aboutsummaryrefslogtreecommitdiffstats
path: root/fs/ocfs2/alloc.c
diff options
context:
space:
mode:
Diffstat (limited to 'fs/ocfs2/alloc.c')
-rw-r--r--fs/ocfs2/alloc.c3037
1 files changed, 2445 insertions, 592 deletions
diff --git a/fs/ocfs2/alloc.c b/fs/ocfs2/alloc.c
index f27e5378caf2..a0c8667caa72 100644
--- a/fs/ocfs2/alloc.c
+++ b/fs/ocfs2/alloc.c
@@ -27,6 +27,7 @@
27#include <linux/types.h> 27#include <linux/types.h>
28#include <linux/slab.h> 28#include <linux/slab.h>
29#include <linux/highmem.h> 29#include <linux/highmem.h>
30#include <linux/swap.h>
30 31
31#define MLOG_MASK_PREFIX ML_DISK_ALLOC 32#define MLOG_MASK_PREFIX ML_DISK_ALLOC
32#include <cluster/masklog.h> 33#include <cluster/masklog.h>
@@ -34,6 +35,7 @@
34#include "ocfs2.h" 35#include "ocfs2.h"
35 36
36#include "alloc.h" 37#include "alloc.h"
38#include "aops.h"
37#include "dlmglue.h" 39#include "dlmglue.h"
38#include "extent_map.h" 40#include "extent_map.h"
39#include "inode.h" 41#include "inode.h"
@@ -47,63 +49,243 @@
47 49
48#include "buffer_head_io.h" 50#include "buffer_head_io.h"
49 51
50static int ocfs2_extent_contig(struct inode *inode, 52static void ocfs2_free_truncate_context(struct ocfs2_truncate_context *tc);
51 struct ocfs2_extent_rec *ext,
52 u64 blkno);
53 53
54static int ocfs2_create_new_meta_bhs(struct ocfs2_super *osb, 54/*
55 handle_t *handle, 55 * Structures which describe a path through a btree, and functions to
56 struct inode *inode, 56 * manipulate them.
57 int wanted, 57 *
58 struct ocfs2_alloc_context *meta_ac, 58 * The idea here is to be as generic as possible with the tree
59 struct buffer_head *bhs[]); 59 * manipulation code.
60 */
61struct ocfs2_path_item {
62 struct buffer_head *bh;
63 struct ocfs2_extent_list *el;
64};
60 65
61static int ocfs2_add_branch(struct ocfs2_super *osb, 66#define OCFS2_MAX_PATH_DEPTH 5
62 handle_t *handle,
63 struct inode *inode,
64 struct buffer_head *fe_bh,
65 struct buffer_head *eb_bh,
66 struct buffer_head *last_eb_bh,
67 struct ocfs2_alloc_context *meta_ac);
68 67
69static int ocfs2_shift_tree_depth(struct ocfs2_super *osb, 68struct ocfs2_path {
70 handle_t *handle, 69 int p_tree_depth;
71 struct inode *inode, 70 struct ocfs2_path_item p_node[OCFS2_MAX_PATH_DEPTH];
72 struct buffer_head *fe_bh, 71};
73 struct ocfs2_alloc_context *meta_ac,
74 struct buffer_head **ret_new_eb_bh);
75 72
76static int ocfs2_do_insert_extent(struct ocfs2_super *osb, 73#define path_root_bh(_path) ((_path)->p_node[0].bh)
77 handle_t *handle, 74#define path_root_el(_path) ((_path)->p_node[0].el)
78 struct inode *inode, 75#define path_leaf_bh(_path) ((_path)->p_node[(_path)->p_tree_depth].bh)
79 struct buffer_head *fe_bh, 76#define path_leaf_el(_path) ((_path)->p_node[(_path)->p_tree_depth].el)
80 u64 blkno, 77#define path_num_items(_path) ((_path)->p_tree_depth + 1)
81 u32 new_clusters);
82 78
83static int ocfs2_find_branch_target(struct ocfs2_super *osb, 79/*
84 struct inode *inode, 80 * Reset the actual path elements so that we can re-use the structure
85 struct buffer_head *fe_bh, 81 * to build another path. Generally, this involves freeing the buffer
86 struct buffer_head **target_bh); 82 * heads.
83 */
84static void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root)
85{
86 int i, start = 0, depth = 0;
87 struct ocfs2_path_item *node;
87 88
88static int ocfs2_find_new_last_ext_blk(struct ocfs2_super *osb, 89 if (keep_root)
89 struct inode *inode, 90 start = 1;
90 struct ocfs2_dinode *fe, 91
91 unsigned int new_i_clusters, 92 for(i = start; i < path_num_items(path); i++) {
92 struct buffer_head *old_last_eb, 93 node = &path->p_node[i];
93 struct buffer_head **new_last_eb); 94
95 brelse(node->bh);
96 node->bh = NULL;
97 node->el = NULL;
98 }
99
100 /*
101 * Tree depth may change during truncate, or insert. If we're
102 * keeping the root extent list, then make sure that our path
103 * structure reflects the proper depth.
104 */
105 if (keep_root)
106 depth = le16_to_cpu(path_root_el(path)->l_tree_depth);
107
108 path->p_tree_depth = depth;
109}
110
111static void ocfs2_free_path(struct ocfs2_path *path)
112{
113 if (path) {
114 ocfs2_reinit_path(path, 0);
115 kfree(path);
116 }
117}
118
119/*
120 * Make the *dest path the same as src and re-initialize src path to
121 * have a root only.
122 */
123static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src)
124{
125 int i;
126
127 BUG_ON(path_root_bh(dest) != path_root_bh(src));
128
129 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
130 brelse(dest->p_node[i].bh);
131
132 dest->p_node[i].bh = src->p_node[i].bh;
133 dest->p_node[i].el = src->p_node[i].el;
134
135 src->p_node[i].bh = NULL;
136 src->p_node[i].el = NULL;
137 }
138}
139
140/*
141 * Insert an extent block at given index.
142 *
143 * This will not take an additional reference on eb_bh.
144 */
145static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index,
146 struct buffer_head *eb_bh)
147{
148 struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data;
149
150 /*
151 * Right now, no root bh is an extent block, so this helps
152 * catch code errors with dinode trees. The assertion can be
153 * safely removed if we ever need to insert extent block
154 * structures at the root.
155 */
156 BUG_ON(index == 0);
157
158 path->p_node[index].bh = eb_bh;
159 path->p_node[index].el = &eb->h_list;
160}
161
162static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh,
163 struct ocfs2_extent_list *root_el)
164{
165 struct ocfs2_path *path;
166
167 BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH);
168
169 path = kzalloc(sizeof(*path), GFP_NOFS);
170 if (path) {
171 path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth);
172 get_bh(root_bh);
173 path_root_bh(path) = root_bh;
174 path_root_el(path) = root_el;
175 }
176
177 return path;
178}
179
180/*
181 * Allocate and initialize a new path based on a disk inode tree.
182 */
183static struct ocfs2_path *ocfs2_new_inode_path(struct buffer_head *di_bh)
184{
185 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
186 struct ocfs2_extent_list *el = &di->id2.i_list;
187
188 return ocfs2_new_path(di_bh, el);
189}
190
191/*
192 * Convenience function to journal all components in a path.
193 */
194static int ocfs2_journal_access_path(struct inode *inode, handle_t *handle,
195 struct ocfs2_path *path)
196{
197 int i, ret = 0;
198
199 if (!path)
200 goto out;
201
202 for(i = 0; i < path_num_items(path); i++) {
203 ret = ocfs2_journal_access(handle, inode, path->p_node[i].bh,
204 OCFS2_JOURNAL_ACCESS_WRITE);
205 if (ret < 0) {
206 mlog_errno(ret);
207 goto out;
208 }
209 }
210
211out:
212 return ret;
213}
214
215enum ocfs2_contig_type {
216 CONTIG_NONE = 0,
217 CONTIG_LEFT,
218 CONTIG_RIGHT
219};
94 220
95static void ocfs2_free_truncate_context(struct ocfs2_truncate_context *tc);
96 221
97static int ocfs2_extent_contig(struct inode *inode, 222/*
98 struct ocfs2_extent_rec *ext, 223 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
99 u64 blkno) 224 * ocfs2_extent_contig only work properly against leaf nodes!
225 */
226static int ocfs2_block_extent_contig(struct super_block *sb,
227 struct ocfs2_extent_rec *ext,
228 u64 blkno)
229{
230 u64 blk_end = le64_to_cpu(ext->e_blkno);
231
232 blk_end += ocfs2_clusters_to_blocks(sb,
233 le16_to_cpu(ext->e_leaf_clusters));
234
235 return blkno == blk_end;
236}
237
238static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
239 struct ocfs2_extent_rec *right)
240{
241 u32 left_range;
242
243 left_range = le32_to_cpu(left->e_cpos) +
244 le16_to_cpu(left->e_leaf_clusters);
245
246 return (left_range == le32_to_cpu(right->e_cpos));
247}
248
249static enum ocfs2_contig_type
250 ocfs2_extent_contig(struct inode *inode,
251 struct ocfs2_extent_rec *ext,
252 struct ocfs2_extent_rec *insert_rec)
100{ 253{
101 return blkno == (le64_to_cpu(ext->e_blkno) + 254 u64 blkno = le64_to_cpu(insert_rec->e_blkno);
102 ocfs2_clusters_to_blocks(inode->i_sb, 255
103 le32_to_cpu(ext->e_clusters))); 256 if (ocfs2_extents_adjacent(ext, insert_rec) &&
257 ocfs2_block_extent_contig(inode->i_sb, ext, blkno))
258 return CONTIG_RIGHT;
259
260 blkno = le64_to_cpu(ext->e_blkno);
261 if (ocfs2_extents_adjacent(insert_rec, ext) &&
262 ocfs2_block_extent_contig(inode->i_sb, insert_rec, blkno))
263 return CONTIG_LEFT;
264
265 return CONTIG_NONE;
104} 266}
105 267
106/* 268/*
269 * NOTE: We can have pretty much any combination of contiguousness and
270 * appending.
271 *
272 * The usefulness of APPEND_TAIL is more in that it lets us know that
273 * we'll have to update the path to that leaf.
274 */
275enum ocfs2_append_type {
276 APPEND_NONE = 0,
277 APPEND_TAIL,
278};
279
280struct ocfs2_insert_type {
281 enum ocfs2_append_type ins_appending;
282 enum ocfs2_contig_type ins_contig;
283 int ins_contig_index;
284 int ins_free_records;
285 int ins_tree_depth;
286};
287
288/*
107 * How many free extents have we got before we need more meta data? 289 * How many free extents have we got before we need more meta data?
108 */ 290 */
109int ocfs2_num_free_extents(struct ocfs2_super *osb, 291int ocfs2_num_free_extents(struct ocfs2_super *osb,
@@ -242,6 +424,28 @@ bail:
242} 424}
243 425
244/* 426/*
427 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
428 *
429 * Returns the sum of the rightmost extent rec logical offset and
430 * cluster count.
431 *
432 * ocfs2_add_branch() uses this to determine what logical cluster
433 * value should be populated into the leftmost new branch records.
434 *
435 * ocfs2_shift_tree_depth() uses this to determine the # clusters
436 * value for the new topmost tree record.
437 */
438static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list *el)
439{
440 int i;
441
442 i = le16_to_cpu(el->l_next_free_rec) - 1;
443
444 return le32_to_cpu(el->l_recs[i].e_cpos) +
445 ocfs2_rec_clusters(el, &el->l_recs[i]);
446}
447
448/*
245 * Add an entire tree branch to our inode. eb_bh is the extent block 449 * Add an entire tree branch to our inode. eb_bh is the extent block
246 * to start at, if we don't want to start the branch at the dinode 450 * to start at, if we don't want to start the branch at the dinode
247 * structure. 451 * structure.
@@ -250,7 +454,7 @@ bail:
250 * for the new last extent block. 454 * for the new last extent block.
251 * 455 *
252 * the new branch will be 'empty' in the sense that every block will 456 * the new branch will be 'empty' in the sense that every block will
253 * contain a single record with e_clusters == 0. 457 * contain a single record with cluster count == 0.
254 */ 458 */
255static int ocfs2_add_branch(struct ocfs2_super *osb, 459static int ocfs2_add_branch(struct ocfs2_super *osb,
256 handle_t *handle, 460 handle_t *handle,
@@ -268,6 +472,7 @@ static int ocfs2_add_branch(struct ocfs2_super *osb,
268 struct ocfs2_extent_block *eb; 472 struct ocfs2_extent_block *eb;
269 struct ocfs2_extent_list *eb_el; 473 struct ocfs2_extent_list *eb_el;
270 struct ocfs2_extent_list *el; 474 struct ocfs2_extent_list *el;
475 u32 new_cpos;
271 476
272 mlog_entry_void(); 477 mlog_entry_void();
273 478
@@ -302,6 +507,9 @@ static int ocfs2_add_branch(struct ocfs2_super *osb,
302 goto bail; 507 goto bail;
303 } 508 }
304 509
510 eb = (struct ocfs2_extent_block *)last_eb_bh->b_data;
511 new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
512
305 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be 513 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
306 * linked with the rest of the tree. 514 * linked with the rest of the tree.
307 * conversly, new_eb_bhs[0] is the new bottommost leaf. 515 * conversly, new_eb_bhs[0] is the new bottommost leaf.
@@ -330,9 +538,18 @@ static int ocfs2_add_branch(struct ocfs2_super *osb,
330 eb->h_next_leaf_blk = 0; 538 eb->h_next_leaf_blk = 0;
331 eb_el->l_tree_depth = cpu_to_le16(i); 539 eb_el->l_tree_depth = cpu_to_le16(i);
332 eb_el->l_next_free_rec = cpu_to_le16(1); 540 eb_el->l_next_free_rec = cpu_to_le16(1);
333 eb_el->l_recs[0].e_cpos = fe->i_clusters; 541 /*
542 * This actually counts as an empty extent as
543 * c_clusters == 0
544 */
545 eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
334 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno); 546 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
335 eb_el->l_recs[0].e_clusters = cpu_to_le32(0); 547 /*
548 * eb_el isn't always an interior node, but even leaf
549 * nodes want a zero'd flags and reserved field so
550 * this gets the whole 32 bits regardless of use.
551 */
552 eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
336 if (!eb_el->l_tree_depth) 553 if (!eb_el->l_tree_depth)
337 new_last_eb_blk = le64_to_cpu(eb->h_blkno); 554 new_last_eb_blk = le64_to_cpu(eb->h_blkno);
338 555
@@ -376,8 +593,8 @@ static int ocfs2_add_branch(struct ocfs2_super *osb,
376 * either be on the fe, or the extent block passed in. */ 593 * either be on the fe, or the extent block passed in. */
377 i = le16_to_cpu(el->l_next_free_rec); 594 i = le16_to_cpu(el->l_next_free_rec);
378 el->l_recs[i].e_blkno = cpu_to_le64(next_blkno); 595 el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
379 el->l_recs[i].e_cpos = fe->i_clusters; 596 el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
380 el->l_recs[i].e_clusters = 0; 597 el->l_recs[i].e_int_clusters = 0;
381 le16_add_cpu(&el->l_next_free_rec, 1); 598 le16_add_cpu(&el->l_next_free_rec, 1);
382 599
383 /* fe needs a new last extent block pointer, as does the 600 /* fe needs a new last extent block pointer, as does the
@@ -425,6 +642,7 @@ static int ocfs2_shift_tree_depth(struct ocfs2_super *osb,
425 struct buffer_head **ret_new_eb_bh) 642 struct buffer_head **ret_new_eb_bh)
426{ 643{
427 int status, i; 644 int status, i;
645 u32 new_clusters;
428 struct buffer_head *new_eb_bh = NULL; 646 struct buffer_head *new_eb_bh = NULL;
429 struct ocfs2_dinode *fe; 647 struct ocfs2_dinode *fe;
430 struct ocfs2_extent_block *eb; 648 struct ocfs2_extent_block *eb;
@@ -461,11 +679,8 @@ static int ocfs2_shift_tree_depth(struct ocfs2_super *osb,
461 /* copy the fe data into the new extent block */ 679 /* copy the fe data into the new extent block */
462 eb_el->l_tree_depth = fe_el->l_tree_depth; 680 eb_el->l_tree_depth = fe_el->l_tree_depth;
463 eb_el->l_next_free_rec = fe_el->l_next_free_rec; 681 eb_el->l_next_free_rec = fe_el->l_next_free_rec;
464 for(i = 0; i < le16_to_cpu(fe_el->l_next_free_rec); i++) { 682 for(i = 0; i < le16_to_cpu(fe_el->l_next_free_rec); i++)
465 eb_el->l_recs[i].e_cpos = fe_el->l_recs[i].e_cpos; 683 eb_el->l_recs[i] = fe_el->l_recs[i];
466 eb_el->l_recs[i].e_clusters = fe_el->l_recs[i].e_clusters;
467 eb_el->l_recs[i].e_blkno = fe_el->l_recs[i].e_blkno;
468 }
469 684
470 status = ocfs2_journal_dirty(handle, new_eb_bh); 685 status = ocfs2_journal_dirty(handle, new_eb_bh);
471 if (status < 0) { 686 if (status < 0) {
@@ -480,16 +695,15 @@ static int ocfs2_shift_tree_depth(struct ocfs2_super *osb,
480 goto bail; 695 goto bail;
481 } 696 }
482 697
698 new_clusters = ocfs2_sum_rightmost_rec(eb_el);
699
483 /* update fe now */ 700 /* update fe now */
484 le16_add_cpu(&fe_el->l_tree_depth, 1); 701 le16_add_cpu(&fe_el->l_tree_depth, 1);
485 fe_el->l_recs[0].e_cpos = 0; 702 fe_el->l_recs[0].e_cpos = 0;
486 fe_el->l_recs[0].e_blkno = eb->h_blkno; 703 fe_el->l_recs[0].e_blkno = eb->h_blkno;
487 fe_el->l_recs[0].e_clusters = fe->i_clusters; 704 fe_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
488 for(i = 1; i < le16_to_cpu(fe_el->l_next_free_rec); i++) { 705 for(i = 1; i < le16_to_cpu(fe_el->l_next_free_rec); i++)
489 fe_el->l_recs[i].e_cpos = 0; 706 memset(&fe_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
490 fe_el->l_recs[i].e_clusters = 0;
491 fe_el->l_recs[i].e_blkno = 0;
492 }
493 fe_el->l_next_free_rec = cpu_to_le16(1); 707 fe_el->l_next_free_rec = cpu_to_le16(1);
494 708
495 /* If this is our 1st tree depth shift, then last_eb_blk 709 /* If this is our 1st tree depth shift, then last_eb_blk
@@ -515,199 +729,6 @@ bail:
515} 729}
516 730
517/* 731/*
518 * Expects the tree to already have room in the rightmost leaf for the
519 * extent. Updates all the extent blocks (and the dinode) on the way
520 * down.
521 */
522static int ocfs2_do_insert_extent(struct ocfs2_super *osb,
523 handle_t *handle,
524 struct inode *inode,
525 struct buffer_head *fe_bh,
526 u64 start_blk,
527 u32 new_clusters)
528{
529 int status, i, num_bhs = 0;
530 u64 next_blkno;
531 u16 next_free;
532 struct buffer_head **eb_bhs = NULL;
533 struct ocfs2_dinode *fe;
534 struct ocfs2_extent_block *eb;
535 struct ocfs2_extent_list *el;
536
537 mlog_entry_void();
538
539 status = ocfs2_journal_access(handle, inode, fe_bh,
540 OCFS2_JOURNAL_ACCESS_WRITE);
541 if (status < 0) {
542 mlog_errno(status);
543 goto bail;
544 }
545
546 fe = (struct ocfs2_dinode *) fe_bh->b_data;
547 el = &fe->id2.i_list;
548 if (el->l_tree_depth) {
549 /* This is another operation where we want to be
550 * careful about our tree updates. An error here means
551 * none of the previous changes we made should roll
552 * forward. As a result, we have to record the buffers
553 * for this part of the tree in an array and reserve a
554 * journal write to them before making any changes. */
555 num_bhs = le16_to_cpu(fe->id2.i_list.l_tree_depth);
556 eb_bhs = kcalloc(num_bhs, sizeof(struct buffer_head *),
557 GFP_KERNEL);
558 if (!eb_bhs) {
559 status = -ENOMEM;
560 mlog_errno(status);
561 goto bail;
562 }
563
564 i = 0;
565 while(el->l_tree_depth) {
566 next_free = le16_to_cpu(el->l_next_free_rec);
567 if (next_free == 0) {
568 ocfs2_error(inode->i_sb,
569 "Dinode %llu has a bad extent list",
570 (unsigned long long)OCFS2_I(inode)->ip_blkno);
571 status = -EIO;
572 goto bail;
573 }
574 next_blkno = le64_to_cpu(el->l_recs[next_free - 1].e_blkno);
575
576 BUG_ON(i >= num_bhs);
577 status = ocfs2_read_block(osb, next_blkno, &eb_bhs[i],
578 OCFS2_BH_CACHED, inode);
579 if (status < 0) {
580 mlog_errno(status);
581 goto bail;
582 }
583 eb = (struct ocfs2_extent_block *) eb_bhs[i]->b_data;
584 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
585 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb,
586 eb);
587 status = -EIO;
588 goto bail;
589 }
590
591 status = ocfs2_journal_access(handle, inode, eb_bhs[i],
592 OCFS2_JOURNAL_ACCESS_WRITE);
593 if (status < 0) {
594 mlog_errno(status);
595 goto bail;
596 }
597
598 el = &eb->h_list;
599 i++;
600 /* When we leave this loop, eb_bhs[num_bhs - 1] will
601 * hold the bottom-most leaf extent block. */
602 }
603 BUG_ON(el->l_tree_depth);
604
605 el = &fe->id2.i_list;
606 /* If we have tree depth, then the fe update is
607 * trivial, and we want to switch el out for the
608 * bottom-most leaf in order to update it with the
609 * actual extent data below. */
610 next_free = le16_to_cpu(el->l_next_free_rec);
611 if (next_free == 0) {
612 ocfs2_error(inode->i_sb,
613 "Dinode %llu has a bad extent list",
614 (unsigned long long)OCFS2_I(inode)->ip_blkno);
615 status = -EIO;
616 goto bail;
617 }
618 le32_add_cpu(&el->l_recs[next_free - 1].e_clusters,
619 new_clusters);
620 /* (num_bhs - 1) to avoid the leaf */
621 for(i = 0; i < (num_bhs - 1); i++) {
622 eb = (struct ocfs2_extent_block *) eb_bhs[i]->b_data;
623 el = &eb->h_list;
624
625 /* finally, make our actual change to the
626 * intermediate extent blocks. */
627 next_free = le16_to_cpu(el->l_next_free_rec);
628 le32_add_cpu(&el->l_recs[next_free - 1].e_clusters,
629 new_clusters);
630
631 status = ocfs2_journal_dirty(handle, eb_bhs[i]);
632 if (status < 0)
633 mlog_errno(status);
634 }
635 BUG_ON(i != (num_bhs - 1));
636 /* note that the leaf block wasn't touched in
637 * the loop above */
638 eb = (struct ocfs2_extent_block *) eb_bhs[num_bhs - 1]->b_data;
639 el = &eb->h_list;
640 BUG_ON(el->l_tree_depth);
641 }
642
643 /* yay, we can finally add the actual extent now! */
644 i = le16_to_cpu(el->l_next_free_rec) - 1;
645 if (le16_to_cpu(el->l_next_free_rec) &&
646 ocfs2_extent_contig(inode, &el->l_recs[i], start_blk)) {
647 le32_add_cpu(&el->l_recs[i].e_clusters, new_clusters);
648 } else if (le16_to_cpu(el->l_next_free_rec) &&
649 (le32_to_cpu(el->l_recs[i].e_clusters) == 0)) {
650 /* having an empty extent at eof is legal. */
651 if (el->l_recs[i].e_cpos != fe->i_clusters) {
652 ocfs2_error(inode->i_sb,
653 "Dinode %llu trailing extent is bad: "
654 "cpos (%u) != number of clusters (%u)",
655 (unsigned long long)OCFS2_I(inode)->ip_blkno,
656 le32_to_cpu(el->l_recs[i].e_cpos),
657 le32_to_cpu(fe->i_clusters));
658 status = -EIO;
659 goto bail;
660 }
661 el->l_recs[i].e_blkno = cpu_to_le64(start_blk);
662 el->l_recs[i].e_clusters = cpu_to_le32(new_clusters);
663 } else {
664 /* No contiguous record, or no empty record at eof, so
665 * we add a new one. */
666
667 BUG_ON(le16_to_cpu(el->l_next_free_rec) >=
668 le16_to_cpu(el->l_count));
669 i = le16_to_cpu(el->l_next_free_rec);
670
671 el->l_recs[i].e_blkno = cpu_to_le64(start_blk);
672 el->l_recs[i].e_clusters = cpu_to_le32(new_clusters);
673 el->l_recs[i].e_cpos = fe->i_clusters;
674 le16_add_cpu(&el->l_next_free_rec, 1);
675 }
676
677 /*
678 * extent_map errors are not fatal, so they are ignored outside
679 * of flushing the thing.
680 */
681 status = ocfs2_extent_map_append(inode, &el->l_recs[i],
682 new_clusters);
683 if (status) {
684 mlog_errno(status);
685 ocfs2_extent_map_drop(inode, le32_to_cpu(fe->i_clusters));
686 }
687
688 status = ocfs2_journal_dirty(handle, fe_bh);
689 if (status < 0)
690 mlog_errno(status);
691 if (fe->id2.i_list.l_tree_depth) {
692 status = ocfs2_journal_dirty(handle, eb_bhs[num_bhs - 1]);
693 if (status < 0)
694 mlog_errno(status);
695 }
696
697 status = 0;
698bail:
699 if (eb_bhs) {
700 for (i = 0; i < num_bhs; i++)
701 if (eb_bhs[i])
702 brelse(eb_bhs[i]);
703 kfree(eb_bhs);
704 }
705
706 mlog_exit(status);
707 return status;
708}
709
710/*
711 * Should only be called when there is no space left in any of the 732 * Should only be called when there is no space left in any of the
712 * leaf nodes. What we want to do is find the lowest tree depth 733 * leaf nodes. What we want to do is find the lowest tree depth
713 * non-leaf extent block with room for new records. There are three 734 * non-leaf extent block with room for new records. There are three
@@ -807,53 +828,1548 @@ bail:
807 return status; 828 return status;
808} 829}
809 830
810/* the caller needs to update fe->i_clusters */ 831/*
811int ocfs2_insert_extent(struct ocfs2_super *osb, 832 * This is only valid for leaf nodes, which are the only ones that can
812 handle_t *handle, 833 * have empty extents anyway.
813 struct inode *inode, 834 */
814 struct buffer_head *fe_bh, 835static inline int ocfs2_is_empty_extent(struct ocfs2_extent_rec *rec)
815 u64 start_blk,
816 u32 new_clusters,
817 struct ocfs2_alloc_context *meta_ac)
818{ 836{
819 int status, i, shift; 837 return !rec->e_leaf_clusters;
820 struct buffer_head *last_eb_bh = NULL; 838}
839
840/*
841 * This function will discard the rightmost extent record.
842 */
843static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
844{
845 int next_free = le16_to_cpu(el->l_next_free_rec);
846 int count = le16_to_cpu(el->l_count);
847 unsigned int num_bytes;
848
849 BUG_ON(!next_free);
850 /* This will cause us to go off the end of our extent list. */
851 BUG_ON(next_free >= count);
852
853 num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
854
855 memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
856}
857
858static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
859 struct ocfs2_extent_rec *insert_rec)
860{
861 int i, insert_index, next_free, has_empty, num_bytes;
862 u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
863 struct ocfs2_extent_rec *rec;
864
865 next_free = le16_to_cpu(el->l_next_free_rec);
866 has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
867
868 BUG_ON(!next_free);
869
870 /* The tree code before us didn't allow enough room in the leaf. */
871 if (el->l_next_free_rec == el->l_count && !has_empty)
872 BUG();
873
874 /*
875 * The easiest way to approach this is to just remove the
876 * empty extent and temporarily decrement next_free.
877 */
878 if (has_empty) {
879 /*
880 * If next_free was 1 (only an empty extent), this
881 * loop won't execute, which is fine. We still want
882 * the decrement above to happen.
883 */
884 for(i = 0; i < (next_free - 1); i++)
885 el->l_recs[i] = el->l_recs[i+1];
886
887 next_free--;
888 }
889
890 /*
891 * Figure out what the new record index should be.
892 */
893 for(i = 0; i < next_free; i++) {
894 rec = &el->l_recs[i];
895
896 if (insert_cpos < le32_to_cpu(rec->e_cpos))
897 break;
898 }
899 insert_index = i;
900
901 mlog(0, "ins %u: index %d, has_empty %d, next_free %d, count %d\n",
902 insert_cpos, insert_index, has_empty, next_free, le16_to_cpu(el->l_count));
903
904 BUG_ON(insert_index < 0);
905 BUG_ON(insert_index >= le16_to_cpu(el->l_count));
906 BUG_ON(insert_index > next_free);
907
908 /*
909 * No need to memmove if we're just adding to the tail.
910 */
911 if (insert_index != next_free) {
912 BUG_ON(next_free >= le16_to_cpu(el->l_count));
913
914 num_bytes = next_free - insert_index;
915 num_bytes *= sizeof(struct ocfs2_extent_rec);
916 memmove(&el->l_recs[insert_index + 1],
917 &el->l_recs[insert_index],
918 num_bytes);
919 }
920
921 /*
922 * Either we had an empty extent, and need to re-increment or
923 * there was no empty extent on a non full rightmost leaf node,
924 * in which case we still need to increment.
925 */
926 next_free++;
927 el->l_next_free_rec = cpu_to_le16(next_free);
928 /*
929 * Make sure none of the math above just messed up our tree.
930 */
931 BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
932
933 el->l_recs[insert_index] = *insert_rec;
934
935}
936
937/*
938 * Create an empty extent record .
939 *
940 * l_next_free_rec may be updated.
941 *
942 * If an empty extent already exists do nothing.
943 */
944static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
945{
946 int next_free = le16_to_cpu(el->l_next_free_rec);
947
948 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
949
950 if (next_free == 0)
951 goto set_and_inc;
952
953 if (ocfs2_is_empty_extent(&el->l_recs[0]))
954 return;
955
956 mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
957 "Asked to create an empty extent in a full list:\n"
958 "count = %u, tree depth = %u",
959 le16_to_cpu(el->l_count),
960 le16_to_cpu(el->l_tree_depth));
961
962 ocfs2_shift_records_right(el);
963
964set_and_inc:
965 le16_add_cpu(&el->l_next_free_rec, 1);
966 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
967}
968
969/*
970 * For a rotation which involves two leaf nodes, the "root node" is
971 * the lowest level tree node which contains a path to both leafs. This
972 * resulting set of information can be used to form a complete "subtree"
973 *
974 * This function is passed two full paths from the dinode down to a
975 * pair of adjacent leaves. It's task is to figure out which path
976 * index contains the subtree root - this can be the root index itself
977 * in a worst-case rotation.
978 *
979 * The array index of the subtree root is passed back.
980 */
981static int ocfs2_find_subtree_root(struct inode *inode,
982 struct ocfs2_path *left,
983 struct ocfs2_path *right)
984{
985 int i = 0;
986
987 /*
988 * Check that the caller passed in two paths from the same tree.
989 */
990 BUG_ON(path_root_bh(left) != path_root_bh(right));
991
992 do {
993 i++;
994
995 /*
996 * The caller didn't pass two adjacent paths.
997 */
998 mlog_bug_on_msg(i > left->p_tree_depth,
999 "Inode %lu, left depth %u, right depth %u\n"
1000 "left leaf blk %llu, right leaf blk %llu\n",
1001 inode->i_ino, left->p_tree_depth,
1002 right->p_tree_depth,
1003 (unsigned long long)path_leaf_bh(left)->b_blocknr,
1004 (unsigned long long)path_leaf_bh(right)->b_blocknr);
1005 } while (left->p_node[i].bh->b_blocknr ==
1006 right->p_node[i].bh->b_blocknr);
1007
1008 return i - 1;
1009}
1010
1011typedef void (path_insert_t)(void *, struct buffer_head *);
1012
1013/*
1014 * Traverse a btree path in search of cpos, starting at root_el.
1015 *
1016 * This code can be called with a cpos larger than the tree, in which
1017 * case it will return the rightmost path.
1018 */
1019static int __ocfs2_find_path(struct inode *inode,
1020 struct ocfs2_extent_list *root_el, u32 cpos,
1021 path_insert_t *func, void *data)
1022{
1023 int i, ret = 0;
1024 u32 range;
1025 u64 blkno;
821 struct buffer_head *bh = NULL; 1026 struct buffer_head *bh = NULL;
822 struct ocfs2_dinode *fe;
823 struct ocfs2_extent_block *eb; 1027 struct ocfs2_extent_block *eb;
824 struct ocfs2_extent_list *el; 1028 struct ocfs2_extent_list *el;
1029 struct ocfs2_extent_rec *rec;
1030 struct ocfs2_inode_info *oi = OCFS2_I(inode);
825 1031
826 mlog_entry_void(); 1032 el = root_el;
1033 while (el->l_tree_depth) {
1034 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1035 ocfs2_error(inode->i_sb,
1036 "Inode %llu has empty extent list at "
1037 "depth %u\n",
1038 (unsigned long long)oi->ip_blkno,
1039 le16_to_cpu(el->l_tree_depth));
1040 ret = -EROFS;
1041 goto out;
827 1042
828 mlog(0, "add %u clusters starting at block %llu to inode %llu\n", 1043 }
829 new_clusters, (unsigned long long)start_blk,
830 (unsigned long long)OCFS2_I(inode)->ip_blkno);
831 1044
832 fe = (struct ocfs2_dinode *) fe_bh->b_data; 1045 for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
833 el = &fe->id2.i_list; 1046 rec = &el->l_recs[i];
1047
1048 /*
1049 * In the case that cpos is off the allocation
1050 * tree, this should just wind up returning the
1051 * rightmost record.
1052 */
1053 range = le32_to_cpu(rec->e_cpos) +
1054 ocfs2_rec_clusters(el, rec);
1055 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1056 break;
1057 }
834 1058
835 if (el->l_tree_depth) { 1059 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
836 /* jump to end of tree */ 1060 if (blkno == 0) {
837 status = ocfs2_read_block(osb, le64_to_cpu(fe->i_last_eb_blk), 1061 ocfs2_error(inode->i_sb,
838 &last_eb_bh, OCFS2_BH_CACHED, inode); 1062 "Inode %llu has bad blkno in extent list "
839 if (status < 0) { 1063 "at depth %u (index %d)\n",
840 mlog_exit(status); 1064 (unsigned long long)oi->ip_blkno,
841 goto bail; 1065 le16_to_cpu(el->l_tree_depth), i);
1066 ret = -EROFS;
1067 goto out;
842 } 1068 }
843 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data; 1069
1070 brelse(bh);
1071 bh = NULL;
1072 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb), blkno,
1073 &bh, OCFS2_BH_CACHED, inode);
1074 if (ret) {
1075 mlog_errno(ret);
1076 goto out;
1077 }
1078
1079 eb = (struct ocfs2_extent_block *) bh->b_data;
844 el = &eb->h_list; 1080 el = &eb->h_list;
1081 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
1082 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
1083 ret = -EIO;
1084 goto out;
1085 }
1086
1087 if (le16_to_cpu(el->l_next_free_rec) >
1088 le16_to_cpu(el->l_count)) {
1089 ocfs2_error(inode->i_sb,
1090 "Inode %llu has bad count in extent list "
1091 "at block %llu (next free=%u, count=%u)\n",
1092 (unsigned long long)oi->ip_blkno,
1093 (unsigned long long)bh->b_blocknr,
1094 le16_to_cpu(el->l_next_free_rec),
1095 le16_to_cpu(el->l_count));
1096 ret = -EROFS;
1097 goto out;
1098 }
1099
1100 if (func)
1101 func(data, bh);
1102 }
1103
1104out:
1105 /*
1106 * Catch any trailing bh that the loop didn't handle.
1107 */
1108 brelse(bh);
1109
1110 return ret;
1111}
1112
1113/*
1114 * Given an initialized path (that is, it has a valid root extent
1115 * list), this function will traverse the btree in search of the path
1116 * which would contain cpos.
1117 *
1118 * The path traveled is recorded in the path structure.
1119 *
1120 * Note that this will not do any comparisons on leaf node extent
1121 * records, so it will work fine in the case that we just added a tree
1122 * branch.
1123 */
1124struct find_path_data {
1125 int index;
1126 struct ocfs2_path *path;
1127};
1128static void find_path_ins(void *data, struct buffer_head *bh)
1129{
1130 struct find_path_data *fp = data;
1131
1132 get_bh(bh);
1133 ocfs2_path_insert_eb(fp->path, fp->index, bh);
1134 fp->index++;
1135}
1136static int ocfs2_find_path(struct inode *inode, struct ocfs2_path *path,
1137 u32 cpos)
1138{
1139 struct find_path_data data;
1140
1141 data.index = 1;
1142 data.path = path;
1143 return __ocfs2_find_path(inode, path_root_el(path), cpos,
1144 find_path_ins, &data);
1145}
1146
1147static void find_leaf_ins(void *data, struct buffer_head *bh)
1148{
1149 struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1150 struct ocfs2_extent_list *el = &eb->h_list;
1151 struct buffer_head **ret = data;
1152
1153 /* We want to retain only the leaf block. */
1154 if (le16_to_cpu(el->l_tree_depth) == 0) {
1155 get_bh(bh);
1156 *ret = bh;
1157 }
1158}
1159/*
1160 * Find the leaf block in the tree which would contain cpos. No
1161 * checking of the actual leaf is done.
1162 *
1163 * Some paths want to call this instead of allocating a path structure
1164 * and calling ocfs2_find_path().
1165 *
1166 * This function doesn't handle non btree extent lists.
1167 */
1168int ocfs2_find_leaf(struct inode *inode, struct ocfs2_extent_list *root_el,
1169 u32 cpos, struct buffer_head **leaf_bh)
1170{
1171 int ret;
1172 struct buffer_head *bh = NULL;
1173
1174 ret = __ocfs2_find_path(inode, root_el, cpos, find_leaf_ins, &bh);
1175 if (ret) {
1176 mlog_errno(ret);
1177 goto out;
1178 }
1179
1180 *leaf_bh = bh;
1181out:
1182 return ret;
1183}
1184
1185/*
1186 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1187 *
1188 * Basically, we've moved stuff around at the bottom of the tree and
1189 * we need to fix up the extent records above the changes to reflect
1190 * the new changes.
1191 *
1192 * left_rec: the record on the left.
1193 * left_child_el: is the child list pointed to by left_rec
1194 * right_rec: the record to the right of left_rec
1195 * right_child_el: is the child list pointed to by right_rec
1196 *
1197 * By definition, this only works on interior nodes.
1198 */
1199static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1200 struct ocfs2_extent_list *left_child_el,
1201 struct ocfs2_extent_rec *right_rec,
1202 struct ocfs2_extent_list *right_child_el)
1203{
1204 u32 left_clusters, right_end;
1205
1206 /*
1207 * Interior nodes never have holes. Their cpos is the cpos of
1208 * the leftmost record in their child list. Their cluster
1209 * count covers the full theoretical range of their child list
1210 * - the range between their cpos and the cpos of the record
1211 * immediately to their right.
1212 */
1213 left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1214 left_clusters -= le32_to_cpu(left_rec->e_cpos);
1215 left_rec->e_int_clusters = cpu_to_le32(left_clusters);
1216
1217 /*
1218 * Calculate the rightmost cluster count boundary before
1219 * moving cpos - we will need to adjust clusters after
1220 * updating e_cpos to keep the same highest cluster count.
1221 */
1222 right_end = le32_to_cpu(right_rec->e_cpos);
1223 right_end += le32_to_cpu(right_rec->e_int_clusters);
1224
1225 right_rec->e_cpos = left_rec->e_cpos;
1226 le32_add_cpu(&right_rec->e_cpos, left_clusters);
1227
1228 right_end -= le32_to_cpu(right_rec->e_cpos);
1229 right_rec->e_int_clusters = cpu_to_le32(right_end);
1230}
1231
1232/*
1233 * Adjust the adjacent root node records involved in a
1234 * rotation. left_el_blkno is passed in as a key so that we can easily
1235 * find it's index in the root list.
1236 */
1237static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
1238 struct ocfs2_extent_list *left_el,
1239 struct ocfs2_extent_list *right_el,
1240 u64 left_el_blkno)
1241{
1242 int i;
1243
1244 BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
1245 le16_to_cpu(left_el->l_tree_depth));
1246
1247 for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
1248 if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
1249 break;
1250 }
1251
1252 /*
1253 * The path walking code should have never returned a root and
1254 * two paths which are not adjacent.
1255 */
1256 BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
1257
1258 ocfs2_adjust_adjacent_records(&root_el->l_recs[i], left_el,
1259 &root_el->l_recs[i + 1], right_el);
1260}
1261
1262/*
1263 * We've changed a leaf block (in right_path) and need to reflect that
1264 * change back up the subtree.
1265 *
1266 * This happens in multiple places:
1267 * - When we've moved an extent record from the left path leaf to the right
1268 * path leaf to make room for an empty extent in the left path leaf.
1269 * - When our insert into the right path leaf is at the leftmost edge
1270 * and requires an update of the path immediately to it's left. This
1271 * can occur at the end of some types of rotation and appending inserts.
1272 */
1273static void ocfs2_complete_edge_insert(struct inode *inode, handle_t *handle,
1274 struct ocfs2_path *left_path,
1275 struct ocfs2_path *right_path,
1276 int subtree_index)
1277{
1278 int ret, i, idx;
1279 struct ocfs2_extent_list *el, *left_el, *right_el;
1280 struct ocfs2_extent_rec *left_rec, *right_rec;
1281 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
1282
1283 /*
1284 * Update the counts and position values within all the
1285 * interior nodes to reflect the leaf rotation we just did.
1286 *
1287 * The root node is handled below the loop.
1288 *
1289 * We begin the loop with right_el and left_el pointing to the
1290 * leaf lists and work our way up.
1291 *
1292 * NOTE: within this loop, left_el and right_el always refer
1293 * to the *child* lists.
1294 */
1295 left_el = path_leaf_el(left_path);
1296 right_el = path_leaf_el(right_path);
1297 for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
1298 mlog(0, "Adjust records at index %u\n", i);
1299
1300 /*
1301 * One nice property of knowing that all of these
1302 * nodes are below the root is that we only deal with
1303 * the leftmost right node record and the rightmost
1304 * left node record.
1305 */
1306 el = left_path->p_node[i].el;
1307 idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
1308 left_rec = &el->l_recs[idx];
1309
1310 el = right_path->p_node[i].el;
1311 right_rec = &el->l_recs[0];
1312
1313 ocfs2_adjust_adjacent_records(left_rec, left_el, right_rec,
1314 right_el);
1315
1316 ret = ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
1317 if (ret)
1318 mlog_errno(ret);
1319
1320 ret = ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
1321 if (ret)
1322 mlog_errno(ret);
1323
1324 /*
1325 * Setup our list pointers now so that the current
1326 * parents become children in the next iteration.
1327 */
1328 left_el = left_path->p_node[i].el;
1329 right_el = right_path->p_node[i].el;
1330 }
1331
1332 /*
1333 * At the root node, adjust the two adjacent records which
1334 * begin our path to the leaves.
1335 */
1336
1337 el = left_path->p_node[subtree_index].el;
1338 left_el = left_path->p_node[subtree_index + 1].el;
1339 right_el = right_path->p_node[subtree_index + 1].el;
1340
1341 ocfs2_adjust_root_records(el, left_el, right_el,
1342 left_path->p_node[subtree_index + 1].bh->b_blocknr);
1343
1344 root_bh = left_path->p_node[subtree_index].bh;
1345
1346 ret = ocfs2_journal_dirty(handle, root_bh);
1347 if (ret)
1348 mlog_errno(ret);
1349}
1350
1351static int ocfs2_rotate_subtree_right(struct inode *inode,
1352 handle_t *handle,
1353 struct ocfs2_path *left_path,
1354 struct ocfs2_path *right_path,
1355 int subtree_index)
1356{
1357 int ret, i;
1358 struct buffer_head *right_leaf_bh;
1359 struct buffer_head *left_leaf_bh = NULL;
1360 struct buffer_head *root_bh;
1361 struct ocfs2_extent_list *right_el, *left_el;
1362 struct ocfs2_extent_rec move_rec;
1363
1364 left_leaf_bh = path_leaf_bh(left_path);
1365 left_el = path_leaf_el(left_path);
1366
1367 if (left_el->l_next_free_rec != left_el->l_count) {
1368 ocfs2_error(inode->i_sb,
1369 "Inode %llu has non-full interior leaf node %llu"
1370 "(next free = %u)",
1371 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1372 (unsigned long long)left_leaf_bh->b_blocknr,
1373 le16_to_cpu(left_el->l_next_free_rec));
1374 return -EROFS;
1375 }
1376
1377 /*
1378 * This extent block may already have an empty record, so we
1379 * return early if so.
1380 */
1381 if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
1382 return 0;
1383
1384 root_bh = left_path->p_node[subtree_index].bh;
1385 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
1386
1387 ret = ocfs2_journal_access(handle, inode, root_bh,
1388 OCFS2_JOURNAL_ACCESS_WRITE);
1389 if (ret) {
1390 mlog_errno(ret);
1391 goto out;
1392 }
1393
1394 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
1395 ret = ocfs2_journal_access(handle, inode,
1396 right_path->p_node[i].bh,
1397 OCFS2_JOURNAL_ACCESS_WRITE);
1398 if (ret) {
1399 mlog_errno(ret);
1400 goto out;
1401 }
1402
1403 ret = ocfs2_journal_access(handle, inode,
1404 left_path->p_node[i].bh,
1405 OCFS2_JOURNAL_ACCESS_WRITE);
1406 if (ret) {
1407 mlog_errno(ret);
1408 goto out;
1409 }
1410 }
1411
1412 right_leaf_bh = path_leaf_bh(right_path);
1413 right_el = path_leaf_el(right_path);
1414
1415 /* This is a code error, not a disk corruption. */
1416 mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
1417 "because rightmost leaf block %llu is empty\n",
1418 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1419 (unsigned long long)right_leaf_bh->b_blocknr);
1420
1421 ocfs2_create_empty_extent(right_el);
1422
1423 ret = ocfs2_journal_dirty(handle, right_leaf_bh);
1424 if (ret) {
1425 mlog_errno(ret);
1426 goto out;
1427 }
1428
1429 /* Do the copy now. */
1430 i = le16_to_cpu(left_el->l_next_free_rec) - 1;
1431 move_rec = left_el->l_recs[i];
1432 right_el->l_recs[0] = move_rec;
1433
1434 /*
1435 * Clear out the record we just copied and shift everything
1436 * over, leaving an empty extent in the left leaf.
1437 *
1438 * We temporarily subtract from next_free_rec so that the
1439 * shift will lose the tail record (which is now defunct).
1440 */
1441 le16_add_cpu(&left_el->l_next_free_rec, -1);
1442 ocfs2_shift_records_right(left_el);
1443 memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1444 le16_add_cpu(&left_el->l_next_free_rec, 1);
1445
1446 ret = ocfs2_journal_dirty(handle, left_leaf_bh);
1447 if (ret) {
1448 mlog_errno(ret);
1449 goto out;
1450 }
1451
1452 ocfs2_complete_edge_insert(inode, handle, left_path, right_path,
1453 subtree_index);
1454
1455out:
1456 return ret;
1457}
1458
1459/*
1460 * Given a full path, determine what cpos value would return us a path
1461 * containing the leaf immediately to the left of the current one.
1462 *
1463 * Will return zero if the path passed in is already the leftmost path.
1464 */
1465static int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
1466 struct ocfs2_path *path, u32 *cpos)
1467{
1468 int i, j, ret = 0;
1469 u64 blkno;
1470 struct ocfs2_extent_list *el;
1471
1472 BUG_ON(path->p_tree_depth == 0);
1473
1474 *cpos = 0;
1475
1476 blkno = path_leaf_bh(path)->b_blocknr;
1477
1478 /* Start at the tree node just above the leaf and work our way up. */
1479 i = path->p_tree_depth - 1;
1480 while (i >= 0) {
1481 el = path->p_node[i].el;
1482
1483 /*
1484 * Find the extent record just before the one in our
1485 * path.
1486 */
1487 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
1488 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
1489 if (j == 0) {
1490 if (i == 0) {
1491 /*
1492 * We've determined that the
1493 * path specified is already
1494 * the leftmost one - return a
1495 * cpos of zero.
1496 */
1497 goto out;
1498 }
1499 /*
1500 * The leftmost record points to our
1501 * leaf - we need to travel up the
1502 * tree one level.
1503 */
1504 goto next_node;
1505 }
1506
1507 *cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
1508 *cpos = *cpos + ocfs2_rec_clusters(el,
1509 &el->l_recs[j - 1]);
1510 *cpos = *cpos - 1;
1511 goto out;
1512 }
1513 }
1514
1515 /*
1516 * If we got here, we never found a valid node where
1517 * the tree indicated one should be.
1518 */
1519 ocfs2_error(sb,
1520 "Invalid extent tree at extent block %llu\n",
1521 (unsigned long long)blkno);
1522 ret = -EROFS;
1523 goto out;
1524
1525next_node:
1526 blkno = path->p_node[i].bh->b_blocknr;
1527 i--;
1528 }
1529
1530out:
1531 return ret;
1532}
1533
1534static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
1535 struct ocfs2_path *path)
1536{
1537 int credits = (path->p_tree_depth - subtree_depth) * 2 + 1;
1538
1539 if (handle->h_buffer_credits < credits)
1540 return ocfs2_extend_trans(handle, credits);
1541
1542 return 0;
1543}
1544
1545/*
1546 * Trap the case where we're inserting into the theoretical range past
1547 * the _actual_ left leaf range. Otherwise, we'll rotate a record
1548 * whose cpos is less than ours into the right leaf.
1549 *
1550 * It's only necessary to look at the rightmost record of the left
1551 * leaf because the logic that calls us should ensure that the
1552 * theoretical ranges in the path components above the leaves are
1553 * correct.
1554 */
1555static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
1556 u32 insert_cpos)
1557{
1558 struct ocfs2_extent_list *left_el;
1559 struct ocfs2_extent_rec *rec;
1560 int next_free;
1561
1562 left_el = path_leaf_el(left_path);
1563 next_free = le16_to_cpu(left_el->l_next_free_rec);
1564 rec = &left_el->l_recs[next_free - 1];
1565
1566 if (insert_cpos > le32_to_cpu(rec->e_cpos))
1567 return 1;
1568 return 0;
1569}
1570
1571/*
1572 * Rotate all the records in a btree right one record, starting at insert_cpos.
1573 *
1574 * The path to the rightmost leaf should be passed in.
1575 *
1576 * The array is assumed to be large enough to hold an entire path (tree depth).
1577 *
1578 * Upon succesful return from this function:
1579 *
1580 * - The 'right_path' array will contain a path to the leaf block
1581 * whose range contains e_cpos.
1582 * - That leaf block will have a single empty extent in list index 0.
1583 * - In the case that the rotation requires a post-insert update,
1584 * *ret_left_path will contain a valid path which can be passed to
1585 * ocfs2_insert_path().
1586 */
1587static int ocfs2_rotate_tree_right(struct inode *inode,
1588 handle_t *handle,
1589 u32 insert_cpos,
1590 struct ocfs2_path *right_path,
1591 struct ocfs2_path **ret_left_path)
1592{
1593 int ret, start;
1594 u32 cpos;
1595 struct ocfs2_path *left_path = NULL;
1596
1597 *ret_left_path = NULL;
1598
1599 left_path = ocfs2_new_path(path_root_bh(right_path),
1600 path_root_el(right_path));
1601 if (!left_path) {
1602 ret = -ENOMEM;
1603 mlog_errno(ret);
1604 goto out;
1605 }
1606
1607 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path, &cpos);
1608 if (ret) {
1609 mlog_errno(ret);
1610 goto out;
1611 }
1612
1613 mlog(0, "Insert: %u, first left path cpos: %u\n", insert_cpos, cpos);
1614
1615 /*
1616 * What we want to do here is:
1617 *
1618 * 1) Start with the rightmost path.
1619 *
1620 * 2) Determine a path to the leaf block directly to the left
1621 * of that leaf.
1622 *
1623 * 3) Determine the 'subtree root' - the lowest level tree node
1624 * which contains a path to both leaves.
1625 *
1626 * 4) Rotate the subtree.
1627 *
1628 * 5) Find the next subtree by considering the left path to be
1629 * the new right path.
1630 *
1631 * The check at the top of this while loop also accepts
1632 * insert_cpos == cpos because cpos is only a _theoretical_
1633 * value to get us the left path - insert_cpos might very well
1634 * be filling that hole.
1635 *
1636 * Stop at a cpos of '0' because we either started at the
1637 * leftmost branch (i.e., a tree with one branch and a
1638 * rotation inside of it), or we've gone as far as we can in
1639 * rotating subtrees.
1640 */
1641 while (cpos && insert_cpos <= cpos) {
1642 mlog(0, "Rotating a tree: ins. cpos: %u, left path cpos: %u\n",
1643 insert_cpos, cpos);
1644
1645 ret = ocfs2_find_path(inode, left_path, cpos);
1646 if (ret) {
1647 mlog_errno(ret);
1648 goto out;
1649 }
1650
1651 mlog_bug_on_msg(path_leaf_bh(left_path) ==
1652 path_leaf_bh(right_path),
1653 "Inode %lu: error during insert of %u "
1654 "(left path cpos %u) results in two identical "
1655 "paths ending at %llu\n",
1656 inode->i_ino, insert_cpos, cpos,
1657 (unsigned long long)
1658 path_leaf_bh(left_path)->b_blocknr);
1659
1660 if (ocfs2_rotate_requires_path_adjustment(left_path,
1661 insert_cpos)) {
1662 mlog(0, "Path adjustment required\n");
1663
1664 /*
1665 * We've rotated the tree as much as we
1666 * should. The rest is up to
1667 * ocfs2_insert_path() to complete, after the
1668 * record insertion. We indicate this
1669 * situation by returning the left path.
1670 *
1671 * The reason we don't adjust the records here
1672 * before the record insert is that an error
1673 * later might break the rule where a parent
1674 * record e_cpos will reflect the actual
1675 * e_cpos of the 1st nonempty record of the
1676 * child list.
1677 */
1678 *ret_left_path = left_path;
1679 goto out_ret_path;
1680 }
1681
1682 start = ocfs2_find_subtree_root(inode, left_path, right_path);
1683
1684 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
1685 start,
1686 (unsigned long long) right_path->p_node[start].bh->b_blocknr,
1687 right_path->p_tree_depth);
1688
1689 ret = ocfs2_extend_rotate_transaction(handle, start,
1690 right_path);
1691 if (ret) {
1692 mlog_errno(ret);
1693 goto out;
1694 }
1695
1696 ret = ocfs2_rotate_subtree_right(inode, handle, left_path,
1697 right_path, start);
1698 if (ret) {
1699 mlog_errno(ret);
1700 goto out;
1701 }
1702
1703 /*
1704 * There is no need to re-read the next right path
1705 * as we know that it'll be our current left
1706 * path. Optimize by copying values instead.
1707 */
1708 ocfs2_mv_path(right_path, left_path);
1709
1710 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path,
1711 &cpos);
1712 if (ret) {
1713 mlog_errno(ret);
1714 goto out;
1715 }
1716 }
1717
1718out:
1719 ocfs2_free_path(left_path);
1720
1721out_ret_path:
1722 return ret;
1723}
1724
1725/*
1726 * Do the final bits of extent record insertion at the target leaf
1727 * list. If this leaf is part of an allocation tree, it is assumed
1728 * that the tree above has been prepared.
1729 */
1730static void ocfs2_insert_at_leaf(struct ocfs2_extent_rec *insert_rec,
1731 struct ocfs2_extent_list *el,
1732 struct ocfs2_insert_type *insert,
1733 struct inode *inode)
1734{
1735 int i = insert->ins_contig_index;
1736 unsigned int range;
1737 struct ocfs2_extent_rec *rec;
1738
1739 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1740
1741 /*
1742 * Contiguous insert - either left or right.
1743 */
1744 if (insert->ins_contig != CONTIG_NONE) {
1745 rec = &el->l_recs[i];
1746 if (insert->ins_contig == CONTIG_LEFT) {
1747 rec->e_blkno = insert_rec->e_blkno;
1748 rec->e_cpos = insert_rec->e_cpos;
1749 }
1750 le16_add_cpu(&rec->e_leaf_clusters,
1751 le16_to_cpu(insert_rec->e_leaf_clusters));
1752 return;
1753 }
1754
1755 /*
1756 * Handle insert into an empty leaf.
1757 */
1758 if (le16_to_cpu(el->l_next_free_rec) == 0 ||
1759 ((le16_to_cpu(el->l_next_free_rec) == 1) &&
1760 ocfs2_is_empty_extent(&el->l_recs[0]))) {
1761 el->l_recs[0] = *insert_rec;
1762 el->l_next_free_rec = cpu_to_le16(1);
1763 return;
1764 }
1765
1766 /*
1767 * Appending insert.
1768 */
1769 if (insert->ins_appending == APPEND_TAIL) {
1770 i = le16_to_cpu(el->l_next_free_rec) - 1;
1771 rec = &el->l_recs[i];
1772 range = le32_to_cpu(rec->e_cpos)
1773 + le16_to_cpu(rec->e_leaf_clusters);
1774 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
1775
1776 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
1777 le16_to_cpu(el->l_count),
1778 "inode %lu, depth %u, count %u, next free %u, "
1779 "rec.cpos %u, rec.clusters %u, "
1780 "insert.cpos %u, insert.clusters %u\n",
1781 inode->i_ino,
1782 le16_to_cpu(el->l_tree_depth),
1783 le16_to_cpu(el->l_count),
1784 le16_to_cpu(el->l_next_free_rec),
1785 le32_to_cpu(el->l_recs[i].e_cpos),
1786 le16_to_cpu(el->l_recs[i].e_leaf_clusters),
1787 le32_to_cpu(insert_rec->e_cpos),
1788 le16_to_cpu(insert_rec->e_leaf_clusters));
1789 i++;
1790 el->l_recs[i] = *insert_rec;
1791 le16_add_cpu(&el->l_next_free_rec, 1);
1792 return;
1793 }
1794
1795 /*
1796 * Ok, we have to rotate.
1797 *
1798 * At this point, it is safe to assume that inserting into an
1799 * empty leaf and appending to a leaf have both been handled
1800 * above.
1801 *
1802 * This leaf needs to have space, either by the empty 1st
1803 * extent record, or by virtue of an l_next_rec < l_count.
1804 */
1805 ocfs2_rotate_leaf(el, insert_rec);
1806}
1807
1808static inline void ocfs2_update_dinode_clusters(struct inode *inode,
1809 struct ocfs2_dinode *di,
1810 u32 clusters)
1811{
1812 le32_add_cpu(&di->i_clusters, clusters);
1813 spin_lock(&OCFS2_I(inode)->ip_lock);
1814 OCFS2_I(inode)->ip_clusters = le32_to_cpu(di->i_clusters);
1815 spin_unlock(&OCFS2_I(inode)->ip_lock);
1816}
1817
1818static int ocfs2_append_rec_to_path(struct inode *inode, handle_t *handle,
1819 struct ocfs2_extent_rec *insert_rec,
1820 struct ocfs2_path *right_path,
1821 struct ocfs2_path **ret_left_path)
1822{
1823 int ret, i, next_free;
1824 struct buffer_head *bh;
1825 struct ocfs2_extent_list *el;
1826 struct ocfs2_path *left_path = NULL;
1827
1828 *ret_left_path = NULL;
1829
1830 /*
1831 * This shouldn't happen for non-trees. The extent rec cluster
1832 * count manipulation below only works for interior nodes.
1833 */
1834 BUG_ON(right_path->p_tree_depth == 0);
1835
1836 /*
1837 * If our appending insert is at the leftmost edge of a leaf,
1838 * then we might need to update the rightmost records of the
1839 * neighboring path.
1840 */
1841 el = path_leaf_el(right_path);
1842 next_free = le16_to_cpu(el->l_next_free_rec);
1843 if (next_free == 0 ||
1844 (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
1845 u32 left_cpos;
1846
1847 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path,
1848 &left_cpos);
1849 if (ret) {
1850 mlog_errno(ret);
1851 goto out;
1852 }
1853
1854 mlog(0, "Append may need a left path update. cpos: %u, "
1855 "left_cpos: %u\n", le32_to_cpu(insert_rec->e_cpos),
1856 left_cpos);
1857
1858 /*
1859 * No need to worry if the append is already in the
1860 * leftmost leaf.
1861 */
1862 if (left_cpos) {
1863 left_path = ocfs2_new_path(path_root_bh(right_path),
1864 path_root_el(right_path));
1865 if (!left_path) {
1866 ret = -ENOMEM;
1867 mlog_errno(ret);
1868 goto out;
1869 }
1870
1871 ret = ocfs2_find_path(inode, left_path, left_cpos);
1872 if (ret) {
1873 mlog_errno(ret);
1874 goto out;
1875 }
1876
1877 /*
1878 * ocfs2_insert_path() will pass the left_path to the
1879 * journal for us.
1880 */
1881 }
1882 }
1883
1884 ret = ocfs2_journal_access_path(inode, handle, right_path);
1885 if (ret) {
1886 mlog_errno(ret);
1887 goto out;
1888 }
1889
1890 el = path_root_el(right_path);
1891 bh = path_root_bh(right_path);
1892 i = 0;
1893 while (1) {
1894 struct ocfs2_extent_rec *rec;
1895
1896 next_free = le16_to_cpu(el->l_next_free_rec);
1897 if (next_free == 0) {
1898 ocfs2_error(inode->i_sb,
1899 "Dinode %llu has a bad extent list",
1900 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1901 ret = -EIO;
1902 goto out;
1903 }
1904
1905 rec = &el->l_recs[next_free - 1];
1906
1907 rec->e_int_clusters = insert_rec->e_cpos;
1908 le32_add_cpu(&rec->e_int_clusters,
1909 le16_to_cpu(insert_rec->e_leaf_clusters));
1910 le32_add_cpu(&rec->e_int_clusters,
1911 -le32_to_cpu(rec->e_cpos));
1912
1913 ret = ocfs2_journal_dirty(handle, bh);
1914 if (ret)
1915 mlog_errno(ret);
1916
1917 /* Don't touch the leaf node */
1918 if (++i >= right_path->p_tree_depth)
1919 break;
1920
1921 bh = right_path->p_node[i].bh;
1922 el = right_path->p_node[i].el;
1923 }
1924
1925 *ret_left_path = left_path;
1926 ret = 0;
1927out:
1928 if (ret != 0)
1929 ocfs2_free_path(left_path);
1930
1931 return ret;
1932}
1933
1934/*
1935 * This function only does inserts on an allocation b-tree. For dinode
1936 * lists, ocfs2_insert_at_leaf() is called directly.
1937 *
1938 * right_path is the path we want to do the actual insert
1939 * in. left_path should only be passed in if we need to update that
1940 * portion of the tree after an edge insert.
1941 */
1942static int ocfs2_insert_path(struct inode *inode,
1943 handle_t *handle,
1944 struct ocfs2_path *left_path,
1945 struct ocfs2_path *right_path,
1946 struct ocfs2_extent_rec *insert_rec,
1947 struct ocfs2_insert_type *insert)
1948{
1949 int ret, subtree_index;
1950 struct buffer_head *leaf_bh = path_leaf_bh(right_path);
1951 struct ocfs2_extent_list *el;
1952
1953 /*
1954 * Pass both paths to the journal. The majority of inserts
1955 * will be touching all components anyway.
1956 */
1957 ret = ocfs2_journal_access_path(inode, handle, right_path);
1958 if (ret < 0) {
1959 mlog_errno(ret);
1960 goto out;
1961 }
1962
1963 if (left_path) {
1964 int credits = handle->h_buffer_credits;
1965
1966 /*
1967 * There's a chance that left_path got passed back to
1968 * us without being accounted for in the
1969 * journal. Extend our transaction here to be sure we
1970 * can change those blocks.
1971 */
1972 credits += left_path->p_tree_depth;
1973
1974 ret = ocfs2_extend_trans(handle, credits);
1975 if (ret < 0) {
1976 mlog_errno(ret);
1977 goto out;
1978 }
1979
1980 ret = ocfs2_journal_access_path(inode, handle, left_path);
1981 if (ret < 0) {
1982 mlog_errno(ret);
1983 goto out;
1984 }
1985 }
1986
1987 el = path_leaf_el(right_path);
1988
1989 ocfs2_insert_at_leaf(insert_rec, el, insert, inode);
1990 ret = ocfs2_journal_dirty(handle, leaf_bh);
1991 if (ret)
1992 mlog_errno(ret);
1993
1994 if (left_path) {
1995 /*
1996 * The rotate code has indicated that we need to fix
1997 * up portions of the tree after the insert.
1998 *
1999 * XXX: Should we extend the transaction here?
2000 */
2001 subtree_index = ocfs2_find_subtree_root(inode, left_path,
2002 right_path);
2003 ocfs2_complete_edge_insert(inode, handle, left_path,
2004 right_path, subtree_index);
2005 }
2006
2007 ret = 0;
2008out:
2009 return ret;
2010}
2011
2012static int ocfs2_do_insert_extent(struct inode *inode,
2013 handle_t *handle,
2014 struct buffer_head *di_bh,
2015 struct ocfs2_extent_rec *insert_rec,
2016 struct ocfs2_insert_type *type)
2017{
2018 int ret, rotate = 0;
2019 u32 cpos;
2020 struct ocfs2_path *right_path = NULL;
2021 struct ocfs2_path *left_path = NULL;
2022 struct ocfs2_dinode *di;
2023 struct ocfs2_extent_list *el;
2024
2025 di = (struct ocfs2_dinode *) di_bh->b_data;
2026 el = &di->id2.i_list;
2027
2028 ret = ocfs2_journal_access(handle, inode, di_bh,
2029 OCFS2_JOURNAL_ACCESS_WRITE);
2030 if (ret) {
2031 mlog_errno(ret);
2032 goto out;
2033 }
2034
2035 if (le16_to_cpu(el->l_tree_depth) == 0) {
2036 ocfs2_insert_at_leaf(insert_rec, el, type, inode);
2037 goto out_update_clusters;
2038 }
2039
2040 right_path = ocfs2_new_inode_path(di_bh);
2041 if (!right_path) {
2042 ret = -ENOMEM;
2043 mlog_errno(ret);
2044 goto out;
2045 }
2046
2047 /*
2048 * Determine the path to start with. Rotations need the
2049 * rightmost path, everything else can go directly to the
2050 * target leaf.
2051 */
2052 cpos = le32_to_cpu(insert_rec->e_cpos);
2053 if (type->ins_appending == APPEND_NONE &&
2054 type->ins_contig == CONTIG_NONE) {
2055 rotate = 1;
2056 cpos = UINT_MAX;
2057 }
2058
2059 ret = ocfs2_find_path(inode, right_path, cpos);
2060 if (ret) {
2061 mlog_errno(ret);
2062 goto out;
2063 }
2064
2065 /*
2066 * Rotations and appends need special treatment - they modify
2067 * parts of the tree's above them.
2068 *
2069 * Both might pass back a path immediate to the left of the
2070 * one being inserted to. This will be cause
2071 * ocfs2_insert_path() to modify the rightmost records of
2072 * left_path to account for an edge insert.
2073 *
2074 * XXX: When modifying this code, keep in mind that an insert
2075 * can wind up skipping both of these two special cases...
2076 */
2077 if (rotate) {
2078 ret = ocfs2_rotate_tree_right(inode, handle,
2079 le32_to_cpu(insert_rec->e_cpos),
2080 right_path, &left_path);
2081 if (ret) {
2082 mlog_errno(ret);
2083 goto out;
2084 }
2085 } else if (type->ins_appending == APPEND_TAIL
2086 && type->ins_contig != CONTIG_LEFT) {
2087 ret = ocfs2_append_rec_to_path(inode, handle, insert_rec,
2088 right_path, &left_path);
2089 if (ret) {
2090 mlog_errno(ret);
2091 goto out;
2092 }
2093 }
2094
2095 ret = ocfs2_insert_path(inode, handle, left_path, right_path,
2096 insert_rec, type);
2097 if (ret) {
2098 mlog_errno(ret);
2099 goto out;
2100 }
2101
2102out_update_clusters:
2103 ocfs2_update_dinode_clusters(inode, di,
2104 le16_to_cpu(insert_rec->e_leaf_clusters));
2105
2106 ret = ocfs2_journal_dirty(handle, di_bh);
2107 if (ret)
2108 mlog_errno(ret);
2109
2110out:
2111 ocfs2_free_path(left_path);
2112 ocfs2_free_path(right_path);
2113
2114 return ret;
2115}
2116
2117static void ocfs2_figure_contig_type(struct inode *inode,
2118 struct ocfs2_insert_type *insert,
2119 struct ocfs2_extent_list *el,
2120 struct ocfs2_extent_rec *insert_rec)
2121{
2122 int i;
2123 enum ocfs2_contig_type contig_type = CONTIG_NONE;
2124
2125 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
2126
2127 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
2128 contig_type = ocfs2_extent_contig(inode, &el->l_recs[i],
2129 insert_rec);
2130 if (contig_type != CONTIG_NONE) {
2131 insert->ins_contig_index = i;
2132 break;
2133 }
2134 }
2135 insert->ins_contig = contig_type;
2136}
2137
2138/*
2139 * This should only be called against the righmost leaf extent list.
2140 *
2141 * ocfs2_figure_appending_type() will figure out whether we'll have to
2142 * insert at the tail of the rightmost leaf.
2143 *
2144 * This should also work against the dinode list for tree's with 0
2145 * depth. If we consider the dinode list to be the rightmost leaf node
2146 * then the logic here makes sense.
2147 */
2148static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
2149 struct ocfs2_extent_list *el,
2150 struct ocfs2_extent_rec *insert_rec)
2151{
2152 int i;
2153 u32 cpos = le32_to_cpu(insert_rec->e_cpos);
2154 struct ocfs2_extent_rec *rec;
2155
2156 insert->ins_appending = APPEND_NONE;
2157
2158 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
2159
2160 if (!el->l_next_free_rec)
2161 goto set_tail_append;
2162
2163 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
2164 /* Were all records empty? */
2165 if (le16_to_cpu(el->l_next_free_rec) == 1)
2166 goto set_tail_append;
845 } 2167 }
846 2168
847 /* Can we allocate without adding/shifting tree bits? */
848 i = le16_to_cpu(el->l_next_free_rec) - 1; 2169 i = le16_to_cpu(el->l_next_free_rec) - 1;
849 if (le16_to_cpu(el->l_next_free_rec) == 0 2170 rec = &el->l_recs[i];
850 || (le16_to_cpu(el->l_next_free_rec) < le16_to_cpu(el->l_count)) 2171
851 || le32_to_cpu(el->l_recs[i].e_clusters) == 0 2172 if (cpos >=
852 || ocfs2_extent_contig(inode, &el->l_recs[i], start_blk)) 2173 (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
853 goto out_add; 2174 goto set_tail_append;
2175
2176 return;
2177
2178set_tail_append:
2179 insert->ins_appending = APPEND_TAIL;
2180}
2181
2182/*
2183 * Helper function called at the begining of an insert.
2184 *
2185 * This computes a few things that are commonly used in the process of
2186 * inserting into the btree:
2187 * - Whether the new extent is contiguous with an existing one.
2188 * - The current tree depth.
2189 * - Whether the insert is an appending one.
2190 * - The total # of free records in the tree.
2191 *
2192 * All of the information is stored on the ocfs2_insert_type
2193 * structure.
2194 */
2195static int ocfs2_figure_insert_type(struct inode *inode,
2196 struct buffer_head *di_bh,
2197 struct buffer_head **last_eb_bh,
2198 struct ocfs2_extent_rec *insert_rec,
2199 struct ocfs2_insert_type *insert)
2200{
2201 int ret;
2202 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2203 struct ocfs2_extent_block *eb;
2204 struct ocfs2_extent_list *el;
2205 struct ocfs2_path *path = NULL;
2206 struct buffer_head *bh = NULL;
2207
2208 el = &di->id2.i_list;
2209 insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
2210
2211 if (el->l_tree_depth) {
2212 /*
2213 * If we have tree depth, we read in the
2214 * rightmost extent block ahead of time as
2215 * ocfs2_figure_insert_type() and ocfs2_add_branch()
2216 * may want it later.
2217 */
2218 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
2219 le64_to_cpu(di->i_last_eb_blk), &bh,
2220 OCFS2_BH_CACHED, inode);
2221 if (ret) {
2222 mlog_exit(ret);
2223 goto out;
2224 }
2225 eb = (struct ocfs2_extent_block *) bh->b_data;
2226 el = &eb->h_list;
2227 }
2228
2229 /*
2230 * Unless we have a contiguous insert, we'll need to know if
2231 * there is room left in our allocation tree for another
2232 * extent record.
2233 *
2234 * XXX: This test is simplistic, we can search for empty
2235 * extent records too.
2236 */
2237 insert->ins_free_records = le16_to_cpu(el->l_count) -
2238 le16_to_cpu(el->l_next_free_rec);
2239
2240 if (!insert->ins_tree_depth) {
2241 ocfs2_figure_contig_type(inode, insert, el, insert_rec);
2242 ocfs2_figure_appending_type(insert, el, insert_rec);
2243 return 0;
2244 }
2245
2246 path = ocfs2_new_inode_path(di_bh);
2247 if (!path) {
2248 ret = -ENOMEM;
2249 mlog_errno(ret);
2250 goto out;
2251 }
2252
2253 /*
2254 * In the case that we're inserting past what the tree
2255 * currently accounts for, ocfs2_find_path() will return for
2256 * us the rightmost tree path. This is accounted for below in
2257 * the appending code.
2258 */
2259 ret = ocfs2_find_path(inode, path, le32_to_cpu(insert_rec->e_cpos));
2260 if (ret) {
2261 mlog_errno(ret);
2262 goto out;
2263 }
2264
2265 el = path_leaf_el(path);
2266
2267 /*
2268 * Now that we have the path, there's two things we want to determine:
2269 * 1) Contiguousness (also set contig_index if this is so)
2270 *
2271 * 2) Are we doing an append? We can trivially break this up
2272 * into two types of appends: simple record append, or a
2273 * rotate inside the tail leaf.
2274 */
2275 ocfs2_figure_contig_type(inode, insert, el, insert_rec);
2276
2277 /*
2278 * The insert code isn't quite ready to deal with all cases of
2279 * left contiguousness. Specifically, if it's an insert into
2280 * the 1st record in a leaf, it will require the adjustment of
2281 * cluster count on the last record of the path directly to it's
2282 * left. For now, just catch that case and fool the layers
2283 * above us. This works just fine for tree_depth == 0, which
2284 * is why we allow that above.
2285 */
2286 if (insert->ins_contig == CONTIG_LEFT &&
2287 insert->ins_contig_index == 0)
2288 insert->ins_contig = CONTIG_NONE;
2289
2290 /*
2291 * Ok, so we can simply compare against last_eb to figure out
2292 * whether the path doesn't exist. This will only happen in
2293 * the case that we're doing a tail append, so maybe we can
2294 * take advantage of that information somehow.
2295 */
2296 if (le64_to_cpu(di->i_last_eb_blk) == path_leaf_bh(path)->b_blocknr) {
2297 /*
2298 * Ok, ocfs2_find_path() returned us the rightmost
2299 * tree path. This might be an appending insert. There are
2300 * two cases:
2301 * 1) We're doing a true append at the tail:
2302 * -This might even be off the end of the leaf
2303 * 2) We're "appending" by rotating in the tail
2304 */
2305 ocfs2_figure_appending_type(insert, el, insert_rec);
2306 }
2307
2308out:
2309 ocfs2_free_path(path);
2310
2311 if (ret == 0)
2312 *last_eb_bh = bh;
2313 else
2314 brelse(bh);
2315 return ret;
2316}
2317
2318/*
2319 * Insert an extent into an inode btree.
2320 *
2321 * The caller needs to update fe->i_clusters
2322 */
2323int ocfs2_insert_extent(struct ocfs2_super *osb,
2324 handle_t *handle,
2325 struct inode *inode,
2326 struct buffer_head *fe_bh,
2327 u32 cpos,
2328 u64 start_blk,
2329 u32 new_clusters,
2330 struct ocfs2_alloc_context *meta_ac)
2331{
2332 int status, shift;
2333 struct buffer_head *last_eb_bh = NULL;
2334 struct buffer_head *bh = NULL;
2335 struct ocfs2_insert_type insert = {0, };
2336 struct ocfs2_extent_rec rec;
2337
2338 mlog(0, "add %u clusters at position %u to inode %llu\n",
2339 new_clusters, cpos, (unsigned long long)OCFS2_I(inode)->ip_blkno);
2340
2341 mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) &&
2342 (OCFS2_I(inode)->ip_clusters != cpos),
2343 "Device %s, asking for sparse allocation: inode %llu, "
2344 "cpos %u, clusters %u\n",
2345 osb->dev_str,
2346 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos,
2347 OCFS2_I(inode)->ip_clusters);
2348
2349 memset(&rec, 0, sizeof(rec));
2350 rec.e_cpos = cpu_to_le32(cpos);
2351 rec.e_blkno = cpu_to_le64(start_blk);
2352 rec.e_leaf_clusters = cpu_to_le16(new_clusters);
2353
2354 status = ocfs2_figure_insert_type(inode, fe_bh, &last_eb_bh, &rec,
2355 &insert);
2356 if (status < 0) {
2357 mlog_errno(status);
2358 goto bail;
2359 }
854 2360
855 mlog(0, "ocfs2_allocate_extent: couldn't do a simple add, traversing " 2361 mlog(0, "Insert.appending: %u, Insert.Contig: %u, "
856 "tree now.\n"); 2362 "Insert.contig_index: %d, Insert.free_records: %d, "
2363 "Insert.tree_depth: %d\n",
2364 insert.ins_appending, insert.ins_contig, insert.ins_contig_index,
2365 insert.ins_free_records, insert.ins_tree_depth);
2366
2367 /*
2368 * Avoid growing the tree unless we're out of records and the
2369 * insert type requres one.
2370 */
2371 if (insert.ins_contig != CONTIG_NONE || insert.ins_free_records)
2372 goto out_add;
857 2373
858 shift = ocfs2_find_branch_target(osb, inode, fe_bh, &bh); 2374 shift = ocfs2_find_branch_target(osb, inode, fe_bh, &bh);
859 if (shift < 0) { 2375 if (shift < 0) {
@@ -866,13 +2382,9 @@ int ocfs2_insert_extent(struct ocfs2_super *osb,
866 * and didn't find room for any more extents - we need to add 2382 * and didn't find room for any more extents - we need to add
867 * another tree level */ 2383 * another tree level */
868 if (shift) { 2384 if (shift) {
869 /* if we hit a leaf, we'd better be empty :) */
870 BUG_ON(le16_to_cpu(el->l_next_free_rec) !=
871 le16_to_cpu(el->l_count));
872 BUG_ON(bh); 2385 BUG_ON(bh);
873 mlog(0, "ocfs2_allocate_extent: need to shift tree depth " 2386 mlog(0, "need to shift tree depth "
874 "(current = %u)\n", 2387 "(current = %d)\n", insert.ins_tree_depth);
875 le16_to_cpu(fe->id2.i_list.l_tree_depth));
876 2388
877 /* ocfs2_shift_tree_depth will return us a buffer with 2389 /* ocfs2_shift_tree_depth will return us a buffer with
878 * the new extent block (so we can pass that to 2390 * the new extent block (so we can pass that to
@@ -883,15 +2395,16 @@ int ocfs2_insert_extent(struct ocfs2_super *osb,
883 mlog_errno(status); 2395 mlog_errno(status);
884 goto bail; 2396 goto bail;
885 } 2397 }
2398 insert.ins_tree_depth++;
886 /* Special case: we have room now if we shifted from 2399 /* Special case: we have room now if we shifted from
887 * tree_depth 0 */ 2400 * tree_depth 0 */
888 if (fe->id2.i_list.l_tree_depth == cpu_to_le16(1)) 2401 if (insert.ins_tree_depth == 1)
889 goto out_add; 2402 goto out_add;
890 } 2403 }
891 2404
892 /* call ocfs2_add_branch to add the final part of the tree with 2405 /* call ocfs2_add_branch to add the final part of the tree with
893 * the new data. */ 2406 * the new data. */
894 mlog(0, "ocfs2_allocate_extent: add branch. bh = %p\n", bh); 2407 mlog(0, "add branch. bh = %p\n", bh);
895 status = ocfs2_add_branch(osb, handle, inode, fe_bh, bh, last_eb_bh, 2408 status = ocfs2_add_branch(osb, handle, inode, fe_bh, bh, last_eb_bh,
896 meta_ac); 2409 meta_ac);
897 if (status < 0) { 2410 if (status < 0) {
@@ -900,11 +2413,12 @@ int ocfs2_insert_extent(struct ocfs2_super *osb,
900 } 2413 }
901 2414
902out_add: 2415out_add:
903 /* Finally, we can add clusters. */ 2416 /* Finally, we can add clusters. This might rotate the tree for us. */
904 status = ocfs2_do_insert_extent(osb, handle, inode, fe_bh, 2417 status = ocfs2_do_insert_extent(inode, handle, fe_bh, &rec, &insert);
905 start_blk, new_clusters);
906 if (status < 0) 2418 if (status < 0)
907 mlog_errno(status); 2419 mlog_errno(status);
2420 else
2421 ocfs2_extent_map_insert_rec(inode, &rec);
908 2422
909bail: 2423bail:
910 if (bh) 2424 if (bh)
@@ -1447,168 +2961,389 @@ int ocfs2_truncate_log_init(struct ocfs2_super *osb)
1447 * block will be deleted, and if it will, what the new last extent 2961 * block will be deleted, and if it will, what the new last extent
1448 * block will be so we can update his h_next_leaf_blk field, as well 2962 * block will be so we can update his h_next_leaf_blk field, as well
1449 * as the dinodes i_last_eb_blk */ 2963 * as the dinodes i_last_eb_blk */
1450static int ocfs2_find_new_last_ext_blk(struct ocfs2_super *osb, 2964static int ocfs2_find_new_last_ext_blk(struct inode *inode,
1451 struct inode *inode, 2965 unsigned int clusters_to_del,
1452 struct ocfs2_dinode *fe, 2966 struct ocfs2_path *path,
1453 u32 new_i_clusters,
1454 struct buffer_head *old_last_eb,
1455 struct buffer_head **new_last_eb) 2967 struct buffer_head **new_last_eb)
1456{ 2968{
1457 int i, status = 0; 2969 int next_free, ret = 0;
1458 u64 block = 0; 2970 u32 cpos;
2971 struct ocfs2_extent_rec *rec;
1459 struct ocfs2_extent_block *eb; 2972 struct ocfs2_extent_block *eb;
1460 struct ocfs2_extent_list *el; 2973 struct ocfs2_extent_list *el;
1461 struct buffer_head *bh = NULL; 2974 struct buffer_head *bh = NULL;
1462 2975
1463 *new_last_eb = NULL; 2976 *new_last_eb = NULL;
1464 2977
1465 if (!OCFS2_IS_VALID_DINODE(fe)) {
1466 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
1467 status = -EIO;
1468 goto bail;
1469 }
1470
1471 /* we have no tree, so of course, no last_eb. */ 2978 /* we have no tree, so of course, no last_eb. */
1472 if (!fe->id2.i_list.l_tree_depth) 2979 if (!path->p_tree_depth)
1473 goto bail; 2980 goto out;
1474 2981
1475 /* trunc to zero special case - this makes tree_depth = 0 2982 /* trunc to zero special case - this makes tree_depth = 0
1476 * regardless of what it is. */ 2983 * regardless of what it is. */
1477 if (!new_i_clusters) 2984 if (OCFS2_I(inode)->ip_clusters == clusters_to_del)
1478 goto bail; 2985 goto out;
1479 2986
1480 eb = (struct ocfs2_extent_block *) old_last_eb->b_data; 2987 el = path_leaf_el(path);
1481 el = &(eb->h_list);
1482 BUG_ON(!el->l_next_free_rec); 2988 BUG_ON(!el->l_next_free_rec);
1483 2989
1484 /* Make sure that this guy will actually be empty after we 2990 /*
1485 * clear away the data. */ 2991 * Make sure that this extent list will actually be empty
1486 if (le32_to_cpu(el->l_recs[0].e_cpos) < new_i_clusters) 2992 * after we clear away the data. We can shortcut out if
1487 goto bail; 2993 * there's more than one non-empty extent in the
2994 * list. Otherwise, a check of the remaining extent is
2995 * necessary.
2996 */
2997 next_free = le16_to_cpu(el->l_next_free_rec);
2998 rec = NULL;
2999 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
3000 if (next_free > 2)
3001 goto out;
1488 3002
1489 /* Ok, at this point, we know that last_eb will definitely 3003 /* We may have a valid extent in index 1, check it. */
1490 * change, so lets traverse the tree and find the second to 3004 if (next_free == 2)
1491 * last extent block. */ 3005 rec = &el->l_recs[1];
1492 el = &(fe->id2.i_list); 3006
1493 /* go down the tree, */ 3007 /*
1494 do { 3008 * Fall through - no more nonempty extents, so we want
1495 for(i = (le16_to_cpu(el->l_next_free_rec) - 1); i >= 0; i--) { 3009 * to delete this leaf.
1496 if (le32_to_cpu(el->l_recs[i].e_cpos) < 3010 */
1497 new_i_clusters) { 3011 } else {
1498 block = le64_to_cpu(el->l_recs[i].e_blkno); 3012 if (next_free > 1)
1499 break; 3013 goto out;
1500 } 3014
3015 rec = &el->l_recs[0];
3016 }
3017
3018 if (rec) {
3019 /*
3020 * Check it we'll only be trimming off the end of this
3021 * cluster.
3022 */
3023 if (le16_to_cpu(rec->e_leaf_clusters) > clusters_to_del)
3024 goto out;
3025 }
3026
3027 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path, &cpos);
3028 if (ret) {
3029 mlog_errno(ret);
3030 goto out;
3031 }
3032
3033 ret = ocfs2_find_leaf(inode, path_root_el(path), cpos, &bh);
3034 if (ret) {
3035 mlog_errno(ret);
3036 goto out;
3037 }
3038
3039 eb = (struct ocfs2_extent_block *) bh->b_data;
3040 el = &eb->h_list;
3041 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
3042 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
3043 ret = -EROFS;
3044 goto out;
3045 }
3046
3047 *new_last_eb = bh;
3048 get_bh(*new_last_eb);
3049 mlog(0, "returning block %llu, (cpos: %u)\n",
3050 (unsigned long long)le64_to_cpu(eb->h_blkno), cpos);
3051out:
3052 brelse(bh);
3053
3054 return ret;
3055}
3056
3057/*
3058 * Trim some clusters off the rightmost edge of a tree. Only called
3059 * during truncate.
3060 *
3061 * The caller needs to:
3062 * - start journaling of each path component.
3063 * - compute and fully set up any new last ext block
3064 */
3065static int ocfs2_trim_tree(struct inode *inode, struct ocfs2_path *path,
3066 handle_t *handle, struct ocfs2_truncate_context *tc,
3067 u32 clusters_to_del, u64 *delete_start)
3068{
3069 int ret, i, index = path->p_tree_depth;
3070 u32 new_edge = 0;
3071 u64 deleted_eb = 0;
3072 struct buffer_head *bh;
3073 struct ocfs2_extent_list *el;
3074 struct ocfs2_extent_rec *rec;
3075
3076 *delete_start = 0;
3077
3078 while (index >= 0) {
3079 bh = path->p_node[index].bh;
3080 el = path->p_node[index].el;
3081
3082 mlog(0, "traveling tree (index = %d, block = %llu)\n",
3083 index, (unsigned long long)bh->b_blocknr);
3084
3085 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
3086
3087 if (index !=
3088 (path->p_tree_depth - le16_to_cpu(el->l_tree_depth))) {
3089 ocfs2_error(inode->i_sb,
3090 "Inode %lu has invalid ext. block %llu",
3091 inode->i_ino,
3092 (unsigned long long)bh->b_blocknr);
3093 ret = -EROFS;
3094 goto out;
1501 } 3095 }
1502 BUG_ON(i < 0);
1503 3096
1504 if (bh) { 3097find_tail_record:
1505 brelse(bh); 3098 i = le16_to_cpu(el->l_next_free_rec) - 1;
1506 bh = NULL; 3099 rec = &el->l_recs[i];
3100
3101 mlog(0, "Extent list before: record %d: (%u, %u, %llu), "
3102 "next = %u\n", i, le32_to_cpu(rec->e_cpos),
3103 ocfs2_rec_clusters(el, rec),
3104 (unsigned long long)le64_to_cpu(rec->e_blkno),
3105 le16_to_cpu(el->l_next_free_rec));
3106
3107 BUG_ON(ocfs2_rec_clusters(el, rec) < clusters_to_del);
3108
3109 if (le16_to_cpu(el->l_tree_depth) == 0) {
3110 /*
3111 * If the leaf block contains a single empty
3112 * extent and no records, we can just remove
3113 * the block.
3114 */
3115 if (i == 0 && ocfs2_is_empty_extent(rec)) {
3116 memset(rec, 0,
3117 sizeof(struct ocfs2_extent_rec));
3118 el->l_next_free_rec = cpu_to_le16(0);
3119
3120 goto delete;
3121 }
3122
3123 /*
3124 * Remove any empty extents by shifting things
3125 * left. That should make life much easier on
3126 * the code below. This condition is rare
3127 * enough that we shouldn't see a performance
3128 * hit.
3129 */
3130 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
3131 le16_add_cpu(&el->l_next_free_rec, -1);
3132
3133 for(i = 0;
3134 i < le16_to_cpu(el->l_next_free_rec); i++)
3135 el->l_recs[i] = el->l_recs[i + 1];
3136
3137 memset(&el->l_recs[i], 0,
3138 sizeof(struct ocfs2_extent_rec));
3139
3140 /*
3141 * We've modified our extent list. The
3142 * simplest way to handle this change
3143 * is to being the search from the
3144 * start again.
3145 */
3146 goto find_tail_record;
3147 }
3148
3149 le16_add_cpu(&rec->e_leaf_clusters, -clusters_to_del);
3150
3151 /*
3152 * We'll use "new_edge" on our way back up the
3153 * tree to know what our rightmost cpos is.
3154 */
3155 new_edge = le16_to_cpu(rec->e_leaf_clusters);
3156 new_edge += le32_to_cpu(rec->e_cpos);
3157
3158 /*
3159 * The caller will use this to delete data blocks.
3160 */
3161 *delete_start = le64_to_cpu(rec->e_blkno)
3162 + ocfs2_clusters_to_blocks(inode->i_sb,
3163 le16_to_cpu(rec->e_leaf_clusters));
3164
3165 /*
3166 * If it's now empty, remove this record.
3167 */
3168 if (le16_to_cpu(rec->e_leaf_clusters) == 0) {
3169 memset(rec, 0,
3170 sizeof(struct ocfs2_extent_rec));
3171 le16_add_cpu(&el->l_next_free_rec, -1);
3172 }
3173 } else {
3174 if (le64_to_cpu(rec->e_blkno) == deleted_eb) {
3175 memset(rec, 0,
3176 sizeof(struct ocfs2_extent_rec));
3177 le16_add_cpu(&el->l_next_free_rec, -1);
3178
3179 goto delete;
3180 }
3181
3182 /* Can this actually happen? */
3183 if (le16_to_cpu(el->l_next_free_rec) == 0)
3184 goto delete;
3185
3186 /*
3187 * We never actually deleted any clusters
3188 * because our leaf was empty. There's no
3189 * reason to adjust the rightmost edge then.
3190 */
3191 if (new_edge == 0)
3192 goto delete;
3193
3194 rec->e_int_clusters = cpu_to_le32(new_edge);
3195 le32_add_cpu(&rec->e_int_clusters,
3196 -le32_to_cpu(rec->e_cpos));
3197
3198 /*
3199 * A deleted child record should have been
3200 * caught above.
3201 */
3202 BUG_ON(le32_to_cpu(rec->e_int_clusters) == 0);
1507 } 3203 }
1508 3204
1509 status = ocfs2_read_block(osb, block, &bh, OCFS2_BH_CACHED, 3205delete:
1510 inode); 3206 ret = ocfs2_journal_dirty(handle, bh);
1511 if (status < 0) { 3207 if (ret) {
1512 mlog_errno(status); 3208 mlog_errno(ret);
1513 goto bail; 3209 goto out;
1514 } 3210 }
1515 eb = (struct ocfs2_extent_block *) bh->b_data; 3211
1516 el = &eb->h_list; 3212 mlog(0, "extent list container %llu, after: record %d: "
1517 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) { 3213 "(%u, %u, %llu), next = %u.\n",
1518 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb); 3214 (unsigned long long)bh->b_blocknr, i,
1519 status = -EIO; 3215 le32_to_cpu(rec->e_cpos), ocfs2_rec_clusters(el, rec),
1520 goto bail; 3216 (unsigned long long)le64_to_cpu(rec->e_blkno),
3217 le16_to_cpu(el->l_next_free_rec));
3218
3219 /*
3220 * We must be careful to only attempt delete of an
3221 * extent block (and not the root inode block).
3222 */
3223 if (index > 0 && le16_to_cpu(el->l_next_free_rec) == 0) {
3224 struct ocfs2_extent_block *eb =
3225 (struct ocfs2_extent_block *)bh->b_data;
3226
3227 /*
3228 * Save this for use when processing the
3229 * parent block.
3230 */
3231 deleted_eb = le64_to_cpu(eb->h_blkno);
3232
3233 mlog(0, "deleting this extent block.\n");
3234
3235 ocfs2_remove_from_cache(inode, bh);
3236
3237 BUG_ON(ocfs2_rec_clusters(el, &el->l_recs[0]));
3238 BUG_ON(le32_to_cpu(el->l_recs[0].e_cpos));
3239 BUG_ON(le64_to_cpu(el->l_recs[0].e_blkno));
3240
3241 if (le16_to_cpu(eb->h_suballoc_slot) == 0) {
3242 /*
3243 * This code only understands how to
3244 * lock the suballocator in slot 0,
3245 * which is fine because allocation is
3246 * only ever done out of that
3247 * suballocator too. A future version
3248 * might change that however, so avoid
3249 * a free if we don't know how to
3250 * handle it. This way an fs incompat
3251 * bit will not be necessary.
3252 */
3253 ret = ocfs2_free_extent_block(handle,
3254 tc->tc_ext_alloc_inode,
3255 tc->tc_ext_alloc_bh,
3256 eb);
3257
3258 /* An error here is not fatal. */
3259 if (ret < 0)
3260 mlog_errno(ret);
3261 }
3262 } else {
3263 deleted_eb = 0;
1521 } 3264 }
1522 } while (el->l_tree_depth);
1523 3265
1524 *new_last_eb = bh; 3266 index--;
1525 get_bh(*new_last_eb); 3267 }
1526 mlog(0, "returning block %llu\n",
1527 (unsigned long long)le64_to_cpu(eb->h_blkno));
1528bail:
1529 if (bh)
1530 brelse(bh);
1531 3268
1532 return status; 3269 ret = 0;
3270out:
3271 return ret;
1533} 3272}
1534 3273
1535static int ocfs2_do_truncate(struct ocfs2_super *osb, 3274static int ocfs2_do_truncate(struct ocfs2_super *osb,
1536 unsigned int clusters_to_del, 3275 unsigned int clusters_to_del,
1537 struct inode *inode, 3276 struct inode *inode,
1538 struct buffer_head *fe_bh, 3277 struct buffer_head *fe_bh,
1539 struct buffer_head *old_last_eb_bh,
1540 handle_t *handle, 3278 handle_t *handle,
1541 struct ocfs2_truncate_context *tc) 3279 struct ocfs2_truncate_context *tc,
3280 struct ocfs2_path *path)
1542{ 3281{
1543 int status, i, depth; 3282 int status;
1544 struct ocfs2_dinode *fe; 3283 struct ocfs2_dinode *fe;
1545 struct ocfs2_extent_block *eb;
1546 struct ocfs2_extent_block *last_eb = NULL; 3284 struct ocfs2_extent_block *last_eb = NULL;
1547 struct ocfs2_extent_list *el; 3285 struct ocfs2_extent_list *el;
1548 struct buffer_head *eb_bh = NULL;
1549 struct buffer_head *last_eb_bh = NULL; 3286 struct buffer_head *last_eb_bh = NULL;
1550 u64 next_eb = 0;
1551 u64 delete_blk = 0; 3287 u64 delete_blk = 0;
1552 3288
1553 fe = (struct ocfs2_dinode *) fe_bh->b_data; 3289 fe = (struct ocfs2_dinode *) fe_bh->b_data;
1554 3290
1555 status = ocfs2_find_new_last_ext_blk(osb, 3291 status = ocfs2_find_new_last_ext_blk(inode, clusters_to_del,
1556 inode, 3292 path, &last_eb_bh);
1557 fe,
1558 le32_to_cpu(fe->i_clusters) -
1559 clusters_to_del,
1560 old_last_eb_bh,
1561 &last_eb_bh);
1562 if (status < 0) { 3293 if (status < 0) {
1563 mlog_errno(status); 3294 mlog_errno(status);
1564 goto bail; 3295 goto bail;
1565 } 3296 }
1566 if (last_eb_bh)
1567 last_eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
1568 3297
1569 status = ocfs2_journal_access(handle, inode, fe_bh, 3298 /*
1570 OCFS2_JOURNAL_ACCESS_WRITE); 3299 * Each component will be touched, so we might as well journal
3300 * here to avoid having to handle errors later.
3301 */
3302 status = ocfs2_journal_access_path(inode, handle, path);
1571 if (status < 0) { 3303 if (status < 0) {
1572 mlog_errno(status); 3304 mlog_errno(status);
1573 goto bail; 3305 goto bail;
1574 } 3306 }
3307
3308 if (last_eb_bh) {
3309 status = ocfs2_journal_access(handle, inode, last_eb_bh,
3310 OCFS2_JOURNAL_ACCESS_WRITE);
3311 if (status < 0) {
3312 mlog_errno(status);
3313 goto bail;
3314 }
3315
3316 last_eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
3317 }
3318
1575 el = &(fe->id2.i_list); 3319 el = &(fe->id2.i_list);
1576 3320
3321 /*
3322 * Lower levels depend on this never happening, but it's best
3323 * to check it up here before changing the tree.
3324 */
3325 if (el->l_tree_depth && el->l_recs[0].e_int_clusters == 0) {
3326 ocfs2_error(inode->i_sb,
3327 "Inode %lu has an empty extent record, depth %u\n",
3328 inode->i_ino, le16_to_cpu(el->l_tree_depth));
3329 status = -EROFS;
3330 goto bail;
3331 }
3332
1577 spin_lock(&OCFS2_I(inode)->ip_lock); 3333 spin_lock(&OCFS2_I(inode)->ip_lock);
1578 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters) - 3334 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters) -
1579 clusters_to_del; 3335 clusters_to_del;
1580 spin_unlock(&OCFS2_I(inode)->ip_lock); 3336 spin_unlock(&OCFS2_I(inode)->ip_lock);
1581 le32_add_cpu(&fe->i_clusters, -clusters_to_del); 3337 le32_add_cpu(&fe->i_clusters, -clusters_to_del);
1582 fe->i_mtime = cpu_to_le64(CURRENT_TIME.tv_sec);
1583 fe->i_mtime_nsec = cpu_to_le32(CURRENT_TIME.tv_nsec);
1584
1585 i = le16_to_cpu(el->l_next_free_rec) - 1;
1586
1587 BUG_ON(le32_to_cpu(el->l_recs[i].e_clusters) < clusters_to_del);
1588 le32_add_cpu(&el->l_recs[i].e_clusters, -clusters_to_del);
1589 /* tree depth zero, we can just delete the clusters, otherwise
1590 * we need to record the offset of the next level extent block
1591 * as we may overwrite it. */
1592 if (!el->l_tree_depth)
1593 delete_blk = le64_to_cpu(el->l_recs[i].e_blkno)
1594 + ocfs2_clusters_to_blocks(osb->sb,
1595 le32_to_cpu(el->l_recs[i].e_clusters));
1596 else
1597 next_eb = le64_to_cpu(el->l_recs[i].e_blkno);
1598 3338
1599 if (!el->l_recs[i].e_clusters) { 3339 status = ocfs2_trim_tree(inode, path, handle, tc,
1600 /* if we deleted the whole extent record, then clear 3340 clusters_to_del, &delete_blk);
1601 * out the other fields and update the extent 3341 if (status) {
1602 * list. For depth > 0 trees, we've already recorded 3342 mlog_errno(status);
1603 * the extent block in 'next_eb' */ 3343 goto bail;
1604 el->l_recs[i].e_cpos = 0;
1605 el->l_recs[i].e_blkno = 0;
1606 BUG_ON(!el->l_next_free_rec);
1607 le16_add_cpu(&el->l_next_free_rec, -1);
1608 } 3344 }
1609 3345
1610 depth = le16_to_cpu(el->l_tree_depth); 3346 if (le32_to_cpu(fe->i_clusters) == 0) {
1611 if (!fe->i_clusters) {
1612 /* trunc to zero is a special case. */ 3347 /* trunc to zero is a special case. */
1613 el->l_tree_depth = 0; 3348 el->l_tree_depth = 0;
1614 fe->i_last_eb_blk = 0; 3349 fe->i_last_eb_blk = 0;
@@ -1625,12 +3360,6 @@ static int ocfs2_do_truncate(struct ocfs2_super *osb,
1625 /* If there will be a new last extent block, then by 3360 /* If there will be a new last extent block, then by
1626 * definition, there cannot be any leaves to the right of 3361 * definition, there cannot be any leaves to the right of
1627 * him. */ 3362 * him. */
1628 status = ocfs2_journal_access(handle, inode, last_eb_bh,
1629 OCFS2_JOURNAL_ACCESS_WRITE);
1630 if (status < 0) {
1631 mlog_errno(status);
1632 goto bail;
1633 }
1634 last_eb->h_next_leaf_blk = 0; 3363 last_eb->h_next_leaf_blk = 0;
1635 status = ocfs2_journal_dirty(handle, last_eb_bh); 3364 status = ocfs2_journal_dirty(handle, last_eb_bh);
1636 if (status < 0) { 3365 if (status < 0) {
@@ -1639,123 +3368,247 @@ static int ocfs2_do_truncate(struct ocfs2_super *osb,
1639 } 3368 }
1640 } 3369 }
1641 3370
1642 /* if our tree depth > 0, update all the tree blocks below us. */ 3371 if (delete_blk) {
1643 while (depth) { 3372 status = ocfs2_truncate_log_append(osb, handle, delete_blk,
1644 mlog(0, "traveling tree (depth = %d, next_eb = %llu)\n", 3373 clusters_to_del);
1645 depth, (unsigned long long)next_eb);
1646 status = ocfs2_read_block(osb, next_eb, &eb_bh,
1647 OCFS2_BH_CACHED, inode);
1648 if (status < 0) { 3374 if (status < 0) {
1649 mlog_errno(status); 3375 mlog_errno(status);
1650 goto bail; 3376 goto bail;
1651 } 3377 }
1652 eb = (struct ocfs2_extent_block *)eb_bh->b_data; 3378 }
1653 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) { 3379 status = 0;
1654 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb); 3380bail:
1655 status = -EIO; 3381
1656 goto bail; 3382 mlog_exit(status);
3383 return status;
3384}
3385
3386static int ocfs2_writeback_zero_func(handle_t *handle, struct buffer_head *bh)
3387{
3388 set_buffer_uptodate(bh);
3389 mark_buffer_dirty(bh);
3390 return 0;
3391}
3392
3393static int ocfs2_ordered_zero_func(handle_t *handle, struct buffer_head *bh)
3394{
3395 set_buffer_uptodate(bh);
3396 mark_buffer_dirty(bh);
3397 return ocfs2_journal_dirty_data(handle, bh);
3398}
3399
3400static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t isize,
3401 struct page **pages, int numpages,
3402 u64 phys, handle_t *handle)
3403{
3404 int i, ret, partial = 0;
3405 void *kaddr;
3406 struct page *page;
3407 unsigned int from, to = PAGE_CACHE_SIZE;
3408 struct super_block *sb = inode->i_sb;
3409
3410 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
3411
3412 if (numpages == 0)
3413 goto out;
3414
3415 from = isize & (PAGE_CACHE_SIZE - 1); /* 1st page offset */
3416 if (PAGE_CACHE_SHIFT > OCFS2_SB(sb)->s_clustersize_bits) {
3417 /*
3418 * Since 'from' has been capped to a value below page
3419 * size, this calculation won't be able to overflow
3420 * 'to'
3421 */
3422 to = ocfs2_align_bytes_to_clusters(sb, from);
3423
3424 /*
3425 * The truncate tail in this case should never contain
3426 * more than one page at maximum. The loop below also
3427 * assumes this.
3428 */
3429 BUG_ON(numpages != 1);
3430 }
3431
3432 for(i = 0; i < numpages; i++) {
3433 page = pages[i];
3434
3435 BUG_ON(from > PAGE_CACHE_SIZE);
3436 BUG_ON(to > PAGE_CACHE_SIZE);
3437
3438 ret = ocfs2_map_page_blocks(page, &phys, inode, from, to, 0);
3439 if (ret)
3440 mlog_errno(ret);
3441
3442 kaddr = kmap_atomic(page, KM_USER0);
3443 memset(kaddr + from, 0, to - from);
3444 kunmap_atomic(kaddr, KM_USER0);
3445
3446 /*
3447 * Need to set the buffers we zero'd into uptodate
3448 * here if they aren't - ocfs2_map_page_blocks()
3449 * might've skipped some
3450 */
3451 if (ocfs2_should_order_data(inode)) {
3452 ret = walk_page_buffers(handle,
3453 page_buffers(page),
3454 from, to, &partial,
3455 ocfs2_ordered_zero_func);
3456 if (ret < 0)
3457 mlog_errno(ret);
3458 } else {
3459 ret = walk_page_buffers(handle, page_buffers(page),
3460 from, to, &partial,
3461 ocfs2_writeback_zero_func);
3462 if (ret < 0)
3463 mlog_errno(ret);
1657 } 3464 }
1658 el = &(eb->h_list);
1659 3465
1660 status = ocfs2_journal_access(handle, inode, eb_bh, 3466 if (!partial)
1661 OCFS2_JOURNAL_ACCESS_WRITE); 3467 SetPageUptodate(page);
1662 if (status < 0) { 3468
1663 mlog_errno(status); 3469 flush_dcache_page(page);
1664 goto bail; 3470
3471 /*
3472 * Every page after the 1st one should be completely zero'd.
3473 */
3474 from = 0;
3475 }
3476out:
3477 if (pages) {
3478 for (i = 0; i < numpages; i++) {
3479 page = pages[i];
3480 unlock_page(page);
3481 mark_page_accessed(page);
3482 page_cache_release(page);
1665 } 3483 }
3484 }
3485}
1666 3486
1667 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0); 3487static int ocfs2_grab_eof_pages(struct inode *inode, loff_t isize, struct page **pages,
1668 BUG_ON(depth != (le16_to_cpu(el->l_tree_depth) + 1)); 3488 int *num, u64 *phys)
3489{
3490 int i, numpages = 0, ret = 0;
3491 unsigned int csize = OCFS2_SB(inode->i_sb)->s_clustersize;
3492 unsigned int ext_flags;
3493 struct super_block *sb = inode->i_sb;
3494 struct address_space *mapping = inode->i_mapping;
3495 unsigned long index;
3496 u64 next_cluster_bytes;
3497
3498 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
3499
3500 /* Cluster boundary, so we don't need to grab any pages. */
3501 if ((isize & (csize - 1)) == 0)
3502 goto out;
1669 3503
1670 i = le16_to_cpu(el->l_next_free_rec) - 1; 3504 ret = ocfs2_extent_map_get_blocks(inode, isize >> sb->s_blocksize_bits,
3505 phys, NULL, &ext_flags);
3506 if (ret) {
3507 mlog_errno(ret);
3508 goto out;
3509 }
1671 3510
1672 mlog(0, "extent block %llu, before: record %d: " 3511 /* Tail is a hole. */
1673 "(%u, %u, %llu), next = %u\n", 3512 if (*phys == 0)
1674 (unsigned long long)le64_to_cpu(eb->h_blkno), i, 3513 goto out;
1675 le32_to_cpu(el->l_recs[i].e_cpos),
1676 le32_to_cpu(el->l_recs[i].e_clusters),
1677 (unsigned long long)le64_to_cpu(el->l_recs[i].e_blkno),
1678 le16_to_cpu(el->l_next_free_rec));
1679 3514
1680 BUG_ON(le32_to_cpu(el->l_recs[i].e_clusters) < clusters_to_del); 3515 /* Tail is marked as unwritten, we can count on write to zero
1681 le32_add_cpu(&el->l_recs[i].e_clusters, -clusters_to_del); 3516 * in that case. */
1682 3517 if (ext_flags & OCFS2_EXT_UNWRITTEN)
1683 next_eb = le64_to_cpu(el->l_recs[i].e_blkno); 3518 goto out;
1684 /* bottom-most block requires us to delete data.*/
1685 if (!el->l_tree_depth)
1686 delete_blk = le64_to_cpu(el->l_recs[i].e_blkno)
1687 + ocfs2_clusters_to_blocks(osb->sb,
1688 le32_to_cpu(el->l_recs[i].e_clusters));
1689 if (!el->l_recs[i].e_clusters) {
1690 el->l_recs[i].e_cpos = 0;
1691 el->l_recs[i].e_blkno = 0;
1692 BUG_ON(!el->l_next_free_rec);
1693 le16_add_cpu(&el->l_next_free_rec, -1);
1694 }
1695 mlog(0, "extent block %llu, after: record %d: "
1696 "(%u, %u, %llu), next = %u\n",
1697 (unsigned long long)le64_to_cpu(eb->h_blkno), i,
1698 le32_to_cpu(el->l_recs[i].e_cpos),
1699 le32_to_cpu(el->l_recs[i].e_clusters),
1700 (unsigned long long)le64_to_cpu(el->l_recs[i].e_blkno),
1701 le16_to_cpu(el->l_next_free_rec));
1702 3519
1703 status = ocfs2_journal_dirty(handle, eb_bh); 3520 next_cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, isize);
1704 if (status < 0) { 3521 index = isize >> PAGE_CACHE_SHIFT;
1705 mlog_errno(status); 3522 do {
1706 goto bail; 3523 pages[numpages] = grab_cache_page(mapping, index);
3524 if (!pages[numpages]) {
3525 ret = -ENOMEM;
3526 mlog_errno(ret);
3527 goto out;
1707 } 3528 }
1708 3529
1709 if (!el->l_next_free_rec) { 3530 numpages++;
1710 mlog(0, "deleting this extent block.\n"); 3531 index++;
1711 3532 } while (index < (next_cluster_bytes >> PAGE_CACHE_SHIFT));
1712 ocfs2_remove_from_cache(inode, eb_bh);
1713 3533
1714 BUG_ON(el->l_recs[0].e_clusters); 3534out:
1715 BUG_ON(el->l_recs[0].e_cpos); 3535 if (ret != 0) {
1716 BUG_ON(el->l_recs[0].e_blkno); 3536 if (pages) {
1717 if (eb->h_suballoc_slot == 0) { 3537 for (i = 0; i < numpages; i++) {
1718 /* 3538 if (pages[i]) {
1719 * This code only understands how to 3539 unlock_page(pages[i]);
1720 * lock the suballocator in slot 0, 3540 page_cache_release(pages[i]);
1721 * which is fine because allocation is
1722 * only ever done out of that
1723 * suballocator too. A future version
1724 * might change that however, so avoid
1725 * a free if we don't know how to
1726 * handle it. This way an fs incompat
1727 * bit will not be necessary.
1728 */
1729 status = ocfs2_free_extent_block(handle,
1730 tc->tc_ext_alloc_inode,
1731 tc->tc_ext_alloc_bh,
1732 eb);
1733 if (status < 0) {
1734 mlog_errno(status);
1735 goto bail;
1736 } 3541 }
1737 } 3542 }
1738 } 3543 }
1739 brelse(eb_bh); 3544 numpages = 0;
1740 eb_bh = NULL;
1741 depth--;
1742 } 3545 }
1743 3546
1744 BUG_ON(!delete_blk); 3547 *num = numpages;
1745 status = ocfs2_truncate_log_append(osb, handle, delete_blk, 3548
1746 clusters_to_del); 3549 return ret;
1747 if (status < 0) { 3550}
1748 mlog_errno(status); 3551
1749 goto bail; 3552/*
3553 * Zero the area past i_size but still within an allocated
3554 * cluster. This avoids exposing nonzero data on subsequent file
3555 * extends.
3556 *
3557 * We need to call this before i_size is updated on the inode because
3558 * otherwise block_write_full_page() will skip writeout of pages past
3559 * i_size. The new_i_size parameter is passed for this reason.
3560 */
3561int ocfs2_zero_tail_for_truncate(struct inode *inode, handle_t *handle,
3562 u64 new_i_size)
3563{
3564 int ret, numpages;
3565 loff_t endbyte;
3566 struct page **pages = NULL;
3567 u64 phys;
3568
3569 /*
3570 * File systems which don't support sparse files zero on every
3571 * extend.
3572 */
3573 if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
3574 return 0;
3575
3576 pages = kcalloc(ocfs2_pages_per_cluster(inode->i_sb),
3577 sizeof(struct page *), GFP_NOFS);
3578 if (pages == NULL) {
3579 ret = -ENOMEM;
3580 mlog_errno(ret);
3581 goto out;
1750 } 3582 }
1751 status = 0; 3583
1752bail: 3584 ret = ocfs2_grab_eof_pages(inode, new_i_size, pages, &numpages, &phys);
1753 if (!status) 3585 if (ret) {
1754 ocfs2_extent_map_trunc(inode, le32_to_cpu(fe->i_clusters)); 3586 mlog_errno(ret);
1755 else 3587 goto out;
1756 ocfs2_extent_map_drop(inode, 0); 3588 }
1757 mlog_exit(status); 3589
1758 return status; 3590 if (numpages == 0)
3591 goto out;
3592
3593 ocfs2_zero_cluster_pages(inode, new_i_size, pages, numpages, phys,
3594 handle);
3595
3596 /*
3597 * Initiate writeout of the pages we zero'd here. We don't
3598 * wait on them - the truncate_inode_pages() call later will
3599 * do that for us.
3600 */
3601 endbyte = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
3602 ret = do_sync_mapping_range(inode->i_mapping, new_i_size,
3603 endbyte - 1, SYNC_FILE_RANGE_WRITE);
3604 if (ret)
3605 mlog_errno(ret);
3606
3607out:
3608 if (pages)
3609 kfree(pages);
3610
3611 return ret;
1759} 3612}
1760 3613
1761/* 3614/*
@@ -1770,82 +3623,90 @@ int ocfs2_commit_truncate(struct ocfs2_super *osb,
1770 struct ocfs2_truncate_context *tc) 3623 struct ocfs2_truncate_context *tc)
1771{ 3624{
1772 int status, i, credits, tl_sem = 0; 3625 int status, i, credits, tl_sem = 0;
1773 u32 clusters_to_del, target_i_clusters; 3626 u32 clusters_to_del, new_highest_cpos, range;
1774 u64 last_eb = 0;
1775 struct ocfs2_dinode *fe;
1776 struct ocfs2_extent_block *eb;
1777 struct ocfs2_extent_list *el; 3627 struct ocfs2_extent_list *el;
1778 struct buffer_head *last_eb_bh;
1779 handle_t *handle = NULL; 3628 handle_t *handle = NULL;
1780 struct inode *tl_inode = osb->osb_tl_inode; 3629 struct inode *tl_inode = osb->osb_tl_inode;
3630 struct ocfs2_path *path = NULL;
1781 3631
1782 mlog_entry_void(); 3632 mlog_entry_void();
1783 3633
1784 down_write(&OCFS2_I(inode)->ip_alloc_sem); 3634 down_write(&OCFS2_I(inode)->ip_alloc_sem);
1785 3635
1786 target_i_clusters = ocfs2_clusters_for_bytes(osb->sb, 3636 new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
1787 i_size_read(inode)); 3637 i_size_read(inode));
1788 3638
1789 last_eb_bh = tc->tc_last_eb_bh; 3639 path = ocfs2_new_inode_path(fe_bh);
1790 tc->tc_last_eb_bh = NULL; 3640 if (!path) {
3641 status = -ENOMEM;
3642 mlog_errno(status);
3643 goto bail;
3644 }
1791 3645
1792 fe = (struct ocfs2_dinode *) fe_bh->b_data; 3646 ocfs2_extent_map_trunc(inode, new_highest_cpos);
1793 3647
1794 if (fe->id2.i_list.l_tree_depth) {
1795 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
1796 el = &eb->h_list;
1797 } else
1798 el = &fe->id2.i_list;
1799 last_eb = le64_to_cpu(fe->i_last_eb_blk);
1800start: 3648start:
1801 mlog(0, "ocfs2_commit_truncate: fe->i_clusters = %u, " 3649 /*
1802 "last_eb = %llu, fe->i_last_eb_blk = %llu, " 3650 * Check that we still have allocation to delete.
1803 "fe->id2.i_list.l_tree_depth = %u last_eb_bh = %p\n", 3651 */
1804 le32_to_cpu(fe->i_clusters), (unsigned long long)last_eb, 3652 if (OCFS2_I(inode)->ip_clusters == 0) {
1805 (unsigned long long)le64_to_cpu(fe->i_last_eb_blk), 3653 status = 0;
1806 le16_to_cpu(fe->id2.i_list.l_tree_depth), last_eb_bh); 3654 goto bail;
1807 3655 }
1808 if (last_eb != le64_to_cpu(fe->i_last_eb_blk)) {
1809 mlog(0, "last_eb changed!\n");
1810 BUG_ON(!fe->id2.i_list.l_tree_depth);
1811 last_eb = le64_to_cpu(fe->i_last_eb_blk);
1812 /* i_last_eb_blk may have changed, read it if
1813 * necessary. We don't have to worry about the
1814 * truncate to zero case here (where there becomes no
1815 * last_eb) because we never loop back after our work
1816 * is done. */
1817 if (last_eb_bh) {
1818 brelse(last_eb_bh);
1819 last_eb_bh = NULL;
1820 }
1821 3656
1822 status = ocfs2_read_block(osb, last_eb, 3657 /*
1823 &last_eb_bh, OCFS2_BH_CACHED, 3658 * Truncate always works against the rightmost tree branch.
1824 inode); 3659 */
1825 if (status < 0) { 3660 status = ocfs2_find_path(inode, path, UINT_MAX);
1826 mlog_errno(status); 3661 if (status) {
1827 goto bail; 3662 mlog_errno(status);
1828 } 3663 goto bail;
1829 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data; 3664 }
1830 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) { 3665
1831 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb); 3666 mlog(0, "inode->ip_clusters = %u, tree_depth = %u\n",
1832 status = -EIO; 3667 OCFS2_I(inode)->ip_clusters, path->p_tree_depth);
1833 goto bail; 3668
1834 } 3669 /*
1835 el = &(eb->h_list); 3670 * By now, el will point to the extent list on the bottom most
3671 * portion of this tree. Only the tail record is considered in
3672 * each pass.
3673 *
3674 * We handle the following cases, in order:
3675 * - empty extent: delete the remaining branch
3676 * - remove the entire record
3677 * - remove a partial record
3678 * - no record needs to be removed (truncate has completed)
3679 */
3680 el = path_leaf_el(path);
3681 if (le16_to_cpu(el->l_next_free_rec) == 0) {
3682 ocfs2_error(inode->i_sb,
3683 "Inode %llu has empty extent block at %llu\n",
3684 (unsigned long long)OCFS2_I(inode)->ip_blkno,
3685 (unsigned long long)path_leaf_bh(path)->b_blocknr);
3686 status = -EROFS;
3687 goto bail;
1836 } 3688 }
1837 3689
1838 /* by now, el will point to the extent list on the bottom most
1839 * portion of this tree. */
1840 i = le16_to_cpu(el->l_next_free_rec) - 1; 3690 i = le16_to_cpu(el->l_next_free_rec) - 1;
1841 if (le32_to_cpu(el->l_recs[i].e_cpos) >= target_i_clusters) 3691 range = le32_to_cpu(el->l_recs[i].e_cpos) +
1842 clusters_to_del = le32_to_cpu(el->l_recs[i].e_clusters); 3692 ocfs2_rec_clusters(el, &el->l_recs[i]);
1843 else 3693 if (i == 0 && ocfs2_is_empty_extent(&el->l_recs[i])) {
1844 clusters_to_del = (le32_to_cpu(el->l_recs[i].e_clusters) + 3694 clusters_to_del = 0;
3695 } else if (le32_to_cpu(el->l_recs[i].e_cpos) >= new_highest_cpos) {
3696 clusters_to_del = ocfs2_rec_clusters(el, &el->l_recs[i]);
3697 } else if (range > new_highest_cpos) {
3698 clusters_to_del = (ocfs2_rec_clusters(el, &el->l_recs[i]) +
1845 le32_to_cpu(el->l_recs[i].e_cpos)) - 3699 le32_to_cpu(el->l_recs[i].e_cpos)) -
1846 target_i_clusters; 3700 new_highest_cpos;
3701 } else {
3702 status = 0;
3703 goto bail;
3704 }
1847 3705
1848 mlog(0, "clusters_to_del = %u in this pass\n", clusters_to_del); 3706 mlog(0, "clusters_to_del = %u in this pass, tail blk=%llu\n",
3707 clusters_to_del, (unsigned long long)path_leaf_bh(path)->b_blocknr);
3708
3709 BUG_ON(clusters_to_del == 0);
1849 3710
1850 mutex_lock(&tl_inode->i_mutex); 3711 mutex_lock(&tl_inode->i_mutex);
1851 tl_sem = 1; 3712 tl_sem = 1;
@@ -1861,7 +3722,8 @@ start:
1861 } 3722 }
1862 3723
1863 credits = ocfs2_calc_tree_trunc_credits(osb->sb, clusters_to_del, 3724 credits = ocfs2_calc_tree_trunc_credits(osb->sb, clusters_to_del,
1864 fe, el); 3725 (struct ocfs2_dinode *)fe_bh->b_data,
3726 el);
1865 handle = ocfs2_start_trans(osb, credits); 3727 handle = ocfs2_start_trans(osb, credits);
1866 if (IS_ERR(handle)) { 3728 if (IS_ERR(handle)) {
1867 status = PTR_ERR(handle); 3729 status = PTR_ERR(handle);
@@ -1870,13 +3732,8 @@ start:
1870 goto bail; 3732 goto bail;
1871 } 3733 }
1872 3734
1873 inode->i_ctime = inode->i_mtime = CURRENT_TIME; 3735 status = ocfs2_do_truncate(osb, clusters_to_del, inode, fe_bh, handle,
1874 status = ocfs2_mark_inode_dirty(handle, inode, fe_bh); 3736 tc, path);
1875 if (status < 0)
1876 mlog_errno(status);
1877
1878 status = ocfs2_do_truncate(osb, clusters_to_del, inode, fe_bh,
1879 last_eb_bh, handle, tc);
1880 if (status < 0) { 3737 if (status < 0) {
1881 mlog_errno(status); 3738 mlog_errno(status);
1882 goto bail; 3739 goto bail;
@@ -1888,9 +3745,14 @@ start:
1888 ocfs2_commit_trans(osb, handle); 3745 ocfs2_commit_trans(osb, handle);
1889 handle = NULL; 3746 handle = NULL;
1890 3747
1891 BUG_ON(le32_to_cpu(fe->i_clusters) < target_i_clusters); 3748 ocfs2_reinit_path(path, 1);
1892 if (le32_to_cpu(fe->i_clusters) > target_i_clusters) 3749
1893 goto start; 3750 /*
3751 * The check above will catch the case where we've truncated
3752 * away all allocation.
3753 */
3754 goto start;
3755
1894bail: 3756bail:
1895 up_write(&OCFS2_I(inode)->ip_alloc_sem); 3757 up_write(&OCFS2_I(inode)->ip_alloc_sem);
1896 3758
@@ -1902,8 +3764,7 @@ bail:
1902 if (handle) 3764 if (handle)
1903 ocfs2_commit_trans(osb, handle); 3765 ocfs2_commit_trans(osb, handle);
1904 3766
1905 if (last_eb_bh) 3767 ocfs2_free_path(path);
1906 brelse(last_eb_bh);
1907 3768
1908 /* This will drop the ext_alloc cluster lock for us */ 3769 /* This will drop the ext_alloc cluster lock for us */
1909 ocfs2_free_truncate_context(tc); 3770 ocfs2_free_truncate_context(tc);
@@ -1912,7 +3773,6 @@ bail:
1912 return status; 3773 return status;
1913} 3774}
1914 3775
1915
1916/* 3776/*
1917 * Expects the inode to already be locked. This will figure out which 3777 * Expects the inode to already be locked. This will figure out which
1918 * inodes need to be locked and will put them on the returned truncate 3778 * inodes need to be locked and will put them on the returned truncate
@@ -1923,7 +3783,7 @@ int ocfs2_prepare_truncate(struct ocfs2_super *osb,
1923 struct buffer_head *fe_bh, 3783 struct buffer_head *fe_bh,
1924 struct ocfs2_truncate_context **tc) 3784 struct ocfs2_truncate_context **tc)
1925{ 3785{
1926 int status, metadata_delete; 3786 int status, metadata_delete, i;
1927 unsigned int new_i_clusters; 3787 unsigned int new_i_clusters;
1928 struct ocfs2_dinode *fe; 3788 struct ocfs2_dinode *fe;
1929 struct ocfs2_extent_block *eb; 3789 struct ocfs2_extent_block *eb;
@@ -1944,21 +3804,6 @@ int ocfs2_prepare_truncate(struct ocfs2_super *osb,
1944 "%llu\n", fe->i_clusters, new_i_clusters, 3804 "%llu\n", fe->i_clusters, new_i_clusters,
1945 (unsigned long long)fe->i_size); 3805 (unsigned long long)fe->i_size);
1946 3806
1947 if (le32_to_cpu(fe->i_clusters) <= new_i_clusters) {
1948 ocfs2_error(inode->i_sb, "Dinode %llu has cluster count "
1949 "%u and size %llu whereas struct inode has "
1950 "cluster count %u and size %llu which caused an "
1951 "invalid truncate to %u clusters.",
1952 (unsigned long long)le64_to_cpu(fe->i_blkno),
1953 le32_to_cpu(fe->i_clusters),
1954 (unsigned long long)le64_to_cpu(fe->i_size),
1955 OCFS2_I(inode)->ip_clusters, i_size_read(inode),
1956 new_i_clusters);
1957 mlog_meta_lvb(ML_ERROR, &OCFS2_I(inode)->ip_meta_lockres);
1958 status = -EIO;
1959 goto bail;
1960 }
1961
1962 *tc = kzalloc(sizeof(struct ocfs2_truncate_context), GFP_KERNEL); 3807 *tc = kzalloc(sizeof(struct ocfs2_truncate_context), GFP_KERNEL);
1963 if (!(*tc)) { 3808 if (!(*tc)) {
1964 status = -ENOMEM; 3809 status = -ENOMEM;
@@ -1986,7 +3831,15 @@ int ocfs2_prepare_truncate(struct ocfs2_super *osb,
1986 goto bail; 3831 goto bail;
1987 } 3832 }
1988 el = &(eb->h_list); 3833 el = &(eb->h_list);
1989 if (le32_to_cpu(el->l_recs[0].e_cpos) >= new_i_clusters) 3834
3835 i = 0;
3836 if (ocfs2_is_empty_extent(&el->l_recs[0]))
3837 i = 1;
3838 /*
3839 * XXX: Should we check that next_free_rec contains
3840 * the extent?
3841 */
3842 if (le32_to_cpu(el->l_recs[i].e_cpos) >= new_i_clusters)
1990 metadata_delete = 1; 3843 metadata_delete = 1;
1991 } 3844 }
1992 3845
generated by cgit v1.2.2 (git 2.25.0) at 2026-01-06 08:42:31 -0500