This should work properly even when the backend
2193 * writes full blocks where the truncation point straddles
2194 * the block because the comparison is against the base
2195 * offset of the record.
2196 */
2198 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2201 }
2203 /*
2204 * Now sync related records. These will typically be directory
2205 * entries or delete-on-disk records.
2206 *
2207 * Not all records will be flushed, but clear XDIRTY anyway. We
2208 * will set it again in the frontend hammer_flush_inode_done()
2209 * if records remain.
2210 */
2218 }
2221 /*
2222 * Re-seek for inode update, assuming our cache hasn't been ripped
2223 * out from under us.
2224 */
2234 }
2236 }
2238 /*
2239 * If we are deleting the inode the frontend had better not have
2240 * any active references on elements making up the inode.
2241 *
2242 * The call to hammer_ip_delete_clean() cleans up auxillary records
2243 * but not DB or DATA records. Those must have already been deleted
2244 * by the normal truncation mechanic.
2245 */
2259 /*
2260 * Set delete_tid in both the frontend and backend
2261 * copy of the inode record. The DELETED flag handles
2262 * this, do not set RDIRTY.
2263 */
2270 /*
2271 * Adjust the inode count in the volume header
2272 */
2276 vol0_stat_inodes);
2279 }
2282 }
2283 }
2290 defer_buffer_flush:
2291 /*
2292 * Now update the inode's on-disk inode-data and/or on-disk record.
2293 * DELETED and ONDISK are managed only in ip->flags.
2294 *
2295 * In the case of a defered buffer flush we still update the on-disk
2296 * inode to satisfy visibility requirements if there happen to be
2297 * directory dependancies.
2298 */
2301 /*
2302 * If deleted and on-disk, don't set any additional flags.
2303 * the delete flag takes care of things.
2304 *
2305 * Clear flags which may have been set by the frontend.
2306 */
2309 HAMMER_INODE_DELETING);
2312 /*
2313 * Take care of the case where a deleted inode was never
2314 * flushed to the disk in the first place.
2315 *
2316 * Clear flags which may have been set by the frontend.
2317 */
2320 HAMMER_INODE_DELETING);
2328 }
2331 /*
2332 * If already on-disk, do not set any additional flags.
2333 */
2336 /*
2337 * If not on-disk and not deleted, set DDIRTY to force
2338 * an initial record to be written.
2339 *
2340 * Also set the create_tid in both the frontend and backend
2341 * copy of the inode record.
2342 */
2349 }
2351 /*
2352 * If RDIRTY or DDIRTY is set, write out a new record. If the inode
2353 * is already on-disk the old record is marked as deleted.
2354 *
2355 * If DELETED is set hammer_update_inode() will delete the existing
2356 * record without writing out a new one.
2357 *
2358 * If *ONLY* the ITIMES flag is set we can update the record in-place.
2359 */
2369 }
2372 done:
2373 /*
2374 * Save the TID we used to sync the inode with to make sure we
2375 * do not improperly reuse it.
2376 */
2380 }
2382 /*
2383 * This routine is called when the OS is no longer actively referencing
2384 * the inode (but might still be keeping it cached), or when releasing
2385 * the last reference to an inode.
2386 *
2387 * At this point if the inode's nlinks count is zero we want to destroy
2388 * it, which may mean destroying it on-media too.
2389 */
2390 void
2392 {
2395 /*
2396 * Set the DELETING flag when the link count drops to 0 and the
2397 * OS no longer has any opens on the inode.
2398 *
2399 * The backend will clear DELETING (a mod flag) and set DELETED
2400 * (a state flag) when it is actually able to perform the
2401 * operation.
2402 */
2412 }
2414 /*
2415 * Final cleanup
2416 */
2420 }
2423 }
2424 }
2425 }
2427 /*
2428 * Re-test an inode when a dependancy had gone away to see if we
2429 * can chain flush it.
2430 */
2431 void
2433 {
2442 }
2444 }
2445 }
2447 /*
2448 * Clear the RECLAIM flag on an inode. This occurs when the inode is
2449 * reassociated with a vp or just before it gets freed.
2450 *
2451 * Wakeup one thread blocked waiting on reclaims to complete. Note that
2452 * the inode the thread is waiting on behalf of is a different inode then
2453 * the inode we are called with. This is to create a pipeline.
2454 */
2455 static void
2457 {
2472 }
2473 }
2475 /*
2476 * Setup our reclaim pipeline. We only let so many detached (and dirty)
2477 * inodes build up before we start blocking.
2478 *
2479 * When we block we don't care *which* inode has finished reclaiming,
2480 * as lone as one does. This is somewhat heuristical... we also put a
2481 * cap on how long we are willing to wait.
2482 */
2483 void
2485 {
2495 }
2499 HAMMER_RECLAIM_WAIT;
2504 }
2505 }