| blob | 8276b331d6fe20a253af44b44e31981e16e08784 |
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 1.92 2008/07/03 04:24:51 dillon Exp $
35 */
38 #include <vm/vm_extern.h>
39 #include <sys/buf.h>
40 #include <sys/buf2.h>
51 #ifdef DEBUG_TRUNCATE
53 #endif
55 /*
56 * RB-Tree support for inode structures
57 */
58 int
60 {
74 }
76 /*
77 * RB-Tree support for inode structures / special LOOKUP_INFO
78 */
79 static int
81 {
95 }
97 /*
98 * Used by hammer_scan_inode_snapshots() to locate all of an object's
99 * snapshots. Note that the asof field is not tested, which we can get
100 * away with because it is the lowest-priority field.
101 */
102 static int
104 {
116 }
118 /*
119 * RB-Tree support for pseudofs structures
120 */
121 static int
123 {
129 }
138 /*
139 * The kernel is not actively referencing this vnode but is still holding
140 * it cached.
141 *
142 * This is called from the frontend.
143 */
144 int
146 {
149 /*
150 * Degenerate case
151 */
155 }
157 /*
158 * If the inode no longer has visibility in the filesystem try to
159 * recycle it immediately, even if the inode is dirty. Recycling
160 * it quickly allows the system to reclaim buffer cache and VM
161 * resources which can matter a lot in a heavily loaded system.
162 *
163 * This can deadlock in vfsync() if we aren't careful.
164 *
165 * Do not queue the inode to the flusher if we still have visibility,
166 * otherwise namespace calls such as chmod will unnecessarily generate
167 * multiple inode updates.
168 */
174 }
176 }
178 /*
179 * Release the vnode association. This is typically (but not always)
180 * the last reference on the inode.
181 *
182 * Once the association is lost we are on our own with regards to
183 * flushing the inode.
184 */
185 int
187 {
189 hammer_mount_t hmp;
206 }
207 }
209 }
211 }
213 /*
214 * Return a locked vnode for the specified inode. The inode must be
215 * referenced but NOT LOCKED on entry and will remain referenced on
216 * return.
217 *
218 * Called from the frontend.
219 */
220 int
222 {
223 hammer_mount_t hmp;
240 }
261 }
263 /*
264 * Only mark as the root vnode if the ip is not
265 * historical, otherwise the VFS cache will get
266 * confused. The other half of the special handling
267 * is in hammer_vop_nlookupdotdot().
268 *
269 * Pseudo-filesystem roots also do not count.
270 */
275 }
278 /* vnode locked by getnewvnode() */
279 /* make related vnode dirty if inode dirty? */
284 }
286 /*
287 * loop if the vget fails (aka races), or if the vp
288 * no longer matches ip->vp.
289 */
294 }
295 }
298 }
300 /*
301 * Locate all copies of the inode for obj_id compatible with the specified
302 * asof, reference, and issue the related call-back. This routine is used
303 * for direct-io invalidation and does not create any new inodes.
304 */
305 void
309 {
311 hammer_inode_info_cmp_all_history,
313 }
315 /*
316 * Acquire a HAMMER inode. The returned inode is not locked. These functions
317 * do not attach or detach the related vnode (use hammer_get_vnode() for
318 * that).
319 *
320 * The flags argument is only applied for newly created inodes, and only
321 * certain flags are inherited.
322 *
323 * Called from the frontend.
324 */
329 {
336 /*
337 * Determine if we already have an inode cached. If we do then
338 * we are golden.
339 */
343 loop:
349 }
351 /*
352 * Allocate a new inode structure and deal with races later.
353 */
372 /*
373 * Locate the on-disk inode.
374 */
375 retry:
386 HAMMER_CURSOR_ASOF;
392 }
394 /*
395 * On success the B-Tree lookup will hold the appropriate
396 * buffer cache buffers and provide a pointer to the requested
397 * information. Copy the information to the in-memory inode
398 * and cache the B-Tree node to improve future operations.
399 */
404 /*
405 * cache[0] tries to cache the location of the object inode.
406 * The assumption is that it is near the directory inode.
407 *
408 * cache[1] tries to cache the location of the object data.
409 * The assumption is that it is near the directory data.
410 */
415 /*
416 * The file should not contain any data past the file size
417 * stored in the inode. Setting save_trunc_off to the
418 * file size instead of max reduces B-Tree lookup overheads
419 * on append by allowing the flusher to avoid checking for
420 * record overwrites.
421 */
424 /*
425 * Locate and assign the pseudofs management structure to
426 * the inode.
427 */
433 }
434 }
436 /*
437 * The inode is placed on the red-black tree and will be synced to
438 * the media when flushed or by the filesystem sync. If this races
439 * another instantiation/lookup the insertion will fail.
440 */
446 }
452 }
456 }
459 }
461 /*
462 * Create a new filesystem object, returning the inode in *ipp. The
463 * returned inode will be referenced.
464 *
465 * The inode is created in-memory.
466 */
467 int
471 {
472 hammer_mount_t hmp;
473 hammer_inode_t ip;
474 uid_t xuid;
475 u_int32_t localization;
480 /*
481 * Assign the localization domain. If if dip is NULL we are creating
482 * a pseudo-fs and must locate an unused localization domain.
483 */
495 }
498 }
501 }
507 /*
508 * Allocate a new object id. If creating a new pseudo-fs the
509 * obj_id is 1.
510 */
513 else
527 /* ip->save_trunc_off = 0; (already zero) */
536 /*
537 * A nohistory designator on the parent directory is inherited by
538 * the child. We will do this even for pseudo-fs creation... the
539 * sysad can turn it off.
540 */
546 HAMMER_LOCALIZE_INODE;
559 /*
560 * Setup the ".." pointer. This only needs to be done for directories
561 * but we do it for all objects as a recovery aid.
562 *
563 * The parent_obj_localization field only applies to pseudo-fs roots.
564 */
570 }
580 }
582 /*
583 * Calculate default uid/gid and overwrite with information from
584 * the vap.
585 */
595 else
602 else
613 }
620 /* not reached */
622 }
625 }
627 /*
628 * Final cleanup / freeing of an inode structure
629 */
630 static void
632 {
644 }
647 }
649 /*
650 * Retrieve pseudo-fs data.
651 */
652 int
654 {
656 hammer_pseudofs_inmem_t pfsm;
661 retry:
669 }
676 HAMMER_LOCALIZE_MISC;
694 }
697 }
706 }
709 /*
710 * Certain aspects of the pseudofs configuration are reflected
711 * in the inode.
712 */
720 }
723 }
725 }
727 /*
728 * Store pseudo-fs data. The backend will automatically delete any prior
729 * on-disk pseudo-fs data but we have to delete in-memory versions.
730 */
731 int
733 {
735 hammer_pseudofs_inmem_t pfsm;
736 hammer_record_t record;
739 retry:
743 HAMMER_LOCALIZE_MISC;
764 }
765 }
771 HAMMER_LOCALIZE_MISC;
777 }
782 /*
783 * Certain aspects of the pseudofs configuration are reflected
784 * in the inode. Note that we cannot mess with the as-of or
785 * clear the read-only state.
786 *
787 * If this inode represented a slave snapshot its asof will
788 * be set to a snapshot tid. When clearing slave mode any
789 * re-access of the inode via the parent directory will
790 * wind up using a different asof and thus will instantiate
791 * a new inode.
792 */
800 }
801 }
803 }
805 void
807 {
812 }
813 }
815 /*
816 * Called by hammer_sync_inode().
817 */
818 static int
820 {
822 hammer_record_t record;
826 retry:
829 /*
830 * If the inode has a presence on-disk then locate it and mark
831 * it deleted, setting DELONDISK.
832 *
833 * The record may or may not be physically deleted, depending on
834 * the retention policy.
835 */
837 HAMMER_INODE_ONDISK) {
840 HAMMER_LOCALIZE_INODE;
858 }
867 }
870 }
873 }
882 }
883 }
885 /*
886 * Ok, write out the initial record or a new record (after deleting
887 * the old one), unless the DELETED flag is set. This routine will
888 * clear DELONDISK if it writes out a record.
889 *
890 * Update our inode statistics if this is the first application of
891 * the inode on-disk.
892 */
894 /*
895 * Generate a record and write it to the media
896 */
907 /*
908 * If this flag is set we cannot sync the new file size
909 * because we haven't finished related truncations. The
910 * inode will be flushed in another flush group to finish
911 * the job.
912 */
919 }
935 }
939 }
941 /*
942 * The record isn't managed by the inode's record tree,
943 * destroy it whether we succeed or fail.
944 */
950 /*
951 * Finish up.
952 */
957 HAMMER_INODE_ATIME |
958 HAMMER_INODE_MTIME);
963 /*
964 * Root volume count of inodes
965 */
969 vol0_stat_inodes);
975 }
976 }
977 }
979 /*
980 * If the inode has been destroyed, clean out any left-over flags
981 * that may have been set by the frontend.
982 */
985 HAMMER_INODE_ATIME |
986 HAMMER_INODE_MTIME);
987 }
989 }
991 /*
992 * Update only the itimes fields.
993 *
994 * ATIME can be updated without generating any UNDO. MTIME is updated
995 * with UNDO so it is guaranteed to be synchronized properly in case of
996 * a crash.
997 *
998 * Neither field is included in the B-Tree leaf element's CRC, which is how
999 * we can get away with updating ATIME the way we do.
1000 */
1001 static int
1003 {
1007 retry:
1009 HAMMER_INODE_ONDISK) {
1011 }
1015 HAMMER_LOCALIZE_INODE;
1033 }
1037 /*
1038 * Updating MTIME requires an UNDO. Just cover
1039 * both atime and mtime.
1040 */
1043 HAMMER_ITIMES_BYTES);
1048 /*
1049 * Updating atime only can be done in-place with
1050 * no UNDO.
1051 */
1056 }
1058 }
1065 }
1067 }
1069 /*
1070 * Release a reference on an inode, flush as requested.
1071 *
1072 * On the last reference we queue the inode to the flusher for its final
1073 * disposition.
1074 */
1075 void
1077 {
1080 /*
1081 * Handle disposition when dropping the last ref.
1082 */
1085 /*
1086 * Determine whether on-disk action is needed for
1087 * the inode's final disposition.
1088 */
1094 HAMMER_FLUSH_SIGNAL);
1097 }
1101 }
1106 /*
1107 * The inode still has multiple refs, try to drop
1108 * one ref.
1109 */
1114 }
1115 }
1116 }
1117 }
1119 /*
1120 * Unload and destroy the specified inode. Must be called with one remaining
1121 * reference. The reference is disposed of.
1122 *
1123 * This can only be called in the context of the flusher.
1124 */
1125 static int
1127 {
1145 }
1147 /*
1148 * Called on mount -u when switching from RW to RO or vise-versa. Adjust
1149 * the read-only flag for cached inodes.
1150 *
1151 * This routine is called from a RB_SCAN().
1152 */
1153 int
1155 {
1160 else
1163 }
1165 /*
1166 * A transaction has modified an inode, requiring updates as specified by
1167 * the passed flags.
1168 *
1169 * HAMMER_INODE_DDIRTY: Inode data has been updated
1170 * HAMMER_INODE_XDIRTY: Dirty in-memory records
1171 * HAMMER_INODE_BUFS: Dirty buffer cache buffers
1172 * HAMMER_INODE_DELETED: Inode record/data must be deleted
1173 * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1174 */
1175 void
1177 {
1185 }
1188 }
1190 /*
1191 * Request that an inode be flushed. This whole mess cannot block and may
1192 * recurse (if not synchronous). Once requested HAMMER will attempt to
1193 * actively flush the inode until the flush can be done.
1194 *
1195 * The inode may already be flushing, or may be in a setup state. We can
1196 * place the inode in a flushing state if it is currently idle and flag it
1197 * to reflush if it is currently flushing.
1198 *
1199 * If the HAMMER_FLUSH_SYNCHRONOUS flag is specified we will attempt to
1200 * flush the indoe synchronously using the caller's context.
1201 */
1202 void
1204 {
1207 /*
1208 * Trivial 'nothing to flush' case. If the inode is ina SETUP
1209 * state we have to put it back into an IDLE state so we can
1210 * drop the extra ref.
1211 */
1216 }
1218 }
1220 /*
1221 * Our flush action will depend on the current state.
1222 */
1225 /*
1226 * We have no dependancies and can flush immediately. Some
1227 * our children may not be flushable so we have to re-test
1228 * with that additional knowledge.
1229 */
1233 /*
1234 * Recurse upwards through dependancies via target_list
1235 * and start their flusher actions going if possible.
1236 *
1237 * 'good' is our connectivity. -1 means we have none and
1238 * can't flush, 0 means there weren't any dependancies, and
1239 * 1 means we have good connectivity.
1240 */
1243 /*
1244 * We can continue if good >= 0. Determine how many records
1245 * under our inode can be flushed (and mark them).
1246 */
1254 }
1255 }
1258 /*
1259 * We are already flushing, flag the inode to reflush
1260 * if needed after it completes its current flush.
1261 */
1267 }
1269 }
1270 }
1272 /*
1273 * Scan ip->target_list, which is a list of records owned by PARENTS to our
1274 * ip which reference our ip.
1275 *
1276 * XXX This is a huge mess of recursive code, but not one bit of it blocks
1277 * so for now do not ref/deref the structures. Note that if we use the
1278 * ref/rel code later, the rel CAN block.
1279 */
1280 static int
1282 {
1283 hammer_record_t depend;
1284 #if 0
1285 hammer_record_t next;
1286 hammer_inode_t pip;
1287 #endif
1299 }
1302 #if 0
1303 retry:
1309 }
1316 }
1322 }
1331 }
1333 #endif
1334 }
1336 /*
1337 * This helper function takes a record representing the dependancy between
1338 * the parent inode and child inode.
1339 *
1340 * record->ip = parent inode
1341 * record->target_ip = child inode
1342 *
1343 * We are asked to recurse upwards and convert the record from SETUP
1344 * to FLUSH if possible.
1345 *
1346 * Return 1 if the record gives us connectivity
1347 *
1348 * Return 0 if the record is not relevant
1349 *
1350 * Return -1 if we can't resolve the dependancy and there is no connectivity.
1351 */
1352 static int
1354 {
1355 hammer_mount_t hmp;
1356 hammer_inode_t pip;
1363 /*
1364 * If the record is already flushing, is it in our flush group?
1365 *
1366 * If it is in our flush group but it is a general record or a
1367 * delete-on-disk, it does not improve our connectivity (return 0),
1368 * and if the target inode is not trying to destroy itself we can't
1369 * allow the operation yet anyway (the second return -1).
1370 */
1375 }
1378 /* GENERAL or DEL */
1380 }
1382 /*
1383 * It must be a setup record. Try to resolve the setup dependancies
1384 * by recursing upwards so we can place ip on the flush list.
1385 */
1390 /*
1391 * We can't flush ip because it has no connectivity (XXX also check
1392 * nlinks for pre-existing connectivity!). Flag it so any resolution
1393 * recurses back down.
1394 */
1398 }
1400 /*
1401 * We are go, place the parent inode in a flushing state so we can
1402 * place its record in a flushing state. Note that the parent
1403 * may already be flushing. The record must be in the same flush
1404 * group as the parent.
1405 */
1411 #if 0
1414 /*
1415 * Regardless of flushing state we cannot sync this path if the
1416 * record represents a delete-on-disk but the target inode
1417 * is not ready to sync its own deletion.
1418 *
1419 * XXX need to count effective nlinks to determine whether
1420 * the flush is ok, otherwise removing a hardlink will
1421 * just leave the DEL record to rot.
1422 */
1426 #endif
1428 /*
1429 * This is the record we wanted to synchronize. If the
1430 * record went into a flush state while we blocked it
1431 * had better be in the correct flush group.
1432 */
1439 }
1443 /*
1444 * A general or delete-on-disk record does not contribute
1445 * to our visibility. We can still flush it, however.
1446 */
1449 /*
1450 * We couldn't resolve the dependancies, request that the
1451 * inode be flushed when the dependancies can be resolved.
1452 */
1455 }
1456 }
1458 /*
1459 * This is the core routine placing an inode into the FST_FLUSH state.
1460 */
1461 static void
1463 {
1466 /*
1467 * Set flush state and prevent the flusher from cycling into
1468 * the next flush group. Do not place the ip on the list yet.
1469 * Inodes not in the idle state get an extra reference.
1470 */
1480 /*
1481 * We need to be able to vfsync/truncate from the backend.
1482 */
1487 }
1489 /*
1490 * Figure out how many in-memory records we can actually flush
1491 * (not including inode meta-data, buffers, etc).
1492 *
1493 * Do not add new records to the flush if this is a recursion or
1494 * if we must still complete a flush from the previous flush cycle.
1495 */
1505 }
1507 /*
1508 * This is a more involved test that includes go_count. If we
1509 * can't flush, flag the inode and return. If go_count is 0 we
1510 * were are unable to flush any records in our rec_tree and
1511 * must ignore the XDIRTY flag.
1512 */
1524 }
1528 }
1532 }
1533 }
1535 /*
1536 * Snapshot the state of the inode for the backend flusher.
1537 *
1538 * We continue to retain save_trunc_off even when all truncations
1539 * have been resolved as an optimization to determine if we can
1540 * skip the B-Tree lookup for overwrite deletions.
1541 *
1542 * NOTE: The DELETING flag is a mod flag, but it is also sticky,
1543 * and stays in ip->flags. Once set, it stays set until the
1544 * inode is destroyed.
1545 *
1546 * NOTE: If a truncation from a previous flush cycle had to be
1547 * continued into this one, the TRUNCATED flag will still be
1548 * set in sync_flags as will WOULDBLOCK. When this occurs
1549 * we CANNOT safely integrate a new truncation from the front-end
1550 * because there may be data records in-memory assigned a flush
1551 * state from the previous cycle that are supposed to be flushed
1552 * before the next frontend truncation.
1553 */
1555 HAMMER_INODE_TRUNCATED) {
1562 /*
1563 * The save_trunc_off used to cache whether the B-Tree
1564 * holds any records past that point is not used until
1565 * after the truncation has succeeded, so we can safely
1566 * set it now.
1567 */
1570 }
1576 #ifdef DEBUG_TRUNCATE
1579 #endif
1581 /*
1582 * The flusher list inherits our inode and reference.
1583 */
1590 }
1591 }
1593 /*
1594 * Callback for scan of ip->rec_tree. Try to include each record in our
1595 * flush. ip->flush_group has been set but the inode has not yet been
1596 * moved into a flushing state.
1597 *
1598 * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
1599 * both inodes.
1600 *
1601 * We return 1 for any record placed or found in FST_FLUSH, which prevents
1602 * the caller from shortcutting the flush.
1603 */
1604 static int
1606 {
1607 hammer_inode_t target_ip;
1608 hammer_inode_t ip;
1611 /*
1612 * Deleted records are ignored. Note that the flush detects deleted
1613 * front-end records at multiple points to deal with races. This is
1614 * just the first line of defense. The only time DELETED_FE cannot
1615 * be set is when HAMMER_RECF_INTERLOCK_BE is set.
1616 *
1617 * Don't get confused between record deletion and, say, directory
1618 * entry deletion. The deletion of a directory entry that is on
1619 * the media has nothing to do with the record deletion flags.
1620 *
1621 * The flush_group for a record already in a flush state must
1622 * be updated. This case can only occur if the inode deleting
1623 * too many records had to be moved to the next flush group.
1624 */
1632 }
1634 }
1636 /*
1637 * If the record is in an idle state it has no dependancies and
1638 * can be flushed.
1639 */
1645 /*
1646 * Record has no setup dependancy, we can flush it.
1647 */
1655 /*
1656 * Record has a setup dependancy. Try to include the
1657 * target ip in the flush.
1658 *
1659 * We have to be careful here, if we do not do the right
1660 * thing we can lose track of dirty inodes and the system
1661 * will lockup trying to allocate buffers.
1662 */
1667 /*
1668 * If the target IP is already flushing in our group
1669 * we are golden, otherwise make sure the target
1670 * reflushes.
1671 */
1679 }
1681 /*
1682 * If the target IP is not flushing we can force
1683 * it to flush, even if it is unable to write out
1684 * any of its own records we have at least one in
1685 * hand that we CAN deal with.
1686 */
1691 HAMMER_FLUSH_RECURSION);
1694 /*
1695 * General or delete-on-disk record.
1696 *
1697 * XXX this needs help. If a delete-on-disk we could
1698 * disconnect the target. If the target has its own
1699 * dependancies they really need to be flushed.
1700 *
1701 * XXX
1702 */
1707 HAMMER_FLUSH_RECURSION);
1709 }
1712 /*
1713 * If the WOULDBLOCK flag is set records may have been left
1714 * over from a previous flush attempt and should be moved
1715 * to the current flush group. If it is not set then all
1716 * such records had better have been flushed already or
1717 * already associated with the current flush group.
1718 */
1723 }
1726 }
1728 }
1730 /*
1731 * This version just moves records already in a flush state to the new
1732 * flush group and that is it.
1733 */
1734 static int
1736 {
1745 }
1749 }
1751 }
1753 /*
1754 * Wait for a previously queued flush to complete. Not only do we need to
1755 * wait for the inode to sync out, we also may have to run the flusher again
1756 * to get it past the UNDO position pertaining to the flush so a crash does
1757 * not 'undo' our flush.
1758 */
1759 void
1761 {
1771 }
1772 /* XXX can we make this != FST_IDLE ? check SETUP depends */
1777 }
1781 }
1782 }
1784 /*
1785 * Called by the backend code when a flush has been completed.
1786 * The inode has already been removed from the flush list.
1787 *
1788 * A pipelined flush can occur, in which case we must re-enter the
1789 * inode on the list and re-copy its fields.
1790 */
1791 void
1793 {
1794 hammer_mount_t hmp;
1801 /*
1802 * Merge left-over flags back into the frontend and fix the state.
1803 * Incomplete truncations are retained by the backend.
1804 */
1808 /*
1809 * The backend may have adjusted nlinks, so if the adjusted nlinks
1810 * does not match the fronttend set the frontend's RDIRTY flag again.
1811 */
1815 /*
1816 * Fix up the dirty buffer status.
1817 */
1820 }
1822 /*
1823 * Re-set the XDIRTY flag if some of the inode's in-memory records
1824 * could not be flushed.
1825 */
1831 /*
1832 * Do not lose track of inodes which no longer have vnode
1833 * assocations, otherwise they may never get flushed again.
1834 */
1838 /*
1839 * Clean up the vnode ref
1840 */
1844 }
1846 /*
1847 * Adjust flush_state. The target state (idle or setup) shouldn't
1848 * be terribly important since we will reflush if we really need
1849 * to do anything.
1850 *
1851 * If the WOULDBLOCK flag is set we must re-flush immediately
1852 * to continue a potentially large deletion. The flag also causes
1853 * the hammer_setup_child_callback() to move records in the old
1854 * flush group to the new one.
1855 */
1868 }
1873 /*
1874 * If the frontend made more changes and requested another flush,
1875 * then try to get it running.
1876 */
1884 }
1885 }
1887 /*
1888 * If the inode is now clean drop the space reservation.
1889 */
1894 }
1896 /*
1897 * Finally, if the frontend is waiting for a flush to complete,
1898 * wake it up.
1899 */
1904 }
1905 }
1908 }
1910 /*
1911 * Called from hammer_sync_inode() to synchronize in-memory records
1912 * to the media.
1913 */
1914 static int
1916 {
1921 /*
1922 * Skip records that do not belong to the current flush.
1923 */
1928 #if 1
1930 kprintf("sync_record %p ip %p bad flush group %d %d\n", record, record->ip, record->flush_group ,record->ip->flush_group);
1933 }
1934 #endif
1937 /*
1938 * Interlock the record using the BE flag. Once BE is set the
1939 * frontend cannot change the state of FE.
1940 *
1941 * NOTE: If FE is set prior to us setting BE we still sync the
1942 * record out, but the flush completion code converts it to
1943 * a delete-on-disk record instead of destroying it.
1944 */
1948 /*
1949 * The backend may have already disposed of the record.
1950 */
1954 }
1956 /*
1957 * If the whole inode is being deleting all on-disk records will
1958 * be deleted very soon, we can't sync any new records to disk
1959 * because they will be deleted in the same transaction they were
1960 * created in (delete_tid == create_tid), which will assert.
1961 *
1962 * XXX There may be a case with RECORD_ADD with DELETED_FE set
1963 * that we currently panic on.
1964 */
1968 /*
1969 * We don't have to do anything, if the record was
1970 * committed the space will have been accounted for
1971 * in the blockmap.
1972 */
1973 /* fall through */
1988 /*
1989 * Follow through and issue the on-disk deletion
1990 */
1992 }
1993 }
1995 /*
1996 * If DELETED_FE is set special handling is needed for directory
1997 * entries. Dependant pieces related to the directory entry may
1998 * have already been synced to disk. If this occurs we have to
1999 * sync the directory entry and then change the in-memory record
2000 * from an ADD to a DELETE to cover the fact that it's been
2001 * deleted by the frontend.
2002 *
2003 * A directory delete covering record (MEM_RECORD_DEL) can never
2004 * be deleted by the frontend.
2005 *
2006 * Any other record type (aka DATA) can be deleted by the frontend.
2007 * XXX At the moment the flusher must skip it because there may
2008 * be another data record in the flush group for the same block,
2009 * meaning that some frontend data changes can leak into the backend's
2010 * synchronization point.
2011 */
2020 }
2021 }
2023 /*
2024 * Assign the create_tid for new records. Deletions already
2025 * have the record's entire key properly set up.
2026 */
2039 }
2048 }
2049 }
2050 done:
2053 }
2055 /*
2056 * XXX error handling
2057 */
2058 int
2060 {
2063 hammer_node_t tmp_node;
2064 hammer_record_t depend;
2065 hammer_record_t next;
2067 u_int64_t nlinks;
2077 /*
2078 * Any directory records referencing this inode which are not in
2079 * our current flush group must adjust our nlink count for the
2080 * purposes of synchronization to disk.
2081 *
2082 * Records which are in our flush group can be unlinked from our
2083 * inode now, potentially allowing the inode to be physically
2084 * deleted.
2085 *
2086 * This cannot block.
2087 */
2094 /*
2095 * If this is an ADD that was deleted by the frontend
2096 * the frontend nlinks count will have already been
2097 * decremented, but the backend is going to sync its
2098 * directory entry and must account for it. The
2099 * record will be converted to a delete-on-disk when
2100 * it gets synced.
2101 *
2102 * If the ADD was not deleted by the frontend we
2103 * can remove the dependancy from our target_list.
2104 */
2109 target_entry);
2111 }
2113 /*
2114 * Not part of our flush group
2115 */
2126 }
2127 }
2128 }
2130 /*
2131 * Set dirty if we had to modify the link count.
2132 */
2137 }
2139 /*
2140 * If there is a trunction queued destroy any data past the (aligned)
2141 * truncation point. Userland will have dealt with the buffer
2142 * containing the truncation point for us.
2143 *
2144 * We don't flush pending frontend data buffers until after we've
2145 * dealt with the truncation.
2146 */
2148 /*
2149 * Interlock trunc_off. The VOP front-end may continue to
2150 * make adjustments to it while we are blocked.
2151 */
2152 off_t trunc_off;
2153 off_t aligned_trunc_off;
2160 /*
2161 * Delete any whole blocks on-media. The front-end has
2162 * already cleaned out any partial block and made it
2163 * pending. The front-end may have updated trunc_off
2164 * while we were blocked so we only use sync_trunc_off.
2165 *
2166 * This operation can blow out the buffer cache, EWOULDBLOCK
2167 * means we were unable to complete the deletion. The
2168 * deletion will update sync_trunc_off in that case.
2169 */
2171 aligned_trunc_off,
2177 }
2182 /*
2183 * Clear the truncation flag on the backend after we have
2184 * complete the deletions. Backend data is now good again
2185 * (including new records we are about to sync, below).
2186 *
2187 * Leave sync_trunc_off intact. As we write additional
2188 * records the backend will update sync_trunc_off. This
2189 * tells the backend whether it can skip the overwrite
2190 * test. This should work properly even when the backend
2191 * writes full blocks where the truncation point straddles
2192 * the block because the comparison is against the base
2193 * offset of the record.
2194 */
2196 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
2199 }
2201 /*
2202 * Now sync related records. These will typically be directory
2203 * entries or delete-on-disk records.
2204 *
2205 * Not all records will be flushed, but clear XDIRTY anyway. We
2206 * will set it again in the frontend hammer_flush_inode_done()
2207 * if records remain.
2208 */
2216 }
2219 /*
2220 * Re-seek for inode update, assuming our cache hasn't been ripped
2221 * out from under us.
2222 */
2232 }
2234 }
2236 /*
2237 * If we are deleting the inode the frontend had better not have
2238 * any active references on elements making up the inode.
2239 *
2240 * The call to hammer_ip_delete_clean() cleans up auxillary records
2241 * but not DB or DATA records. Those must have already been deleted
2242 * by the normal truncation mechanic.
2243 */
2257 /*
2258 * Set delete_tid in both the frontend and backend
2259 * copy of the inode record. The DELETED flag handles
2260 * this, do not set RDIRTY.
2261 */
2268 /*
2269 * Adjust the inode count in the volume header
2270 */
2274 vol0_stat_inodes);
2277 }
2280 }
2281 }
2288 defer_buffer_flush:
2289 /*
2290 * Now update the inode's on-disk inode-data and/or on-disk record.
2291 * DELETED and ONDISK are managed only in ip->flags.
2292 *
2293 * In the case of a defered buffer flush we still update the on-disk
2294 * inode to satisfy visibility requirements if there happen to be
2295 * directory dependancies.
2296 */
2299 /*
2300 * If deleted and on-disk, don't set any additional flags.
2301 * the delete flag takes care of things.
2302 *
2303 * Clear flags which may have been set by the frontend.
2304 */
2307 HAMMER_INODE_DELETING);
2310 /*
2311 * Take care of the case where a deleted inode was never
2312 * flushed to the disk in the first place.
2313 *
2314 * Clear flags which may have been set by the frontend.
2315 */
2318 HAMMER_INODE_DELETING);
2326 }
2329 /*
2330 * If already on-disk, do not set any additional flags.
2331 */
2334 /*
2335 * If not on-disk and not deleted, set DDIRTY to force
2336 * an initial record to be written.
2337 *
2338 * Also set the create_tid in both the frontend and backend
2339 * copy of the inode record.
2340 */
2347 }
2349 /*
2350 * If RDIRTY or DDIRTY is set, write out a new record. If the inode
2351 * is already on-disk the old record is marked as deleted.
2352 *
2353 * If DELETED is set hammer_update_inode() will delete the existing
2354 * record without writing out a new one.
2355 *
2356 * If *ONLY* the ITIMES flag is set we can update the record in-place.
2357 */
2367 }
2370 done:
2371 /*
2372 * Save the TID we used to sync the inode with to make sure we
2373 * do not improperly reuse it.
2374 */
2378 }
2380 /*
2381 * This routine is called when the OS is no longer actively referencing
2382 * the inode (but might still be keeping it cached), or when releasing
2383 * the last reference to an inode.
2384 *
2385 * At this point if the inode's nlinks count is zero we want to destroy
2386 * it, which may mean destroying it on-media too.
2387 */
2388 void
2390 {
2393 /*
2394 * Set the DELETING flag when the link count drops to 0 and the
2395 * OS no longer has any opens on the inode.
2396 *
2397 * The backend will clear DELETING (a mod flag) and set DELETED
2398 * (a state flag) when it is actually able to perform the
2399 * operation.
2400 */
2410 }
2412 /*
2413 * Final cleanup
2414 */
2418 }
2421 }
2422 }
2423 }
2425 /*
2426 * Re-test an inode when a dependancy had gone away to see if we
2427 * can chain flush it.
2428 */
2429 void
2431 {
2440 }
2442 }
2443 }
2445 /*
2446 * Clear the RECLAIM flag on an inode. This occurs when the inode is
2447 * reassociated with a vp or just before it gets freed.
2448 *
2449 * Wakeup one thread blocked waiting on reclaims to complete. Note that
2450 * the inode the thread is waiting on behalf of is a different inode then
2451 * the inode we are called with. This is to create a pipeline.
2452 */
2453 static void
2455 {
2470 }
2471 }
2473 /*
2474 * Setup our reclaim pipeline. We only let so many detached (and dirty)
2475 * inodes build up before we start blocking.
2476 *
2477 * When we block we don't care *which* inode has finished reclaiming,
2478 * as lone as one does. This is somewhat heuristical... we also put a
2479 * cap on how long we are willing to wait.
2480 */
2481 void
2483 {
2493 }
2497 HAMMER_RECLAIM_WAIT;
2502 }
2503 }