| blob | 0f3ec742a77e07718578e457d1c2675e78e42465 |
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 1.99 2008/07/11 01:22:29 dillon Exp $
35 */
38 #include <vm/vm_extern.h>
39 #include <sys/buf.h>
40 #include <sys/buf2.h>
51 #ifdef DEBUG_TRUNCATE
53 #endif
55 /*
56 * RB-Tree support for inode structures
57 */
58 int
60 {
74 }
76 /*
77 * RB-Tree support for inode structures / special LOOKUP_INFO
78 */
79 static int
81 {
95 }
97 /*
98 * Used by hammer_scan_inode_snapshots() to locate all of an object's
99 * snapshots. Note that the asof field is not tested, which we can get
100 * away with because it is the lowest-priority field.
101 */
102 static int
104 {
116 }
118 /*
119 * RB-Tree support for pseudofs structures
120 */
121 static int
123 {
129 }
138 /*
139 * The kernel is not actively referencing this vnode but is still holding
140 * it cached.
141 *
142 * This is called from the frontend.
143 */
144 int
146 {
149 /*
150 * Degenerate case
151 */
155 }
157 /*
158 * If the inode no longer has visibility in the filesystem try to
159 * recycle it immediately, even if the inode is dirty. Recycling
160 * it quickly allows the system to reclaim buffer cache and VM
161 * resources which can matter a lot in a heavily loaded system.
162 *
163 * This can deadlock in vfsync() if we aren't careful.
164 *
165 * Do not queue the inode to the flusher if we still have visibility,
166 * otherwise namespace calls such as chmod will unnecessarily generate
167 * multiple inode updates.
168 */
174 }
176 }
178 /*
179 * Release the vnode association. This is typically (but not always)
180 * the last reference on the inode.
181 *
182 * Once the association is lost we are on our own with regards to
183 * flushing the inode.
184 */
185 int
187 {
189 hammer_mount_t hmp;
206 }
207 }
209 }
211 }
213 /*
214 * Return a locked vnode for the specified inode. The inode must be
215 * referenced but NOT LOCKED on entry and will remain referenced on
216 * return.
217 *
218 * Called from the frontend.
219 */
220 int
222 {
223 hammer_mount_t hmp;
226 u_int8_t obj_type;
241 }
263 }
265 /*
266 * Only mark as the root vnode if the ip is not
267 * historical, otherwise the VFS cache will get
268 * confused. The other half of the special handling
269 * is in hammer_vop_nlookupdotdot().
270 *
271 * Pseudo-filesystem roots also do not count.
272 */
277 }
280 /* vnode locked by getnewvnode() */
281 /* make related vnode dirty if inode dirty? */
286 }
288 /*
289 * loop if the vget fails (aka races), or if the vp
290 * no longer matches ip->vp.
291 */
296 }
297 }
300 }
302 /*
303 * Locate all copies of the inode for obj_id compatible with the specified
304 * asof, reference, and issue the related call-back. This routine is used
305 * for direct-io invalidation and does not create any new inodes.
306 */
307 void
311 {
313 hammer_inode_info_cmp_all_history,
315 }
317 /*
318 * Acquire a HAMMER inode. The returned inode is not locked. These functions
319 * do not attach or detach the related vnode (use hammer_get_vnode() for
320 * that).
321 *
322 * The flags argument is only applied for newly created inodes, and only
323 * certain flags are inherited.
324 *
325 * Called from the frontend.
326 */
331 {
338 /*
339 * Determine if we already have an inode cached. If we do then
340 * we are golden.
341 */
345 loop:
351 }
353 /*
354 * Allocate a new inode structure and deal with races later.
355 */
374 /*
375 * Locate the on-disk inode. If this is a PFS root we always
376 * access the current version of the root inode and (if it is not
377 * a master) always access information under it with a snapshot
378 * TID.
379 */
380 retry:
392 HAMMER_CURSOR_ASOF;
398 }
400 /*
401 * On success the B-Tree lookup will hold the appropriate
402 * buffer cache buffers and provide a pointer to the requested
403 * information. Copy the information to the in-memory inode
404 * and cache the B-Tree node to improve future operations.
405 */
410 /*
411 * cache[0] tries to cache the location of the object inode.
412 * The assumption is that it is near the directory inode.
413 *
414 * cache[1] tries to cache the location of the object data.
415 * The assumption is that it is near the directory data.
416 */
421 /*
422 * The file should not contain any data past the file size
423 * stored in the inode. Setting save_trunc_off to the
424 * file size instead of max reduces B-Tree lookup overheads
425 * on append by allowing the flusher to avoid checking for
426 * record overwrites.
427 */
430 /*
431 * Locate and assign the pseudofs management structure to
432 * the inode.
433 */
440 errorp);
442 }
443 }
445 /*
446 * The inode is placed on the red-black tree and will be synced to
447 * the media when flushed or by the filesystem sync. If this races
448 * another instantiation/lookup the insertion will fail.
449 */
455 }
461 }
465 }
468 }
470 /*
471 * Create a new filesystem object, returning the inode in *ipp. The
472 * returned inode will be referenced. The inode is created in-memory.
473 *
474 * If pfsm is non-NULL the caller wishes to create the root inode for
475 * a master PFS.
476 */
477 int
481 {
482 hammer_mount_t hmp;
483 hammer_inode_t ip;
484 uid_t xuid;
501 }
513 /* ip->save_trunc_off = 0; (already zero) */
522 /*
523 * A nohistory designator on the parent directory is inherited by
524 * the child. We will do this even for pseudo-fs creation... the
525 * sysad can turn it off.
526 */
530 }
534 HAMMER_LOCALIZE_INODE;
547 /*
548 * Setup the ".." pointer. This only needs to be done for directories
549 * but we do it for all objects as a recovery aid.
550 */
553 #if 0
554 /*
555 * The parent_obj_localization field only applies to pseudo-fs roots.
556 * XXX this is no longer applicable, PFSs are no longer directly
557 * tied into the parent's directory structure.
558 */
563 }
564 #endif
574 }
576 /*
577 * Calculate default uid/gid and overwrite with information from
578 * the vap.
579 */
586 }
593 else
618 }
625 /* not reached */
627 }
630 }
632 /*
633 * Final cleanup / freeing of an inode structure
634 */
635 static void
637 {
649 }
652 }
654 /*
655 * Retrieve pseudo-fs data. NULL will never be returned.
656 *
657 * If an error occurs *errorp will be set and a default template is returned,
658 * otherwise *errorp is set to 0. Typically when an error occurs it will
659 * be ENOENT.
660 */
661 hammer_pseudofs_inmem_t
664 {
666 hammer_inode_t ip;
667 hammer_pseudofs_inmem_t pfsm;
671 retry:
677 }
679 /*
680 * PFS records are stored in the root inode (not the PFS root inode,
681 * but the real root). Avoid an infinite recursion if loading
682 * the PFS for the real root.
683 */
686 HAMMER_MAX_TID,
690 }
699 HAMMER_LOCALIZE_MISC;
711 else
720 }
721 }
731 }
733 }
735 /*
736 * Store pseudo-fs data. The backend will automatically delete any prior
737 * on-disk pseudo-fs data but we have to delete in-memory versions.
738 */
739 int
741 {
743 hammer_record_t record;
744 hammer_inode_t ip;
749 retry:
753 HAMMER_LOCALIZE_MISC;
774 }
775 }
781 HAMMER_LOCALIZE_MISC;
787 }
793 }
795 /*
796 * Create a root directory for a PFS if one does not alredy exist.
797 *
798 * The PFS root stands alone so we must also bump the nlinks count
799 * to prevent it from being destroyed on release.
800 */
801 int
803 hammer_pseudofs_inmem_t pfsm)
804 {
805 hammer_inode_t ip;
819 }
820 }
824 }
826 /*
827 * Release a reference on a PFS
828 */
829 void
831 {
836 }
837 }
839 /*
840 * Called by hammer_sync_inode().
841 */
842 static int
844 {
846 hammer_record_t record;
850 retry:
853 /*
854 * If the inode has a presence on-disk then locate it and mark
855 * it deleted, setting DELONDISK.
856 *
857 * The record may or may not be physically deleted, depending on
858 * the retention policy.
859 */
861 HAMMER_INODE_ONDISK) {
864 HAMMER_LOCALIZE_INODE;
882 }
891 }
894 }
897 }
906 }
907 }
909 /*
910 * Ok, write out the initial record or a new record (after deleting
911 * the old one), unless the DELETED flag is set. This routine will
912 * clear DELONDISK if it writes out a record.
913 *
914 * Update our inode statistics if this is the first application of
915 * the inode on-disk.
916 */
918 /*
919 * Generate a record and write it to the media
920 */
931 /*
932 * If this flag is set we cannot sync the new file size
933 * because we haven't finished related truncations. The
934 * inode will be flushed in another flush group to finish
935 * the job.
936 */
943 }
959 }
963 }
965 /*
966 * The record isn't managed by the inode's record tree,
967 * destroy it whether we succeed or fail.
968 */
974 /*
975 * Finish up.
976 */
981 HAMMER_INODE_ATIME |
982 HAMMER_INODE_MTIME);
987 /*
988 * Root volume count of inodes
989 */
994 vol0_stat_inodes);
1000 }
1002 }
1003 }
1005 /*
1006 * If the inode has been destroyed, clean out any left-over flags
1007 * that may have been set by the frontend.
1008 */
1011 HAMMER_INODE_ATIME |
1012 HAMMER_INODE_MTIME);
1013 }
1015 }
1017 /*
1018 * Update only the itimes fields.
1019 *
1020 * ATIME can be updated without generating any UNDO. MTIME is updated
1021 * with UNDO so it is guaranteed to be synchronized properly in case of
1022 * a crash.
1023 *
1024 * Neither field is included in the B-Tree leaf element's CRC, which is how
1025 * we can get away with updating ATIME the way we do.
1026 */
1027 static int
1029 {
1033 retry:
1035 HAMMER_INODE_ONDISK) {
1037 }
1041 HAMMER_LOCALIZE_INODE;
1059 }
1063 /*
1064 * Updating MTIME requires an UNDO. Just cover
1065 * both atime and mtime.
1066 */
1070 HAMMER_ITIMES_BYTES);
1076 /*
1077 * Updating atime only can be done in-place with
1078 * no UNDO.
1079 */
1086 }
1088 }
1095 }
1097 }
1099 /*
1100 * Release a reference on an inode, flush as requested.
1101 *
1102 * On the last reference we queue the inode to the flusher for its final
1103 * disposition.
1104 */
1105 void
1107 {
1110 /*
1111 * Handle disposition when dropping the last ref.
1112 */
1115 /*
1116 * Determine whether on-disk action is needed for
1117 * the inode's final disposition.
1118 */
1124 HAMMER_FLUSH_SIGNAL);
1127 }
1131 }
1136 /*
1137 * The inode still has multiple refs, try to drop
1138 * one ref.
1139 */
1144 }
1145 }
1146 }
1147 }
1149 /*
1150 * Unload and destroy the specified inode. Must be called with one remaining
1151 * reference. The reference is disposed of.
1152 *
1153 * This can only be called in the context of the flusher.
1154 */
1155 static int
1157 {
1175 }
1177 /*
1178 * Called on mount -u when switching from RW to RO or vise-versa. Adjust
1179 * the read-only flag for cached inodes.
1180 *
1181 * This routine is called from a RB_SCAN().
1182 */
1183 int
1185 {
1190 else
1193 }
1195 /*
1196 * A transaction has modified an inode, requiring updates as specified by
1197 * the passed flags.
1198 *
1199 * HAMMER_INODE_DDIRTY: Inode data has been updated
1200 * HAMMER_INODE_XDIRTY: Dirty in-memory records
1201 * HAMMER_INODE_BUFS: Dirty buffer cache buffers
1202 * HAMMER_INODE_DELETED: Inode record/data must be deleted
1203 * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1204 */
1205 void
1207 {
1215 }
1218 }
1220 /*
1221 * Request that an inode be flushed. This whole mess cannot block and may
1222 * recurse (if not synchronous). Once requested HAMMER will attempt to
1223 * actively flush the inode until the flush can be done.
1224 *
1225 * The inode may already be flushing, or may be in a setup state. We can
1226 * place the inode in a flushing state if it is currently idle and flag it
1227 * to reflush if it is currently flushing.
1228 *
1229 * If the HAMMER_FLUSH_SYNCHRONOUS flag is specified we will attempt to
1230 * flush the indoe synchronously using the caller's context.
1231 */
1232 void
1234 {
1237 /*
1238 * Trivial 'nothing to flush' case. If the inode is ina SETUP
1239 * state we have to put it back into an IDLE state so we can
1240 * drop the extra ref.
1241 */
1246 }
1248 }
1250 /*
1251 * Our flush action will depend on the current state.
1252 */
1255 /*
1256 * We have no dependancies and can flush immediately. Some
1257 * our children may not be flushable so we have to re-test
1258 * with that additional knowledge.
1259 */
1263 /*
1264 * Recurse upwards through dependancies via target_list
1265 * and start their flusher actions going if possible.
1266 *
1267 * 'good' is our connectivity. -1 means we have none and
1268 * can't flush, 0 means there weren't any dependancies, and
1269 * 1 means we have good connectivity.
1270 */
1273 /*
1274 * We can continue if good >= 0. Determine how many records
1275 * under our inode can be flushed (and mark them).
1276 */
1284 }
1285 }
1288 /*
1289 * We are already flushing, flag the inode to reflush
1290 * if needed after it completes its current flush.
1291 */
1297 }
1299 }
1300 }
1302 /*
1303 * Scan ip->target_list, which is a list of records owned by PARENTS to our
1304 * ip which reference our ip.
1305 *
1306 * XXX This is a huge mess of recursive code, but not one bit of it blocks
1307 * so for now do not ref/deref the structures. Note that if we use the
1308 * ref/rel code later, the rel CAN block.
1309 */
1310 static int
1312 {
1313 hammer_record_t depend;
1314 #if 0
1315 hammer_record_t next;
1316 hammer_inode_t pip;
1317 #endif
1329 }
1332 #if 0
1333 retry:
1339 }
1346 }
1352 }
1361 }
1363 #endif
1364 }
1366 /*
1367 * This helper function takes a record representing the dependancy between
1368 * the parent inode and child inode.
1369 *
1370 * record->ip = parent inode
1371 * record->target_ip = child inode
1372 *
1373 * We are asked to recurse upwards and convert the record from SETUP
1374 * to FLUSH if possible.
1375 *
1376 * Return 1 if the record gives us connectivity
1377 *
1378 * Return 0 if the record is not relevant
1379 *
1380 * Return -1 if we can't resolve the dependancy and there is no connectivity.
1381 */
1382 static int
1384 {
1385 hammer_mount_t hmp;
1386 hammer_inode_t pip;
1393 /*
1394 * If the record is already flushing, is it in our flush group?
1395 *
1396 * If it is in our flush group but it is a general record or a
1397 * delete-on-disk, it does not improve our connectivity (return 0),
1398 * and if the target inode is not trying to destroy itself we can't
1399 * allow the operation yet anyway (the second return -1).
1400 */
1405 }
1408 /* GENERAL or DEL */
1410 }
1412 /*
1413 * It must be a setup record. Try to resolve the setup dependancies
1414 * by recursing upwards so we can place ip on the flush list.
1415 */
1420 /*
1421 * We can't flush ip because it has no connectivity (XXX also check
1422 * nlinks for pre-existing connectivity!). Flag it so any resolution
1423 * recurses back down.
1424 */
1428 }
1430 /*
1431 * We are go, place the parent inode in a flushing state so we can
1432 * place its record in a flushing state. Note that the parent
1433 * may already be flushing. The record must be in the same flush
1434 * group as the parent.
1435 */
1441 #if 0
1444 /*
1445 * Regardless of flushing state we cannot sync this path if the
1446 * record represents a delete-on-disk but the target inode
1447 * is not ready to sync its own deletion.
1448 *
1449 * XXX need to count effective nlinks to determine whether
1450 * the flush is ok, otherwise removing a hardlink will
1451 * just leave the DEL record to rot.
1452 */
1456 #endif
1458 /*
1459 * This is the record we wanted to synchronize. If the
1460 * record went into a flush state while we blocked it
1461 * had better be in the correct flush group.
1462 */
1469 }
1473 /*
1474 * A general or delete-on-disk record does not contribute
1475 * to our visibility. We can still flush it, however.
1476 */
1479 /*
1480 * We couldn't resolve the dependancies, request that the
1481 * inode be flushed when the dependancies can be resolved.
1482 */
1485 }
1486 }
1488 /*
1489 * This is the core routine placing an inode into the FST_FLUSH state.
1490 */
1491 static void
1493 {
1496 /*
1497 * Set flush state and prevent the flusher from cycling into
1498 * the next flush group. Do not place the ip on the list yet.
1499 * Inodes not in the idle state get an extra reference.
1500 */
1510 /*
1511 * We need to be able to vfsync/truncate from the backend.
1512 */
1517 }
1519 /*
1520 * Figure out how many in-memory records we can actually flush
1521 * (not including inode meta-data, buffers, etc).
1522 *
1523 * Do not add new records to the flush if this is a recursion or
1524 * if we must still complete a flush from the previous flush cycle.
1525 */
1535 }
1537 /*
1538 * This is a more involved test that includes go_count. If we
1539 * can't flush, flag the inode and return. If go_count is 0 we
1540 * were are unable to flush any records in our rec_tree and
1541 * must ignore the XDIRTY flag.
1542 */
1554 }
1558 }
1562 }
1563 }
1565 /*
1566 * Snapshot the state of the inode for the backend flusher.
1567 *
1568 * We continue to retain save_trunc_off even when all truncations
1569 * have been resolved as an optimization to determine if we can
1570 * skip the B-Tree lookup for overwrite deletions.
1571 *
1572 * NOTE: The DELETING flag is a mod flag, but it is also sticky,
1573 * and stays in ip->flags. Once set, it stays set until the
1574 * inode is destroyed.
1575 *
1576 * NOTE: If a truncation from a previous flush cycle had to be
1577 * continued into this one, the TRUNCATED flag will still be
1578 * set in sync_flags as will WOULDBLOCK. When this occurs
1579 * we CANNOT safely integrate a new truncation from the front-end
1580 * because there may be data records in-memory assigned a flush
1581 * state from the previous cycle that are supposed to be flushed
1582 * before the next frontend truncation.
1583 */
1585 HAMMER_INODE_TRUNCATED) {
1592 /*
1593 * The save_trunc_off used to cache whether the B-Tree
1594 * holds any records past that point is not used until
1595 * after the truncation has succeeded, so we can safely
1596 * set it now.
1597 */
1600 }
1606 #ifdef DEBUG_TRUNCATE
1609 #endif
1611 /*
1612 * The flusher list inherits our inode and reference.
1613 */
1620 }
1621 }
1623 /*
1624 * Callback for scan of ip->rec_tree. Try to include each record in our
1625 * flush. ip->flush_group has been set but the inode has not yet been
1626 * moved into a flushing state.
1627 *
1628 * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1629 * both inodes.
1630 *
1631 * We return 1 for any record placed or found in FST_FLUSH, which prevents
1632 * the caller from shortcutting the flush.
1633 */
1634 static int
1636 {
1637 hammer_inode_t target_ip;
1638 hammer_inode_t ip;
1641 /*
1642 * Deleted records are ignored. Note that the flush detects deleted
1643 * front-end records at multiple points to deal with races. This is
1644 * just the first line of defense. The only time DELETED_FE cannot
1645 * be set is when HAMMER_RECF_INTERLOCK_BE is set.
1646 *
1647 * Don't get confused between record deletion and, say, directory
1648 * entry deletion. The deletion of a directory entry that is on
1649 * the media has nothing to do with the record deletion flags.
1650 *
1651 * The flush_group for a record already in a flush state must
1652 * be updated. This case can only occur if the inode deleting
1653 * too many records had to be moved to the next flush group.
1654 */
1662 }
1664 }
1666 /*
1667 * If the record is in an idle state it has no dependancies and
1668 * can be flushed.
1669 */
1675 /*
1676 * Record has no setup dependancy, we can flush it.
1677 */
1685 /*
1686 * Record has a setup dependancy. Try to include the
1687 * target ip in the flush.
1688 *
1689 * We have to be careful here, if we do not do the right
1690 * thing we can lose track of dirty inodes and the system
1691 * will lockup trying to allocate buffers.
1692 */
1697 /*
1698 * If the target IP is already flushing in our group
1699 * we are golden, otherwise make sure the target
1700 * reflushes.
1701 */
1709 }
1711 /*
1712 * If the target IP is not flushing we can force
1713 * it to flush, even if it is unable to write out
1714 * any of its own records we have at least one in
1715 * hand that we CAN deal with.
1716 */
1721 HAMMER_FLUSH_RECURSION);
1724 /*
1725 * General or delete-on-disk record.
1726 *
1727 * XXX this needs help. If a delete-on-disk we could
1728 * disconnect the target. If the target has its own
1729 * dependancies they really need to be flushed.
1730 *
1731 * XXX
1732 */
1737 HAMMER_FLUSH_RECURSION);
1739 }
1742 /*
1743 * If the WOULDBLOCK flag is set records may have been left
1744 * over from a previous flush attempt and should be moved
1745 * to the current flush group. If it is not set then all
1746 * such records had better have been flushed already or
1747 * already associated with the current flush group.
1748 */
1753 }
1756 }
1758 }
1760 /*
1761 * This version just moves records already in a flush state to the new
1762 * flush group and that is it.
1763 */
1764 static int
1766 {
1775 }
1779 }
1781 }
1783 /*
1784 * Wait for a previously queued flush to complete. Not only do we need to
1785 * wait for the inode to sync out, we also may have to run the flusher again
1786 * to get it past the UNDO position pertaining to the flush so a crash does
1787 * not 'undo' our flush.
1788 */
1789 void
1791 {
1801 }
1802 /* XXX can we make this != FST_IDLE ? check SETUP depends */
1807 }
1811 }
1812 }
1814 /*
1815 * Called by the backend code when a flush has been completed.
1816 * The inode has already been removed from the flush list.
1817 *
1818 * A pipelined flush can occur, in which case we must re-enter the
1819 * inode on the list and re-copy its fields.
1820 */
1821 void
1823 {
1824 hammer_mount_t hmp;
1831 /*
1832 * Merge left-over flags back into the frontend and fix the state.
1833 * Incomplete truncations are retained by the backend.
1834 */
1838 /*
1839 * The backend may have adjusted nlinks, so if the adjusted nlinks
1840 * does not match the fronttend set the frontend's RDIRTY flag again.
1841 */
1845 /*
1846 * Fix up the dirty buffer status.
1847 */
1850 }
1852 /*
1853 * Re-set the XDIRTY flag if some of the inode's in-memory records
1854 * could not be flushed.
1855 */
1861 /*
1862 * Do not lose track of inodes which no longer have vnode
1863 * assocations, otherwise they may never get flushed again.
1864 */
1868 /*
1869 * Clean up the vnode ref
1870 */
1874 }
1876 /*
1877 * Adjust flush_state. The target state (idle or setup) shouldn't
1878 * be terribly important since we will reflush if we really need
1879 * to do anything.
1880 *
1881 * If the WOULDBLOCK flag is set we must re-flush immediately
1882 * to continue a potentially large deletion. The flag also causes
1883 * the hammer_setup_child_callback() to move records in the old
1884 * flush group to the new one.
1885 */
1897 }
1902 /*
1903 * If the frontend made more changes and requested another flush,
1904 * then try to get it running.
1905 */
1913 }
1914 }
1916 /*
1917 * If the inode is now clean drop the space reservation.
1918 */
1923 }
1925 /*
1926 * Finally, if the frontend is waiting for a flush to complete,
1927 * wake it up.
1928 */
1933 }
1934 }
1937 }
1939 /*
1940 * Called from hammer_sync_inode() to synchronize in-memory records
1941 * to the media.
1942 */
1943 static int
1945 {
1951 /*
1952 * Skip records that do not belong to the current flush.
1953 */
1958 #if 1
1960 kprintf("sync_record %p ip %p bad flush group %d %d\n", record, record->ip, record->flush_group ,record->ip->flush_group);
1963 }
1964 #endif
1967 /*
1968 * Interlock the record using the BE flag. Once BE is set the
1969 * frontend cannot change the state of FE.
1970 *
1971 * NOTE: If FE is set prior to us setting BE we still sync the
1972 * record out, but the flush completion code converts it to
1973 * a delete-on-disk record instead of destroying it.
1974 */
1978 /*
1979 * The backend may have already disposed of the record.
1980 */
1984 }
1986 /*
1987 * If the whole inode is being deleting all on-disk records will
1988 * be deleted very soon, we can't sync any new records to disk
1989 * because they will be deleted in the same transaction they were
1990 * created in (delete_tid == create_tid), which will assert.
1991 *
1992 * XXX There may be a case with RECORD_ADD with DELETED_FE set
1993 * that we currently panic on.
1994 */
1998 /*
1999 * We don't have to do anything, if the record was
2000 * committed the space will have been accounted for
2001 * in the blockmap.
2002 */
2003 /* fall through */
2018 /*
2019 * Follow through and issue the on-disk deletion
2020 */
2022 }
2023 }
2025 /*
2026 * If DELETED_FE is set special handling is needed for directory
2027 * entries. Dependant pieces related to the directory entry may
2028 * have already been synced to disk. If this occurs we have to
2029 * sync the directory entry and then change the in-memory record
2030 * from an ADD to a DELETE to cover the fact that it's been
2031 * deleted by the frontend.
2032 *
2033 * A directory delete covering record (MEM_RECORD_DEL) can never
2034 * be deleted by the frontend.
2035 *
2036 * Any other record type (aka DATA) can be deleted by the frontend.
2037 * XXX At the moment the flusher must skip it because there may
2038 * be another data record in the flush group for the same block,
2039 * meaning that some frontend data changes can leak into the backend's
2040 * synchronization point.
2041 */
2050 }
2051 }
2053 /*
2054 * Assign the create_tid for new records. Deletions already
2055 * have the record's entire key properly set up.
2056 */
2069 }
2078 }
2079 }
2080 done:
2083 /*
2084 * Do partial finalization if we have built up too many dirty
2085 * buffers. Otherwise a buffer cache deadlock can occur when
2086 * doing things like creating tens of thousands of tiny files.
2087 *
2088 * The finalization lock is already being held by virtue of the
2089 * flusher calling us.
2090 */
2095 }
2097 /*
2098 * XXX error handling
2099 */
2100 int
2102 {
2105 hammer_node_t tmp_node;
2106 hammer_record_t depend;
2107 hammer_record_t next;
2109 u_int64_t nlinks;
2119 /*
2120 * Any directory records referencing this inode which are not in
2121 * our current flush group must adjust our nlink count for the
2122 * purposes of synchronization to disk.
2123 *
2124 * Records which are in our flush group can be unlinked from our
2125 * inode now, potentially allowing the inode to be physically
2126 * deleted.
2127 *
2128 * This cannot block.
2129 */
2136 /*
2137 * If this is an ADD that was deleted by the frontend
2138 * the frontend nlinks count will have already been
2139 * decremented, but the backend is going to sync its
2140 * directory entry and must account for it. The
2141 * record will be converted to a delete-on-disk when
2142 * it gets synced.
2143 *
2144 * If the ADD was not deleted by the frontend we
2145 * can remove the dependancy from our target_list.
2146 */
2151 target_entry);
2153 }
2155 /*
2156 * Not part of our flush group
2157 */
2168 }
2169 }
2170 }
2172 /*
2173 * Set dirty if we had to modify the link count.
2174 */
2179 }
2181 /*
2182 * If there is a trunction queued destroy any data past the (aligned)
2183 * truncation point. Userland will have dealt with the buffer
2184 * containing the truncation point for us.
2185 *
2186 * We don't flush pending frontend data buffers until after we've
2187 * dealt with the truncation.
2188 */
2190 /*
2191 * Interlock trunc_off. The VOP front-end may continue to
2192 * make adjustments to it while we are blocked.
2193 */
2194 off_t trunc_off;
2195 off_t aligned_trunc_off;
2202 /*
2203 * Delete any whole blocks on-media. The front-end has
2204 * already cleaned out any partial block and made it
2205 * pending. The front-end may have updated trunc_off
2206 * while we were blocked so we only use sync_trunc_off.
2207 *
2208 * This operation can blow out the buffer cache, EWOULDBLOCK
2209 * means we were unable to complete the deletion. The
2210 * deletion will update sync_trunc_off in that case.
2211 */
2213 aligned_trunc_off,
2219 }
2224 /*
2225 * Clear the truncation flag on the backend after we have
2226 * complete the deletions. Backend data is now good again
2227 * (including new records we are about to sync, below).
2228 *
2229 * Leave sync_trunc_off intact. As we write additional
2230 * records the backend will update sync_trunc_off. This
2231 * tells the backend whether it can skip the overwrite
2232 * test. This should work properly even when the backend
2233 * writes full blocks where the truncation point straddles
2234 * the block because the comparison is against the base
2235 * offset of the record.
2236 */
2238 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2241 }
2243 /*
2244 * Now sync related records. These will typically be directory
2245 * entries, records tracking direct-writes, or delete-on-disk records.
2246 */
2254 }
2257 /*
2258 * Re-seek for inode update, assuming our cache hasn't been ripped
2259 * out from under us.
2260 */
2270 }
2272 }
2274 /*
2275 * If we are deleting the inode the frontend had better not have
2276 * any active references on elements making up the inode.
2277 *
2278 * The call to hammer_ip_delete_clean() cleans up auxillary records
2279 * but not DB or DATA records. Those must have already been deleted
2280 * by the normal truncation mechanic.
2281 */
2295 /*
2296 * Set delete_tid in both the frontend and backend
2297 * copy of the inode record. The DELETED flag handles
2298 * this, do not set RDIRTY.
2299 */
2306 /*
2307 * Adjust the inode count in the volume header
2308 */
2313 vol0_stat_inodes);
2316 }
2320 }
2321 }
2328 defer_buffer_flush:
2329 /*
2330 * Now update the inode's on-disk inode-data and/or on-disk record.
2331 * DELETED and ONDISK are managed only in ip->flags.
2332 *
2333 * In the case of a defered buffer flush we still update the on-disk
2334 * inode to satisfy visibility requirements if there happen to be
2335 * directory dependancies.
2336 */
2339 /*
2340 * If deleted and on-disk, don't set any additional flags.
2341 * the delete flag takes care of things.
2342 *
2343 * Clear flags which may have been set by the frontend.
2344 */
2347 HAMMER_INODE_DELETING);
2350 /*
2351 * Take care of the case where a deleted inode was never
2352 * flushed to the disk in the first place.
2353 *
2354 * Clear flags which may have been set by the frontend.
2355 */
2358 HAMMER_INODE_DELETING);
2366 }
2369 /*
2370 * If already on-disk, do not set any additional flags.
2371 */
2374 /*
2375 * If not on-disk and not deleted, set DDIRTY to force
2376 * an initial record to be written.
2377 *
2378 * Also set the create_tid in both the frontend and backend
2379 * copy of the inode record.
2380 */
2387 }
2389 /*
2390 * If RDIRTY or DDIRTY is set, write out a new record. If the inode
2391 * is already on-disk the old record is marked as deleted.
2392 *
2393 * If DELETED is set hammer_update_inode() will delete the existing
2394 * record without writing out a new one.
2395 *
2396 * If *ONLY* the ITIMES flag is set we can update the record in-place.
2397 */
2407 }
2410 done:
2411 /*
2412 * Save the TID we used to sync the inode with to make sure we
2413 * do not improperly reuse it.
2414 */
2418 }
2420 /*
2421 * This routine is called when the OS is no longer actively referencing
2422 * the inode (but might still be keeping it cached), or when releasing
2423 * the last reference to an inode.
2424 *
2425 * At this point if the inode's nlinks count is zero we want to destroy
2426 * it, which may mean destroying it on-media too.
2427 */
2428 void
2430 {
2433 /*
2434 * Set the DELETING flag when the link count drops to 0 and the
2435 * OS no longer has any opens on the inode.
2436 *
2437 * The backend will clear DELETING (a mod flag) and set DELETED
2438 * (a state flag) when it is actually able to perform the
2439 * operation.
2440 */
2450 }
2452 /*
2453 * Final cleanup
2454 */
2458 }
2461 }
2462 }
2463 }
2465 /*
2466 * Re-test an inode when a dependancy had gone away to see if we
2467 * can chain flush it.
2468 */
2469 void
2471 {
2480 }
2482 }
2483 }
2485 /*
2486 * Clear the RECLAIM flag on an inode. This occurs when the inode is
2487 * reassociated with a vp or just before it gets freed.
2488 *
2489 * Wakeup one thread blocked waiting on reclaims to complete. Note that
2490 * the inode the thread is waiting on behalf of is a different inode then
2491 * the inode we are called with. This is to create a pipeline.
2492 */
2493 static void
2495 {
2510 }
2511 }
2513 /*
2514 * Setup our reclaim pipeline. We only let so many detached (and dirty)
2515 * inodes build up before we start blocking.
2516 *
2517 * When we block we don't care *which* inode has finished reclaiming,
2518 * as lone as one does. This is somewhat heuristical... we also put a
2519 * cap on how long we are willing to wait.
2520 */
2521 void
2523 {
2533 }
2537 HAMMER_RECLAIM_WAIT;
2542 }
2543 }