| blob | 8fa958b1a039ef81f09ae570fd4f5b1e5d738ab5 |
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 1.103.2.2 2008/07/18 00:21:09 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;
224 /*
225 * Poke the flusher. If we don't do this programs
226 * will start to stall on the reclaiming count.
227 */
231 }
232 }
234 }
236 }
238 /*
239 * Return a locked vnode for the specified inode. The inode must be
240 * referenced but NOT LOCKED on entry and will remain referenced on
241 * return.
242 *
243 * Called from the frontend.
244 */
245 int
247 {
248 hammer_mount_t hmp;
251 u_int8_t obj_type;
266 }
288 }
290 /*
291 * Only mark as the root vnode if the ip is not
292 * historical, otherwise the VFS cache will get
293 * confused. The other half of the special handling
294 * is in hammer_vop_nlookupdotdot().
295 *
296 * Pseudo-filesystem roots also do not count.
297 */
302 }
305 /* vnode locked by getnewvnode() */
306 /* make related vnode dirty if inode dirty? */
311 }
313 /*
314 * loop if the vget fails (aka races), or if the vp
315 * no longer matches ip->vp.
316 */
321 }
322 }
325 }
327 /*
328 * Locate all copies of the inode for obj_id compatible with the specified
329 * asof, reference, and issue the related call-back. This routine is used
330 * for direct-io invalidation and does not create any new inodes.
331 */
332 void
336 {
338 hammer_inode_info_cmp_all_history,
340 }
342 /*
343 * Acquire a HAMMER inode. The returned inode is not locked. These functions
344 * do not attach or detach the related vnode (use hammer_get_vnode() for
345 * that).
346 *
347 * The flags argument is only applied for newly created inodes, and only
348 * certain flags are inherited.
349 *
350 * Called from the frontend.
351 */
356 {
363 /*
364 * Determine if we already have an inode cached. If we do then
365 * we are golden.
366 */
370 loop:
376 }
378 /*
379 * Allocate a new inode structure and deal with races later.
380 */
399 /*
400 * Locate the on-disk inode. If this is a PFS root we always
401 * access the current version of the root inode and (if it is not
402 * a master) always access information under it with a snapshot
403 * TID.
404 */
405 retry:
417 HAMMER_CURSOR_ASOF;
423 }
425 /*
426 * On success the B-Tree lookup will hold the appropriate
427 * buffer cache buffers and provide a pointer to the requested
428 * information. Copy the information to the in-memory inode
429 * and cache the B-Tree node to improve future operations.
430 */
435 /*
436 * cache[0] tries to cache the location of the object inode.
437 * The assumption is that it is near the directory inode.
438 *
439 * cache[1] tries to cache the location of the object data.
440 * The assumption is that it is near the directory data.
441 */
446 /*
447 * The file should not contain any data past the file size
448 * stored in the inode. Setting save_trunc_off to the
449 * file size instead of max reduces B-Tree lookup overheads
450 * on append by allowing the flusher to avoid checking for
451 * record overwrites.
452 */
455 /*
456 * Locate and assign the pseudofs management structure to
457 * the inode.
458 */
465 errorp);
467 }
468 }
470 /*
471 * The inode is placed on the red-black tree and will be synced to
472 * the media when flushed or by the filesystem sync. If this races
473 * another instantiation/lookup the insertion will fail.
474 */
480 }
486 }
490 }
493 }
495 /*
496 * Create a new filesystem object, returning the inode in *ipp. The
497 * returned inode will be referenced. The inode is created in-memory.
498 *
499 * If pfsm is non-NULL the caller wishes to create the root inode for
500 * a master PFS.
501 */
502 int
506 {
507 hammer_mount_t hmp;
508 hammer_inode_t ip;
509 uid_t xuid;
526 }
538 /* ip->save_trunc_off = 0; (already zero) */
547 /*
548 * A nohistory designator on the parent directory is inherited by
549 * the child. We will do this even for pseudo-fs creation... the
550 * sysad can turn it off.
551 */
555 }
559 HAMMER_LOCALIZE_INODE;
572 /*
573 * Setup the ".." pointer. This only needs to be done for directories
574 * but we do it for all objects as a recovery aid.
575 */
578 #if 0
579 /*
580 * The parent_obj_localization field only applies to pseudo-fs roots.
581 * XXX this is no longer applicable, PFSs are no longer directly
582 * tied into the parent's directory structure.
583 */
588 }
589 #endif
599 }
601 /*
602 * Calculate default uid/gid and overwrite with information from
603 * the vap.
604 */
611 }
618 else
643 }
650 /* not reached */
652 }
655 }
657 /*
658 * Final cleanup / freeing of an inode structure
659 */
660 static void
662 {
674 }
677 }
679 /*
680 * Retrieve pseudo-fs data. NULL will never be returned.
681 *
682 * If an error occurs *errorp will be set and a default template is returned,
683 * otherwise *errorp is set to 0. Typically when an error occurs it will
684 * be ENOENT.
685 */
686 hammer_pseudofs_inmem_t
689 {
691 hammer_inode_t ip;
692 hammer_pseudofs_inmem_t pfsm;
696 retry:
702 }
704 /*
705 * PFS records are stored in the root inode (not the PFS root inode,
706 * but the real root). Avoid an infinite recursion if loading
707 * the PFS for the real root.
708 */
711 HAMMER_MAX_TID,
715 }
724 HAMMER_LOCALIZE_MISC;
736 else
742 HAMMER_PFSD_DELETED) {
749 }
750 }
751 }
761 }
763 }
765 /*
766 * Store pseudo-fs data. The backend will automatically delete any prior
767 * on-disk pseudo-fs data but we have to delete in-memory versions.
768 */
769 int
771 {
773 hammer_record_t record;
774 hammer_inode_t ip;
779 retry:
783 HAMMER_LOCALIZE_MISC;
804 }
805 }
811 HAMMER_LOCALIZE_MISC;
817 }
823 }
825 /*
826 * Create a root directory for a PFS if one does not alredy exist.
827 *
828 * The PFS root stands alone so we must also bump the nlinks count
829 * to prevent it from being destroyed on release.
830 */
831 int
833 hammer_pseudofs_inmem_t pfsm)
834 {
835 hammer_inode_t ip;
849 }
850 }
854 }
856 /*
857 * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
858 * if we are unable to disassociate all the inodes.
859 */
860 static
861 int
863 {
871 else
875 }
877 int
879 {
885 hammer_inode_pfs_cmp,
886 hammer_unload_pseudofs_callback,
891 }
895 }
898 /*
899 * Release a reference on a PFS
900 */
901 void
903 {
908 }
909 }
911 /*
912 * Called by hammer_sync_inode().
913 */
914 static int
916 {
918 hammer_record_t record;
922 retry:
925 /*
926 * If the inode has a presence on-disk then locate it and mark
927 * it deleted, setting DELONDISK.
928 *
929 * The record may or may not be physically deleted, depending on
930 * the retention policy.
931 */
933 HAMMER_INODE_ONDISK) {
936 HAMMER_LOCALIZE_INODE;
958 }
961 }
970 }
971 }
973 /*
974 * Ok, write out the initial record or a new record (after deleting
975 * the old one), unless the DELETED flag is set. This routine will
976 * clear DELONDISK if it writes out a record.
977 *
978 * Update our inode statistics if this is the first application of
979 * the inode on-disk.
980 */
982 /*
983 * Generate a record and write it to the media. We clean-up
984 * the state before releasing so we do not have to set-up
985 * a flush_group.
986 */
997 /*
998 * If this flag is set we cannot sync the new file size
999 * because we haven't finished related truncations. The
1000 * inode will be flushed in another flush group to finish
1001 * the job.
1002 */
1009 }
1025 }
1027 /*
1028 * The record isn't managed by the inode's record tree,
1029 * destroy it whether we succeed or fail.
1030 */
1036 /*
1037 * Finish up.
1038 */
1043 HAMMER_INODE_ATIME |
1044 HAMMER_INODE_MTIME);
1049 /*
1050 * Root volume count of inodes
1051 */
1056 vol0_stat_inodes);
1062 }
1064 }
1065 }
1067 /*
1068 * If the inode has been destroyed, clean out any left-over flags
1069 * that may have been set by the frontend.
1070 */
1073 HAMMER_INODE_ATIME |
1074 HAMMER_INODE_MTIME);
1075 }
1077 }
1079 /*
1080 * Update only the itimes fields.
1081 *
1082 * ATIME can be updated without generating any UNDO. MTIME is updated
1083 * with UNDO so it is guaranteed to be synchronized properly in case of
1084 * a crash.
1085 *
1086 * Neither field is included in the B-Tree leaf element's CRC, which is how
1087 * we can get away with updating ATIME the way we do.
1088 */
1089 static int
1091 {
1095 retry:
1097 HAMMER_INODE_ONDISK) {
1099 }
1103 HAMMER_LOCALIZE_INODE;
1121 /*
1122 * Updating MTIME requires an UNDO. Just cover
1123 * both atime and mtime.
1124 */
1128 HAMMER_ITIMES_BYTES);
1134 /*
1135 * Updating atime only can be done in-place with
1136 * no UNDO.
1137 */
1144 }
1146 }
1153 }
1155 }
1157 /*
1158 * Release a reference on an inode, flush as requested.
1159 *
1160 * On the last reference we queue the inode to the flusher for its final
1161 * disposition.
1162 */
1163 void
1165 {
1168 /*
1169 * Handle disposition when dropping the last ref.
1170 */
1173 /*
1174 * Determine whether on-disk action is needed for
1175 * the inode's final disposition.
1176 */
1182 HAMMER_FLUSH_SIGNAL);
1185 }
1189 }
1194 /*
1195 * The inode still has multiple refs, try to drop
1196 * one ref.
1197 */
1202 }
1203 }
1204 }
1205 }
1207 /*
1208 * Unload and destroy the specified inode. Must be called with one remaining
1209 * reference. The reference is disposed of.
1210 *
1211 * The inode must be completely clean.
1212 */
1213 static int
1215 {
1233 }
1235 /*
1236 * Called during unmounting if a critical error occured. The in-memory
1237 * inode and all related structures are destroyed.
1238 *
1239 * If a critical error did not occur the unmount code calls the standard
1240 * release and asserts that the inode is gone.
1241 */
1242 int
1244 {
1245 hammer_record_t rec;
1247 /*
1248 * Get rid of the inodes in-memory records, regardless of their
1249 * state, and clear the mod-mask.
1250 */
1256 }
1260 else
1268 }
1273 /*
1274 * Remove the inode from any flush group, force it idle. FLUSH
1275 * and SETUP states have an inode ref.
1276 */
1282 /* fall through */
1286 /* fall through */
1289 }
1291 /*
1292 * There shouldn't be any associated vnode. The unload needs at
1293 * least one ref, if we do have a vp steal its ip ref.
1294 */
1302 }
1305 }
1307 /*
1308 * Called on mount -u when switching from RW to RO or vise-versa. Adjust
1309 * the read-only flag for cached inodes.
1310 *
1311 * This routine is called from a RB_SCAN().
1312 */
1313 int
1315 {
1320 else
1323 }
1325 /*
1326 * A transaction has modified an inode, requiring updates as specified by
1327 * the passed flags.
1328 *
1329 * HAMMER_INODE_DDIRTY: Inode data has been updated
1330 * HAMMER_INODE_XDIRTY: Dirty in-memory records
1331 * HAMMER_INODE_BUFS: Dirty buffer cache buffers
1332 * HAMMER_INODE_DELETED: Inode record/data must be deleted
1333 * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1334 */
1335 void
1337 {
1338 /*
1339 * ronly of 0 or 2 does not trigger assertion.
1340 * 2 is a special error state
1341 */
1349 }
1352 }
1354 /*
1355 * Request that an inode be flushed. This whole mess cannot block and may
1356 * recurse (if not synchronous). Once requested HAMMER will attempt to
1357 * actively flush the inode until the flush can be done.
1358 *
1359 * The inode may already be flushing, or may be in a setup state. We can
1360 * place the inode in a flushing state if it is currently idle and flag it
1361 * to reflush if it is currently flushing.
1362 *
1363 * If the HAMMER_FLUSH_SYNCHRONOUS flag is specified we will attempt to
1364 * flush the indoe synchronously using the caller's context.
1365 */
1366 void
1368 {
1369 hammer_mount_t hmp;
1370 hammer_flush_group_t flg;
1373 /*
1374 * next_flush_group is the first flush group we can place the inode
1375 * in. It may be NULL. If it becomes full we append a new flush
1376 * group and make that the next_flush_group.
1377 */
1385 }
1391 }
1393 /*
1394 * Trivial 'nothing to flush' case. If the inode is in a SETUP
1395 * state we have to put it back into an IDLE state so we can
1396 * drop the extra ref.
1397 *
1398 * If we have a parent dependancy we must still fall through
1399 * so we can run it.
1400 */
1406 }
1409 }
1411 /*
1412 * Our flush action will depend on the current state.
1413 */
1416 /*
1417 * We have no dependancies and can flush immediately. Some
1418 * our children may not be flushable so we have to re-test
1419 * with that additional knowledge.
1420 */
1424 /*
1425 * Recurse upwards through dependancies via target_list
1426 * and start their flusher actions going if possible.
1427 *
1428 * 'good' is our connectivity. -1 means we have none and
1429 * can't flush, 0 means there weren't any dependancies, and
1430 * 1 means we have good connectivity.
1431 */
1435 /*
1436 * We can continue if good >= 0. Determine how
1437 * many records under our inode can be flushed (and
1438 * mark them).
1439 */
1442 /*
1443 * parent has no connectivity, tell it to flush
1444 * us as soon as it does.
1445 */
1450 }
1451 }
1454 /*
1455 * We are already flushing, flag the inode to reflush
1456 * if needed after it completes its current flush.
1457 */
1463 }
1465 }
1466 }
1468 /*
1469 * Scan ip->target_list, which is a list of records owned by PARENTS to our
1470 * ip which reference our ip.
1471 *
1472 * XXX This is a huge mess of recursive code, but not one bit of it blocks
1473 * so for now do not ref/deref the structures. Note that if we use the
1474 * ref/rel code later, the rel CAN block.
1475 */
1476 static int
1478 {
1479 hammer_record_t depend;
1491 }
1493 }
1495 /*
1496 * This helper function takes a record representing the dependancy between
1497 * the parent inode and child inode.
1498 *
1499 * record->ip = parent inode
1500 * record->target_ip = child inode
1501 *
1502 * We are asked to recurse upwards and convert the record from SETUP
1503 * to FLUSH if possible.
1504 *
1505 * Return 1 if the record gives us connectivity
1506 *
1507 * Return 0 if the record is not relevant
1508 *
1509 * Return -1 if we can't resolve the dependancy and there is no connectivity.
1510 */
1511 static int
1513 hammer_flush_group_t flg)
1514 {
1515 hammer_mount_t hmp;
1516 hammer_inode_t pip;
1523 /*
1524 * If the record is already flushing, is it in our flush group?
1525 *
1526 * If it is in our flush group but it is a general record or a
1527 * delete-on-disk, it does not improve our connectivity (return 0),
1528 * and if the target inode is not trying to destroy itself we can't
1529 * allow the operation yet anyway (the second return -1).
1530 */
1532 /*
1533 * If not in our flush group ask the parent to reflush
1534 * us as soon as possible.
1535 */
1540 }
1542 /*
1543 * If in our flush group everything is already set up,
1544 * just return whether the record will improve our
1545 * visibility or not.
1546 */
1550 }
1552 /*
1553 * It must be a setup record. Try to resolve the setup dependancies
1554 * by recursing upwards so we can place ip on the flush list.
1555 */
1560 /*
1561 * If good < 0 the parent has no connectivity and we cannot safely
1562 * flush the directory entry, which also means we can't flush our
1563 * ip. Flag the parent and us for downward recursion once the
1564 * parent's connectivity is resolved.
1565 */
1567 /* pip->flags |= HAMMER_INODE_CONN_DOWN; set by recursion */
1570 }
1572 /*
1573 * We are go, place the parent inode in a flushing state so we can
1574 * place its record in a flushing state. Note that the parent
1575 * may already be flushing. The record must be in the same flush
1576 * group as the parent.
1577 */
1583 #if 0
1586 /*
1587 * Regardless of flushing state we cannot sync this path if the
1588 * record represents a delete-on-disk but the target inode
1589 * is not ready to sync its own deletion.
1590 *
1591 * XXX need to count effective nlinks to determine whether
1592 * the flush is ok, otherwise removing a hardlink will
1593 * just leave the DEL record to rot.
1594 */
1598 #endif
1600 /*
1601 * If the parent is in the same flush group as us we can
1602 * just set the record to a flushing state and we are
1603 * done.
1604 */
1610 /*
1611 * A general directory-add contributes to our visibility.
1612 *
1613 * Otherwise it is probably a directory-delete or
1614 * delete-on-disk record and does not contribute to our
1615 * visbility (but we can still flush it).
1616 */
1621 /*
1622 * If the parent is not in our flush group we cannot
1623 * flush this record yet, there is no visibility.
1624 * We tell the parent to reflush and mark ourselves
1625 * so the parent knows it should flush us too.
1626 */
1630 }
1631 }
1633 /*
1634 * This is the core routine placing an inode into the FST_FLUSH state.
1635 */
1636 static void
1638 {
1641 /*
1642 * Set flush state and prevent the flusher from cycling into
1643 * the next flush group. Do not place the ip on the list yet.
1644 * Inodes not in the idle state get an extra reference.
1645 */
1656 /*
1657 * We need to be able to vfsync/truncate from the backend.
1658 */
1663 }
1665 /*
1666 * Figure out how many in-memory records we can actually flush
1667 * (not including inode meta-data, buffers, etc).
1668 */
1670 /*
1671 * If this is a upwards recursion we do not want to
1672 * recurse down again!
1673 */
1676 /*
1677 * No new records are added if we must complete a flush
1678 * from a previous cycle, but we do have to move the records
1679 * from the previous cycle to the current one.
1680 */
1681 #if 0
1684 #endif
1687 /*
1688 * Normal flush, scan records and bring them into the flush.
1689 * Directory adds and deletes are usually skipped (they are
1690 * grouped with the related inode rather then with the
1691 * directory).
1692 *
1693 * go_count can be negative, which means the scan aborted
1694 * due to the flush group being over-full and we should
1695 * flush what we have.
1696 */
1699 }
1701 /*
1702 * This is a more involved test that includes go_count. If we
1703 * can't flush, flag the inode and return. If go_count is 0 we
1704 * were are unable to flush any records in our rec_tree and
1705 * must ignore the XDIRTY flag.
1706 */
1719 }
1723 }
1727 }
1728 }
1730 /*
1731 * Snapshot the state of the inode for the backend flusher.
1732 *
1733 * We continue to retain save_trunc_off even when all truncations
1734 * have been resolved as an optimization to determine if we can
1735 * skip the B-Tree lookup for overwrite deletions.
1736 *
1737 * NOTE: The DELETING flag is a mod flag, but it is also sticky,
1738 * and stays in ip->flags. Once set, it stays set until the
1739 * inode is destroyed.
1740 *
1741 * NOTE: If a truncation from a previous flush cycle had to be
1742 * continued into this one, the TRUNCATED flag will still be
1743 * set in sync_flags as will WOULDBLOCK. When this occurs
1744 * we CANNOT safely integrate a new truncation from the front-end
1745 * because there may be data records in-memory assigned a flush
1746 * state from the previous cycle that are supposed to be flushed
1747 * before the next frontend truncation.
1748 */
1750 HAMMER_INODE_TRUNCATED) {
1757 /*
1758 * The save_trunc_off used to cache whether the B-Tree
1759 * holds any records past that point is not used until
1760 * after the truncation has succeeded, so we can safely
1761 * set it now.
1762 */
1765 }
1771 #ifdef DEBUG_TRUNCATE
1774 #endif
1776 /*
1777 * The flusher list inherits our inode and reference.
1778 */
1786 }
1787 }
1789 /*
1790 * Callback for scan of ip->rec_tree. Try to include each record in our
1791 * flush. ip->flush_group has been set but the inode has not yet been
1792 * moved into a flushing state.
1793 *
1794 * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1795 * both inodes.
1796 *
1797 * We return 1 for any record placed or found in FST_FLUSH, which prevents
1798 * the caller from shortcutting the flush.
1799 */
1800 static int
1802 {
1803 hammer_flush_group_t flg;
1804 hammer_inode_t target_ip;
1805 hammer_inode_t ip;
1808 /*
1809 * Deleted records are ignored. Note that the flush detects deleted
1810 * front-end records at multiple points to deal with races. This is
1811 * just the first line of defense. The only time DELETED_FE cannot
1812 * be set is when HAMMER_RECF_INTERLOCK_BE is set.
1813 *
1814 * Don't get confused between record deletion and, say, directory
1815 * entry deletion. The deletion of a directory entry that is on
1816 * the media has nothing to do with the record deletion flags.
1817 */
1824 }
1826 }
1828 /*
1829 * If the record is in an idle state it has no dependancies and
1830 * can be flushed.
1831 */
1838 /*
1839 * The record has no setup dependancy, we can flush it.
1840 */
1849 /*
1850 * The record has a setup dependancy. These are typically
1851 * directory entry adds and deletes. Such entries will be
1852 * flushed when their inodes are flushed so we do not
1853 * usually have to add them to the flush here. However,
1854 * if the target_ip has set HAMMER_INODE_CONN_DOWN then
1855 * it is asking us to flush this record (and it).
1856 */
1861 /*
1862 * If the target IP is already flushing in our group
1863 * we are golden, otherwise make sure the target
1864 * reflushes.
1865 */
1875 }
1877 }
1879 /*
1880 * Target IP is not yet flushing. This can get complex
1881 * because we have to be careful about the recursion.
1882 *
1883 * Directories create an issue for us in that if a flush
1884 * of a directory is requested the expectation is to flush
1885 * any pending directory entries, but this will cause the
1886 * related inodes to recursively flush as well. We can't
1887 * really defer the operation so just get as many as we
1888 * can and
1889 */
1890 #if 0
1893 /*
1894 * We aren't reclaiming and the target ip was not
1895 * previously prevented from flushing due to this
1896 * record dependancy. Do not flush this record.
1897 */
1898 /*r = 0;*/
1900 #endif
1903 /*
1904 * Our flush group is over-full and we risk blowing
1905 * out the UNDO FIFO. Stop the scan, flush what we
1906 * have, then reflush the directory.
1907 *
1908 * The directory may be forced through multiple
1909 * flush groups before it can be completely
1910 * flushed.
1911 */
1916 /*
1917 * If the target IP is not flushing we can force
1918 * it to flush, even if it is unable to write out
1919 * any of its own records we have at least one in
1920 * hand that we CAN deal with.
1921 */
1927 HAMMER_FLUSH_RECURSION);
1930 /*
1931 * General or delete-on-disk record.
1932 *
1933 * XXX this needs help. If a delete-on-disk we could
1934 * disconnect the target. If the target has its own
1935 * dependancies they really need to be flushed.
1936 *
1937 * XXX
1938 */
1944 HAMMER_FLUSH_RECURSION);
1946 }
1949 /*
1950 * If the WOULDBLOCK flag is set records may have been left
1951 * over from a previous flush attempt. The flush group will
1952 * have been left intact - we are probably reflushing it
1953 * now.
1954 *
1955 * If a flush error occured ip->error will be non-zero.
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;
1999 }
2004 }
2005 }
2006 }
2008 /*
2009 * Called by the backend code when a flush has been completed.
2010 * The inode has already been removed from the flush list.
2011 *
2012 * A pipelined flush can occur, in which case we must re-enter the
2013 * inode on the list and re-copy its fields.
2014 */
2015 void
2017 {
2018 hammer_mount_t hmp;
2025 /*
2026 * Merge left-over flags back into the frontend and fix the state.
2027 * Incomplete truncations are retained by the backend.
2028 */
2033 /*
2034 * The backend may have adjusted nlinks, so if the adjusted nlinks
2035 * does not match the fronttend set the frontend's RDIRTY flag again.
2036 */
2040 /*
2041 * Fix up the dirty buffer status.
2042 */
2045 }
2047 /*
2048 * Re-set the XDIRTY flag if some of the inode's in-memory records
2049 * could not be flushed.
2050 */
2056 /*
2057 * Do not lose track of inodes which no longer have vnode
2058 * assocations, otherwise they may never get flushed again.
2059 */
2063 /*
2064 * Adjust the flush state.
2065 */
2067 /*
2068 * We were unable to flush out all our records, leave the
2069 * inode in a flush state and in the current flush group.
2070 *
2071 * This occurs if the UNDO block gets too full
2072 * or there is too much dirty meta-data and allows the
2073 * flusher to finalize the UNDO block and then re-flush.
2074 */
2078 /*
2079 * Remove from the flush_group
2080 */
2084 /*
2085 * Clean up the vnode ref and tracking counts.
2086 */
2090 }
2094 /*
2095 * And adjust the state.
2096 */
2103 }
2105 /*
2106 * If the frontend is waiting for a flush to complete,
2107 * wake it up.
2108 */
2112 }
2113 }
2115 /*
2116 * If the frontend made more changes and requested another flush,
2117 * then try to get it running.
2118 *
2119 * Reflushes are aborted when the inode is errored out.
2120 */
2128 }
2129 }
2131 /*
2132 * If we have no parent dependancies we can clear CONN_DOWN
2133 */
2137 /*
2138 * If the inode is now clean drop the space reservation.
2139 */
2144 }
2148 }
2150 /*
2151 * Called from hammer_sync_inode() to synchronize in-memory records
2152 * to the media.
2153 */
2154 static int
2156 {
2162 /*
2163 * Skip records that do not belong to the current flush.
2164 */
2169 #if 1
2171 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
2174 }
2175 #endif
2178 /*
2179 * Interlock the record using the BE flag. Once BE is set the
2180 * frontend cannot change the state of FE.
2181 *
2182 * NOTE: If FE is set prior to us setting BE we still sync the
2183 * record out, but the flush completion code converts it to
2184 * a delete-on-disk record instead of destroying it.
2185 */
2189 /*
2190 * The backend may have already disposed of the record.
2191 */
2195 }
2197 /*
2198 * If the whole inode is being deleting all on-disk records will
2199 * be deleted very soon, we can't sync any new records to disk
2200 * because they will be deleted in the same transaction they were
2201 * created in (delete_tid == create_tid), which will assert.
2202 *
2203 * XXX There may be a case with RECORD_ADD with DELETED_FE set
2204 * that we currently panic on.
2205 */
2209 /*
2210 * We don't have to do anything, if the record was
2211 * committed the space will have been accounted for
2212 * in the blockmap.
2213 */
2214 /* fall through */
2229 /*
2230 * Follow through and issue the on-disk deletion
2231 */
2233 }
2234 }
2236 /*
2237 * If DELETED_FE is set special handling is needed for directory
2238 * entries. Dependant pieces related to the directory entry may
2239 * have already been synced to disk. If this occurs we have to
2240 * sync the directory entry and then change the in-memory record
2241 * from an ADD to a DELETE to cover the fact that it's been
2242 * deleted by the frontend.
2243 *
2244 * A directory delete covering record (MEM_RECORD_DEL) can never
2245 * be deleted by the frontend.
2246 *
2247 * Any other record type (aka DATA) can be deleted by the frontend.
2248 * XXX At the moment the flusher must skip it because there may
2249 * be another data record in the flush group for the same block,
2250 * meaning that some frontend data changes can leak into the backend's
2251 * synchronization point.
2252 */
2261 }
2262 }
2264 /*
2265 * Assign the create_tid for new records. Deletions already
2266 * have the record's entire key properly set up.
2267 */
2280 }
2285 done:
2288 /*
2289 * Do partial finalization if we have built up too many dirty
2290 * buffers. Otherwise a buffer cache deadlock can occur when
2291 * doing things like creating tens of thousands of tiny files.
2292 *
2293 * We must release our cursor lock to avoid a 3-way deadlock
2294 * due to the exclusive sync lock the finalizer must get.
2295 */
2300 }
2303 }
2305 /*
2306 * XXX error handling
2307 */
2308 int
2310 {
2312 hammer_node_t tmp_node;
2313 hammer_record_t depend;
2314 hammer_record_t next;
2316 u_int64_t nlinks;
2325 /*
2326 * Any directory records referencing this inode which are not in
2327 * our current flush group must adjust our nlink count for the
2328 * purposes of synchronization to disk.
2329 *
2330 * Records which are in our flush group can be unlinked from our
2331 * inode now, potentially allowing the inode to be physically
2332 * deleted.
2333 *
2334 * This cannot block.
2335 */
2342 /*
2343 * If this is an ADD that was deleted by the frontend
2344 * the frontend nlinks count will have already been
2345 * decremented, but the backend is going to sync its
2346 * directory entry and must account for it. The
2347 * record will be converted to a delete-on-disk when
2348 * it gets synced.
2349 *
2350 * If the ADD was not deleted by the frontend we
2351 * can remove the dependancy from our target_list.
2352 */
2357 target_entry);
2359 }
2361 /*
2362 * Not part of our flush group
2363 */
2374 }
2375 }
2376 }
2378 /*
2379 * Set dirty if we had to modify the link count.
2380 */
2385 }
2387 /*
2388 * If there is a trunction queued destroy any data past the (aligned)
2389 * truncation point. Userland will have dealt with the buffer
2390 * containing the truncation point for us.
2391 *
2392 * We don't flush pending frontend data buffers until after we've
2393 * dealt with the truncation.
2394 */
2396 /*
2397 * Interlock trunc_off. The VOP front-end may continue to
2398 * make adjustments to it while we are blocked.
2399 */
2400 off_t trunc_off;
2401 off_t aligned_trunc_off;
2408 /*
2409 * Delete any whole blocks on-media. The front-end has
2410 * already cleaned out any partial block and made it
2411 * pending. The front-end may have updated trunc_off
2412 * while we were blocked so we only use sync_trunc_off.
2413 *
2414 * This operation can blow out the buffer cache, EWOULDBLOCK
2415 * means we were unable to complete the deletion. The
2416 * deletion will update sync_trunc_off in that case.
2417 */
2419 aligned_trunc_off,
2425 }
2430 /*
2431 * Clear the truncation flag on the backend after we have
2432 * complete the deletions. Backend data is now good again
2433 * (including new records we are about to sync, below).
2434 *
2435 * Leave sync_trunc_off intact. As we write additional
2436 * records the backend will update sync_trunc_off. This
2437 * tells the backend whether it can skip the overwrite
2438 * test. This should work properly even when the backend
2439 * writes full blocks where the truncation point straddles
2440 * the block because the comparison is against the base
2441 * offset of the record.
2442 */
2444 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2447 }
2449 /*
2450 * Now sync related records. These will typically be directory
2451 * entries, records tracking direct-writes, or delete-on-disk records.
2452 */
2460 }
2463 /*
2464 * Re-seek for inode update, assuming our cache hasn't been ripped
2465 * out from under us.
2466 */
2476 }
2478 }
2480 /*
2481 * If we are deleting the inode the frontend had better not have
2482 * any active references on elements making up the inode.
2483 *
2484 * The call to hammer_ip_delete_clean() cleans up auxillary records
2485 * but not DB or DATA records. Those must have already been deleted
2486 * by the normal truncation mechanic.
2487 */
2501 /*
2502 * Set delete_tid in both the frontend and backend
2503 * copy of the inode record. The DELETED flag handles
2504 * this, do not set RDIRTY.
2505 */
2512 /*
2513 * Adjust the inode count in the volume header
2514 */
2519 vol0_stat_inodes);
2522 }
2524 }
2525 }
2531 defer_buffer_flush:
2532 /*
2533 * Now update the inode's on-disk inode-data and/or on-disk record.
2534 * DELETED and ONDISK are managed only in ip->flags.
2535 *
2536 * In the case of a defered buffer flush we still update the on-disk
2537 * inode to satisfy visibility requirements if there happen to be
2538 * directory dependancies.
2539 */
2542 /*
2543 * If deleted and on-disk, don't set any additional flags.
2544 * the delete flag takes care of things.
2545 *
2546 * Clear flags which may have been set by the frontend.
2547 */
2550 HAMMER_INODE_DELETING);
2553 /*
2554 * Take care of the case where a deleted inode was never
2555 * flushed to the disk in the first place.
2556 *
2557 * Clear flags which may have been set by the frontend.
2558 */
2561 HAMMER_INODE_DELETING);
2569 }
2572 /*
2573 * If already on-disk, do not set any additional flags.
2574 */
2577 /*
2578 * If not on-disk and not deleted, set DDIRTY to force
2579 * an initial record to be written.
2580 *
2581 * Also set the create_tid in both the frontend and backend
2582 * copy of the inode record.
2583 */
2590 }
2592 /*
2593 * If RDIRTY or DDIRTY is set, write out a new record. If the inode
2594 * is already on-disk the old record is marked as deleted.
2595 *
2596 * If DELETED is set hammer_update_inode() will delete the existing
2597 * record without writing out a new one.
2598 *
2599 * If *ONLY* the ITIMES flag is set we can update the record in-place.
2600 */
2610 }
2611 done:
2615 }
2618 }
2620 /*
2621 * This routine is called when the OS is no longer actively referencing
2622 * the inode (but might still be keeping it cached), or when releasing
2623 * the last reference to an inode.
2624 *
2625 * At this point if the inode's nlinks count is zero we want to destroy
2626 * it, which may mean destroying it on-media too.
2627 */
2628 void
2630 {
2633 /*
2634 * Set the DELETING flag when the link count drops to 0 and the
2635 * OS no longer has any opens on the inode.
2636 *
2637 * The backend will clear DELETING (a mod flag) and set DELETED
2638 * (a state flag) when it is actually able to perform the
2639 * operation.
2640 */
2650 }
2652 /*
2653 * Final cleanup
2654 */
2658 }
2661 }
2662 }
2663 }
2665 /*
2666 * After potentially resolving a dependancy the inode is tested
2667 * to determine whether it needs to be reflushed.
2668 */
2669 void
2671 {
2680 }
2682 }
2683 }
2685 /*
2686 * Clear the RECLAIM flag on an inode. This occurs when the inode is
2687 * reassociated with a vp or just before it gets freed.
2688 *
2689 * Wakeup one thread blocked waiting on reclaims to complete. Note that
2690 * the inode the thread is waiting on behalf of is a different inode then
2691 * the inode we are called with. This is to create a pipeline.
2692 */
2693 static void
2695 {
2710 }
2711 }
2713 /*
2714 * Setup our reclaim pipeline. We only let so many detached (and dirty)
2715 * inodes build up before we start blocking.
2716 *
2717 * When we block we don't care *which* inode has finished reclaiming,
2718 * as lone as one does. This is somewhat heuristical... we also put a
2719 * cap on how long we are willing to wait.
2720 */
2721 void
2723 {
2733 }
2737 HAMMER_RECLAIM_WAIT;
2742 }
2743 }