| blob | ed553c60de827e0ebad24e3501e0e00d21c82cfc |
1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * alloc.c
5 *
6 * Extent allocs and frees
7 *
8 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 */
26 #include <linux/fs.h>
27 #include <linux/types.h>
28 #include <linux/slab.h>
29 #include <linux/highmem.h>
30 #include <linux/swap.h>
31 #include <linux/quotaops.h>
32 #include <linux/blkdev.h>
34 #include <cluster/masklog.h>
59 CONTIG_LEFT,
60 CONTIG_RIGHT,
61 CONTIG_LEFTRIGHT,
62 };
64 static enum ocfs2_contig_type
68 /*
69 * Operations for a specific extent tree type.
70 *
71 * To implement an on-disk btree (extent tree) type in ocfs2, add
72 * an ocfs2_extent_tree_operations structure and the matching
73 * ocfs2_init_<thingy>_extent_tree() function. That's pretty much it
74 * for the allocation portion of the extent tree.
75 */
77 /*
78 * last_eb_blk is the block number of the right most leaf extent
79 * block. Most on-disk structures containing an extent tree store
80 * this value for fast access. The ->eo_set_last_eb_blk() and
81 * ->eo_get_last_eb_blk() operations access this value. They are
82 * both required.
83 */
85 u64 blkno);
88 /*
89 * The on-disk structure usually keeps track of how many total
90 * clusters are stored in this extent tree. This function updates
91 * that value. new_clusters is the delta, and must be
92 * added to the total. Required.
93 */
95 u32 new_clusters);
97 /*
98 * If this extent tree is supported by an extent map, insert
99 * a record into the map.
100 */
104 /*
105 * If this extent tree is supported by an extent map, truncate the
106 * map to clusters,
107 */
109 u32 clusters);
111 /*
112 * If ->eo_insert_check() exists, it is called before rec is
113 * inserted into the extent tree. It is optional.
114 */
119 /*
120 * --------------------------------------------------------------
121 * The remaining are internal to ocfs2_extent_tree and don't have
122 * accessor functions
123 */
125 /*
126 * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el.
127 * It is required.
128 */
131 /*
132 * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if
133 * it exists. If it does not, et->et_max_leaf_clusters is set
134 * to 0 (unlimited). Optional.
135 */
138 /*
139 * ->eo_extent_contig test whether the 2 ocfs2_extent_rec
140 * are contiguous or not. Optional. Don't need to set it if use
141 * ocfs2_extent_rec as the tree leaf.
142 */
143 enum ocfs2_contig_type
147 };
150 /*
151 * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check
152 * in the methods.
153 */
156 u64 blkno);
158 u32 clusters);
162 u32 clusters);
176 };
179 u64 blkno)
180 {
185 }
188 {
193 }
196 u32 clusters)
197 {
205 }
209 {
213 }
216 u32 clusters)
217 {
221 }
225 {
232 "Device %s, asking for sparse allocation: inode %llu, "
239 }
242 {
249 }
252 {
256 }
260 {
264 }
267 u64 blkno)
268 {
272 }
275 {
279 }
282 u32 clusters)
283 {
287 }
294 };
297 {
301 }
304 {
308 }
311 u64 blkno)
312 {
317 }
320 {
325 }
328 u32 clusters)
329 {
333 }
341 };
344 u64 blkno)
345 {
349 }
352 {
356 }
359 u32 clusters)
360 {
364 }
367 {
373 }
376 {
380 }
388 };
391 {
395 }
398 u64 blkno)
399 {
403 }
406 {
410 }
413 u32 clusters)
414 {
418 }
420 static enum ocfs2_contig_type
424 {
426 }
434 };
439 ocfs2_journal_access_func access,
442 {
454 else
456 }
461 {
464 }
469 {
472 }
477 {
480 }
485 {
488 }
493 {
496 }
499 u64 new_last_eb_blk)
500 {
502 }
505 {
507 }
510 u32 clusters)
511 {
513 }
517 {
520 }
523 u32 clusters)
524 {
527 }
532 {
534 type);
535 }
541 {
548 }
552 {
558 }
561 {
567 }
575 /*
576 * Reset the actual path elements so that we can re-use the structure
577 * to build another path. Generally, this involves freeing the buffer
578 * heads.
579 */
581 {
594 }
596 /*
597 * Tree depth may change during truncate, or insert. If we're
598 * keeping the root extent list, then make sure that our path
599 * structure reflects the proper depth.
600 */
603 else
607 }
610 {
614 }
615 }
617 /*
618 * All the elements of src into dest. After this call, src could be freed
619 * without affecting dest.
620 *
621 * Both paths should have the same root. Any non-root elements of dest
622 * will be freed.
623 */
625 {
640 }
641 }
643 /*
644 * Make the *dest path the same as src and re-initialize src path to
645 * have a root only.
646 */
648 {
662 }
663 }
665 /*
666 * Insert an extent block at given index.
667 *
668 * This will not take an additional reference on eb_bh.
669 */
672 {
675 /*
676 * Right now, no root bh is an extent block, so this helps
677 * catch code errors with dinode trees. The assertion can be
678 * safely removed if we ever need to insert extent block
679 * structures at the root.
680 */
685 }
689 ocfs2_journal_access_func access)
690 {
702 }
705 }
708 {
711 }
714 {
717 }
719 /*
720 * Journal the buffer at depth idx. All idx>0 are extent_blocks,
721 * otherwise it's the root_access function.
722 *
723 * I don't like the way this function's name looks next to
724 * ocfs2_journal_access_path(), but I don't have a better one.
725 */
730 {
740 OCFS2_JOURNAL_ACCESS_WRITE);
741 }
743 /*
744 * Convenience function to journal all components in a path.
745 */
749 {
760 }
761 }
763 out:
765 }
767 /*
768 * Return the index of the extent record which contains cluster #v_cluster.
769 * -1 is returned if it was not found.
770 *
771 * Should work fine on interior and exterior nodes.
772 */
774 {
791 }
792 }
795 }
797 /*
798 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
799 * ocfs2_extent_rec_contig only work properly against leaf nodes!
800 */
803 u64 blkno)
804 {
811 }
815 {
816 u32 left_range;
822 }
824 static enum ocfs2_contig_type
828 {
831 /*
832 * Refuse to coalesce extent records with different flag
833 * fields - we don't want to mix unwritten extents with user
834 * data.
835 */
849 }
851 /*
852 * NOTE: We can have pretty much any combination of contiguousness and
853 * appending.
854 *
855 * The usefulness of APPEND_TAIL is more in that it lets us know that
856 * we'll have to update the path to that leaf.
857 */
860 APPEND_TAIL,
861 };
865 SPLIT_LEFT,
866 SPLIT_RIGHT,
867 };
875 };
881 };
885 {
894 /*
895 * If the ecc fails, we return the error but otherwise
896 * leave the filesystem running. We know any error is
897 * local to this block.
898 */
904 }
906 /*
907 * Errors after here are fatal.
908 */
916 }
920 "Extent block #%llu has an invalid h_blkno "
925 }
929 "Extent block #%llu has an invalid "
934 }
937 }
941 {
946 ocfs2_validate_extent_block);
948 /* If ocfs2_read_block() got us a new bh, pass it up. */
953 }
956 /*
957 * How many free extents have we got before we need more meta data?
958 */
961 {
977 }
980 }
985 bail:
990 }
992 /* expects array to already be allocated
993 *
994 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
995 * l_count for you
996 */
1002 {
1004 u16 suballoc_bit_start;
1005 u32 num_got;
1014 meta_ac,
1023 }
1031 }
1036 OCFS2_JOURNAL_ACCESS_CREATE);
1040 }
1044 /* Ok, setup the minimal stuff here. */
1055 suballoc_bit_start++;
1056 first_blkno++;
1058 /* We'll also be dirtied by the caller, so
1059 * this isn't absolutely necessary. */
1061 }
1064 }
1067 bail:
1072 }
1074 }
1076 }
1078 /*
1079 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
1080 *
1081 * Returns the sum of the rightmost extent rec logical offset and
1082 * cluster count.
1083 *
1084 * ocfs2_add_branch() uses this to determine what logical cluster
1085 * value should be populated into the leftmost new branch records.
1086 *
1087 * ocfs2_shift_tree_depth() uses this to determine the # clusters
1088 * value for the new topmost tree record.
1089 */
1091 {
1098 }
1100 /*
1101 * Change range of the branches in the right most path according to the leaf
1102 * extent block's rightmost record.
1103 */
1106 {
1116 }
1122 }
1128 }
1134 }
1141 out:
1144 }
1146 /*
1147 * Add an entire tree branch to our inode. eb_bh is the extent block
1148 * to start at, if we don't want to start the branch at the root
1149 * structure.
1150 *
1151 * last_eb_bh is required as we have to update it's next_leaf pointer
1152 * for the new last extent block.
1153 *
1154 * the new branch will be 'empty' in the sense that every block will
1155 * contain a single record with cluster count == 0.
1156 */
1162 {
1180 /* we never add a branch to a leaf. */
1189 /*
1190 * If there is a gap before the root end and the real end
1191 * of the righmost leaf block, we need to remove the gap
1192 * between new_cpos and root_end first so that the tree
1193 * is consistent after we add a new branch(it will start
1194 * from new_cpos).
1195 */
1206 }
1207 }
1209 /* allocate the number of new eb blocks we need */
1211 GFP_KERNEL);
1216 }
1223 }
1225 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
1226 * linked with the rest of the tree.
1227 * conversly, new_eb_bhs[0] is the new bottommost leaf.
1228 *
1229 * when we leave the loop, new_last_eb_blk will point to the
1230 * newest leaf, and next_blkno will point to the topmost extent
1231 * block. */
1236 /* ocfs2_create_new_meta_bhs() should create it right! */
1241 OCFS2_JOURNAL_ACCESS_CREATE);
1245 }
1250 /*
1251 * This actually counts as an empty extent as
1252 * c_clusters == 0
1253 */
1256 /*
1257 * eb_el isn't always an interior node, but even leaf
1258 * nodes want a zero'd flags and reserved field so
1259 * this gets the whole 32 bits regardless of use.
1260 */
1267 }
1269 /* This is a bit hairy. We want to update up to three blocks
1270 * here without leaving any of them in an inconsistent state
1271 * in case of error. We don't have to worry about
1272 * journal_dirty erroring as it won't unless we've aborted the
1273 * handle (in which case we would never be here) so reserving
1274 * the write with journal_access is all we need to do. */
1276 OCFS2_JOURNAL_ACCESS_WRITE);
1280 }
1282 OCFS2_JOURNAL_ACCESS_WRITE);
1286 }
1289 OCFS2_JOURNAL_ACCESS_WRITE);
1293 }
1294 }
1296 /* Link the new branch into the rest of the tree (el will
1297 * either be on the root_bh, or the extent block passed in. */
1304 /* fe needs a new last extent block pointer, as does the
1305 * next_leaf on the previously last-extent-block. */
1316 /*
1317 * Some callers want to track the rightmost leaf so pass it
1318 * back here.
1319 */
1325 bail:
1330 }
1333 }
1335 /*
1336 * adds another level to the allocation tree.
1337 * returns back the new extent block so you can add a branch to it
1338 * after this call.
1339 */
1344 {
1346 u32 new_clusters;
1357 }
1360 /* ocfs2_create_new_meta_bhs() should create it right! */
1367 OCFS2_JOURNAL_ACCESS_CREATE);
1371 }
1373 /* copy the root extent list data into the new extent block */
1382 OCFS2_JOURNAL_ACCESS_WRITE);
1386 }
1390 /* update root_bh now */
1399 /* If this is our 1st tree depth shift, then last_eb_blk
1400 * becomes the allocated extent block */
1409 bail:
1413 }
1415 /*
1416 * Should only be called when there is no space left in any of the
1417 * leaf nodes. What we want to do is find the lowest tree depth
1418 * non-leaf extent block with room for new records. There are three
1419 * valid results of this search:
1420 *
1421 * 1) a lowest extent block is found, then we pass it back in
1422 * *lowest_eb_bh and return '0'
1423 *
1424 * 2) the search fails to find anything, but the root_el has room. We
1425 * pass NULL back in *lowest_eb_bh, but still return '0'
1426 *
1427 * 3) the search fails to find anything AND the root_el is full, in
1428 * which case we return > 0
1429 *
1430 * return status < 0 indicates an error.
1431 */
1434 {
1436 u64 blkno;
1449 "Owner %llu has empty "
1454 }
1459 "Owner %llu has extent "
1460 "list where extent # %d has no physical "
1465 }
1474 }
1484 }
1485 }
1487 /* If we didn't find one and the fe doesn't have any room,
1488 * then return '1' */
1494 bail:
1498 }
1500 /*
1501 * Grow a b-tree so that it has more records.
1502 *
1503 * We might shift the tree depth in which case existing paths should
1504 * be considered invalid.
1505 *
1506 * Tree depth after the grow is returned via *final_depth.
1507 *
1508 * *last_eb_bh will be updated by ocfs2_add_branch().
1509 */
1513 {
1526 }
1528 /* We traveled all the way to the bottom of the allocation tree
1529 * and didn't find room for any more extents - we need to add
1530 * another tree level */
1536 depth);
1538 /* ocfs2_shift_tree_depth will return us a buffer with
1539 * the new extent block (so we can pass that to
1540 * ocfs2_add_branch). */
1545 }
1546 depth++;
1548 /*
1549 * Special case: we have room now if we shifted from
1550 * tree_depth 0, so no more work needs to be done.
1551 *
1552 * We won't be calling add_branch, so pass
1553 * back *last_eb_bh as the new leaf. At depth
1554 * zero, it should always be null so there's
1555 * no reason to brelse.
1556 */
1561 }
1562 }
1564 /* call ocfs2_add_branch to add the final part of the tree with
1565 * the new data. */
1567 meta_ac);
1571 }
1573 out:
1578 }
1580 /*
1581 * This function will discard the rightmost extent record.
1582 */
1584 {
1590 /* This will cause us to go off the end of our extent list. */
1596 }
1600 {
1610 /* The tree code before us didn't allow enough room in the leaf. */
1613 /*
1614 * The easiest way to approach this is to just remove the
1615 * empty extent and temporarily decrement next_free.
1616 */
1618 /*
1619 * If next_free was 1 (only an empty extent), this
1620 * loop won't execute, which is fine. We still want
1621 * the decrement above to happen.
1622 */
1626 next_free--;
1627 }
1629 /*
1630 * Figure out what the new record index should be.
1631 */
1637 }
1648 /*
1649 * No need to memmove if we're just adding to the tail.
1650 */
1658 num_bytes);
1659 }
1661 /*
1662 * Either we had an empty extent, and need to re-increment or
1663 * there was no empty extent on a non full rightmost leaf node,
1664 * in which case we still need to increment.
1665 */
1666 next_free++;
1668 /*
1669 * Make sure none of the math above just messed up our tree.
1670 */
1675 }
1678 {
1684 num_recs--;
1690 }
1691 }
1693 /*
1694 * Create an empty extent record .
1695 *
1696 * l_next_free_rec may be updated.
1697 *
1698 * If an empty extent already exists do nothing.
1699 */
1701 {
1720 set_and_inc:
1723 }
1725 /*
1726 * For a rotation which involves two leaf nodes, the "root node" is
1727 * the lowest level tree node which contains a path to both leafs. This
1728 * resulting set of information can be used to form a complete "subtree"
1729 *
1730 * This function is passed two full paths from the dinode down to a
1731 * pair of adjacent leaves. It's task is to figure out which path
1732 * index contains the subtree root - this can be the root index itself
1733 * in a worst-case rotation.
1734 *
1735 * The array index of the subtree root is passed back.
1736 */
1740 {
1743 /*
1744 * Check that the caller passed in two paths from the same tree.
1745 */
1749 i++;
1751 /*
1752 * The caller didn't pass two adjacent paths.
1753 */
1765 }
1769 /*
1770 * Traverse a btree path in search of cpos, starting at root_el.
1771 *
1772 * This code can be called with a cpos larger than the tree, in which
1773 * case it will return the rightmost path.
1774 */
1778 {
1780 u32 range;
1781 u64 blkno;
1791 "Owner %llu has empty extent list at "
1798 }
1803 /*
1804 * In the case that cpos is off the allocation
1805 * tree, this should just wind up returning the
1806 * rightmost record.
1807 */
1812 }
1817 "Owner %llu has bad blkno in extent list "
1823 }
1831 }
1839 "Owner %llu has bad count in extent list "
1847 }
1851 }
1853 out:
1854 /*
1855 * Catch any trailing bh that the loop didn't handle.
1856 */
1860 }
1862 /*
1863 * Given an initialized path (that is, it has a valid root extent
1864 * list), this function will traverse the btree in search of the path
1865 * which would contain cpos.
1866 *
1867 * The path traveled is recorded in the path structure.
1868 *
1869 * Note that this will not do any comparisons on leaf node extent
1870 * records, so it will work fine in the case that we just added a tree
1871 * branch.
1872 */
1876 };
1878 {
1884 }
1887 {
1894 }
1897 {
1902 /* We want to retain only the leaf block. */
1906 }
1907 }
1908 /*
1909 * Find the leaf block in the tree which would contain cpos. No
1910 * checking of the actual leaf is done.
1911 *
1912 * Some paths want to call this instead of allocating a path structure
1913 * and calling ocfs2_find_path().
1914 *
1915 * This function doesn't handle non btree extent lists.
1916 */
1920 {
1928 }
1931 out:
1933 }
1935 /*
1936 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1937 *
1938 * Basically, we've moved stuff around at the bottom of the tree and
1939 * we need to fix up the extent records above the changes to reflect
1940 * the new changes.
1941 *
1942 * left_rec: the record on the left.
1943 * left_child_el: is the child list pointed to by left_rec
1944 * right_rec: the record to the right of left_rec
1945 * right_child_el: is the child list pointed to by right_rec
1946 *
1947 * By definition, this only works on interior nodes.
1948 */
1953 {
1956 /*
1957 * Interior nodes never have holes. Their cpos is the cpos of
1958 * the leftmost record in their child list. Their cluster
1959 * count covers the full theoretical range of their child list
1960 * - the range between their cpos and the cpos of the record
1961 * immediately to their right.
1962 */
1968 }
1972 /*
1973 * Calculate the rightmost cluster count boundary before
1974 * moving cpos - we will need to adjust clusters after
1975 * updating e_cpos to keep the same highest cluster count.
1976 */
1985 }
1987 /*
1988 * Adjust the adjacent root node records involved in a
1989 * rotation. left_el_blkno is passed in as a key so that we can easily
1990 * find it's index in the root list.
1991 */
1995 u64 left_el_blkno)
1996 {
2005 }
2007 /*
2008 * The path walking code should have never returned a root and
2009 * two paths which are not adjacent.
2010 */
2015 }
2017 /*
2018 * We've changed a leaf block (in right_path) and need to reflect that
2019 * change back up the subtree.
2020 *
2021 * This happens in multiple places:
2022 * - When we've moved an extent record from the left path leaf to the right
2023 * path leaf to make room for an empty extent in the left path leaf.
2024 * - When our insert into the right path leaf is at the leftmost edge
2025 * and requires an update of the path immediately to it's left. This
2026 * can occur at the end of some types of rotation and appending inserts.
2027 * - When we've adjusted the last extent record in the left path leaf and the
2028 * 1st extent record in the right path leaf during cross extent block merge.
2029 */
2034 {
2040 /*
2041 * Update the counts and position values within all the
2042 * interior nodes to reflect the leaf rotation we just did.
2043 *
2044 * The root node is handled below the loop.
2045 *
2046 * We begin the loop with right_el and left_el pointing to the
2047 * leaf lists and work our way up.
2048 *
2049 * NOTE: within this loop, left_el and right_el always refer
2050 * to the *child* lists.
2051 */
2057 /*
2058 * One nice property of knowing that all of these
2059 * nodes are below the root is that we only deal with
2060 * the leftmost right node record and the rightmost
2061 * left node record.
2062 */
2071 right_el);
2076 /*
2077 * Setup our list pointers now so that the current
2078 * parents become children in the next iteration.
2079 */
2082 }
2084 /*
2085 * At the root node, adjust the two adjacent records which
2086 * begin our path to the leaves.
2087 */
2099 }
2106 {
2119 "Inode %llu has non-full interior leaf node %llu"
2125 }
2127 /*
2128 * This extent block may already have an empty record, so we
2129 * return early if so.
2130 */
2138 subtree_index);
2142 }
2150 }
2157 }
2158 }
2163 /* This is a code error, not a disk corruption. */
2173 /* Do the copy now. */
2178 /*
2179 * Clear out the record we just copied and shift everything
2180 * over, leaving an empty extent in the left leaf.
2181 *
2182 * We temporarily subtract from next_free_rec so that the
2183 * shift will lose the tail record (which is now defunct).
2184 */
2193 subtree_index);
2195 out:
2197 }
2199 /*
2200 * Given a full path, determine what cpos value would return us a path
2201 * containing the leaf immediately to the left of the current one.
2202 *
2203 * Will return zero if the path passed in is already the leftmost path.
2204 */
2207 {
2209 u64 blkno;
2218 /* Start at the tree node just above the leaf and work our way up. */
2223 /*
2224 * Find the extent record just before the one in our
2225 * path.
2226 */
2231 /*
2232 * We've determined that the
2233 * path specified is already
2234 * the leftmost one - return a
2235 * cpos of zero.
2236 */
2238 }
2239 /*
2240 * The leftmost record points to our
2241 * leaf - we need to travel up the
2242 * tree one level.
2243 */
2245 }
2252 }
2253 }
2255 /*
2256 * If we got here, we never found a valid node where
2257 * the tree indicated one should be.
2258 */
2265 next_node:
2267 i--;
2268 }
2270 out:
2272 }
2274 /*
2275 * Extend the transaction by enough credits to complete the rotation,
2276 * and still leave at least the original number of credits allocated
2277 * to this transaction.
2278 */
2282 {
2291 }
2293 /*
2294 * Trap the case where we're inserting into the theoretical range past
2295 * the _actual_ left leaf range. Otherwise, we'll rotate a record
2296 * whose cpos is less than ours into the right leaf.
2297 *
2298 * It's only necessary to look at the rightmost record of the left
2299 * leaf because the logic that calls us should ensure that the
2300 * theoretical ranges in the path components above the leaves are
2301 * correct.
2302 */
2304 u32 insert_cpos)
2305 {
2317 }
2320 {
2330 /* Empty list. */
2334 }
2340 }
2342 /*
2343 * Rotate all the records in a btree right one record, starting at insert_cpos.
2344 *
2345 * The path to the rightmost leaf should be passed in.
2346 *
2347 * The array is assumed to be large enough to hold an entire path (tree depth).
2348 *
2349 * Upon successful return from this function:
2350 *
2351 * - The 'right_path' array will contain a path to the leaf block
2352 * whose range contains e_cpos.
2353 * - That leaf block will have a single empty extent in list index 0.
2354 * - In the case that the rotation requires a post-insert update,
2355 * *ret_left_path will contain a valid path which can be passed to
2356 * ocfs2_insert_path().
2357 */
2361 u32 insert_cpos,
2364 {
2366 u32 cpos;
2377 }
2383 }
2389 /*
2390 * What we want to do here is:
2391 *
2392 * 1) Start with the rightmost path.
2393 *
2394 * 2) Determine a path to the leaf block directly to the left
2395 * of that leaf.
2396 *
2397 * 3) Determine the 'subtree root' - the lowest level tree node
2398 * which contains a path to both leaves.
2399 *
2400 * 4) Rotate the subtree.
2401 *
2402 * 5) Find the next subtree by considering the left path to be
2403 * the new right path.
2404 *
2405 * The check at the top of this while loop also accepts
2406 * insert_cpos == cpos because cpos is only a _theoretical_
2407 * value to get us the left path - insert_cpos might very well
2408 * be filling that hole.
2409 *
2410 * Stop at a cpos of '0' because we either started at the
2411 * leftmost branch (i.e., a tree with one branch and a
2412 * rotation inside of it), or we've gone as far as we can in
2413 * rotating subtrees.
2414 */
2425 }
2429 "Owner %llu: error during insert of %u "
2430 "(left path cpos %u) results in two identical "
2439 insert_cpos)) {
2441 /*
2442 * We've rotated the tree as much as we
2443 * should. The rest is up to
2444 * ocfs2_insert_path() to complete, after the
2445 * record insertion. We indicate this
2446 * situation by returning the left path.
2447 *
2448 * The reason we don't adjust the records here
2449 * before the record insert is that an error
2450 * later might break the rule where a parent
2451 * record e_cpos will reflect the actual
2452 * e_cpos of the 1st nonempty record of the
2453 * child list.
2454 */
2457 }
2471 }
2478 }
2482 insert_cpos)) {
2483 /*
2484 * A rotate moves the rightmost left leaf
2485 * record over to the leftmost right leaf
2486 * slot. If we're doing an extent split
2487 * instead of a real insert, then we have to
2488 * check that the extent to be split wasn't
2489 * just moved over. If it was, then we can
2490 * exit here, passing left_path back -
2491 * ocfs2_split_extent() is smart enough to
2492 * search both leaves.
2493 */
2496 }
2498 /*
2499 * There is no need to re-read the next right path
2500 * as we know that it'll be our current left
2501 * path. Optimize by copying values instead.
2502 */
2509 }
2510 }
2512 out:
2515 out_ret_path:
2517 }
2522 {
2527 u32 range;
2529 /*
2530 * In normal tree rotation process, we will never touch the
2531 * tree branch above subtree_index and ocfs2_extend_rotate_transaction
2532 * doesn't reserve the credits for them either.
2533 *
2534 * But we do have a special case here which will update the rightmost
2535 * records for all the bh in the path.
2536 * So we have to allocate extra credits and access them.
2537 */
2542 }
2548 }
2550 /* Path should always be rightmost. */
2569 }
2570 out:
2572 }
2578 {
2588 /*
2589 * Not all nodes might have had their final count
2590 * decremented by the caller - handle this here.
2591 */
2595 "Inode %llu, attempted to remove extent block "
2604 }
2616 }
2617 }
2625 {
2653 }
2662 {
2680 /*
2681 * It's legal for us to proceed if the right leaf is
2682 * the rightmost one and it has an empty extent. There
2683 * are two cases to handle - whether the leaf will be
2684 * empty after removal or not. If the leaf isn't empty
2685 * then just remove the empty extent up front. The
2686 * next block will handle empty leaves by flagging
2687 * them for unlink.
2688 *
2689 * Non rightmost leaves will throw -EAGAIN and the
2690 * caller can manually move the subtree and retry.
2691 */
2699 OCFS2_JOURNAL_ACCESS_WRITE);
2703 }
2708 }
2712 /*
2713 * We have to update i_last_eb_blk during the meta
2714 * data delete.
2715 */
2717 OCFS2_JOURNAL_ACCESS_WRITE);
2721 }
2724 }
2726 /*
2727 * Getting here with an empty extent in the right path implies
2728 * that it's the rightmost path and will be deleted.
2729 */
2733 subtree_index);
2737 }
2745 }
2752 }
2753 }
2756 /*
2757 * Only do this if we're moving a real
2758 * record. Otherwise, the action is delayed until
2759 * after removal of the right path in which case we
2760 * can do a simple shift to remove the empty extent.
2761 */
2765 }
2767 /*
2768 * Move recs over to get rid of empty extent, decrease
2769 * next_free. This is allowed to remove the last
2770 * extent in our leaf (setting l_next_free_rec to
2771 * zero) - the delete code below won't care.
2772 */
2774 }
2783 left_path);
2787 }
2792 /*
2793 * Removal of the extent in the left leaf was skipped
2794 * above so we could delete the right path
2795 * 1st.
2796 */
2805 subtree_index);
2807 out:
2809 }
2811 /*
2812 * Given a full path, determine what cpos value would return us a path
2813 * containing the leaf immediately to the right of the current one.
2814 *
2815 * Will return zero if the path passed in is already the rightmost path.
2816 *
2817 * This looks similar, but is subtly different to
2818 * ocfs2_find_cpos_for_left_leaf().
2819 */
2822 {
2824 u64 blkno;
2834 /* Start at the tree node just above the leaf and work our way up. */
2841 /*
2842 * Find the extent record just after the one in our
2843 * path.
2844 */
2850 /*
2851 * We've determined that the
2852 * path specified is already
2853 * the rightmost one - return a
2854 * cpos of zero.
2855 */
2857 }
2858 /*
2859 * The rightmost record points to our
2860 * leaf - we need to travel up the
2861 * tree one level.
2862 */
2864 }
2868 }
2869 }
2871 /*
2872 * If we got here, we never found a valid node where
2873 * the tree indicated one should be.
2874 */
2881 next_node:
2883 i--;
2884 }
2886 out:
2888 }
2893 {
2906 }
2911 out:
2913 }
2921 {
2923 u32 right_cpos;
2936 }
2943 }
2952 }
2959 }
2962 right_path);
2974 }
2976 /*
2977 * Caller might still want to make changes to the
2978 * tree root, so re-add it to the journal here.
2979 */
2985 }
2991 /*
2992 * The rotation has to temporarily stop due to
2993 * the right subtree having an empty
2994 * extent. Pass it back to the caller for a
2995 * fixup.
2996 */
3000 }
3004 }
3006 /*
3007 * The subtree rotate might have removed records on
3008 * the rightmost edge. If so, then rotation is
3009 * complete.
3010 */
3021 }
3022 }
3024 out:
3029 }
3035 {
3037 u32 cpos;
3046 /*
3047 * There's two ways we handle this depending on
3048 * whether path is the only existing one.
3049 */
3052 path);
3056 }
3062 }
3069 }
3072 /*
3073 * We have a path to the left of this one - it needs
3074 * an update too.
3075 */
3081 }
3087 }
3093 }
3100 left_path);
3104 }
3109 /*
3110 * 'path' is also the leftmost path which
3111 * means it must be the only one. This gets
3112 * handled differently because we want to
3113 * revert the root back to having extents
3114 * in-line.
3115 */
3124 }
3128 out:
3131 }
3133 /*
3134 * Left rotation of btree records.
3135 *
3136 * In many ways, this is (unsurprisingly) the opposite of right
3137 * rotation. We start at some non-rightmost path containing an empty
3138 * extent in the leaf block. The code works its way to the rightmost
3139 * path by rotating records to the left in every subtree.
3140 *
3141 * This is used by any code which reduces the number of extent records
3142 * in a leaf. After removal, an empty record should be placed in the
3143 * leftmost list position.
3144 *
3145 * This won't handle a length update of the rightmost path records if
3146 * the rightmost tree leaf record is removed so the caller is
3147 * responsible for detecting and correcting that.
3148 */
3153 {
3164 rightmost_no_delete:
3165 /*
3166 * Inline extents. This is trivially handled, so do
3167 * it up front.
3168 */
3173 }
3175 /*
3176 * Handle rightmost branch now. There's several cases:
3177 * 1) simple rotation leaving records in there. That's trivial.
3178 * 2) rotation requiring a branch delete - there's no more
3179 * records left. Two cases of this:
3180 * a) There are branches to the left.
3181 * b) This is also the leftmost (the only) branch.
3182 *
3183 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
3184 * 2a) we need the left branch so that we can update it with the unlink
3185 * 2b) we need to bring the root back to inline extents.
3186 */
3191 /*
3192 * This gets a bit tricky if we're going to delete the
3193 * rightmost path. Get the other cases out of the way
3194 * 1st.
3195 */
3206 }
3208 /*
3209 * XXX: The caller can not trust "path" any more after
3210 * this as it will have been deleted. What do we do?
3211 *
3212 * In theory the rotate-for-merge code will never get
3213 * here because it'll always ask for a rotate in a
3214 * nonempty list.
3215 */
3218 dealloc);
3222 }
3224 /*
3225 * Now we can loop, remembering the path we get from -EAGAIN
3226 * and restarting from there.
3227 */
3228 try_rotate:
3234 }
3246 }
3253 }
3255 out:
3259 }
3263 {
3268 /*
3269 * We consumed all of the merged-from record. An empty
3270 * extent cannot exist anywhere but the 1st array
3271 * position, so move things over if the merged-from
3272 * record doesn't occupy that position.
3273 *
3274 * This creates a new empty extent so the caller
3275 * should be smart enough to have removed any existing
3276 * ones.
3277 */
3282 }
3284 /*
3285 * Always memset - the caller doesn't check whether it
3286 * created an empty extent, so there could be junk in
3287 * the other fields.
3288 */
3290 }
3291 }
3296 {
3298 u32 right_cpos;
3304 /* This function shouldn't be called for non-trees. */
3315 }
3317 /* This function shouldn't be called for the rightmost leaf. */
3325 }
3331 }
3334 out:
3338 }
3340 /*
3341 * Remove split_rec clusters from the record at index and merge them
3342 * onto the beginning of the record "next" to it.
3343 * For index < l_count - 1, the next means the extent rec at index + 1.
3344 * For index == l_count - 1, the "next" means the 1st extent rec of the
3345 * next extent block.
3346 */
3352 {
3369 /* we meet with a cross extent block merge. */
3374 }
3383 }
3390 right_path);
3394 right_path);
3398 }
3404 subtree_index);
3408 }
3417 }
3424 }
3425 }
3430 }
3437 }
3444 split_clusters));
3453 subtree_index);
3454 }
3455 out:
3459 }
3464 {
3466 u32 left_cpos;
3471 /* This function shouldn't be called for non-trees. */
3479 }
3481 /* This function shouldn't be called for the leftmost leaf. */
3489 }
3495 }
3498 out:
3502 }
3504 /*
3505 * Remove split_rec clusters from the record at index and merge them
3506 * onto the tail of the record "before" it.
3507 * For index > 0, the "before" means the extent rec at index - 1.
3508 *
3509 * For index == 0, the "before" means the last record of the previous
3510 * extent block. And there is also a situation that we may need to
3511 * remove the rightmost leaf extent block in the right_path and change
3512 * the right path to indicate the new rightmost path.
3513 */
3520 {
3535 /* we meet with a cross extent block merge. */
3540 }
3553 right_path);
3557 left_path);
3561 }
3567 subtree_index);
3571 }
3580 }
3587 }
3588 }
3593 }
3600 }
3603 /*
3604 * The easy case - we can just plop the record right in.
3605 */
3615 split_clusters));
3624 /*
3625 * In the situation that the right_rec is empty and the extent
3626 * block is empty also, ocfs2_complete_edge_insert can't handle
3627 * it and we need to delete the right extent block.
3628 */
3633 right_path,
3634 dealloc);
3638 }
3640 /* Now the rightmost extent block has been deleted.
3641 * So we use the new rightmost path.
3642 */
3648 }
3649 out:
3653 }
3662 {
3670 /*
3671 * The merge code will need to create an empty
3672 * extent to take the place of the newly
3673 * emptied slot. Remove any pre-existing empty
3674 * extents - having more than one in a leaf is
3675 * illegal.
3676 */
3681 }
3682 split_index--;
3684 }
3687 /*
3688 * Left-right contig implies this.
3689 */
3692 /*
3693 * Since the leftright insert always covers the entire
3694 * extent, this call will delete the insert record
3695 * entirely, resulting in an empty extent record added to
3696 * the extent block.
3697 *
3698 * Since the adding of an empty extent shifts
3699 * everything back to the right, there's no need to
3700 * update split_index here.
3701 *
3702 * When the split_index is zero, we need to merge it to the
3703 * prevoius extent block. It is more efficient and easier
3704 * if we do merge_right first and merge_left later.
3705 */
3707 split_index);
3711 }
3713 /*
3714 * We can only get this from logic error above.
3715 */
3718 /* The merge left us with an empty extent, remove it. */
3723 }
3727 /*
3728 * Note that we don't pass split_rec here on purpose -
3729 * we've merged it into the rec already.
3730 */
3737 }
3740 /*
3741 * Error from this last rotate is not critical, so
3742 * print but don't bubble it up.
3743 */
3748 /*
3749 * Merge a record to the left or right.
3750 *
3751 * 'contig_type' is relative to the existing record,
3752 * so for example, if we're "right contig", it's to
3753 * the record on the left (hence the left merge).
3754 */
3758 split_index);
3762 }
3766 split_index);
3770 }
3771 }
3774 /*
3775 * The merge may have left an empty extent in
3776 * our leaf. Try to rotate it away.
3777 */
3779 dealloc);
3783 }
3784 }
3786 out:
3788 }
3794 {
3795 u64 len_blocks;
3801 /*
3802 * Region is on the left edge of the existing
3803 * record.
3804 */
3811 /*
3812 * Region is on the right edge of the existing
3813 * record.
3814 */
3817 }
3818 }
3820 /*
3821 * Do the final bits of extent record insertion at the target leaf
3822 * list. If this leaf is part of an allocation tree, it is assumed
3823 * that the tree above has been prepared.
3824 */
3829 {
3842 insert_rec);
3844 }
3846 /*
3847 * Contiguous insert - either left or right.
3848 */
3854 }
3858 }
3860 /*
3861 * Handle insert into an empty leaf.
3862 */
3869 }
3871 /*
3872 * Appending insert.
3873 */
3883 "owner %llu, depth %u, count %u, next free %u, "
3884 "rec.cpos %u, rec.clusters %u, "
3894 i++;
3898 }
3900 rotate:
3901 /*
3902 * Ok, we have to rotate.
3903 *
3904 * At this point, it is safe to assume that inserting into an
3905 * empty leaf and appending to a leaf have both been handled
3906 * above.
3907 *
3908 * This leaf needs to have space, either by the empty 1st
3909 * extent record, or by virtue of an l_next_rec < l_count.
3910 */
3912 }
3918 {
3924 /*
3925 * Update everything except the leaf block.
3926 */
3938 }
3949 }
3950 }
3957 {
3964 /*
3965 * This shouldn't happen for non-trees. The extent rec cluster
3966 * count manipulation below only works for interior nodes.
3967 */
3970 /*
3971 * If our appending insert is at the leftmost edge of a leaf,
3972 * then we might need to update the rightmost records of the
3973 * neighboring path.
3974 */
3979 u32 left_cpos;
3986 }
3992 left_cpos);
3994 /*
3995 * No need to worry if the append is already in the
3996 * leftmost leaf.
3997 */
4004 }
4007 left_cpos);
4011 }
4013 /*
4014 * ocfs2_insert_path() will pass the left_path to the
4015 * journal for us.
4016 */
4017 }
4018 }
4024 }
4030 out:
4035 }
4042 {
4059 /*
4060 * This typically means that the record
4061 * started in the left path but moved to the
4062 * right as a result of rotation. We either
4063 * move the existing record to the left, or we
4064 * do the later insert there.
4065 *
4066 * In this case, the left path should always
4067 * exist as the rotate code will have passed
4068 * it back for a post-insert update.
4069 */
4072 /*
4073 * It's a left split. Since we know
4074 * that the rotate code gave us an
4075 * empty extent in the left path, we
4076 * can just do the insert there.
4077 */
4080 /*
4081 * Right split - we have to move the
4082 * existing record over to the left
4083 * leaf. The insert will be into the
4084 * newly created empty extent in the
4085 * right leaf.
4086 */
4094 }
4095 }
4099 /*
4100 * Left path is easy - we can just allow the insert to
4101 * happen.
4102 */
4107 }
4113 }
4115 /*
4116 * This function only does inserts on an allocation b-tree. For tree
4117 * depth = 0, ocfs2_insert_at_leaf() is called directly.
4118 *
4119 * right_path is the path we want to do the actual insert
4120 * in. left_path should only be passed in if we need to update that
4121 * portion of the tree after an edge insert.
4122 */
4129 {
4134 /*
4135 * There's a chance that left_path got passed back to
4136 * us without being accounted for in the
4137 * journal. Extend our transaction here to be sure we
4138 * can change those blocks.
4139 */
4144 }
4150 }
4151 }
4153 /*
4154 * Pass both paths to the journal. The majority of inserts
4155 * will be touching all components anyway.
4156 */
4161 }
4164 /*
4165 * We could call ocfs2_insert_at_leaf() for some types
4166 * of splits, but it's easier to just let one separate
4167 * function sort it all out.
4168 */
4172 /*
4173 * Split might have modified either leaf and we don't
4174 * have a guarantee that the later edge insert will
4175 * dirty this for us.
4176 */
4182 insert);
4187 /*
4188 * The rotate code has indicated that we need to fix
4189 * up portions of the tree after the insert.
4190 *
4191 * XXX: Should we extend the transaction here?
4192 */
4194 right_path);
4196 subtree_index);
4197 }
4200 out:
4202 }
4208 {
4210 u32 cpos;
4218 OCFS2_JOURNAL_ACCESS_WRITE);
4222 }
4227 }
4234 }
4236 /*
4237 * Determine the path to start with. Rotations need the
4238 * rightmost path, everything else can go directly to the
4239 * target leaf.
4240 */
4246 }
4252 }
4254 /*
4255 * Rotations and appends need special treatment - they modify
4256 * parts of the tree's above them.
4257 *
4258 * Both might pass back a path immediate to the left of the
4259 * one being inserted to. This will be cause
4260 * ocfs2_insert_path() to modify the rightmost records of
4261 * left_path to account for an edge insert.
4262 *
4263 * XXX: When modifying this code, keep in mind that an insert
4264 * can wind up skipping both of these two special cases...
4265 */
4273 }
4275 /*
4276 * ocfs2_rotate_tree_right() might have extended the
4277 * transaction without re-journaling our tree root.
4278 */
4280 OCFS2_JOURNAL_ACCESS_WRITE);
4284 }
4292 }
4293 }
4300 }
4302 out_update_clusters:
4309 out:
4314 }
4316 static enum ocfs2_contig_type
4321 {
4345 left_cpos);
4356 "Extent block #%llu has an "
4357 "invalid l_next_free_rec of "
4358 "%d. It should have "
4365 }
4368 }
4369 }
4371 /*
4372 * We're careful to check for an empty extent record here -
4373 * the merge code will know what to do if it sees one.
4374 */
4381 }
4382 }
4411 "Extent block #%llu has an "
4417 }
4419 }
4420 }
4431 }
4433 out:
4440 }
4446 {
4454 insert_rec);
4458 }
4459 }
4468 /*
4469 * Caller might want us to limit the size of extents, don't
4470 * calculate contiguousness if we might exceed that limit.
4471 */
4475 }
4476 }
4478 /*
4479 * This should only be called against the righmost leaf extent list.
4480 *
4481 * ocfs2_figure_appending_type() will figure out whether we'll have to
4482 * insert at the tail of the rightmost leaf.
4483 *
4484 * This should also work against the root extent list for tree's with 0
4485 * depth. If we consider the root extent list to be the rightmost leaf node
4486 * then the logic here makes sense.
4487 */
4491 {
4504 /* Were all records empty? */
4507 }
4518 set_tail_append:
4520 }
4522 /*
4523 * Helper function called at the beginning of an insert.
4524 *
4525 * This computes a few things that are commonly used in the process of
4526 * inserting into the btree:
4527 * - Whether the new extent is contiguous with an existing one.
4528 * - The current tree depth.
4529 * - Whether the insert is an appending one.
4530 * - The total # of free records in the tree.
4531 *
4532 * All of the information is stored on the ocfs2_insert_type
4533 * structure.
4534 */
4540 {
4553 /*
4554 * If we have tree depth, we read in the
4555 * rightmost extent block ahead of time as
4556 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4557 * may want it later.
4558 */
4565 }
4568 }
4570 /*
4571 * Unless we have a contiguous insert, we'll need to know if
4572 * there is room left in our allocation tree for another
4573 * extent record.
4574 *
4575 * XXX: This test is simplistic, we can search for empty
4576 * extent records too.
4577 */
4585 }
4592 }
4594 /*
4595 * In the case that we're inserting past what the tree
4596 * currently accounts for, ocfs2_find_path() will return for
4597 * us the rightmost tree path. This is accounted for below in
4598 * the appending code.
4599 */
4604 }
4608 /*
4609 * Now that we have the path, there's two things we want to determine:
4610 * 1) Contiguousness (also set contig_index if this is so)
4611 *
4612 * 2) Are we doing an append? We can trivially break this up
4613 * into two types of appends: simple record append, or a
4614 * rotate inside the tail leaf.
4615 */
4618 /*
4619 * The insert code isn't quite ready to deal with all cases of
4620 * left contiguousness. Specifically, if it's an insert into
4621 * the 1st record in a leaf, it will require the adjustment of
4622 * cluster count on the last record of the path directly to it's
4623 * left. For now, just catch that case and fool the layers
4624 * above us. This works just fine for tree_depth == 0, which
4625 * is why we allow that above.
4626 */
4631 /*
4632 * Ok, so we can simply compare against last_eb to figure out
4633 * whether the path doesn't exist. This will only happen in
4634 * the case that we're doing a tail append, so maybe we can
4635 * take advantage of that information somehow.
4636 */
4639 /*
4640 * Ok, ocfs2_find_path() returned us the rightmost
4641 * tree path. This might be an appending insert. There are
4642 * two cases:
4643 * 1) We're doing a true append at the tail:
4644 * -This might even be off the end of the leaf
4645 * 2) We're "appending" by rotating in the tail
4646 */
4648 }
4650 out:
4655 else
4658 }
4660 /*
4661 * Insert an extent into a btree.
4662 *
4663 * The caller needs to update the owning btree's cluster count.
4664 */
4667 u32 cpos,
4668 u64 start_blk,
4669 u32 new_clusters,
4670 u8 flags,
4672 {
4692 }
4699 }
4708 meta_ac);
4712 }
4713 }
4715 /* Finally, we can add clusters. This might rotate the tree for us. */
4719 else
4722 bail:
4726 }
4728 /*
4729 * Allcate and add clusters into the extent b-tree.
4730 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4731 * The extent b-tree's root is specified by et, and
4732 * it is not limited to the file storage. Any extent tree can use this
4733 * function if it implements the proper ocfs2_extent_tree.
4734 */
4738 u32 clusters_to_add,
4743 {
4748 u64 block;
4763 }
4765 /* there are two cases which could cause us to EAGAIN in the
4766 * we-need-more-metadata case:
4767 * 1) we haven't reserved *any*
4768 * 2) we are so fragmented, we've needed to add metadata too
4769 * many times. */
4782 }
4790 }
4794 /* reserve our write early -- insert_extent may update the tree root */
4796 OCFS2_JOURNAL_ACCESS_WRITE);
4800 }
4811 }
4822 }
4824 leave:
4829 }
4833 u32 cpos,
4835 {
4849 }
4858 {
4868 leftright:
4869 /*
4870 * Store a copy of the record on the stack - it might move
4871 * around as the tree is manipulated below.
4872 */
4882 }
4891 }
4892 }
4909 /*
4910 * Left/right split. We fake this as a right split
4911 * first and then make a second pass as a left split.
4912 */
4922 }
4928 }
4931 u32 cpos;
4934 do_leftright++;
4944 }
4949 }
4950 out:
4953 }
4961 {
4969 }
4974 out:
4976 }
4978 /*
4979 * Split part or all of the extent record at split_index in the leaf
4980 * pointed to by path. Merge with the contiguous extent record if needed.
4981 *
4982 * Care is taken to handle contiguousness so as to not grow the tree.
4983 *
4984 * meta_ac is not strictly necessary - we only truly need it if growth
4985 * of the tree is required. All other cases will degrade into a less
4986 * optimal tree layout.
4987 *
4988 * last_eb_bh should be the rightmost leaf block for any extent
4989 * btree. Since a split may grow the tree or a merge might shrink it,
4990 * the caller cannot trust the contents of that buffer after this call.
4991 *
4992 * This code is optimized for readability - several passes might be
4993 * made over certain portions of the tree. All of those blocks will
4994 * have been brought into cache (and pinned via the journal), so the
4995 * extra overhead is not expressed in terms of disk reads.
4996 */
5004 {
5018 }
5021 split_index,
5022 split_rec);
5024 /*
5025 * The core merge / split code wants to know how much room is
5026 * left in this allocation tree, so we pass the
5027 * rightmost extent list.
5028 */
5038 }
5048 else
5061 else
5073 }
5075 out:
5078 }
5080 /*
5081 * Change the flags of the already-existing extent at cpos for len clusters.
5082 *
5083 * new_flags: the flags we want to set.
5084 * clear_flags: the flags we want to clear.
5085 * phys: the new physical offset we want this new extent starts from.
5086 *
5087 * If the existing extent is larger than the request, initiate a
5088 * split. An attempt will be made at merging with adjacent extents.
5089 *
5090 * The caller is responsible for passing down meta_ac if we'll need it.
5091 */
5098 {
5112 }
5118 }
5124 "Owner %llu has an extent at cpos %u which can no "
5130 }
5138 new_flags);
5140 }
5146 clear_flags);
5148 }
5162 dealloc);
5166 out:
5170 }
5172 /*
5173 * Mark the already-existing extent at cpos as written for len clusters.
5174 * This removes the unwritten extent flag.
5175 *
5176 * If the existing extent is larger than the request, initiate a
5177 * split. An attempt will be made at merging with adjacent extents.
5178 *
5179 * The caller is responsible for passing down meta_ac if we'll need it.
5180 */
5186 {
5195 "that are being written to, but the feature bit "
5200 }
5202 /*
5203 * XXX: This should be fixed up so that we just re-insert the
5204 * next extent records.
5205 */
5214 out:
5216 }
5222 {
5231 /*
5232 * Setup the record to split before we grow the tree.
5233 */
5247 }
5260 }
5265 meta_ac);
5269 }
5270 }
5282 out:
5285 }
5292 {
5307 }
5309 index--;
5310 }
5314 /*
5315 * Check whether this is the rightmost tree record. If
5316 * we remove all of this record or part of its right
5317 * edge then an update of the record lengths above it
5318 * will be required.
5319 */
5323 }
5328 /*
5329 * Changing the leftmost offset (via partial or whole
5330 * record truncate) of an interior (or rightmost) path
5331 * means we have to update the subtree that is formed
5332 * by this leaf and the one to it's left.
5333 *
5334 * There are two cases we can skip:
5335 * 1) Path is the leftmost one in our btree.
5336 * 2) The leaf is rightmost and will be empty after
5337 * we remove the extent record - the rotate code
5338 * knows how to update the newly formed edge.
5339 */
5345 }
5353 }
5356 left_cpos);
5360 }
5361 }
5362 }
5366 path);
5370 }
5376 }
5382 }
5396 /*
5397 * We skip the edge update if this path will
5398 * be deleted by the rotate code.
5399 */
5402 rec);
5403 }
5405 /* Remove leftmost portion of the record. */
5410 /* Remove rightmost portion of the record */
5415 /* Caller should have trapped this. */
5422 }
5429 subtree_index);
5430 }
5438 }
5440 out:
5443 }
5450 {
5457 /*
5458 * XXX: Why are we truncating to 0 instead of wherever this
5459 * affects us?
5460 */
5468 }
5474 }
5480 "Owner %llu has an extent at cpos %u which can no "
5483 cpos);
5486 }
5488 /*
5489 * We have 3 cases of extent removal:
5490 * 1) Range covers the entire extent rec
5491 * 2) Range begins or ends on one edge of the extent rec
5492 * 3) Range is in the middle of the extent rec (no shared edges)
5493 *
5494 * For case 1 we remove the extent rec and left rotate to
5495 * fill the hole.
5496 *
5497 * For case 2 we just shrink the existing extent rec, with a
5498 * tree update if the shrinking edge is also the edge of an
5499 * extent block.
5500 *
5501 * For case 3 we do a right split to turn the extent rec into
5502 * something case 2 can handle.
5503 */
5521 }
5528 }
5530 /*
5531 * The split could have manipulated the tree enough to
5532 * move the record location, so we have to look for it again.
5533 */
5540 }
5548 cpos);
5551 }
5553 /*
5554 * Double check our values here. If anything is fishy,
5555 * it's easier to catch it at the top level.
5556 */
5562 "Owner %llu: error after split at cpos %u"
5569 }
5576 }
5577 }
5579 out:
5582 }
5584 /*
5585 * ocfs2_reserve_blocks_for_rec_trunc() would look basically the
5586 * same as ocfs2_lock_alloctors(), except for it accepts a blocks
5587 * number to reserve some extra blocks, and it only handles meta
5588 * data allocations.
5589 *
5590 * Currently, only ocfs2_remove_btree_range() uses it for truncating
5591 * and punching holes.
5592 */
5595 u32 extents_to_split,
5598 {
5610 }
5622 }
5623 }
5625 out:
5630 }
5631 }
5634 }
5640 u64 refcount_loc)
5641 {
5652 OCFS2_HAS_REFCOUNT_FL));
5659 }
5662 refcount_loc,
5663 phys_blkno,
5664 len,
5670 }
5671 }
5674 extra_blocks);
5678 }
5687 }
5688 }
5696 }
5699 OCFS2_JOURNAL_ACCESS_WRITE);
5703 }
5712 }
5722 phys_blkno),
5725 else
5731 }
5733 out_commit:
5735 out:
5745 }
5748 {
5757 "slot %d, invalid truncate log parameters: used = "
5761 }
5765 {
5769 /* No records, nothing to coalesce */
5778 }
5782 u64 start_blk,
5784 {
5798 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5799 * by the underlying call to ocfs2_read_inode_block(), so any
5800 * corruption is a code bug */
5807 "Truncate record count on #%llu invalid "
5813 /* Caller should have known to flush before calling us. */
5819 }
5822 OCFS2_JOURNAL_ACCESS_WRITE);
5826 }
5832 /*
5833 * Move index back to the record we are coalescing with.
5834 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5835 */
5836 index--;
5842 num_clusters);
5846 }
5852 bail:
5854 }
5860 {
5864 u64 start_blk;
5875 /* Caller has given us at least enough credits to
5876 * update the truncate log dinode */
5878 OCFS2_JOURNAL_ACCESS_WRITE);
5882 }
5888 /* TODO: Perhaps we can calculate the bulk of the
5889 * credits up front rather than extending like
5890 * this. */
5892 OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
5896 }
5903 /* if start_blk is not set, we ignore the record as
5904 * invalid. */
5912 num_clusters);
5916 }
5917 }
5918 i--;
5919 }
5923 bail:
5925 }
5927 /* Expects you to already be holding tl_inode->i_mutex */
5929 {
5944 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5945 * by the underlying call to ocfs2_read_inode_block(), so any
5946 * corruption is a code bug */
5953 num_to_flush);
5957 }
5960 GLOBAL_BITMAP_SYSTEM_INODE,
5961 OCFS2_INVALID_SLOT);
5966 }
5974 }
5981 }
5984 data_alloc_bh);
5990 out_unlock:
5994 out_mutex:
5998 out:
6000 }
6003 {
6012 }
6015 {
6024 else
6026 }
6028 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
6031 {
6033 /* We want to push off log flushes while truncates are
6034 * still running. */
6039 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
6040 }
6041 }
6047 {
6053 TRUNCATE_LOG_SYSTEM_INODE,
6054 slot_num);
6059 }
6066 }
6070 bail:
6072 }
6074 /* called during the 1st stage of node recovery. we stamp a clean
6075 * truncate log and pass back a copy for processing later. if the
6076 * truncate log does not require processing, a *tl_copy is set to
6077 * NULL. */
6081 {
6096 }
6100 /* tl_bh is loaded from ocfs2_get_truncate_log_info(). It's
6101 * validated by the underlying call to ocfs2_read_inode_block(),
6102 * so any corruption is a code bug */
6114 }
6116 /* Assuming the write-out below goes well, this copy
6117 * will be passed back to recovery for processing. */
6120 /* All we need to do to clear the truncate log is set
6121 * tl_used. */
6129 }
6130 }
6132 bail:
6141 }
6144 }
6148 {
6152 u64 start_blk;
6160 }
6166 num_recs);
6175 }
6176 }
6183 }
6195 }
6196 }
6198 bail_up:
6202 }
6205 {
6219 }
6220 }
6223 {
6235 /* ocfs2_truncate_log_shutdown keys on the existence of
6236 * osb->osb_tl_inode so we don't set any of the osb variables
6237 * until we're sure all is well. */
6239 ocfs2_truncate_log_worker);
6244 }
6246 /*
6247 * Delayed de-allocation of suballocator blocks.
6248 *
6249 * Some sets of block de-allocations might involve multiple suballocator inodes.
6250 *
6251 * The locking for this can get extremely complicated, especially when
6252 * the suballocator inodes to delete from aren't known until deep
6253 * within an unrelated codepath.
6254 *
6255 * ocfs2_extent_block structures are a good example of this - an inode
6256 * btree could have been grown by any number of nodes each allocating
6257 * out of their own suballoc inode.
6258 *
6259 * These structures allow the delay of block de-allocation until a
6260 * later time, when locking of multiple cluster inodes won't cause
6261 * deadlock.
6262 */
6264 /*
6265 * Describe a single bit freed from a suballocator. For the block
6266 * suballocators, it represents one block. For the global cluster
6267 * allocator, it represents some clusters and free_bit indicates
6268 * clusters number.
6269 */
6272 u64 free_bg;
6273 u64 free_blk;
6275 };
6282 };
6288 {
6290 u64 bg_blkno;
6301 }
6309 }
6316 }
6321 else
6332 }
6338 }
6343 }
6345 out_journal:
6348 out_unlock:
6351 out_mutex:
6354 out:
6356 /* Premature exit may have left some dangling items. */
6360 }
6363 }
6367 {
6376 }
6386 }
6390 {
6404 }
6405 }
6412 }
6425 }
6426 }
6431 /* Premature exit may have left some dangling items. */
6435 }
6438 }
6442 {
6463 }
6467 }
6478 }
6481 }
6487 {
6495 }
6505 }
6507 }
6512 {
6522 }
6529 }
6543 out:
6545 }
6549 {
6555 }
6558 {
6562 }
6567 {
6577 /*
6578 * Need to set the buffers we zero'd into uptodate
6579 * here if they aren't - ocfs2_map_page_blocks()
6580 * might've skipped some
6581 */
6584 ocfs2_zero_func);
6591 }
6597 }
6602 {
6628 }
6629 out:
6632 }
6636 {
6640 loff_t last_page_bytes;
6653 }
6655 numpages++;
6656 index++;
6659 out:
6664 }
6669 }
6673 {
6680 }
6682 /*
6683 * Zero the area past i_size but still within an allocated
6684 * cluster. This avoids exposing nonzero data on subsequent file
6685 * extends.
6686 *
6687 * We need to call this before i_size is updated on the inode because
6688 * otherwise block_write_full_page() will skip writeout of pages past
6689 * i_size. The new_i_size parameter is passed for this reason.
6690 */
6693 {
6696 u64 phys;
6700 /*
6701 * File systems which don't support sparse files zero on every
6702 * extend.
6703 */
6713 }
6724 }
6726 /*
6727 * Tail is a hole, or is marked unwritten. In either case, we
6728 * can count on read and write to return/push zero's.
6729 */
6738 }
6743 /*
6744 * Initiate writeout of the pages we zero'd here. We don't
6745 * wait on them - the truncate_inode_pages() call later will
6746 * do that for us.
6747 */
6753 out:
6758 }
6762 {
6769 xattrsize);
6770 else
6773 }
6777 {
6783 }
6786 {
6795 /*
6796 * We clear the entire i_data structure here so that all
6797 * fields can be properly initialized.
6798 */
6803 }
6807 {
6829 }
6835 }
6836 }
6844 }
6847 OCFS2_JOURNAL_ACCESS_WRITE);
6851 }
6856 u64 phys;
6871 }
6873 /*
6874 * Save two copies, one for insert, and one that can
6875 * be changed by ocfs2_map_and_dirty_page() below.
6876 */
6879 /*
6880 * Non sparse file systems zero on extend, so no need
6881 * to do that now.
6882 */
6891 }
6893 /*
6894 * This should populate the 1st page for us and mark
6895 * it up to date.
6896 */
6901 }
6910 }
6922 /*
6923 * An error at this point should be extremely rare. If
6924 * this proves to be false, we could always re-build
6925 * the in-inode data from our pages.
6926 */
6932 }
6935 }
6937 out_commit:
6944 out_unlock:
6948 out:
6952 }
6955 }
6957 /*
6958 * It is expected, that by the time you call this function,
6959 * inode->i_size and fe->i_size have been adjusted.
6960 *
6961 * WARNING: This will kfree the truncate context
6962 */
6966 {
6986 ocfs2_journal_access_di);
6991 }
6995 start:
6996 /*
6997 * Check that we still have allocation to delete.
6998 */
7002 }
7004 /*
7005 * Truncate always works against the rightmost tree branch.
7006 */
7011 }
7015 new_highest_cpos,
7019 /*
7020 * By now, el will point to the extent list on the bottom most
7021 * portion of this tree. Only the tail record is considered in
7022 * each pass.
7023 *
7024 * We handle the following cases, in order:
7025 * - empty extent: delete the remaining branch
7026 * - remove the entire record
7027 * - remove a partial record
7028 * - no record needs to be removed (truncate has completed)
7029 */
7038 }
7046 /*
7047 * Lower levels depend on this never happening, but it's best
7048 * to check it up here before changing the tree.
7049 */
7056 }
7061 /*
7062 * Truncate entire record.
7063 */
7068 /*
7069 * Partial truncate. it also should be
7070 * the last truncate we're doing.
7071 */
7078 /*
7079 * Truncate completed, leave happily.
7080 */
7083 }
7089 refcount_loc);
7093 }
7097 /*
7098 * The check above will catch the case where we've truncated
7099 * away all allocation.
7100 */
7103 bail:
7112 }
7114 /*
7115 * 'start' is inclusive, 'end' is not.
7116 */
7119 {
7136 "Inline data flags for inode %llu don't agree! "
7144 }
7151 }
7154 OCFS2_JOURNAL_ACCESS_WRITE);
7158 }
7163 /*
7164 * No need to worry about the data page here - it's been
7165 * truncated already and inline data doesn't need it for
7166 * pushing zero's to disk, so we'll let readpage pick it up
7167 * later.
7168 */
7172 }
7182 out_commit:
7185 out:
7187 }
7192 {
7202 }
7207 {
7229 }
7231 }
7237 }
7241 }
7247 }
7250 {
7269 }
7275 GLOBAL_BITMAP_SYSTEM_INODE,
7276 OCFS2_INVALID_SLOT);
7281 }
7289 }
7295 }
7302 /* Determine first and last group to examine based on start and len */
7306 else
7314 else
7323 }
7333 }
7340 else
7342 }
7344 out_unlock:
7347 out_mutex:
7350 out:
7352 }