| blob | 18abae74189b4d6f622efd817f2efd6421e14dde |
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 1.114 2008/09/24 00:53:51 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 * If we are running version 2 or greater we use dirhash algorithm #1
570 * which is semi-sorted. Algorithm #0 was just a pure crc.
571 */
575 }
576 }
578 /*
579 * Setup the ".." pointer. This only needs to be done for directories
580 * but we do it for all objects as a recovery aid.
581 */
584 #if 0
585 /*
586 * The parent_obj_localization field only applies to pseudo-fs roots.
587 * XXX this is no longer applicable, PFSs are no longer directly
588 * tied into the parent's directory structure.
589 */
594 }
595 #endif
605 }
607 /*
608 * Calculate default uid/gid and overwrite with information from
609 * the vap.
610 */
617 }
624 else
649 }
656 /* not reached */
658 }
661 }
663 /*
664 * Final cleanup / freeing of an inode structure
665 */
666 static void
668 {
683 }
686 }
688 /*
689 * Retrieve pseudo-fs data. NULL will never be returned.
690 *
691 * If an error occurs *errorp will be set and a default template is returned,
692 * otherwise *errorp is set to 0. Typically when an error occurs it will
693 * be ENOENT.
694 */
695 hammer_pseudofs_inmem_t
698 {
700 hammer_inode_t ip;
701 hammer_pseudofs_inmem_t pfsm;
705 retry:
711 }
713 /*
714 * PFS records are stored in the root inode (not the PFS root inode,
715 * but the real root). Avoid an infinite recursion if loading
716 * the PFS for the real root.
717 */
720 HAMMER_MAX_TID,
724 }
733 HAMMER_LOCALIZE_MISC;
745 else
751 HAMMER_PFSD_DELETED) {
758 }
759 }
760 }
770 }
772 }
774 /*
775 * Store pseudo-fs data. The backend will automatically delete any prior
776 * on-disk pseudo-fs data but we have to delete in-memory versions.
777 */
778 int
780 {
782 hammer_record_t record;
783 hammer_inode_t ip;
788 retry:
792 HAMMER_LOCALIZE_MISC;
813 }
814 }
820 HAMMER_LOCALIZE_MISC;
826 }
832 }
834 /*
835 * Create a root directory for a PFS if one does not alredy exist.
836 *
837 * The PFS root stands alone so we must also bump the nlinks count
838 * to prevent it from being destroyed on release.
839 */
840 int
842 hammer_pseudofs_inmem_t pfsm)
843 {
844 hammer_inode_t ip;
858 }
859 }
863 }
865 /*
866 * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
867 * if we are unable to disassociate all the inodes.
868 */
869 static
870 int
872 {
880 else
884 }
886 int
888 {
894 hammer_inode_pfs_cmp,
895 hammer_unload_pseudofs_callback,
900 }
904 }
907 /*
908 * Release a reference on a PFS
909 */
910 void
912 {
917 }
918 }
920 /*
921 * Called by hammer_sync_inode().
922 */
923 static int
925 {
927 hammer_record_t record;
931 retry:
934 /*
935 * If the inode has a presence on-disk then locate it and mark
936 * it deleted, setting DELONDISK.
937 *
938 * The record may or may not be physically deleted, depending on
939 * the retention policy.
940 */
942 HAMMER_INODE_ONDISK) {
945 HAMMER_LOCALIZE_INODE;
967 }
970 }
979 }
980 }
982 /*
983 * Ok, write out the initial record or a new record (after deleting
984 * the old one), unless the DELETED flag is set. This routine will
985 * clear DELONDISK if it writes out a record.
986 *
987 * Update our inode statistics if this is the first application of
988 * the inode on-disk.
989 */
991 /*
992 * Generate a record and write it to the media. We clean-up
993 * the state before releasing so we do not have to set-up
994 * a flush_group.
995 */
1006 /*
1007 * If this flag is set we cannot sync the new file size
1008 * because we haven't finished related truncations. The
1009 * inode will be flushed in another flush group to finish
1010 * the job.
1011 */
1018 }
1034 }
1036 /*
1037 * The record isn't managed by the inode's record tree,
1038 * destroy it whether we succeed or fail.
1039 */
1045 /*
1046 * Finish up.
1047 */
1052 HAMMER_INODE_ATIME |
1053 HAMMER_INODE_MTIME);
1058 /*
1059 * Root volume count of inodes
1060 */
1065 vol0_stat_inodes);
1071 }
1073 }
1074 }
1076 /*
1077 * If the inode has been destroyed, clean out any left-over flags
1078 * that may have been set by the frontend.
1079 */
1082 HAMMER_INODE_ATIME |
1083 HAMMER_INODE_MTIME);
1084 }
1086 }
1088 /*
1089 * Update only the itimes fields.
1090 *
1091 * ATIME can be updated without generating any UNDO. MTIME is updated
1092 * with UNDO so it is guaranteed to be synchronized properly in case of
1093 * a crash.
1094 *
1095 * Neither field is included in the B-Tree leaf element's CRC, which is how
1096 * we can get away with updating ATIME the way we do.
1097 */
1098 static int
1100 {
1104 retry:
1106 HAMMER_INODE_ONDISK) {
1108 }
1112 HAMMER_LOCALIZE_INODE;
1130 /*
1131 * Updating MTIME requires an UNDO. Just cover
1132 * both atime and mtime.
1133 */
1137 HAMMER_ITIMES_BYTES);
1143 /*
1144 * Updating atime only can be done in-place with
1145 * no UNDO.
1146 */
1153 }
1155 }
1162 }
1164 }
1166 /*
1167 * Release a reference on an inode, flush as requested.
1168 *
1169 * On the last reference we queue the inode to the flusher for its final
1170 * disposition.
1171 */
1172 void
1174 {
1175 /*hammer_mount_t hmp = ip->hmp;*/
1177 /*
1178 * Handle disposition when dropping the last ref.
1179 */
1182 /*
1183 * Determine whether on-disk action is needed for
1184 * the inode's final disposition.
1185 */
1193 }
1198 /*
1199 * The inode still has multiple refs, try to drop
1200 * one ref.
1201 */
1206 }
1207 }
1208 }
1209 }
1211 /*
1212 * Unload and destroy the specified inode. Must be called with one remaining
1213 * reference. The reference is disposed of.
1214 *
1215 * The inode must be completely clean.
1216 */
1217 static int
1219 {
1237 }
1239 /*
1240 * Called during unmounting if a critical error occured. The in-memory
1241 * inode and all related structures are destroyed.
1242 *
1243 * If a critical error did not occur the unmount code calls the standard
1244 * release and asserts that the inode is gone.
1245 */
1246 int
1248 {
1249 hammer_record_t rec;
1251 /*
1252 * Get rid of the inodes in-memory records, regardless of their
1253 * state, and clear the mod-mask.
1254 */
1260 }
1264 else
1272 }
1277 /*
1278 * Remove the inode from any flush group, force it idle. FLUSH
1279 * and SETUP states have an inode ref.
1280 */
1286 /* fall through */
1290 /* fall through */
1293 }
1295 /*
1296 * There shouldn't be any associated vnode. The unload needs at
1297 * least one ref, if we do have a vp steal its ip ref.
1298 */
1306 }
1309 }
1311 /*
1312 * Called on mount -u when switching from RW to RO or vise-versa. Adjust
1313 * the read-only flag for cached inodes.
1314 *
1315 * This routine is called from a RB_SCAN().
1316 */
1317 int
1319 {
1324 else
1327 }
1329 /*
1330 * A transaction has modified an inode, requiring updates as specified by
1331 * the passed flags.
1332 *
1333 * HAMMER_INODE_DDIRTY: Inode data has been updated
1334 * HAMMER_INODE_XDIRTY: Dirty in-memory records
1335 * HAMMER_INODE_BUFS: Dirty buffer cache buffers
1336 * HAMMER_INODE_DELETED: Inode record/data must be deleted
1337 * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1338 */
1339 void
1341 {
1342 /*
1343 * ronly of 0 or 2 does not trigger assertion.
1344 * 2 is a special error state
1345 */
1353 }
1356 }
1358 /*
1359 * Request that an inode be flushed. This whole mess cannot block and may
1360 * recurse (if not synchronous). Once requested HAMMER will attempt to
1361 * actively flush the inode until the flush can be done.
1362 *
1363 * The inode may already be flushing, or may be in a setup state. We can
1364 * place the inode in a flushing state if it is currently idle and flag it
1365 * to reflush if it is currently flushing.
1366 *
1367 * Upon return if the inode could not be flushed due to a setup
1368 * dependancy, then it will be automatically flushed when the dependancy
1369 * is satisfied.
1370 */
1371 void
1373 {
1374 hammer_mount_t hmp;
1375 hammer_flush_group_t flg;
1378 /*
1379 * next_flush_group is the first flush group we can place the inode
1380 * in. It may be NULL. If it becomes full we append a new flush
1381 * group and make that the next_flush_group.
1382 */
1390 }
1396 }
1398 /*
1399 * Trivial 'nothing to flush' case. If the inode is in a SETUP
1400 * state we have to put it back into an IDLE state so we can
1401 * drop the extra ref.
1402 *
1403 * If we have a parent dependancy we must still fall through
1404 * so we can run it.
1405 */
1411 }
1414 }
1416 /*
1417 * Our flush action will depend on the current state.
1418 */
1421 /*
1422 * We have no dependancies and can flush immediately. Some
1423 * our children may not be flushable so we have to re-test
1424 * with that additional knowledge.
1425 */
1429 /*
1430 * Recurse upwards through dependancies via target_list
1431 * and start their flusher actions going if possible.
1432 *
1433 * 'good' is our connectivity. -1 means we have none and
1434 * can't flush, 0 means there weren't any dependancies, and
1435 * 1 means we have good connectivity.
1436 */
1440 /*
1441 * We can continue if good >= 0. Determine how
1442 * many records under our inode can be flushed (and
1443 * mark them).
1444 */
1447 /*
1448 * Parent has no connectivity, tell it to flush
1449 * us as soon as it does.
1450 *
1451 * The REFLUSH flag is also needed to trigger
1452 * dependancy wakeups.
1453 */
1455 HAMMER_INODE_REFLUSH;
1459 }
1460 }
1463 /*
1464 * We are already flushing, flag the inode to reflush
1465 * if needed after it completes its current flush.
1466 *
1467 * The REFLUSH flag is also needed to trigger
1468 * dependancy wakeups.
1469 */
1475 }
1477 }
1478 }
1480 /*
1481 * Scan ip->target_list, which is a list of records owned by PARENTS to our
1482 * ip which reference our ip.
1483 *
1484 * XXX This is a huge mess of recursive code, but not one bit of it blocks
1485 * so for now do not ref/deref the structures. Note that if we use the
1486 * ref/rel code later, the rel CAN block.
1487 */
1488 static int
1490 {
1491 hammer_record_t depend;
1503 }
1505 }
1507 /*
1508 * This helper function takes a record representing the dependancy between
1509 * the parent inode and child inode.
1510 *
1511 * record->ip = parent inode
1512 * record->target_ip = child inode
1513 *
1514 * We are asked to recurse upwards and convert the record from SETUP
1515 * to FLUSH if possible.
1516 *
1517 * Return 1 if the record gives us connectivity
1518 *
1519 * Return 0 if the record is not relevant
1520 *
1521 * Return -1 if we can't resolve the dependancy and there is no connectivity.
1522 */
1523 static int
1525 hammer_flush_group_t flg)
1526 {
1527 hammer_mount_t hmp;
1528 hammer_inode_t pip;
1535 /*
1536 * If the record is already flushing, is it in our flush group?
1537 *
1538 * If it is in our flush group but it is a general record or a
1539 * delete-on-disk, it does not improve our connectivity (return 0),
1540 * and if the target inode is not trying to destroy itself we can't
1541 * allow the operation yet anyway (the second return -1).
1542 */
1544 /*
1545 * If not in our flush group ask the parent to reflush
1546 * us as soon as possible.
1547 */
1552 }
1554 /*
1555 * If in our flush group everything is already set up,
1556 * just return whether the record will improve our
1557 * visibility or not.
1558 */
1562 }
1564 /*
1565 * It must be a setup record. Try to resolve the setup dependancies
1566 * by recursing upwards so we can place ip on the flush list.
1567 */
1572 /*
1573 * If good < 0 the parent has no connectivity and we cannot safely
1574 * flush the directory entry, which also means we can't flush our
1575 * ip. Flag the parent and us for downward recursion once the
1576 * parent's connectivity is resolved.
1577 */
1579 /* pip->flags |= HAMMER_INODE_CONN_DOWN; set by recursion */
1582 }
1584 /*
1585 * We are go, place the parent inode in a flushing state so we can
1586 * place its record in a flushing state. Note that the parent
1587 * may already be flushing. The record must be in the same flush
1588 * group as the parent.
1589 */
1595 #if 0
1598 /*
1599 * Regardless of flushing state we cannot sync this path if the
1600 * record represents a delete-on-disk but the target inode
1601 * is not ready to sync its own deletion.
1602 *
1603 * XXX need to count effective nlinks to determine whether
1604 * the flush is ok, otherwise removing a hardlink will
1605 * just leave the DEL record to rot.
1606 */
1610 #endif
1612 /*
1613 * Because we have not calculated nlinks yet we can just
1614 * set records to the flush state if the parent is in
1615 * the same flush group as we are.
1616 */
1622 /*
1623 * A general directory-add contributes to our visibility.
1624 *
1625 * Otherwise it is probably a directory-delete or
1626 * delete-on-disk record and does not contribute to our
1627 * visbility (but we can still flush it).
1628 */
1633 /*
1634 * If the parent is not in our flush group we cannot
1635 * flush this record yet, there is no visibility.
1636 * We tell the parent to reflush and mark ourselves
1637 * so the parent knows it should flush us too.
1638 */
1642 }
1643 }
1645 /*
1646 * This is the core routine placing an inode into the FST_FLUSH state.
1647 */
1648 static void
1650 {
1653 /*
1654 * Set flush state and prevent the flusher from cycling into
1655 * the next flush group. Do not place the ip on the list yet.
1656 * Inodes not in the idle state get an extra reference.
1657 */
1668 /*
1669 * If the flush group reaches the autoflush limit we want to signal
1670 * the flusher. This is particularly important for remove()s.
1671 */
1675 /*
1676 * We need to be able to vfsync/truncate from the backend.
1677 */
1682 }
1684 /*
1685 * Figure out how many in-memory records we can actually flush
1686 * (not including inode meta-data, buffers, etc).
1687 */
1690 /*
1691 * If this is a upwards recursion we do not want to
1692 * recurse down again!
1693 */
1695 #if 0
1697 /*
1698 * No new records are added if we must complete a flush
1699 * from a previous cycle, but we do have to move the records
1700 * from the previous cycle to the current one.
1701 */
1702 #if 0
1705 #endif
1707 #endif
1709 /*
1710 * Normal flush, scan records and bring them into the flush.
1711 * Directory adds and deletes are usually skipped (they are
1712 * grouped with the related inode rather then with the
1713 * directory).
1714 *
1715 * go_count can be negative, which means the scan aborted
1716 * due to the flush group being over-full and we should
1717 * flush what we have.
1718 */
1721 }
1723 /*
1724 * This is a more involved test that includes go_count. If we
1725 * can't flush, flag the inode and return. If go_count is 0 we
1726 * were are unable to flush any records in our rec_tree and
1727 * must ignore the XDIRTY flag.
1728 */
1740 }
1742 /*
1743 * REFLUSH is needed to trigger dependancy wakeups
1744 * when an inode is in SETUP.
1745 */
1750 }
1754 }
1755 }
1757 /*
1758 * Snapshot the state of the inode for the backend flusher.
1759 *
1760 * We continue to retain save_trunc_off even when all truncations
1761 * have been resolved as an optimization to determine if we can
1762 * skip the B-Tree lookup for overwrite deletions.
1763 *
1764 * NOTE: The DELETING flag is a mod flag, but it is also sticky,
1765 * and stays in ip->flags. Once set, it stays set until the
1766 * inode is destroyed.
1767 */
1775 /*
1776 * The save_trunc_off used to cache whether the B-Tree
1777 * holds any records past that point is not used until
1778 * after the truncation has succeeded, so we can safely
1779 * set it now.
1780 */
1783 }
1789 #ifdef DEBUG_TRUNCATE
1792 #endif
1794 /*
1795 * The flusher list inherits our inode and reference.
1796 */
1804 }
1805 }
1807 /*
1808 * Callback for scan of ip->rec_tree. Try to include each record in our
1809 * flush. ip->flush_group has been set but the inode has not yet been
1810 * moved into a flushing state.
1811 *
1812 * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1813 * both inodes.
1814 *
1815 * We return 1 for any record placed or found in FST_FLUSH, which prevents
1816 * the caller from shortcutting the flush.
1817 */
1818 static int
1820 {
1821 hammer_flush_group_t flg;
1822 hammer_inode_t target_ip;
1823 hammer_inode_t ip;
1826 /*
1827 * Deleted records are ignored. Note that the flush detects deleted
1828 * front-end records at multiple points to deal with races. This is
1829 * just the first line of defense. The only time DELETED_FE cannot
1830 * be set is when HAMMER_RECF_INTERLOCK_BE is set.
1831 *
1832 * Don't get confused between record deletion and, say, directory
1833 * entry deletion. The deletion of a directory entry that is on
1834 * the media has nothing to do with the record deletion flags.
1835 */
1842 }
1844 }
1846 /*
1847 * If the record is in an idle state it has no dependancies and
1848 * can be flushed.
1849 */
1856 /*
1857 * The record has no setup dependancy, we can flush it.
1858 */
1867 /*
1868 * The record has a setup dependancy. These are typically
1869 * directory entry adds and deletes. Such entries will be
1870 * flushed when their inodes are flushed so we do not
1871 * usually have to add them to the flush here. However,
1872 * if the target_ip has set HAMMER_INODE_CONN_DOWN then
1873 * it is asking us to flush this record (and it).
1874 */
1879 /*
1880 * If the target IP is already flushing in our group
1881 * we could associate the record, but target_ip has
1882 * already synced ino_data to sync_ino_data and we
1883 * would also have to adjust nlinks. Plus there are
1884 * ordering issues for adds and deletes.
1885 *
1886 * Reflush downward if this is an ADD, and upward if
1887 * this is a DEL.
1888 */
1892 else
1895 }
1897 /*
1898 * Target IP is not yet flushing. This can get complex
1899 * because we have to be careful about the recursion.
1900 *
1901 * Directories create an issue for us in that if a flush
1902 * of a directory is requested the expectation is to flush
1903 * any pending directory entries, but this will cause the
1904 * related inodes to recursively flush as well. We can't
1905 * really defer the operation so just get as many as we
1906 * can and
1907 */
1908 #if 0
1911 /*
1912 * We aren't reclaiming and the target ip was not
1913 * previously prevented from flushing due to this
1914 * record dependancy. Do not flush this record.
1915 */
1916 /*r = 0;*/
1918 #endif
1921 /*
1922 * Our flush group is over-full and we risk blowing
1923 * out the UNDO FIFO. Stop the scan, flush what we
1924 * have, then reflush the directory.
1925 *
1926 * The directory may be forced through multiple
1927 * flush groups before it can be completely
1928 * flushed.
1929 */
1931 HAMMER_INODE_REFLUSH;
1934 /*
1935 * If the target IP is not flushing we can force
1936 * it to flush, even if it is unable to write out
1937 * any of its own records we have at least one in
1938 * hand that we CAN deal with.
1939 */
1945 HAMMER_FLUSH_RECURSION);
1948 /*
1949 * General or delete-on-disk record.
1950 *
1951 * XXX this needs help. If a delete-on-disk we could
1952 * disconnect the target. If the target has its own
1953 * dependancies they really need to be flushed.
1954 *
1955 * XXX
1956 */
1962 HAMMER_FLUSH_RECURSION);
1964 }
1967 /*
1968 * The flush_group should already match.
1969 */
1973 }
1975 }
1977 #if 0
1978 /*
1979 * This version just moves records already in a flush state to the new
1980 * flush group and that is it.
1981 */
1982 static int
1984 {
1993 }
1995 }
1996 #endif
1998 /*
1999 * Wait for a previously queued flush to complete.
2000 *
2001 * If a critical error occured we don't try to wait.
2002 */
2003 void
2005 {
2006 hammer_flush_group_t flg;
2017 }
2018 }
2019 }
2020 }
2022 /*
2023 * Called by the backend code when a flush has been completed.
2024 * The inode has already been removed from the flush list.
2025 *
2026 * A pipelined flush can occur, in which case we must re-enter the
2027 * inode on the list and re-copy its fields.
2028 */
2029 void
2031 {
2032 hammer_mount_t hmp;
2039 /*
2040 * Merge left-over flags back into the frontend and fix the state.
2041 * Incomplete truncations are retained by the backend.
2042 */
2047 /*
2048 * The backend may have adjusted nlinks, so if the adjusted nlinks
2049 * does not match the fronttend set the frontend's RDIRTY flag again.
2050 */
2054 /*
2055 * Fix up the dirty buffer status.
2056 */
2059 }
2061 /*
2062 * Re-set the XDIRTY flag if some of the inode's in-memory records
2063 * could not be flushed.
2064 */
2070 /*
2071 * Do not lose track of inodes which no longer have vnode
2072 * assocations, otherwise they may never get flushed again.
2073 *
2074 * The reflush flag can be set superfluously, causing extra pain
2075 * for no reason. If the inode is no longer modified it no longer
2076 * needs to be flushed.
2077 */
2083 }
2085 /*
2086 * Adjust the flush state.
2087 */
2089 /*
2090 * We were unable to flush out all our records, leave the
2091 * inode in a flush state and in the current flush group.
2092 * The flush group will be re-run.
2093 *
2094 * This occurs if the UNDO block gets too full or there is
2095 * too much dirty meta-data and allows the flusher to
2096 * finalize the UNDO block and then re-flush.
2097 */
2101 /*
2102 * Remove from the flush_group
2103 */
2107 /*
2108 * Clean up the vnode ref and tracking counts.
2109 */
2113 }
2117 /*
2118 * And adjust the state.
2119 */
2126 }
2128 /*
2129 * If the frontend is waiting for a flush to complete,
2130 * wake it up.
2131 */
2135 }
2137 /*
2138 * If the frontend made more changes and requested another
2139 * flush, then try to get it running.
2140 *
2141 * Reflushes are aborted when the inode is errored out.
2142 */
2150 }
2151 }
2152 }
2154 /*
2155 * If we have no parent dependancies we can clear CONN_DOWN
2156 */
2160 /*
2161 * If the inode is now clean drop the space reservation.
2162 */
2167 }
2171 }
2173 /*
2174 * Called from hammer_sync_inode() to synchronize in-memory records
2175 * to the media.
2176 */
2177 static int
2179 {
2185 /*
2186 * Skip records that do not belong to the current flush.
2187 */
2192 #if 1
2194 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2197 }
2198 #endif
2201 /*
2202 * Interlock the record using the BE flag. Once BE is set the
2203 * frontend cannot change the state of FE.
2204 *
2205 * NOTE: If FE is set prior to us setting BE we still sync the
2206 * record out, but the flush completion code converts it to
2207 * a delete-on-disk record instead of destroying it.
2208 */
2212 /*
2213 * The backend may have already disposed of the record.
2214 */
2218 }
2220 /*
2221 * If the whole inode is being deleting all on-disk records will
2222 * be deleted very soon, we can't sync any new records to disk
2223 * because they will be deleted in the same transaction they were
2224 * created in (delete_tid == create_tid), which will assert.
2225 *
2226 * XXX There may be a case with RECORD_ADD with DELETED_FE set
2227 * that we currently panic on.
2228 */
2232 /*
2233 * We don't have to do anything, if the record was
2234 * committed the space will have been accounted for
2235 * in the blockmap.
2236 */
2237 /* fall through */
2252 /*
2253 * Follow through and issue the on-disk deletion
2254 */
2256 }
2257 }
2259 /*
2260 * If DELETED_FE is set special handling is needed for directory
2261 * entries. Dependant pieces related to the directory entry may
2262 * have already been synced to disk. If this occurs we have to
2263 * sync the directory entry and then change the in-memory record
2264 * from an ADD to a DELETE to cover the fact that it's been
2265 * deleted by the frontend.
2266 *
2267 * A directory delete covering record (MEM_RECORD_DEL) can never
2268 * be deleted by the frontend.
2269 *
2270 * Any other record type (aka DATA) can be deleted by the frontend.
2271 * XXX At the moment the flusher must skip it because there may
2272 * be another data record in the flush group for the same block,
2273 * meaning that some frontend data changes can leak into the backend's
2274 * synchronization point.
2275 */
2284 }
2285 }
2287 /*
2288 * Assign the create_tid for new records. Deletions already
2289 * have the record's entire key properly set up.
2290 */
2303 }
2308 done:
2311 /*
2312 * Do partial finalization if we have built up too many dirty
2313 * buffers. Otherwise a buffer cache deadlock can occur when
2314 * doing things like creating tens of thousands of tiny files.
2315 *
2316 * We must release our cursor lock to avoid a 3-way deadlock
2317 * due to the exclusive sync lock the finalizer must get.
2318 */
2323 }
2326 }
2328 /*
2329 * Backend function called by the flusher to sync an inode to media.
2330 */
2331 int
2333 {
2335 hammer_node_t tmp_node;
2336 hammer_record_t depend;
2337 hammer_record_t next;
2339 u_int64_t nlinks;
2348 /*
2349 * Any directory records referencing this inode which are not in
2350 * our current flush group must adjust our nlink count for the
2351 * purposes of synchronization to disk.
2352 *
2353 * Records which are in our flush group can be unlinked from our
2354 * inode now, potentially allowing the inode to be physically
2355 * deleted.
2356 *
2357 * This cannot block.
2358 */
2365 /*
2366 * If this is an ADD that was deleted by the frontend
2367 * the frontend nlinks count will have already been
2368 * decremented, but the backend is going to sync its
2369 * directory entry and must account for it. The
2370 * record will be converted to a delete-on-disk when
2371 * it gets synced.
2372 *
2373 * If the ADD was not deleted by the frontend we
2374 * can remove the dependancy from our target_list.
2375 */
2380 target_entry);
2382 }
2384 /*
2385 * Not part of our flush group
2386 */
2397 }
2398 }
2399 }
2401 /*
2402 * Set dirty if we had to modify the link count.
2403 */
2408 }
2410 /*
2411 * If there is a trunction queued destroy any data past the (aligned)
2412 * truncation point. Userland will have dealt with the buffer
2413 * containing the truncation point for us.
2414 *
2415 * We don't flush pending frontend data buffers until after we've
2416 * dealt with the truncation.
2417 */
2419 /*
2420 * Interlock trunc_off. The VOP front-end may continue to
2421 * make adjustments to it while we are blocked.
2422 */
2423 off_t trunc_off;
2424 off_t aligned_trunc_off;
2431 /*
2432 * Delete any whole blocks on-media. The front-end has
2433 * already cleaned out any partial block and made it
2434 * pending. The front-end may have updated trunc_off
2435 * while we were blocked so we only use sync_trunc_off.
2436 *
2437 * This operation can blow out the buffer cache, EWOULDBLOCK
2438 * means we were unable to complete the deletion. The
2439 * deletion will update sync_trunc_off in that case.
2440 */
2442 aligned_trunc_off,
2448 }
2453 /*
2454 * Clear the truncation flag on the backend after we have
2455 * complete the deletions. Backend data is now good again
2456 * (including new records we are about to sync, below).
2457 *
2458 * Leave sync_trunc_off intact. As we write additional
2459 * records the backend will update sync_trunc_off. This
2460 * tells the backend whether it can skip the overwrite
2461 * test. This should work properly even when the backend
2462 * writes full blocks where the truncation point straddles
2463 * the block because the comparison is against the base
2464 * offset of the record.
2465 */
2467 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2470 }
2472 /*
2473 * Now sync related records. These will typically be directory
2474 * entries, records tracking direct-writes, or delete-on-disk records.
2475 */
2483 }
2486 /*
2487 * Re-seek for inode update, assuming our cache hasn't been ripped
2488 * out from under us.
2489 */
2499 }
2501 }
2503 /*
2504 * If we are deleting the inode the frontend had better not have
2505 * any active references on elements making up the inode.
2506 *
2507 * The call to hammer_ip_delete_clean() cleans up auxillary records
2508 * but not DB or DATA records. Those must have already been deleted
2509 * by the normal truncation mechanic.
2510 */
2524 /*
2525 * Set delete_tid in both the frontend and backend
2526 * copy of the inode record. The DELETED flag handles
2527 * this, do not set RDIRTY.
2528 */
2535 /*
2536 * Adjust the inode count in the volume header
2537 */
2542 vol0_stat_inodes);
2545 }
2547 }
2548 }
2554 defer_buffer_flush:
2555 /*
2556 * Now update the inode's on-disk inode-data and/or on-disk record.
2557 * DELETED and ONDISK are managed only in ip->flags.
2558 *
2559 * In the case of a defered buffer flush we still update the on-disk
2560 * inode to satisfy visibility requirements if there happen to be
2561 * directory dependancies.
2562 */
2565 /*
2566 * If deleted and on-disk, don't set any additional flags.
2567 * the delete flag takes care of things.
2568 *
2569 * Clear flags which may have been set by the frontend.
2570 */
2573 HAMMER_INODE_DELETING);
2576 /*
2577 * Take care of the case where a deleted inode was never
2578 * flushed to the disk in the first place.
2579 *
2580 * Clear flags which may have been set by the frontend.
2581 */
2584 HAMMER_INODE_DELETING);
2592 }
2595 /*
2596 * If already on-disk, do not set any additional flags.
2597 */
2600 /*
2601 * If not on-disk and not deleted, set DDIRTY to force
2602 * an initial record to be written.
2603 *
2604 * Also set the create_tid in both the frontend and backend
2605 * copy of the inode record.
2606 */
2613 }
2615 /*
2616 * If RDIRTY or DDIRTY is set, write out a new record. If the inode
2617 * is already on-disk the old record is marked as deleted.
2618 *
2619 * If DELETED is set hammer_update_inode() will delete the existing
2620 * record without writing out a new one.
2621 *
2622 * If *ONLY* the ITIMES flag is set we can update the record in-place.
2623 */
2633 }
2634 done:
2638 }
2641 }
2643 /*
2644 * This routine is called when the OS is no longer actively referencing
2645 * the inode (but might still be keeping it cached), or when releasing
2646 * the last reference to an inode.
2647 *
2648 * At this point if the inode's nlinks count is zero we want to destroy
2649 * it, which may mean destroying it on-media too.
2650 */
2651 void
2653 {
2656 /*
2657 * Set the DELETING flag when the link count drops to 0 and the
2658 * OS no longer has any opens on the inode.
2659 *
2660 * The backend will clear DELETING (a mod flag) and set DELETED
2661 * (a state flag) when it is actually able to perform the
2662 * operation.
2663 *
2664 * Don't reflag the deletion if the flusher is currently syncing
2665 * one that was already flagged. A previously set DELETING flag
2666 * may bounce around flags and sync_flags until the operation is
2667 * completely done.
2668 */
2678 }
2680 /*
2681 * Final cleanup
2682 */
2686 }
2689 }
2690 }
2691 }
2693 /*
2694 * After potentially resolving a dependancy the inode is tested
2695 * to determine whether it needs to be reflushed.
2696 */
2697 void
2699 {
2708 }
2710 }
2711 }
2713 /*
2714 * Clear the RECLAIM flag on an inode. This occurs when the inode is
2715 * reassociated with a vp or just before it gets freed.
2716 *
2717 * Wakeup one thread blocked waiting on reclaims to complete. Note that
2718 * the inode the thread is waiting on behalf of is a different inode then
2719 * the inode we are called with. This is to create a pipeline.
2720 */
2721 static void
2723 {
2738 }
2739 }
2741 /*
2742 * Setup our reclaim pipeline. We only let so many detached (and dirty)
2743 * inodes build up before we start blocking.
2744 *
2745 * When we block we don't care *which* inode has finished reclaiming,
2746 * as lone as one does. This is somewhat heuristical... we also put a
2747 * cap on how long we are willing to wait.
2748 */
2749 void
2751 {
2761 }
2770 }
2771 }