| blob | e0657d217d0f51739ab58a696d33c9a755361c33 |
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 1.103.2.6 2008/09/25 01:42:52 dillon Exp $
35 */
38 #include <vm/vm_extern.h>
39 #include <sys/buf.h>
40 #include <sys/buf2.h>
47 #if 0
49 #endif
51 hammer_flush_group_t flg);
53 hammer_flush_group_t flg);
56 #ifdef DEBUG_TRUNCATE
58 #endif
60 /*
61 * RB-Tree support for inode structures
62 */
63 int
65 {
79 }
81 /*
82 * RB-Tree support for inode structures / special LOOKUP_INFO
83 */
84 static int
86 {
100 }
102 /*
103 * Used by hammer_scan_inode_snapshots() to locate all of an object's
104 * snapshots. Note that the asof field is not tested, which we can get
105 * away with because it is the lowest-priority field.
106 */
107 static int
109 {
121 }
123 /*
124 * Used by hammer_unload_pseudofs() to locate all inodes associated with
125 * a particular PFS.
126 */
127 static int
129 {
136 }
138 /*
139 * RB-Tree support for pseudofs structures
140 */
141 static int
143 {
149 }
158 /*
159 * The kernel is not actively referencing this vnode but is still holding
160 * it cached.
161 *
162 * This is called from the frontend.
163 */
164 int
166 {
169 /*
170 * Degenerate case
171 */
175 }
177 /*
178 * If the inode no longer has visibility in the filesystem try to
179 * recycle it immediately, even if the inode is dirty. Recycling
180 * it quickly allows the system to reclaim buffer cache and VM
181 * resources which can matter a lot in a heavily loaded system.
182 *
183 * This can deadlock in vfsync() if we aren't careful.
184 *
185 * Do not queue the inode to the flusher if we still have visibility,
186 * otherwise namespace calls such as chmod will unnecessarily generate
187 * multiple inode updates.
188 */
194 }
196 }
198 /*
199 * Release the vnode association. This is typically (but not always)
200 * the last reference on the inode.
201 *
202 * Once the association is lost we are on our own with regards to
203 * flushing the inode.
204 */
205 int
207 {
209 hammer_mount_t hmp;
223 }
225 }
227 }
229 /*
230 * Return a locked vnode for the specified inode. The inode must be
231 * referenced but NOT LOCKED on entry and will remain referenced on
232 * return.
233 *
234 * Called from the frontend.
235 */
236 int
238 {
239 hammer_mount_t hmp;
242 u_int8_t obj_type;
257 }
279 }
281 /*
282 * Only mark as the root vnode if the ip is not
283 * historical, otherwise the VFS cache will get
284 * confused. The other half of the special handling
285 * is in hammer_vop_nlookupdotdot().
286 *
287 * Pseudo-filesystem roots can be accessed via
288 * non-root filesystem paths and setting VROOT may
289 * confuse the namecache. Set VPFSROOT instead.
290 */
295 else
297 }
300 /* vnode locked by getnewvnode() */
301 /* make related vnode dirty if inode dirty? */
306 }
308 /*
309 * loop if the vget fails (aka races), or if the vp
310 * no longer matches ip->vp.
311 */
316 }
317 }
320 }
322 /*
323 * Locate all copies of the inode for obj_id compatible with the specified
324 * asof, reference, and issue the related call-back. This routine is used
325 * for direct-io invalidation and does not create any new inodes.
326 */
327 void
331 {
333 hammer_inode_info_cmp_all_history,
335 }
337 /*
338 * Acquire a HAMMER inode. The returned inode is not locked. These functions
339 * do not attach or detach the related vnode (use hammer_get_vnode() for
340 * that).
341 *
342 * The flags argument is only applied for newly created inodes, and only
343 * certain flags are inherited.
344 *
345 * Called from the frontend.
346 */
351 {
358 /*
359 * Determine if we already have an inode cached. If we do then
360 * we are golden.
361 */
365 loop:
371 }
373 /*
374 * Allocate a new inode structure and deal with races later.
375 */
394 /*
395 * Locate the on-disk inode. If this is a PFS root we always
396 * access the current version of the root inode and (if it is not
397 * a master) always access information under it with a snapshot
398 * TID.
399 */
400 retry:
412 HAMMER_CURSOR_ASOF;
418 }
420 /*
421 * On success the B-Tree lookup will hold the appropriate
422 * buffer cache buffers and provide a pointer to the requested
423 * information. Copy the information to the in-memory inode
424 * and cache the B-Tree node to improve future operations.
425 */
430 /*
431 * cache[0] tries to cache the location of the object inode.
432 * The assumption is that it is near the directory inode.
433 *
434 * cache[1] tries to cache the location of the object data.
435 * The assumption is that it is near the directory data.
436 */
441 /*
442 * The file should not contain any data past the file size
443 * stored in the inode. Setting save_trunc_off to the
444 * file size instead of max reduces B-Tree lookup overheads
445 * on append by allowing the flusher to avoid checking for
446 * record overwrites.
447 */
450 /*
451 * Locate and assign the pseudofs management structure to
452 * the inode.
453 */
460 errorp);
462 }
463 }
465 /*
466 * The inode is placed on the red-black tree and will be synced to
467 * the media when flushed or by the filesystem sync. If this races
468 * another instantiation/lookup the insertion will fail.
469 */
475 }
481 }
485 }
489 }
491 /*
492 * Create a new filesystem object, returning the inode in *ipp. The
493 * returned inode will be referenced. The inode is created in-memory.
494 *
495 * If pfsm is non-NULL the caller wishes to create the root inode for
496 * a master PFS.
497 */
498 int
502 {
503 hammer_mount_t hmp;
504 hammer_inode_t ip;
505 uid_t xuid;
522 }
534 /* ip->save_trunc_off = 0; (already zero) */
543 /*
544 * A nohistory designator on the parent directory is inherited by
545 * the child. We will do this even for pseudo-fs creation... the
546 * sysad can turn it off.
547 */
551 }
555 HAMMER_LOCALIZE_INODE;
568 /*
569 * Setup the ".." pointer. This only needs to be done for directories
570 * but we do it for all objects as a recovery aid.
571 */
574 #if 0
575 /*
576 * The parent_obj_localization field only applies to pseudo-fs roots.
577 * XXX this is no longer applicable, PFSs are no longer directly
578 * tied into the parent's directory structure.
579 */
584 }
585 #endif
595 }
597 /*
598 * Calculate default uid/gid and overwrite with information from
599 * the vap.
600 */
607 }
614 else
639 }
646 /* not reached */
648 }
651 }
653 /*
654 * Final cleanup / freeing of an inode structure
655 */
656 static void
658 {
670 }
673 }
675 /*
676 * Retrieve pseudo-fs data. NULL will never be returned.
677 *
678 * If an error occurs *errorp will be set and a default template is returned,
679 * otherwise *errorp is set to 0. Typically when an error occurs it will
680 * be ENOENT.
681 */
682 hammer_pseudofs_inmem_t
685 {
687 hammer_inode_t ip;
688 hammer_pseudofs_inmem_t pfsm;
692 retry:
698 }
700 /*
701 * PFS records are stored in the root inode (not the PFS root inode,
702 * but the real root). Avoid an infinite recursion if loading
703 * the PFS for the real root.
704 */
707 HAMMER_MAX_TID,
711 }
720 HAMMER_LOCALIZE_MISC;
732 else
738 HAMMER_PFSD_DELETED) {
745 }
746 }
747 }
757 }
759 }
761 /*
762 * Store pseudo-fs data. The backend will automatically delete any prior
763 * on-disk pseudo-fs data but we have to delete in-memory versions.
764 */
765 int
767 {
769 hammer_record_t record;
770 hammer_inode_t ip;
775 retry:
779 HAMMER_LOCALIZE_MISC;
800 }
801 }
807 HAMMER_LOCALIZE_MISC;
813 }
819 }
821 /*
822 * Create a root directory for a PFS if one does not alredy exist.
823 *
824 * The PFS root stands alone so we must also bump the nlinks count
825 * to prevent it from being destroyed on release.
826 */
827 int
829 hammer_pseudofs_inmem_t pfsm)
830 {
831 hammer_inode_t ip;
845 }
846 }
850 }
852 /*
853 * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
854 * if we are unable to disassociate all the inodes.
855 */
856 static
857 int
859 {
867 else
871 }
873 int
875 {
881 hammer_inode_pfs_cmp,
882 hammer_unload_pseudofs_callback,
887 }
891 }
894 /*
895 * Release a reference on a PFS
896 */
897 void
899 {
904 }
905 }
907 /*
908 * Called by hammer_sync_inode().
909 */
910 static int
912 {
914 hammer_record_t record;
918 retry:
921 /*
922 * If the inode has a presence on-disk then locate it and mark
923 * it deleted, setting DELONDISK.
924 *
925 * The record may or may not be physically deleted, depending on
926 * the retention policy.
927 */
929 HAMMER_INODE_ONDISK) {
932 HAMMER_LOCALIZE_INODE;
954 }
957 }
966 }
967 }
969 /*
970 * Ok, write out the initial record or a new record (after deleting
971 * the old one), unless the DELETED flag is set. This routine will
972 * clear DELONDISK if it writes out a record.
973 *
974 * Update our inode statistics if this is the first application of
975 * the inode on-disk.
976 */
978 /*
979 * Generate a record and write it to the media. We clean-up
980 * the state before releasing so we do not have to set-up
981 * a flush_group.
982 */
993 /*
994 * If this flag is set we cannot sync the new file size
995 * because we haven't finished related truncations. The
996 * inode will be flushed in another flush group to finish
997 * the job.
998 */
1005 }
1021 }
1023 /*
1024 * The record isn't managed by the inode's record tree,
1025 * destroy it whether we succeed or fail.
1026 */
1032 /*
1033 * Finish up.
1034 */
1039 HAMMER_INODE_ATIME |
1040 HAMMER_INODE_MTIME);
1045 /*
1046 * Root volume count of inodes
1047 */
1052 vol0_stat_inodes);
1058 }
1060 }
1061 }
1063 /*
1064 * If the inode has been destroyed, clean out any left-over flags
1065 * that may have been set by the frontend.
1066 */
1069 HAMMER_INODE_ATIME |
1070 HAMMER_INODE_MTIME);
1071 }
1073 }
1075 /*
1076 * Update only the itimes fields.
1077 *
1078 * ATIME can be updated without generating any UNDO. MTIME is updated
1079 * with UNDO so it is guaranteed to be synchronized properly in case of
1080 * a crash.
1081 *
1082 * Neither field is included in the B-Tree leaf element's CRC, which is how
1083 * we can get away with updating ATIME the way we do.
1084 */
1085 static int
1087 {
1091 retry:
1093 HAMMER_INODE_ONDISK) {
1095 }
1099 HAMMER_LOCALIZE_INODE;
1117 /*
1118 * Updating MTIME requires an UNDO. Just cover
1119 * both atime and mtime.
1120 */
1124 HAMMER_ITIMES_BYTES);
1130 /*
1131 * Updating atime only can be done in-place with
1132 * no UNDO.
1133 */
1140 }
1142 }
1149 }
1151 }
1153 /*
1154 * Release a reference on an inode, flush as requested.
1155 *
1156 * On the last reference we queue the inode to the flusher for its final
1157 * disposition.
1158 */
1159 void
1161 {
1162 /*hammer_mount_t hmp = ip->hmp;*/
1164 /*
1165 * Handle disposition when dropping the last ref.
1166 */
1169 /*
1170 * Determine whether on-disk action is needed for
1171 * the inode's final disposition.
1172 */
1180 }
1185 /*
1186 * The inode still has multiple refs, try to drop
1187 * one ref.
1188 */
1193 }
1194 }
1195 }
1196 }
1198 /*
1199 * Unload and destroy the specified inode. Must be called with one remaining
1200 * reference. The reference is disposed of.
1201 *
1202 * The inode must be completely clean.
1203 */
1204 static int
1206 {
1224 }
1226 /*
1227 * Called during unmounting if a critical error occured. The in-memory
1228 * inode and all related structures are destroyed.
1229 *
1230 * If a critical error did not occur the unmount code calls the standard
1231 * release and asserts that the inode is gone.
1232 */
1233 int
1235 {
1236 hammer_record_t rec;
1238 /*
1239 * Get rid of the inodes in-memory records, regardless of their
1240 * state, and clear the mod-mask.
1241 */
1247 }
1251 else
1259 }
1264 /*
1265 * Remove the inode from any flush group, force it idle. FLUSH
1266 * and SETUP states have an inode ref.
1267 */
1273 /* fall through */
1277 /* fall through */
1280 }
1282 /*
1283 * There shouldn't be any associated vnode. The unload needs at
1284 * least one ref, if we do have a vp steal its ip ref.
1285 */
1293 }
1296 }
1298 /*
1299 * Called on mount -u when switching from RW to RO or vise-versa. Adjust
1300 * the read-only flag for cached inodes.
1301 *
1302 * This routine is called from a RB_SCAN().
1303 */
1304 int
1306 {
1311 else
1314 }
1316 /*
1317 * A transaction has modified an inode, requiring updates as specified by
1318 * the passed flags.
1319 *
1320 * HAMMER_INODE_DDIRTY: Inode data has been updated
1321 * HAMMER_INODE_XDIRTY: Dirty in-memory records
1322 * HAMMER_INODE_BUFS: Dirty buffer cache buffers
1323 * HAMMER_INODE_DELETED: Inode record/data must be deleted
1324 * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1325 */
1326 void
1328 {
1329 /*
1330 * ronly of 0 or 2 does not trigger assertion.
1331 * 2 is a special error state
1332 */
1340 }
1343 }
1345 /*
1346 * Request that an inode be flushed. This whole mess cannot block and may
1347 * recurse (if not synchronous). Once requested HAMMER will attempt to
1348 * actively flush the inode until the flush can be done.
1349 *
1350 * The inode may already be flushing, or may be in a setup state. We can
1351 * place the inode in a flushing state if it is currently idle and flag it
1352 * to reflush if it is currently flushing.
1353 *
1354 * Upon return if the inode could not be flushed due to a setup
1355 * dependancy, then it will be automatically flushed when the dependancy
1356 * is satisfied.
1357 */
1358 void
1360 {
1361 hammer_mount_t hmp;
1362 hammer_flush_group_t flg;
1365 /*
1366 * next_flush_group is the first flush group we can place the inode
1367 * in. It may be NULL. If it becomes full we append a new flush
1368 * group and make that the next_flush_group.
1369 */
1377 }
1383 }
1385 /*
1386 * Trivial 'nothing to flush' case. If the inode is in a SETUP
1387 * state we have to put it back into an IDLE state so we can
1388 * drop the extra ref.
1389 *
1390 * If we have a parent dependancy we must still fall through
1391 * so we can run it.
1392 */
1398 }
1401 }
1403 /*
1404 * Our flush action will depend on the current state.
1405 */
1408 /*
1409 * We have no dependancies and can flush immediately. Some
1410 * our children may not be flushable so we have to re-test
1411 * with that additional knowledge.
1412 */
1416 /*
1417 * Recurse upwards through dependancies via target_list
1418 * and start their flusher actions going if possible.
1419 *
1420 * 'good' is our connectivity. -1 means we have none and
1421 * can't flush, 0 means there weren't any dependancies, and
1422 * 1 means we have good connectivity.
1423 */
1427 /*
1428 * We can continue if good >= 0. Determine how
1429 * many records under our inode can be flushed (and
1430 * mark them).
1431 */
1434 /*
1435 * Parent has no connectivity, tell it to flush
1436 * us as soon as it does.
1437 *
1438 * The REFLUSH flag is also needed to trigger
1439 * dependancy wakeups.
1440 */
1442 HAMMER_INODE_REFLUSH;
1446 }
1447 }
1450 /*
1451 * We are already flushing, flag the inode to reflush
1452 * if needed after it completes its current flush.
1453 *
1454 * The REFLUSH flag is also needed to trigger
1455 * dependancy wakeups.
1456 */
1462 }
1464 }
1465 }
1467 /*
1468 * Scan ip->target_list, which is a list of records owned by PARENTS to our
1469 * ip which reference our ip.
1470 *
1471 * XXX This is a huge mess of recursive code, but not one bit of it blocks
1472 * so for now do not ref/deref the structures. Note that if we use the
1473 * ref/rel code later, the rel CAN block.
1474 */
1475 static int
1477 {
1478 hammer_record_t depend;
1490 }
1492 }
1494 /*
1495 * This helper function takes a record representing the dependancy between
1496 * the parent inode and child inode.
1497 *
1498 * record->ip = parent inode
1499 * record->target_ip = child inode
1500 *
1501 * We are asked to recurse upwards and convert the record from SETUP
1502 * to FLUSH if possible.
1503 *
1504 * Return 1 if the record gives us connectivity
1505 *
1506 * Return 0 if the record is not relevant
1507 *
1508 * Return -1 if we can't resolve the dependancy and there is no connectivity.
1509 */
1510 static int
1512 hammer_flush_group_t flg)
1513 {
1514 hammer_mount_t hmp;
1515 hammer_inode_t pip;
1522 /*
1523 * If the record is already flushing, is it in our flush group?
1524 *
1525 * If it is in our flush group but it is a general record or a
1526 * delete-on-disk, it does not improve our connectivity (return 0),
1527 * and if the target inode is not trying to destroy itself we can't
1528 * allow the operation yet anyway (the second return -1).
1529 */
1531 /*
1532 * If not in our flush group ask the parent to reflush
1533 * us as soon as possible.
1534 */
1539 }
1541 /*
1542 * If in our flush group everything is already set up,
1543 * just return whether the record will improve our
1544 * visibility or not.
1545 */
1549 }
1551 /*
1552 * It must be a setup record. Try to resolve the setup dependancies
1553 * by recursing upwards so we can place ip on the flush list.
1554 */
1559 /*
1560 * If good < 0 the parent has no connectivity and we cannot safely
1561 * flush the directory entry, which also means we can't flush our
1562 * ip. Flag the parent and us for downward recursion once the
1563 * parent's connectivity is resolved.
1564 */
1566 /* pip->flags |= HAMMER_INODE_CONN_DOWN; set by recursion */
1569 }
1571 /*
1572 * We are go, place the parent inode in a flushing state so we can
1573 * place its record in a flushing state. Note that the parent
1574 * may already be flushing. The record must be in the same flush
1575 * group as the parent.
1576 */
1582 #if 0
1585 /*
1586 * Regardless of flushing state we cannot sync this path if the
1587 * record represents a delete-on-disk but the target inode
1588 * is not ready to sync its own deletion.
1589 *
1590 * XXX need to count effective nlinks to determine whether
1591 * the flush is ok, otherwise removing a hardlink will
1592 * just leave the DEL record to rot.
1593 */
1597 #endif
1599 /*
1600 * Because we have not calculated nlinks yet we can just
1601 * set records to the flush state if the parent is in
1602 * the same flush group as we are.
1603 */
1609 /*
1610 * A general directory-add contributes to our visibility.
1611 *
1612 * Otherwise it is probably a directory-delete or
1613 * delete-on-disk record and does not contribute to our
1614 * visbility (but we can still flush it).
1615 */
1620 /*
1621 * If the parent is not in our flush group we cannot
1622 * flush this record yet, there is no visibility.
1623 * We tell the parent to reflush and mark ourselves
1624 * so the parent knows it should flush us too.
1625 */
1629 }
1630 }
1632 /*
1633 * This is the core routine placing an inode into the FST_FLUSH state.
1634 */
1635 static void
1637 {
1640 /*
1641 * Set flush state and prevent the flusher from cycling into
1642 * the next flush group. Do not place the ip on the list yet.
1643 * Inodes not in the idle state get an extra reference.
1644 */
1655 /*
1656 * If the flush group reaches the autoflush limit we want to signal
1657 * the flusher. This is particularly important for remove()s.
1658 */
1662 /*
1663 * We need to be able to vfsync/truncate from the backend.
1664 */
1669 }
1671 /*
1672 * Figure out how many in-memory records we can actually flush
1673 * (not including inode meta-data, buffers, etc).
1674 */
1677 /*
1678 * If this is a upwards recursion we do not want to
1679 * recurse down again!
1680 */
1682 #if 0
1684 /*
1685 * No new records are added if we must complete a flush
1686 * from a previous cycle, but we do have to move the records
1687 * from the previous cycle to the current one.
1688 */
1689 #if 0
1692 #endif
1694 #endif
1696 /*
1697 * Normal flush, scan records and bring them into the flush.
1698 * Directory adds and deletes are usually skipped (they are
1699 * grouped with the related inode rather then with the
1700 * directory).
1701 *
1702 * go_count can be negative, which means the scan aborted
1703 * due to the flush group being over-full and we should
1704 * flush what we have.
1705 */
1708 }
1710 /*
1711 * This is a more involved test that includes go_count. If we
1712 * can't flush, flag the inode and return. If go_count is 0 we
1713 * were are unable to flush any records in our rec_tree and
1714 * must ignore the XDIRTY flag.
1715 */
1727 }
1729 /*
1730 * REFLUSH is needed to trigger dependancy wakeups
1731 * when an inode is in SETUP.
1732 */
1737 }
1741 }
1742 }
1744 /*
1745 * Snapshot the state of the inode for the backend flusher.
1746 *
1747 * We continue to retain save_trunc_off even when all truncations
1748 * have been resolved as an optimization to determine if we can
1749 * skip the B-Tree lookup for overwrite deletions.
1750 *
1751 * NOTE: The DELETING flag is a mod flag, but it is also sticky,
1752 * and stays in ip->flags. Once set, it stays set until the
1753 * inode is destroyed.
1754 */
1762 /*
1763 * The save_trunc_off used to cache whether the B-Tree
1764 * holds any records past that point is not used until
1765 * after the truncation has succeeded, so we can safely
1766 * set it now.
1767 */
1770 }
1776 #ifdef DEBUG_TRUNCATE
1779 #endif
1781 /*
1782 * The flusher list inherits our inode and reference.
1783 */
1791 }
1792 }
1794 /*
1795 * Callback for scan of ip->rec_tree. Try to include each record in our
1796 * flush. ip->flush_group has been set but the inode has not yet been
1797 * moved into a flushing state.
1798 *
1799 * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1800 * both inodes.
1801 *
1802 * We return 1 for any record placed or found in FST_FLUSH, which prevents
1803 * the caller from shortcutting the flush.
1804 */
1805 static int
1807 {
1808 hammer_flush_group_t flg;
1809 hammer_inode_t target_ip;
1810 hammer_inode_t ip;
1813 /*
1814 * Deleted records are ignored. Note that the flush detects deleted
1815 * front-end records at multiple points to deal with races. This is
1816 * just the first line of defense. The only time DELETED_FE cannot
1817 * be set is when HAMMER_RECF_INTERLOCK_BE is set.
1818 *
1819 * Don't get confused between record deletion and, say, directory
1820 * entry deletion. The deletion of a directory entry that is on
1821 * the media has nothing to do with the record deletion flags.
1822 */
1829 }
1831 }
1833 /*
1834 * If the record is in an idle state it has no dependancies and
1835 * can be flushed.
1836 */
1843 /*
1844 * The record has no setup dependancy, we can flush it.
1845 */
1854 /*
1855 * The record has a setup dependancy. These are typically
1856 * directory entry adds and deletes. Such entries will be
1857 * flushed when their inodes are flushed so we do not
1858 * usually have to add them to the flush here. However,
1859 * if the target_ip has set HAMMER_INODE_CONN_DOWN then
1860 * it is asking us to flush this record (and it).
1861 */
1866 /*
1867 * If the target IP is already flushing in our group
1868 * we could associate the record, but target_ip has
1869 * already synced ino_data to sync_ino_data and we
1870 * would also have to adjust nlinks. Plus there are
1871 * ordering issues for adds and deletes.
1872 *
1873 * Reflush downward if this is an ADD, and upward if
1874 * this is a DEL.
1875 */
1879 else
1882 }
1884 /*
1885 * Target IP is not yet flushing. This can get complex
1886 * because we have to be careful about the recursion.
1887 *
1888 * Directories create an issue for us in that if a flush
1889 * of a directory is requested the expectation is to flush
1890 * any pending directory entries, but this will cause the
1891 * related inodes to recursively flush as well. We can't
1892 * really defer the operation so just get as many as we
1893 * can and
1894 */
1895 #if 0
1898 /*
1899 * We aren't reclaiming and the target ip was not
1900 * previously prevented from flushing due to this
1901 * record dependancy. Do not flush this record.
1902 */
1903 /*r = 0;*/
1905 #endif
1908 /*
1909 * Our flush group is over-full and we risk blowing
1910 * out the UNDO FIFO. Stop the scan, flush what we
1911 * have, then reflush the directory.
1912 *
1913 * The directory may be forced through multiple
1914 * flush groups before it can be completely
1915 * flushed.
1916 */
1918 HAMMER_INODE_REFLUSH;
1921 /*
1922 * If the target IP is not flushing we can force
1923 * it to flush, even if it is unable to write out
1924 * any of its own records we have at least one in
1925 * hand that we CAN deal with.
1926 */
1932 HAMMER_FLUSH_RECURSION);
1935 /*
1936 * General or delete-on-disk record.
1937 *
1938 * XXX this needs help. If a delete-on-disk we could
1939 * disconnect the target. If the target has its own
1940 * dependancies they really need to be flushed.
1941 *
1942 * XXX
1943 */
1949 HAMMER_FLUSH_RECURSION);
1951 }
1954 /*
1955 * The flush_group should already match.
1956 */
1960 }
1962 }
1964 #if 0
1965 /*
1966 * This version just moves records already in a flush state to the new
1967 * flush group and that is it.
1968 */
1969 static int
1971 {
1980 }
1982 }
1983 #endif
1985 /*
1986 * Wait for a previously queued flush to complete.
1987 *
1988 * If a critical error occured we don't try to wait.
1989 */
1990 void
1992 {
1993 hammer_flush_group_t flg;
2004 }
2005 }
2006 }
2007 }
2009 /*
2010 * Called by the backend code when a flush has been completed.
2011 * The inode has already been removed from the flush list.
2012 *
2013 * A pipelined flush can occur, in which case we must re-enter the
2014 * inode on the list and re-copy its fields.
2015 */
2016 void
2018 {
2019 hammer_mount_t hmp;
2026 /*
2027 * Merge left-over flags back into the frontend and fix the state.
2028 * Incomplete truncations are retained by the backend.
2029 */
2034 /*
2035 * The backend may have adjusted nlinks, so if the adjusted nlinks
2036 * does not match the fronttend set the frontend's RDIRTY flag again.
2037 */
2041 /*
2042 * Fix up the dirty buffer status.
2043 */
2046 }
2048 /*
2049 * Re-set the XDIRTY flag if some of the inode's in-memory records
2050 * could not be flushed.
2051 */
2057 /*
2058 * Do not lose track of inodes which no longer have vnode
2059 * assocations, otherwise they may never get flushed again.
2060 *
2061 * The reflush flag can be set superfluously, causing extra pain
2062 * for no reason. If the inode is no longer modified it no longer
2063 * needs to be flushed.
2064 */
2070 }
2072 /*
2073 * Adjust the flush state.
2074 */
2076 /*
2077 * We were unable to flush out all our records, leave the
2078 * inode in a flush state and in the current flush group.
2079 * The flush group will be re-run.
2080 *
2081 * This occurs if the UNDO block gets too full or there is
2082 * too much dirty meta-data and allows the flusher to
2083 * finalize the UNDO block and then re-flush.
2084 */
2088 /*
2089 * Remove from the flush_group
2090 */
2094 /*
2095 * Clean up the vnode ref and tracking counts.
2096 */
2100 }
2104 /*
2105 * And adjust the state.
2106 */
2113 }
2115 /*
2116 * If the frontend is waiting for a flush to complete,
2117 * wake it up.
2118 */
2122 }
2124 /*
2125 * If the frontend made more changes and requested another
2126 * flush, then try to get it running.
2127 *
2128 * Reflushes are aborted when the inode is errored out.
2129 */
2137 }
2138 }
2139 }
2141 /*
2142 * If we have no parent dependancies we can clear CONN_DOWN
2143 */
2147 /*
2148 * If the inode is now clean drop the space reservation.
2149 */
2154 }
2158 }
2160 /*
2161 * Called from hammer_sync_inode() to synchronize in-memory records
2162 * to the media.
2163 */
2164 static int
2166 {
2172 /*
2173 * Skip records that do not belong to the current flush.
2174 */
2179 #if 1
2181 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2184 }
2185 #endif
2188 /*
2189 * Interlock the record using the BE flag. Once BE is set the
2190 * frontend cannot change the state of FE.
2191 *
2192 * NOTE: If FE is set prior to us setting BE we still sync the
2193 * record out, but the flush completion code converts it to
2194 * a delete-on-disk record instead of destroying it.
2195 */
2199 /*
2200 * The backend may have already disposed of the record.
2201 */
2205 }
2207 /*
2208 * If the whole inode is being deleting all on-disk records will
2209 * be deleted very soon, we can't sync any new records to disk
2210 * because they will be deleted in the same transaction they were
2211 * created in (delete_tid == create_tid), which will assert.
2212 *
2213 * XXX There may be a case with RECORD_ADD with DELETED_FE set
2214 * that we currently panic on.
2215 */
2219 /*
2220 * We don't have to do anything, if the record was
2221 * committed the space will have been accounted for
2222 * in the blockmap.
2223 */
2224 /* fall through */
2239 /*
2240 * Follow through and issue the on-disk deletion
2241 */
2243 }
2244 }
2246 /*
2247 * If DELETED_FE is set special handling is needed for directory
2248 * entries. Dependant pieces related to the directory entry may
2249 * have already been synced to disk. If this occurs we have to
2250 * sync the directory entry and then change the in-memory record
2251 * from an ADD to a DELETE to cover the fact that it's been
2252 * deleted by the frontend.
2253 *
2254 * A directory delete covering record (MEM_RECORD_DEL) can never
2255 * be deleted by the frontend.
2256 *
2257 * Any other record type (aka DATA) can be deleted by the frontend.
2258 * XXX At the moment the flusher must skip it because there may
2259 * be another data record in the flush group for the same block,
2260 * meaning that some frontend data changes can leak into the backend's
2261 * synchronization point.
2262 */
2271 }
2272 }
2274 /*
2275 * Assign the create_tid for new records. Deletions already
2276 * have the record's entire key properly set up.
2277 */
2290 }
2295 done:
2298 /*
2299 * Do partial finalization if we have built up too many dirty
2300 * buffers. Otherwise a buffer cache deadlock can occur when
2301 * doing things like creating tens of thousands of tiny files.
2302 *
2303 * We must release our cursor lock to avoid a 3-way deadlock
2304 * due to the exclusive sync lock the finalizer must get.
2305 */
2310 }
2313 }
2315 /*
2316 * Backend function called by the flusher to sync an inode to media.
2317 */
2318 int
2320 {
2322 hammer_node_t tmp_node;
2323 hammer_record_t depend;
2324 hammer_record_t next;
2326 u_int64_t nlinks;
2335 /*
2336 * Any directory records referencing this inode which are not in
2337 * our current flush group must adjust our nlink count for the
2338 * purposes of synchronization to disk.
2339 *
2340 * Records which are in our flush group can be unlinked from our
2341 * inode now, potentially allowing the inode to be physically
2342 * deleted.
2343 *
2344 * This cannot block.
2345 */
2352 /*
2353 * If this is an ADD that was deleted by the frontend
2354 * the frontend nlinks count will have already been
2355 * decremented, but the backend is going to sync its
2356 * directory entry and must account for it. The
2357 * record will be converted to a delete-on-disk when
2358 * it gets synced.
2359 *
2360 * If the ADD was not deleted by the frontend we
2361 * can remove the dependancy from our target_list.
2362 */
2367 target_entry);
2369 }
2371 /*
2372 * Not part of our flush group
2373 */
2384 }
2385 }
2386 }
2388 /*
2389 * Set dirty if we had to modify the link count.
2390 */
2395 }
2397 /*
2398 * If there is a trunction queued destroy any data past the (aligned)
2399 * truncation point. Userland will have dealt with the buffer
2400 * containing the truncation point for us.
2401 *
2402 * We don't flush pending frontend data buffers until after we've
2403 * dealt with the truncation.
2404 */
2406 /*
2407 * Interlock trunc_off. The VOP front-end may continue to
2408 * make adjustments to it while we are blocked.
2409 */
2410 off_t trunc_off;
2411 off_t aligned_trunc_off;
2418 /*
2419 * Delete any whole blocks on-media. The front-end has
2420 * already cleaned out any partial block and made it
2421 * pending. The front-end may have updated trunc_off
2422 * while we were blocked so we only use sync_trunc_off.
2423 *
2424 * This operation can blow out the buffer cache, EWOULDBLOCK
2425 * means we were unable to complete the deletion. The
2426 * deletion will update sync_trunc_off in that case.
2427 */
2429 aligned_trunc_off,
2435 }
2440 /*
2441 * Clear the truncation flag on the backend after we have
2442 * complete the deletions. Backend data is now good again
2443 * (including new records we are about to sync, below).
2444 *
2445 * Leave sync_trunc_off intact. As we write additional
2446 * records the backend will update sync_trunc_off. This
2447 * tells the backend whether it can skip the overwrite
2448 * test. This should work properly even when the backend
2449 * writes full blocks where the truncation point straddles
2450 * the block because the comparison is against the base
2451 * offset of the record.
2452 */
2454 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2457 }
2459 /*
2460 * Now sync related records. These will typically be directory
2461 * entries, records tracking direct-writes, or delete-on-disk records.
2462 */
2470 }
2473 /*
2474 * Re-seek for inode update, assuming our cache hasn't been ripped
2475 * out from under us.
2476 */
2486 }
2488 }
2490 /*
2491 * If we are deleting the inode the frontend had better not have
2492 * any active references on elements making up the inode.
2493 *
2494 * The call to hammer_ip_delete_clean() cleans up auxillary records
2495 * but not DB or DATA records. Those must have already been deleted
2496 * by the normal truncation mechanic.
2497 */
2511 /*
2512 * Set delete_tid in both the frontend and backend
2513 * copy of the inode record. The DELETED flag handles
2514 * this, do not set RDIRTY.
2515 */
2522 /*
2523 * Adjust the inode count in the volume header
2524 */
2529 vol0_stat_inodes);
2532 }
2534 }
2535 }
2541 defer_buffer_flush:
2542 /*
2543 * Now update the inode's on-disk inode-data and/or on-disk record.
2544 * DELETED and ONDISK are managed only in ip->flags.
2545 *
2546 * In the case of a defered buffer flush we still update the on-disk
2547 * inode to satisfy visibility requirements if there happen to be
2548 * directory dependancies.
2549 */
2552 /*
2553 * If deleted and on-disk, don't set any additional flags.
2554 * the delete flag takes care of things.
2555 *
2556 * Clear flags which may have been set by the frontend.
2557 */
2560 HAMMER_INODE_DELETING);
2563 /*
2564 * Take care of the case where a deleted inode was never
2565 * flushed to the disk in the first place.
2566 *
2567 * Clear flags which may have been set by the frontend.
2568 */
2571 HAMMER_INODE_DELETING);
2579 }
2582 /*
2583 * If already on-disk, do not set any additional flags.
2584 */
2587 /*
2588 * If not on-disk and not deleted, set DDIRTY to force
2589 * an initial record to be written.
2590 *
2591 * Also set the create_tid in both the frontend and backend
2592 * copy of the inode record.
2593 */
2600 }
2602 /*
2603 * If RDIRTY or DDIRTY is set, write out a new record. If the inode
2604 * is already on-disk the old record is marked as deleted.
2605 *
2606 * If DELETED is set hammer_update_inode() will delete the existing
2607 * record without writing out a new one.
2608 *
2609 * If *ONLY* the ITIMES flag is set we can update the record in-place.
2610 */
2620 }
2621 done:
2625 }
2628 }
2630 /*
2631 * This routine is called when the OS is no longer actively referencing
2632 * the inode (but might still be keeping it cached), or when releasing
2633 * the last reference to an inode.
2634 *
2635 * At this point if the inode's nlinks count is zero we want to destroy
2636 * it, which may mean destroying it on-media too.
2637 */
2638 void
2640 {
2643 /*
2644 * Set the DELETING flag when the link count drops to 0 and the
2645 * OS no longer has any opens on the inode.
2646 *
2647 * The backend will clear DELETING (a mod flag) and set DELETED
2648 * (a state flag) when it is actually able to perform the
2649 * operation.
2650 *
2651 * Don't reflag the deletion if the flusher is currently syncing
2652 * one that was already flagged. A previously set DELETING flag
2653 * may bounce around flags and sync_flags until the operation is
2654 * completely done.
2655 */
2665 }
2667 /*
2668 * Final cleanup
2669 */
2673 }
2676 }
2677 }
2678 }
2680 /*
2681 * After potentially resolving a dependancy the inode is tested
2682 * to determine whether it needs to be reflushed.
2683 */
2684 void
2686 {
2695 }
2697 }
2698 }
2700 /*
2701 * Clear the RECLAIM flag on an inode. This occurs when the inode is
2702 * reassociated with a vp or just before it gets freed.
2703 *
2704 * Wakeup one thread blocked waiting on reclaims to complete. Note that
2705 * the inode the thread is waiting on behalf of is a different inode then
2706 * the inode we are called with. This is to create a pipeline.
2707 */
2708 static void
2710 {
2725 }
2726 }
2728 /*
2729 * Setup our reclaim pipeline. We only let so many detached (and dirty)
2730 * inodes build up before we start blocking.
2731 *
2732 * When we block we don't care *which* inode has finished reclaiming,
2733 * as lone as one does. This is somewhat heuristical... we also put a
2734 * cap on how long we are willing to wait.
2735 */
2736 void
2738 {
2748 }
2757 }
2758 }