aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorLinus Torvalds <torvalds@woody.linux-foundation.org>2007-04-27 13:29:56 -0400
committerLinus Torvalds <torvalds@woody.linux-foundation.org>2007-04-27 13:29:56 -0400
commitea6db58f3ea55f413c882095d2afaea8137f4f8c (patch)
tree9f7509b5dfe0fdd422b3e2b3a98ed8321d796c66
parentc58b8e4a25a1ba347a0e5d21984c97bd296f1691 (diff)
parent83418978827324918a8cd25ce5227312de1d4468 (diff)
Merge branch 'upstream-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mfasheh/ocfs2
* 'upstream-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mfasheh/ocfs2: (27 commits) ocfs2: Cache extent records ocfs2: Remember rw lock level during direct io ocfs2: Fix up i_blocks calculation to know about holes ocfs2: Fix extent lookup to return true size of holes ocfs2: Read from an unwritten extent returns zeros ocfs2: make room for unwritten extents flag ocfs2: Use own splice write actor ocfs2: Use do_sync_mapping_range() in ocfs2_zero_tail_for_truncate() [PATCH] Turn do_sync_file_range() into do_sync_mapping_range() ocfs2: zero tail of sparse files on truncate ocfs2: Teach ocfs2_get_block() about holes ocfs2: remove ocfs2_prepare_write() and ocfs2_commit_write() ocfs2: teach ocfs2_file_aio_write() about sparse files ocfs2: Turn off shared writeable mmap for local files systems with holes. ocfs2: abstract out allocation locking ocfs2: teach extend/truncate about sparse files ocfs2: temporarily remove extent map caching ocfs2: sparse b-tree support ocfs2: small cleanup of ocfs2_request_delete() ocfs2: remove unused code ...
Diffstat
-rw-r--r--fs/ocfs2/alloc.c3037
-rw-r--r--fs/ocfs2/alloc.h27
-rw-r--r--fs/ocfs2/aops.c1011
-rw-r--r--fs/ocfs2/aops.h77
-rw-r--r--fs/ocfs2/cluster/quorum.c5
-rw-r--r--fs/ocfs2/cluster/tcp_internal.h5
-rw-r--r--fs/ocfs2/dir.c15
-rw-r--r--fs/ocfs2/dlm/dlmdomain.c5
-rw-r--r--fs/ocfs2/dlm/dlmrecovery.c2
-rw-r--r--fs/ocfs2/dlmglue.c143
-rw-r--r--fs/ocfs2/dlmglue.h3
-rw-r--r--fs/ocfs2/extent_map.c1233
-rw-r--r--fs/ocfs2/extent_map.h39
-rw-r--r--fs/ocfs2/file.c637
-rw-r--r--fs/ocfs2/file.h5
-rw-r--r--fs/ocfs2/inode.c199
-rw-r--r--fs/ocfs2/inode.h23
-rw-r--r--fs/ocfs2/journal.c24
-rw-r--r--fs/ocfs2/journal.h2
-rw-r--r--fs/ocfs2/mmap.c7
-rw-r--r--fs/ocfs2/namei.c23
-rw-r--r--fs/ocfs2/ocfs2.h55
-rw-r--r--fs/ocfs2/ocfs2_fs.h31
-rw-r--r--fs/ocfs2/ocfs2_lockid.h5
-rw-r--r--fs/ocfs2/slot_map.c2
-rw-r--r--fs/ocfs2/suballoc.c3
-rw-r--r--fs/ocfs2/super.c7
-rw-r--r--fs/ocfs2/vote.c289
-rw-r--r--fs/ocfs2/vote.h3
-rw-r--r--fs/sync.c8
-rw-r--r--include/linux/fs.h9
31 files changed, 4697 insertions, 2237 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
diff --git a/fs/ocfs2/alloc.h b/fs/ocfs2/alloc.h
index 0b82e8044325..fbcb5934a081 100644
--- a/fs/ocfs2/alloc.h
+++ b/fs/ocfs2/alloc.h
@@ -31,7 +31,8 @@ int ocfs2_insert_extent(struct ocfs2_super *osb,
31 handle_t *handle, 31 handle_t *handle,
32 struct inode *inode, 32 struct inode *inode,
33 struct buffer_head *fe_bh, 33 struct buffer_head *fe_bh,
34 u64 blkno, 34 u32 cpos,
35 u64 start_blk,
35 u32 new_clusters, 36 u32 new_clusters,
36 struct ocfs2_alloc_context *meta_ac); 37 struct ocfs2_alloc_context *meta_ac);
37int ocfs2_num_free_extents(struct ocfs2_super *osb, 38int ocfs2_num_free_extents(struct ocfs2_super *osb,
@@ -70,6 +71,8 @@ struct ocfs2_truncate_context {
70 struct buffer_head *tc_last_eb_bh; 71 struct buffer_head *tc_last_eb_bh;
71}; 72};
72 73
74int ocfs2_zero_tail_for_truncate(struct inode *inode, handle_t *handle,
75 u64 new_i_size);
73int ocfs2_prepare_truncate(struct ocfs2_super *osb, 76int ocfs2_prepare_truncate(struct ocfs2_super *osb,
74 struct inode *inode, 77 struct inode *inode,
75 struct buffer_head *fe_bh, 78 struct buffer_head *fe_bh,
@@ -79,4 +82,26 @@ int ocfs2_commit_truncate(struct ocfs2_super *osb,
79 struct buffer_head *fe_bh, 82 struct buffer_head *fe_bh,
80 struct ocfs2_truncate_context *tc); 83 struct ocfs2_truncate_context *tc);
81 84
85int ocfs2_find_leaf(struct inode *inode, struct ocfs2_extent_list *root_el,
86 u32 cpos, struct buffer_head **leaf_bh);
87
88/*
89 * Helper function to look at the # of clusters in an extent record.
90 */
91static inline unsigned int ocfs2_rec_clusters(struct ocfs2_extent_list *el,
92 struct ocfs2_extent_rec *rec)
93{
94 /*
95 * Cluster count in extent records is slightly different
96 * between interior nodes and leaf nodes. This is to support
97 * unwritten extents which need a flags field in leaf node
98 * records, thus shrinking the available space for a clusters
99 * field.
100 */
101 if (el->l_tree_depth)
102 return le32_to_cpu(rec->e_int_clusters);
103 else
104 return le16_to_cpu(rec->e_leaf_clusters);
105}
106
82#endif /* OCFS2_ALLOC_H */ 107#endif /* OCFS2_ALLOC_H */
diff --git a/fs/ocfs2/aops.c b/fs/ocfs2/aops.c
index 875c11443817..56963e6c46c0 100644
--- a/fs/ocfs2/aops.c
+++ b/fs/ocfs2/aops.c
@@ -24,6 +24,8 @@
24#include <linux/highmem.h> 24#include <linux/highmem.h>
25#include <linux/pagemap.h> 25#include <linux/pagemap.h>
26#include <asm/byteorder.h> 26#include <asm/byteorder.h>
27#include <linux/swap.h>
28#include <linux/pipe_fs_i.h>
27 29
28#define MLOG_MASK_PREFIX ML_FILE_IO 30#define MLOG_MASK_PREFIX ML_FILE_IO
29#include <cluster/masklog.h> 31#include <cluster/masklog.h>
@@ -37,6 +39,7 @@
37#include "file.h" 39#include "file.h"
38#include "inode.h" 40#include "inode.h"
39#include "journal.h" 41#include "journal.h"
42#include "suballoc.h"
40#include "super.h" 43#include "super.h"
41#include "symlink.h" 44#include "symlink.h"
42 45
@@ -134,7 +137,9 @@ static int ocfs2_get_block(struct inode *inode, sector_t iblock,
134 struct buffer_head *bh_result, int create) 137 struct buffer_head *bh_result, int create)
135{ 138{
136 int err = 0; 139 int err = 0;
140 unsigned int ext_flags;
137 u64 p_blkno, past_eof; 141 u64 p_blkno, past_eof;
142 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
138 143
139 mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode, 144 mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
140 (unsigned long long)iblock, bh_result, create); 145 (unsigned long long)iblock, bh_result, create);
@@ -149,17 +154,8 @@ static int ocfs2_get_block(struct inode *inode, sector_t iblock,
149 goto bail; 154 goto bail;
150 } 155 }
151 156
152 /* this can happen if another node truncs after our extend! */ 157 err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, NULL,
153 spin_lock(&OCFS2_I(inode)->ip_lock); 158 &ext_flags);
154 if (iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
155 OCFS2_I(inode)->ip_clusters))
156 err = -EIO;
157 spin_unlock(&OCFS2_I(inode)->ip_lock);
158 if (err)
159 goto bail;
160
161 err = ocfs2_extent_map_get_blocks(inode, iblock, 1, &p_blkno,
162 NULL);
163 if (err) { 159 if (err) {
164 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, " 160 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
165 "%llu, NULL)\n", err, inode, (unsigned long long)iblock, 161 "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
@@ -167,22 +163,39 @@ static int ocfs2_get_block(struct inode *inode, sector_t iblock,
167 goto bail; 163 goto bail;
168 } 164 }
169 165
170 map_bh(bh_result, inode->i_sb, p_blkno); 166 /*
171 167 * ocfs2 never allocates in this function - the only time we
172 if (bh_result->b_blocknr == 0) { 168 * need to use BH_New is when we're extending i_size on a file
173 err = -EIO; 169 * system which doesn't support holes, in which case BH_New
174 mlog(ML_ERROR, "iblock = %llu p_blkno = %llu blkno=(%llu)\n", 170 * allows block_prepare_write() to zero.
175 (unsigned long long)iblock, 171 */
176 (unsigned long long)p_blkno, 172 mlog_bug_on_msg(create && p_blkno == 0 && ocfs2_sparse_alloc(osb),
177 (unsigned long long)OCFS2_I(inode)->ip_blkno); 173 "ino %lu, iblock %llu\n", inode->i_ino,
178 } 174 (unsigned long long)iblock);
175
176 /* Treat the unwritten extent as a hole for zeroing purposes. */
177 if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
178 map_bh(bh_result, inode->i_sb, p_blkno);
179
180 if (!ocfs2_sparse_alloc(osb)) {
181 if (p_blkno == 0) {
182 err = -EIO;
183 mlog(ML_ERROR,
184 "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
185 (unsigned long long)iblock,
186 (unsigned long long)p_blkno,
187 (unsigned long long)OCFS2_I(inode)->ip_blkno);
188 mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
189 dump_stack();
190 }
179 191
180 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode)); 192 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
181 mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino, 193 mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino,
182 (unsigned long long)past_eof); 194 (unsigned long long)past_eof);
183 195
184 if (create && (iblock >= past_eof)) 196 if (create && (iblock >= past_eof))
185 set_buffer_new(bh_result); 197 set_buffer_new(bh_result);
198 }
186 199
187bail: 200bail:
188 if (err < 0) 201 if (err < 0)
@@ -276,8 +289,11 @@ static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
276 return ret; 289 return ret;
277} 290}
278 291
279/* This can also be called from ocfs2_write_zero_page() which has done 292/*
280 * it's own cluster locking. */ 293 * This is called from ocfs2_write_zero_page() which has handled it's
294 * own cluster locking and has ensured allocation exists for those
295 * blocks to be written.
296 */
281int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page, 297int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
282 unsigned from, unsigned to) 298 unsigned from, unsigned to)
283{ 299{
@@ -292,44 +308,17 @@ int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
292 return ret; 308 return ret;
293} 309}
294 310
295/*
296 * ocfs2_prepare_write() can be an outer-most ocfs2 call when it is called
297 * from loopback. It must be able to perform its own locking around
298 * ocfs2_get_block().
299 */
300static int ocfs2_prepare_write(struct file *file, struct page *page,
301 unsigned from, unsigned to)
302{
303 struct inode *inode = page->mapping->host;
304 int ret;
305
306 mlog_entry("(0x%p, 0x%p, %u, %u)\n", file, page, from, to);
307
308 ret = ocfs2_meta_lock_with_page(inode, NULL, 0, page);
309 if (ret != 0) {
310 mlog_errno(ret);
311 goto out;
312 }
313
314 ret = ocfs2_prepare_write_nolock(inode, page, from, to);
315
316 ocfs2_meta_unlock(inode, 0);
317out:
318 mlog_exit(ret);
319 return ret;
320}
321
322/* Taken from ext3. We don't necessarily need the full blown 311/* Taken from ext3. We don't necessarily need the full blown
323 * functionality yet, but IMHO it's better to cut and paste the whole 312 * functionality yet, but IMHO it's better to cut and paste the whole
324 * thing so we can avoid introducing our own bugs (and easily pick up 313 * thing so we can avoid introducing our own bugs (and easily pick up
325 * their fixes when they happen) --Mark */ 314 * their fixes when they happen) --Mark */
326static int walk_page_buffers( handle_t *handle, 315int walk_page_buffers( handle_t *handle,
327 struct buffer_head *head, 316 struct buffer_head *head,
328 unsigned from, 317 unsigned from,
329 unsigned to, 318 unsigned to,
330 int *partial, 319 int *partial,
331 int (*fn)( handle_t *handle, 320 int (*fn)( handle_t *handle,
332 struct buffer_head *bh)) 321 struct buffer_head *bh))
333{ 322{
334 struct buffer_head *bh; 323 struct buffer_head *bh;
335 unsigned block_start, block_end; 324 unsigned block_start, block_end;
@@ -388,95 +377,6 @@ out:
388 return handle; 377 return handle;
389} 378}
390 379
391static int ocfs2_commit_write(struct file *file, struct page *page,
392 unsigned from, unsigned to)
393{
394 int ret;
395 struct buffer_head *di_bh = NULL;
396 struct inode *inode = page->mapping->host;
397 handle_t *handle = NULL;
398 struct ocfs2_dinode *di;
399
400 mlog_entry("(0x%p, 0x%p, %u, %u)\n", file, page, from, to);
401
402 /* NOTE: ocfs2_file_aio_write has ensured that it's safe for
403 * us to continue here without rechecking the I/O against
404 * changed inode values.
405 *
406 * 1) We're currently holding the inode alloc lock, so no
407 * nodes can change it underneath us.
408 *
409 * 2) We've had to take the metadata lock at least once
410 * already to check for extending writes, suid removal, etc.
411 * The meta data update code then ensures that we don't get a
412 * stale inode allocation image (i_size, i_clusters, etc).
413 */
414
415 ret = ocfs2_meta_lock_with_page(inode, &di_bh, 1, page);
416 if (ret != 0) {
417 mlog_errno(ret);
418 goto out;
419 }
420
421 ret = ocfs2_data_lock_with_page(inode, 1, page);
422 if (ret != 0) {
423 mlog_errno(ret);
424 goto out_unlock_meta;
425 }
426
427 handle = ocfs2_start_walk_page_trans(inode, page, from, to);
428 if (IS_ERR(handle)) {
429 ret = PTR_ERR(handle);
430 goto out_unlock_data;
431 }
432
433 /* Mark our buffer early. We'd rather catch this error up here
434 * as opposed to after a successful commit_write which would
435 * require us to set back inode->i_size. */
436 ret = ocfs2_journal_access(handle, inode, di_bh,
437 OCFS2_JOURNAL_ACCESS_WRITE);
438 if (ret < 0) {
439 mlog_errno(ret);
440 goto out_commit;
441 }
442
443 /* might update i_size */
444 ret = generic_commit_write(file, page, from, to);
445 if (ret < 0) {
446 mlog_errno(ret);
447 goto out_commit;
448 }
449
450 di = (struct ocfs2_dinode *)di_bh->b_data;
451
452 /* ocfs2_mark_inode_dirty() is too heavy to use here. */
453 inode->i_mtime = inode->i_ctime = CURRENT_TIME;
454 di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
455 di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
456
457 inode->i_blocks = ocfs2_align_bytes_to_sectors((u64)(i_size_read(inode)));
458 di->i_size = cpu_to_le64((u64)i_size_read(inode));
459
460 ret = ocfs2_journal_dirty(handle, di_bh);
461 if (ret < 0) {
462 mlog_errno(ret);
463 goto out_commit;
464 }
465
466out_commit:
467 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
468out_unlock_data:
469 ocfs2_data_unlock(inode, 1);
470out_unlock_meta:
471 ocfs2_meta_unlock(inode, 1);
472out:
473 if (di_bh)
474 brelse(di_bh);
475
476 mlog_exit(ret);
477 return ret;
478}
479
480static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block) 380static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
481{ 381{
482 sector_t status; 382 sector_t status;
@@ -499,8 +399,7 @@ static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
499 down_read(&OCFS2_I(inode)->ip_alloc_sem); 399 down_read(&OCFS2_I(inode)->ip_alloc_sem);
500 } 400 }
501 401
502 err = ocfs2_extent_map_get_blocks(inode, block, 1, &p_blkno, 402 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL, NULL);
503 NULL);
504 403
505 if (!INODE_JOURNAL(inode)) { 404 if (!INODE_JOURNAL(inode)) {
506 up_read(&OCFS2_I(inode)->ip_alloc_sem); 405 up_read(&OCFS2_I(inode)->ip_alloc_sem);
@@ -540,8 +439,8 @@ static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
540 struct buffer_head *bh_result, int create) 439 struct buffer_head *bh_result, int create)
541{ 440{
542 int ret; 441 int ret;
543 u64 p_blkno, inode_blocks; 442 u64 p_blkno, inode_blocks, contig_blocks;
544 int contig_blocks; 443 unsigned int ext_flags;
545 unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits; 444 unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
546 unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits; 445 unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
547 446
@@ -549,33 +448,20 @@ static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
549 * nicely aligned and of the right size, so there's no need 448 * nicely aligned and of the right size, so there's no need
550 * for us to check any of that. */ 449 * for us to check any of that. */
551 450
552 spin_lock(&OCFS2_I(inode)->ip_lock); 451 inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
553 inode_blocks = ocfs2_clusters_to_blocks(inode->i_sb,
554 OCFS2_I(inode)->ip_clusters);
555
556 /*
557 * For a read which begins past the end of file, we return a hole.
558 */
559 if (!create && (iblock >= inode_blocks)) {
560 spin_unlock(&OCFS2_I(inode)->ip_lock);
561 ret = 0;
562 goto bail;
563 }
564 452
565 /* 453 /*
566 * Any write past EOF is not allowed because we'd be extending. 454 * Any write past EOF is not allowed because we'd be extending.
567 */ 455 */
568 if (create && (iblock + max_blocks) > inode_blocks) { 456 if (create && (iblock + max_blocks) > inode_blocks) {
569 spin_unlock(&OCFS2_I(inode)->ip_lock);
570 ret = -EIO; 457 ret = -EIO;
571 goto bail; 458 goto bail;
572 } 459 }
573 spin_unlock(&OCFS2_I(inode)->ip_lock);
574 460
575 /* This figures out the size of the next contiguous block, and 461 /* This figures out the size of the next contiguous block, and
576 * our logical offset */ 462 * our logical offset */
577 ret = ocfs2_extent_map_get_blocks(inode, iblock, 1, &p_blkno, 463 ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
578 &contig_blocks); 464 &contig_blocks, &ext_flags);
579 if (ret) { 465 if (ret) {
580 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n", 466 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
581 (unsigned long long)iblock); 467 (unsigned long long)iblock);
@@ -583,7 +469,37 @@ static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
583 goto bail; 469 goto bail;
584 } 470 }
585 471
586 map_bh(bh_result, inode->i_sb, p_blkno); 472 if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)) && !p_blkno) {
473 ocfs2_error(inode->i_sb,
474 "Inode %llu has a hole at block %llu\n",
475 (unsigned long long)OCFS2_I(inode)->ip_blkno,
476 (unsigned long long)iblock);
477 ret = -EROFS;
478 goto bail;
479 }
480
481 /*
482 * get_more_blocks() expects us to describe a hole by clearing
483 * the mapped bit on bh_result().
484 *
485 * Consider an unwritten extent as a hole.
486 */
487 if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
488 map_bh(bh_result, inode->i_sb, p_blkno);
489 else {
490 /*
491 * ocfs2_prepare_inode_for_write() should have caught
492 * the case where we'd be filling a hole and triggered
493 * a buffered write instead.
494 */
495 if (create) {
496 ret = -EIO;
497 mlog_errno(ret);
498 goto bail;
499 }
500
501 clear_buffer_mapped(bh_result);
502 }
587 503
588 /* make sure we don't map more than max_blocks blocks here as 504 /* make sure we don't map more than max_blocks blocks here as
589 that's all the kernel will handle at this point. */ 505 that's all the kernel will handle at this point. */
@@ -606,12 +522,17 @@ static void ocfs2_dio_end_io(struct kiocb *iocb,
606 void *private) 522 void *private)
607{ 523{
608 struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode; 524 struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
525 int level;
609 526
610 /* this io's submitter should not have unlocked this before we could */ 527 /* this io's submitter should not have unlocked this before we could */
611 BUG_ON(!ocfs2_iocb_is_rw_locked(iocb)); 528 BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
529
612 ocfs2_iocb_clear_rw_locked(iocb); 530 ocfs2_iocb_clear_rw_locked(iocb);
613 up_read(&inode->i_alloc_sem); 531
614 ocfs2_rw_unlock(inode, 0); 532 level = ocfs2_iocb_rw_locked_level(iocb);
533 if (!level)
534 up_read(&inode->i_alloc_sem);
535 ocfs2_rw_unlock(inode, level);
615} 536}
616 537
617/* 538/*
@@ -647,23 +568,27 @@ static ssize_t ocfs2_direct_IO(int rw,
647 568
648 mlog_entry_void(); 569 mlog_entry_void();
649 570
650 /* 571 if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
651 * We get PR data locks even for O_DIRECT. This allows 572 /*
652 * concurrent O_DIRECT I/O but doesn't let O_DIRECT with 573 * We get PR data locks even for O_DIRECT. This
653 * extending and buffered zeroing writes race. If they did 574 * allows concurrent O_DIRECT I/O but doesn't let
654 * race then the buffered zeroing could be written back after 575 * O_DIRECT with extending and buffered zeroing writes
655 * the O_DIRECT I/O. It's one thing to tell people not to mix 576 * race. If they did race then the buffered zeroing
656 * buffered and O_DIRECT writes, but expecting them to 577 * could be written back after the O_DIRECT I/O. It's
657 * understand that file extension is also an implicit buffered 578 * one thing to tell people not to mix buffered and
658 * write is too much. By getting the PR we force writeback of 579 * O_DIRECT writes, but expecting them to understand
659 * the buffered zeroing before proceeding. 580 * that file extension is also an implicit buffered
660 */ 581 * write is too much. By getting the PR we force
661 ret = ocfs2_data_lock(inode, 0); 582 * writeback of the buffered zeroing before
662 if (ret < 0) { 583 * proceeding.
663 mlog_errno(ret); 584 */
664 goto out; 585 ret = ocfs2_data_lock(inode, 0);
586 if (ret < 0) {
587 mlog_errno(ret);
588 goto out;
589 }
590 ocfs2_data_unlock(inode, 0);
665 } 591 }
666 ocfs2_data_unlock(inode, 0);
667 592
668 ret = blockdev_direct_IO_no_locking(rw, iocb, inode, 593 ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
669 inode->i_sb->s_bdev, iov, offset, 594 inode->i_sb->s_bdev, iov, offset,
@@ -675,11 +600,715 @@ out:
675 return ret; 600 return ret;
676} 601}
677 602
603static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
604 u32 cpos,
605 unsigned int *start,
606 unsigned int *end)
607{
608 unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
609
610 if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
611 unsigned int cpp;
612
613 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
614
615 cluster_start = cpos % cpp;
616 cluster_start = cluster_start << osb->s_clustersize_bits;
617
618 cluster_end = cluster_start + osb->s_clustersize;
619 }
620
621 BUG_ON(cluster_start > PAGE_SIZE);
622 BUG_ON(cluster_end > PAGE_SIZE);
623
624 if (start)
625 *start = cluster_start;
626 if (end)
627 *end = cluster_end;
628}
629
630/*
631 * 'from' and 'to' are the region in the page to avoid zeroing.
632 *
633 * If pagesize > clustersize, this function will avoid zeroing outside
634 * of the cluster boundary.
635 *
636 * from == to == 0 is code for "zero the entire cluster region"
637 */
638static void ocfs2_clear_page_regions(struct page *page,
639 struct ocfs2_super *osb, u32 cpos,
640 unsigned from, unsigned to)
641{
642 void *kaddr;
643 unsigned int cluster_start, cluster_end;
644
645 ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
646
647 kaddr = kmap_atomic(page, KM_USER0);
648
649 if (from || to) {
650 if (from > cluster_start)
651 memset(kaddr + cluster_start, 0, from - cluster_start);
652 if (to < cluster_end)
653 memset(kaddr + to, 0, cluster_end - to);
654 } else {
655 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
656 }
657
658 kunmap_atomic(kaddr, KM_USER0);
659}
660
661/*
662 * Some of this taken from block_prepare_write(). We already have our
663 * mapping by now though, and the entire write will be allocating or
664 * it won't, so not much need to use BH_New.
665 *
666 * This will also skip zeroing, which is handled externally.
667 */
668int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
669 struct inode *inode, unsigned int from,
670 unsigned int to, int new)
671{
672 int ret = 0;
673 struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
674 unsigned int block_end, block_start;
675 unsigned int bsize = 1 << inode->i_blkbits;
676
677 if (!page_has_buffers(page))
678 create_empty_buffers(page, bsize, 0);
679
680 head = page_buffers(page);
681 for (bh = head, block_start = 0; bh != head || !block_start;
682 bh = bh->b_this_page, block_start += bsize) {
683 block_end = block_start + bsize;
684
685 /*
686 * Ignore blocks outside of our i/o range -
687 * they may belong to unallocated clusters.
688 */
689 if (block_start >= to || block_end <= from) {
690 if (PageUptodate(page))
691 set_buffer_uptodate(bh);
692 continue;
693 }
694
695 /*
696 * For an allocating write with cluster size >= page
697 * size, we always write the entire page.
698 */
699
700 if (buffer_new(bh))
701 clear_buffer_new(bh);
702
703 if (!buffer_mapped(bh)) {
704 map_bh(bh, inode->i_sb, *p_blkno);
705 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
706 }
707
708 if (PageUptodate(page)) {
709 if (!buffer_uptodate(bh))
710 set_buffer_uptodate(bh);
711 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
712 (block_start < from || block_end > to)) {
713 ll_rw_block(READ, 1, &bh);
714 *wait_bh++=bh;
715 }
716
717 *p_blkno = *p_blkno + 1;
718 }
719
720 /*
721 * If we issued read requests - let them complete.
722 */
723 while(wait_bh > wait) {
724 wait_on_buffer(*--wait_bh);
725 if (!buffer_uptodate(*wait_bh))
726 ret = -EIO;
727 }
728
729 if (ret == 0 || !new)
730 return ret;
731
732 /*
733 * If we get -EIO above, zero out any newly allocated blocks
734 * to avoid exposing stale data.
735 */
736 bh = head;
737 block_start = 0;
738 do {
739 void *kaddr;
740
741 block_end = block_start + bsize;
742 if (block_end <= from)
743 goto next_bh;
744 if (block_start >= to)
745 break;
746
747 kaddr = kmap_atomic(page, KM_USER0);
748 memset(kaddr+block_start, 0, bh->b_size);
749 flush_dcache_page(page);
750 kunmap_atomic(kaddr, KM_USER0);
751 set_buffer_uptodate(bh);
752 mark_buffer_dirty(bh);
753
754next_bh:
755 block_start = block_end;
756 bh = bh->b_this_page;
757 } while (bh != head);
758
759 return ret;
760}
761
762/*
763 * This will copy user data from the buffer page in the splice
764 * context.
765 *
766 * For now, we ignore SPLICE_F_MOVE as that would require some extra
767 * communication out all the way to ocfs2_write().
768 */
769int ocfs2_map_and_write_splice_data(struct inode *inode,
770 struct ocfs2_write_ctxt *wc, u64 *p_blkno,
771 unsigned int *ret_from, unsigned int *ret_to)
772{
773 int ret;
774 unsigned int to, from, cluster_start, cluster_end;
775 char *src, *dst;
776 struct ocfs2_splice_write_priv *sp = wc->w_private;
777 struct pipe_buffer *buf = sp->s_buf;
778 unsigned long bytes, src_from;
779 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
780
781 ocfs2_figure_cluster_boundaries(osb, wc->w_cpos, &cluster_start,
782 &cluster_end);
783
784 from = sp->s_offset;
785 src_from = sp->s_buf_offset;
786 bytes = wc->w_count;
787
788 if (wc->w_large_pages) {
789 /*
790 * For cluster size < page size, we have to
791 * calculate pos within the cluster and obey
792 * the rightmost boundary.
793 */
794 bytes = min(bytes, (unsigned long)(osb->s_clustersize
795 - (wc->w_pos & (osb->s_clustersize - 1))));
796 }
797 to = from + bytes;
798
799 if (wc->w_this_page_new)
800 ret = ocfs2_map_page_blocks(wc->w_this_page, p_blkno, inode,
801 cluster_start, cluster_end, 1);
802 else
803 ret = ocfs2_map_page_blocks(wc->w_this_page, p_blkno, inode,
804 from, to, 0);
805 if (ret) {
806 mlog_errno(ret);
807 goto out;
808 }
809
810 BUG_ON(from > PAGE_CACHE_SIZE);
811 BUG_ON(to > PAGE_CACHE_SIZE);
812 BUG_ON(from > osb->s_clustersize);
813 BUG_ON(to > osb->s_clustersize);
814
815 src = buf->ops->map(sp->s_pipe, buf, 1);
816 dst = kmap_atomic(wc->w_this_page, KM_USER1);
817 memcpy(dst + from, src + src_from, bytes);
818 kunmap_atomic(wc->w_this_page, KM_USER1);
819 buf->ops->unmap(sp->s_pipe, buf, src);
820
821 wc->w_finished_copy = 1;
822
823 *ret_from = from;
824 *ret_to = to;
825out:
826
827 return bytes ? (unsigned int)bytes : ret;
828}
829
830/*
831 * This will copy user data from the iovec in the buffered write
832 * context.
833 */
834int ocfs2_map_and_write_user_data(struct inode *inode,
835 struct ocfs2_write_ctxt *wc, u64 *p_blkno,
836 unsigned int *ret_from, unsigned int *ret_to)
837{
838 int ret;
839 unsigned int to, from, cluster_start, cluster_end;
840 unsigned long bytes, src_from;
841 char *dst;
842 struct ocfs2_buffered_write_priv *bp = wc->w_private;
843 const struct iovec *cur_iov = bp->b_cur_iov;
844 char __user *buf;
845 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
846
847 ocfs2_figure_cluster_boundaries(osb, wc->w_cpos, &cluster_start,
848 &cluster_end);
849
850 buf = cur_iov->iov_base + bp->b_cur_off;
851 src_from = (unsigned long)buf & ~PAGE_CACHE_MASK;
852
853 from = wc->w_pos & (PAGE_CACHE_SIZE - 1);
854
855 /*
856 * This is a lot of comparisons, but it reads quite
857 * easily, which is important here.
858 */
859 /* Stay within the src page */
860 bytes = PAGE_SIZE - src_from;
861 /* Stay within the vector */
862 bytes = min(bytes,
863 (unsigned long)(cur_iov->iov_len - bp->b_cur_off));
864 /* Stay within count */
865 bytes = min(bytes, (unsigned long)wc->w_count);
866 /*
867 * For clustersize > page size, just stay within
868 * target page, otherwise we have to calculate pos
869 * within the cluster and obey the rightmost
870 * boundary.
871 */
872 if (wc->w_large_pages) {
873 /*
874 * For cluster size < page size, we have to
875 * calculate pos within the cluster and obey
876 * the rightmost boundary.
877 */
878 bytes = min(bytes, (unsigned long)(osb->s_clustersize
879 - (wc->w_pos & (osb->s_clustersize - 1))));
880 } else {
881 /*
882 * cluster size > page size is the most common
883 * case - we just stay within the target page
884 * boundary.
885 */
886 bytes = min(bytes, PAGE_CACHE_SIZE - from);
887 }
888
889 to = from + bytes;
890
891 if (wc->w_this_page_new)
892 ret = ocfs2_map_page_blocks(wc->w_this_page, p_blkno, inode,
893 cluster_start, cluster_end, 1);
894 else
895 ret = ocfs2_map_page_blocks(wc->w_this_page, p_blkno, inode,
896 from, to, 0);
897 if (ret) {
898 mlog_errno(ret);
899 goto out;
900 }
901
902 BUG_ON(from > PAGE_CACHE_SIZE);
903 BUG_ON(to > PAGE_CACHE_SIZE);
904 BUG_ON(from > osb->s_clustersize);
905 BUG_ON(to > osb->s_clustersize);
906
907 dst = kmap(wc->w_this_page);
908 memcpy(dst + from, bp->b_src_buf + src_from, bytes);
909 kunmap(wc->w_this_page);
910
911 /*
912 * XXX: This is slow, but simple. The caller of
913 * ocfs2_buffered_write_cluster() is responsible for
914 * passing through the iovecs, so it's difficult to
915 * predict what our next step is in here after our
916 * initial write. A future version should be pushing
917 * that iovec manipulation further down.
918 *
919 * By setting this, we indicate that a copy from user
920 * data was done, and subsequent calls for this
921 * cluster will skip copying more data.
922 */
923 wc->w_finished_copy = 1;
924
925 *ret_from = from;
926 *ret_to = to;
927out:
928
929 return bytes ? (unsigned int)bytes : ret;
930}
931
932/*
933 * Map, fill and write a page to disk.
934 *
935 * The work of copying data is done via callback. Newly allocated
936 * pages which don't take user data will be zero'd (set 'new' to
937 * indicate an allocating write)
938 *
939 * Returns a negative error code or the number of bytes copied into
940 * the page.
941 */
942int ocfs2_write_data_page(struct inode *inode, handle_t *handle,
943 u64 *p_blkno, struct page *page,
944 struct ocfs2_write_ctxt *wc, int new)
945{
946 int ret, copied = 0;
947 unsigned int from = 0, to = 0;
948 unsigned int cluster_start, cluster_end;
949 unsigned int zero_from = 0, zero_to = 0;
950
951 ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), wc->w_cpos,
952 &cluster_start, &cluster_end);
953
954 if ((wc->w_pos >> PAGE_CACHE_SHIFT) == page->index
955 && !wc->w_finished_copy) {
956
957 wc->w_this_page = page;
958 wc->w_this_page_new = new;
959 ret = wc->w_write_data_page(inode, wc, p_blkno, &from, &to);
960 if (ret < 0) {
961 mlog_errno(ret);
962 goto out;
963 }
964
965 copied = ret;
966
967 zero_from = from;
968 zero_to = to;
969 if (new) {
970 from = cluster_start;
971 to = cluster_end;
972 }
973 } else {
974 /*
975 * If we haven't allocated the new page yet, we
976 * shouldn't be writing it out without copying user
977 * data. This is likely a math error from the caller.
978 */
979 BUG_ON(!new);
980
981 from = cluster_start;
982 to = cluster_end;
983
984 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
985 cluster_start, cluster_end, 1);
986 if (ret) {
987 mlog_errno(ret);
988 goto out;
989 }
990 }
991
992 /*
993 * Parts of newly allocated pages need to be zero'd.
994 *
995 * Above, we have also rewritten 'to' and 'from' - as far as
996 * the rest of the function is concerned, the entire cluster
997 * range inside of a page needs to be written.
998 *
999 * We can skip this if the page is up to date - it's already
1000 * been zero'd from being read in as a hole.
1001 */
1002 if (new && !PageUptodate(page))
1003 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1004 wc->w_cpos, zero_from, zero_to);
1005
1006 flush_dcache_page(page);
1007
1008 if (ocfs2_should_order_data(inode)) {
1009 ret = walk_page_buffers(handle,
1010 page_buffers(page),
1011 from, to, NULL,
1012 ocfs2_journal_dirty_data);
1013 if (ret < 0)
1014 mlog_errno(ret);
1015 }
1016
1017 /*
1018 * We don't use generic_commit_write() because we need to
1019 * handle our own i_size update.
1020 */
1021 ret = block_commit_write(page, from, to);
1022 if (ret)
1023 mlog_errno(ret);
1024out:
1025
1026 return copied ? copied : ret;
1027}
1028
1029/*
1030 * Do the actual write of some data into an inode. Optionally allocate
1031 * in order to fulfill the write.
1032 *
1033 * cpos is the logical cluster offset within the file to write at
1034 *
1035 * 'phys' is the physical mapping of that offset. a 'phys' value of
1036 * zero indicates that allocation is required. In this case, data_ac
1037 * and meta_ac should be valid (meta_ac can be null if metadata
1038 * allocation isn't required).
1039 */
1040static ssize_t ocfs2_write(struct file *file, u32 phys, handle_t *handle,
1041 struct buffer_head *di_bh,
1042 struct ocfs2_alloc_context *data_ac,
1043 struct ocfs2_alloc_context *meta_ac,
1044 struct ocfs2_write_ctxt *wc)
1045{
1046 int ret, i, numpages = 1, new;
1047 unsigned int copied = 0;
1048 u32 tmp_pos;
1049 u64 v_blkno, p_blkno;
1050 struct address_space *mapping = file->f_mapping;
1051 struct inode *inode = mapping->host;
1052 unsigned long index, start;
1053 struct page **cpages;
1054
1055 new = phys == 0 ? 1 : 0;
1056
1057 /*
1058 * Figure out how many pages we'll be manipulating here. For
1059 * non allocating write, we just change the one
1060 * page. Otherwise, we'll need a whole clusters worth.
1061 */
1062 if (new)
1063 numpages = ocfs2_pages_per_cluster(inode->i_sb);
1064
1065 cpages = kzalloc(sizeof(*cpages) * numpages, GFP_NOFS);
1066 if (!cpages) {
1067 ret = -ENOMEM;
1068 mlog_errno(ret);
1069 return ret;
1070 }
1071
1072 /*
1073 * Fill our page array first. That way we've grabbed enough so
1074 * that we can zero and flush if we error after adding the
1075 * extent.
1076 */
1077 if (new) {
1078 start = ocfs2_align_clusters_to_page_index(inode->i_sb,
1079 wc->w_cpos);
1080 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, wc->w_cpos);
1081 } else {
1082 start = wc->w_pos >> PAGE_CACHE_SHIFT;
1083 v_blkno = wc->w_pos >> inode->i_sb->s_blocksize_bits;
1084 }
1085
1086 for(i = 0; i < numpages; i++) {
1087 index = start + i;
1088
1089 cpages[i] = grab_cache_page(mapping, index);
1090 if (!cpages[i]) {
1091 ret = -ENOMEM;
1092 mlog_errno(ret);
1093 goto out;
1094 }
1095 }
1096
1097 if (new) {
1098 /*
1099 * This is safe to call with the page locks - it won't take
1100 * any additional semaphores or cluster locks.
1101 */
1102 tmp_pos = wc->w_cpos;
1103 ret = ocfs2_do_extend_allocation(OCFS2_SB(inode->i_sb), inode,
1104 &tmp_pos, 1, di_bh, handle,
1105 data_ac, meta_ac, NULL);
1106 /*
1107 * This shouldn't happen because we must have already
1108 * calculated the correct meta data allocation required. The
1109 * internal tree allocation code should know how to increase
1110 * transaction credits itself.
1111 *
1112 * If need be, we could handle -EAGAIN for a
1113 * RESTART_TRANS here.
1114 */
1115 mlog_bug_on_msg(ret == -EAGAIN,
1116 "Inode %llu: EAGAIN return during allocation.\n",
1117 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1118 if (ret < 0) {
1119 mlog_errno(ret);
1120 goto out;
1121 }
1122 }
1123
1124 ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1125 NULL);
1126 if (ret < 0) {
1127
1128 /*
1129 * XXX: Should we go readonly here?
1130 */
1131
1132 mlog_errno(ret);
1133 goto out;
1134 }
1135
1136 BUG_ON(p_blkno == 0);
1137
1138 for(i = 0; i < numpages; i++) {
1139 ret = ocfs2_write_data_page(inode, handle, &p_blkno, cpages[i],
1140 wc, new);
1141 if (ret < 0) {
1142 mlog_errno(ret);
1143 goto out;
1144 }
1145
1146 copied += ret;
1147 }
1148
1149out:
1150 for(i = 0; i < numpages; i++) {
1151 unlock_page(cpages[i]);
1152 mark_page_accessed(cpages[i]);
1153 page_cache_release(cpages[i]);
1154 }
1155 kfree(cpages);
1156
1157 return copied ? copied : ret;
1158}
1159
1160static void ocfs2_write_ctxt_init(struct ocfs2_write_ctxt *wc,
1161 struct ocfs2_super *osb, loff_t pos,
1162 size_t count, ocfs2_page_writer *cb,
1163 void *cb_priv)
1164{
1165 wc->w_count = count;
1166 wc->w_pos = pos;
1167 wc->w_cpos = wc->w_pos >> osb->s_clustersize_bits;
1168 wc->w_finished_copy = 0;
1169
1170 if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
1171 wc->w_large_pages = 1;
1172 else
1173 wc->w_large_pages = 0;
1174
1175 wc->w_write_data_page = cb;
1176 wc->w_private = cb_priv;
1177}
1178
1179/*
1180 * Write a cluster to an inode. The cluster may not be allocated yet,
1181 * in which case it will be. This only exists for buffered writes -
1182 * O_DIRECT takes a more "traditional" path through the kernel.
1183 *
1184 * The caller is responsible for incrementing pos, written counts, etc
1185 *
1186 * For file systems that don't support sparse files, pre-allocation
1187 * and page zeroing up until cpos should be done prior to this
1188 * function call.
1189 *
1190 * Callers should be holding i_sem, and the rw cluster lock.
1191 *
1192 * Returns the number of user bytes written, or less than zero for
1193 * error.
1194 */
1195ssize_t ocfs2_buffered_write_cluster(struct file *file, loff_t pos,
1196 size_t count, ocfs2_page_writer *actor,
1197 void *priv)
1198{
1199 int ret, credits = OCFS2_INODE_UPDATE_CREDITS;
1200 ssize_t written = 0;
1201 u32 phys;
1202 struct inode *inode = file->f_mapping->host;
1203 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1204 struct buffer_head *di_bh = NULL;
1205 struct ocfs2_dinode *di;
1206 struct ocfs2_alloc_context *data_ac = NULL;
1207 struct ocfs2_alloc_context *meta_ac = NULL;
1208 handle_t *handle;
1209 struct ocfs2_write_ctxt wc;
1210
1211 ocfs2_write_ctxt_init(&wc, osb, pos, count, actor, priv);
1212
1213 ret = ocfs2_meta_lock(inode, &di_bh, 1);
1214 if (ret) {
1215 mlog_errno(ret);
1216 goto out;
1217 }
1218 di = (struct ocfs2_dinode *)di_bh->b_data;
1219
1220 /*
1221 * Take alloc sem here to prevent concurrent lookups. That way
1222 * the mapping, zeroing and tree manipulation within
1223 * ocfs2_write() will be safe against ->readpage(). This
1224 * should also serve to lock out allocation from a shared
1225 * writeable region.
1226 */
1227 down_write(&OCFS2_I(inode)->ip_alloc_sem);
1228
1229 ret = ocfs2_get_clusters(inode, wc.w_cpos, &phys, NULL, NULL);
1230 if (ret) {
1231 mlog_errno(ret);
1232 goto out_meta;
1233 }
1234
1235 /* phys == 0 means that allocation is required. */
1236 if (phys == 0) {
1237 ret = ocfs2_lock_allocators(inode, di, 1, &data_ac, &meta_ac);
1238 if (ret) {
1239 mlog_errno(ret);
1240 goto out_meta;
1241 }
1242
1243 credits = ocfs2_calc_extend_credits(inode->i_sb, di, 1);
1244 }
1245
1246 ret = ocfs2_data_lock(inode, 1);
1247 if (ret) {
1248 mlog_errno(ret);
1249 goto out_meta;
1250 }
1251
1252 handle = ocfs2_start_trans(osb, credits);
1253 if (IS_ERR(handle)) {
1254 ret = PTR_ERR(handle);
1255 mlog_errno(ret);
1256 goto out_data;
1257 }
1258
1259 written = ocfs2_write(file, phys, handle, di_bh, data_ac,
1260 meta_ac, &wc);
1261 if (written < 0) {
1262 ret = written;
1263 mlog_errno(ret);
1264 goto out_commit;
1265 }
1266
1267 ret = ocfs2_journal_access(handle, inode, di_bh,
1268 OCFS2_JOURNAL_ACCESS_WRITE);
1269 if (ret) {
1270 mlog_errno(ret);
1271 goto out_commit;
1272 }
1273
1274 pos += written;
1275 if (pos > inode->i_size) {
1276 i_size_write(inode, pos);
1277 mark_inode_dirty(inode);
1278 }
1279 inode->i_blocks = ocfs2_inode_sector_count(inode);
1280 di->i_size = cpu_to_le64((u64)i_size_read(inode));
1281 inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1282 di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
1283 di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
1284
1285 ret = ocfs2_journal_dirty(handle, di_bh);
1286 if (ret)
1287 mlog_errno(ret);
1288
1289out_commit:
1290 ocfs2_commit_trans(osb, handle);
1291
1292out_data:
1293 ocfs2_data_unlock(inode, 1);
1294
1295out_meta:
1296 up_write(&OCFS2_I(inode)->ip_alloc_sem);
1297 ocfs2_meta_unlock(inode, 1);
1298
1299out:
1300 brelse(di_bh);
1301 if (data_ac)
1302 ocfs2_free_alloc_context(data_ac);
1303 if (meta_ac)
1304 ocfs2_free_alloc_context(meta_ac);
1305
1306 return written ? written : ret;
1307}
1308
678const struct address_space_operations ocfs2_aops = { 1309const struct address_space_operations ocfs2_aops = {
679 .readpage = ocfs2_readpage, 1310 .readpage = ocfs2_readpage,
680 .writepage = ocfs2_writepage, 1311 .writepage = ocfs2_writepage,
681 .prepare_write = ocfs2_prepare_write,
682 .commit_write = ocfs2_commit_write,
683 .bmap = ocfs2_bmap, 1312 .bmap = ocfs2_bmap,
684 .sync_page = block_sync_page, 1313 .sync_page = block_sync_page,
685 .direct_IO = ocfs2_direct_IO, 1314 .direct_IO = ocfs2_direct_IO,
diff --git a/fs/ocfs2/aops.h b/fs/ocfs2/aops.h
index f446a15eab88..45821d479b5a 100644
--- a/fs/ocfs2/aops.h
+++ b/fs/ocfs2/aops.h
@@ -30,12 +30,83 @@ handle_t *ocfs2_start_walk_page_trans(struct inode *inode,
30 unsigned from, 30 unsigned from,
31 unsigned to); 31 unsigned to);
32 32
33int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
34 struct inode *inode, unsigned int from,
35 unsigned int to, int new);
36
37int walk_page_buffers( handle_t *handle,
38 struct buffer_head *head,
39 unsigned from,
40 unsigned to,
41 int *partial,
42 int (*fn)( handle_t *handle,
43 struct buffer_head *bh));
44
45struct ocfs2_write_ctxt;
46typedef int (ocfs2_page_writer)(struct inode *, struct ocfs2_write_ctxt *,
47 u64 *, unsigned int *, unsigned int *);
48
49ssize_t ocfs2_buffered_write_cluster(struct file *file, loff_t pos,
50 size_t count, ocfs2_page_writer *actor,
51 void *priv);
52
53struct ocfs2_write_ctxt {
54 size_t w_count;
55 loff_t w_pos;
56 u32 w_cpos;
57 unsigned int w_finished_copy;
58
59 /* This is true if page_size > cluster_size */
60 unsigned int w_large_pages;
61
62 /* Filler callback and private data */
63 ocfs2_page_writer *w_write_data_page;
64 void *w_private;
65
66 /* Only valid for the filler callback */
67 struct page *w_this_page;
68 unsigned int w_this_page_new;
69};
70
71struct ocfs2_buffered_write_priv {
72 char *b_src_buf;
73 const struct iovec *b_cur_iov; /* Current iovec */
74 size_t b_cur_off; /* Offset in the
75 * current iovec */
76};
77int ocfs2_map_and_write_user_data(struct inode *inode,
78 struct ocfs2_write_ctxt *wc,
79 u64 *p_blkno,
80 unsigned int *ret_from,
81 unsigned int *ret_to);
82
83struct ocfs2_splice_write_priv {
84 struct splice_desc *s_sd;
85 struct pipe_buffer *s_buf;
86 struct pipe_inode_info *s_pipe;
87 /* Neither offset value is ever larger than one page */
88 unsigned int s_offset;
89 unsigned int s_buf_offset;
90};
91int ocfs2_map_and_write_splice_data(struct inode *inode,
92 struct ocfs2_write_ctxt *wc,
93 u64 *p_blkno,
94 unsigned int *ret_from,
95 unsigned int *ret_to);
96
33/* all ocfs2_dio_end_io()'s fault */ 97/* all ocfs2_dio_end_io()'s fault */
34#define ocfs2_iocb_is_rw_locked(iocb) \ 98#define ocfs2_iocb_is_rw_locked(iocb) \
35 test_bit(0, (unsigned long *)&iocb->private) 99 test_bit(0, (unsigned long *)&iocb->private)
36#define ocfs2_iocb_set_rw_locked(iocb) \ 100static inline void ocfs2_iocb_set_rw_locked(struct kiocb *iocb, int level)
37 set_bit(0, (unsigned long *)&iocb->private) 101{
102 set_bit(0, (unsigned long *)&iocb->private);
103 if (level)
104 set_bit(1, (unsigned long *)&iocb->private);
105 else
106 clear_bit(1, (unsigned long *)&iocb->private);
107}
38#define ocfs2_iocb_clear_rw_locked(iocb) \ 108#define ocfs2_iocb_clear_rw_locked(iocb) \
39 clear_bit(0, (unsigned long *)&iocb->private) 109 clear_bit(0, (unsigned long *)&iocb->private)
40 110#define ocfs2_iocb_rw_locked_level(iocb) \
111 test_bit(1, (unsigned long *)&iocb->private)
41#endif /* OCFS2_FILE_H */ 112#endif /* OCFS2_FILE_H */
diff --git a/fs/ocfs2/cluster/quorum.c b/fs/ocfs2/cluster/quorum.c
index 4705d659fe57..bbacf7da48a4 100644
--- a/fs/ocfs2/cluster/quorum.c
+++ b/fs/ocfs2/cluster/quorum.c
@@ -46,6 +46,7 @@
46#include <linux/kernel.h> 46#include <linux/kernel.h>
47#include <linux/slab.h> 47#include <linux/slab.h>
48#include <linux/workqueue.h> 48#include <linux/workqueue.h>
49#include <linux/reboot.h>
49 50
50#include "heartbeat.h" 51#include "heartbeat.h"
51#include "nodemanager.h" 52#include "nodemanager.h"
@@ -72,7 +73,9 @@ static void o2quo_fence_self(void)
72 /* panic spins with interrupts enabled. with preempt 73 /* panic spins with interrupts enabled. with preempt
73 * threads can still schedule, etc, etc */ 74 * threads can still schedule, etc, etc */
74 o2hb_stop_all_regions(); 75 o2hb_stop_all_regions();
75 panic("ocfs2 is very sorry to be fencing this system by panicing\n"); 76
77 printk("ocfs2 is very sorry to be fencing this system by restarting\n");
78 emergency_restart();
76} 79}
77 80
78/* Indicate that a timeout occured on a hearbeat region write. The 81/* Indicate that a timeout occured on a hearbeat region write. The
diff --git a/fs/ocfs2/cluster/tcp_internal.h b/fs/ocfs2/cluster/tcp_internal.h
index 4dae5df5e467..9606111fe89d 100644
--- a/fs/ocfs2/cluster/tcp_internal.h
+++ b/fs/ocfs2/cluster/tcp_internal.h
@@ -38,6 +38,9 @@
38 * locking semantics of the file system using the protocol. It should 38 * locking semantics of the file system using the protocol. It should
39 * be somewhere else, I'm sure, but right now it isn't. 39 * be somewhere else, I'm sure, but right now it isn't.
40 * 40 *
41 * New in version 8:
42 * - Replace delete inode votes with a cluster lock
43 *
41 * New in version 7: 44 * New in version 7:
42 * - DLM join domain includes the live nodemap 45 * - DLM join domain includes the live nodemap
43 * 46 *
@@ -57,7 +60,7 @@
57 * - full 64 bit i_size in the metadata lock lvbs 60 * - full 64 bit i_size in the metadata lock lvbs
58 * - introduction of "rw" lock and pushing meta/data locking down 61 * - introduction of "rw" lock and pushing meta/data locking down
59 */ 62 */
60#define O2NET_PROTOCOL_VERSION 7ULL 63#define O2NET_PROTOCOL_VERSION 8ULL
61struct o2net_handshake { 64struct o2net_handshake {
62 __be64 protocol_version; 65 __be64 protocol_version;
63 __be64 connector_id; 66 __be64 connector_id;
diff --git a/fs/ocfs2/dir.c b/fs/ocfs2/dir.c
index 66821e178167..67e6866a2a4f 100644
--- a/fs/ocfs2/dir.c
+++ b/fs/ocfs2/dir.c
@@ -358,15 +358,17 @@ int ocfs2_do_extend_dir(struct super_block *sb,
358{ 358{
359 int status; 359 int status;
360 int extend; 360 int extend;
361 u64 p_blkno; 361 u64 p_blkno, v_blkno;
362 362
363 spin_lock(&OCFS2_I(dir)->ip_lock); 363 spin_lock(&OCFS2_I(dir)->ip_lock);
364 extend = (i_size_read(dir) == ocfs2_clusters_to_bytes(sb, OCFS2_I(dir)->ip_clusters)); 364 extend = (i_size_read(dir) == ocfs2_clusters_to_bytes(sb, OCFS2_I(dir)->ip_clusters));
365 spin_unlock(&OCFS2_I(dir)->ip_lock); 365 spin_unlock(&OCFS2_I(dir)->ip_lock);
366 366
367 if (extend) { 367 if (extend) {
368 status = ocfs2_do_extend_allocation(OCFS2_SB(sb), dir, 1, 368 u32 offset = OCFS2_I(dir)->ip_clusters;
369 parent_fe_bh, handle, 369
370 status = ocfs2_do_extend_allocation(OCFS2_SB(sb), dir, &offset,
371 1, parent_fe_bh, handle,
370 data_ac, meta_ac, NULL); 372 data_ac, meta_ac, NULL);
371 BUG_ON(status == -EAGAIN); 373 BUG_ON(status == -EAGAIN);
372 if (status < 0) { 374 if (status < 0) {
@@ -375,9 +377,8 @@ int ocfs2_do_extend_dir(struct super_block *sb,
375 } 377 }
376 } 378 }
377 379
378 status = ocfs2_extent_map_get_blocks(dir, (dir->i_blocks >> 380 v_blkno = ocfs2_blocks_for_bytes(sb, i_size_read(dir));
379 (sb->s_blocksize_bits - 9)), 381 status = ocfs2_extent_map_get_blocks(dir, v_blkno, &p_blkno, NULL, NULL);
380 1, &p_blkno, NULL);
381 if (status < 0) { 382 if (status < 0) {
382 mlog_errno(status); 383 mlog_errno(status);
383 goto bail; 384 goto bail;
@@ -486,7 +487,7 @@ static int ocfs2_extend_dir(struct ocfs2_super *osb,
486 487
487 dir_i_size += dir->i_sb->s_blocksize; 488 dir_i_size += dir->i_sb->s_blocksize;
488 i_size_write(dir, dir_i_size); 489 i_size_write(dir, dir_i_size);
489 dir->i_blocks = ocfs2_align_bytes_to_sectors(dir_i_size); 490 dir->i_blocks = ocfs2_inode_sector_count(dir);
490 status = ocfs2_mark_inode_dirty(handle, dir, parent_fe_bh); 491 status = ocfs2_mark_inode_dirty(handle, dir, parent_fe_bh);
491 if (status < 0) { 492 if (status < 0) {
492 mlog_errno(status); 493 mlog_errno(status);
diff --git a/fs/ocfs2/dlm/dlmdomain.c b/fs/ocfs2/dlm/dlmdomain.c
index c558442a0b44..d836b98dd99a 100644
--- a/fs/ocfs2/dlm/dlmdomain.c
+++ b/fs/ocfs2/dlm/dlmdomain.c
@@ -430,11 +430,10 @@ redo_bucket:
430 430
431 dlm_lockres_put(res); 431 dlm_lockres_put(res);
432 432
433 cond_resched_lock(&dlm->spinlock);
434
435 if (dropped) 433 if (dropped)
436 goto redo_bucket; 434 goto redo_bucket;
437 } 435 }
436 cond_resched_lock(&dlm->spinlock);
438 num += n; 437 num += n;
439 mlog(0, "%s: touched %d lockreses in bucket %d " 438 mlog(0, "%s: touched %d lockreses in bucket %d "
440 "(tot=%d)\n", dlm->name, n, i, num); 439 "(tot=%d)\n", dlm->name, n, i, num);
@@ -1035,7 +1034,7 @@ static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
1035{ 1034{
1036 int status = 0, tmpstat, node; 1035 int status = 0, tmpstat, node;
1037 struct domain_join_ctxt *ctxt; 1036 struct domain_join_ctxt *ctxt;
1038 enum dlm_query_join_response response; 1037 enum dlm_query_join_response response = JOIN_DISALLOW;
1039 1038
1040 mlog_entry("%p", dlm); 1039 mlog_entry("%p", dlm);
1041 1040
diff --git a/fs/ocfs2/dlm/dlmrecovery.c b/fs/ocfs2/dlm/dlmrecovery.c
index 6d4a83d50152..c1807a42c49f 100644
--- a/fs/ocfs2/dlm/dlmrecovery.c
+++ b/fs/ocfs2/dlm/dlmrecovery.c
@@ -611,6 +611,7 @@ static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node)
611 } 611 }
612 } while (status != 0); 612 } while (status != 0);
613 613
614 spin_lock(&dlm_reco_state_lock);
614 switch (ndata->state) { 615 switch (ndata->state) {
615 case DLM_RECO_NODE_DATA_INIT: 616 case DLM_RECO_NODE_DATA_INIT:
616 case DLM_RECO_NODE_DATA_FINALIZE_SENT: 617 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
@@ -641,6 +642,7 @@ static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node)
641 ndata->node_num, dead_node); 642 ndata->node_num, dead_node);
642 break; 643 break;
643 } 644 }
645 spin_unlock(&dlm_reco_state_lock);
644 } 646 }
645 647
646 mlog(0, "done requesting all lock info\n"); 648 mlog(0, "done requesting all lock info\n");
diff --git a/fs/ocfs2/dlmglue.c b/fs/ocfs2/dlmglue.c
index e335541727f9..27e43b0c0eae 100644
--- a/fs/ocfs2/dlmglue.c
+++ b/fs/ocfs2/dlmglue.c
@@ -225,11 +225,17 @@ static struct ocfs2_lock_res_ops ocfs2_dentry_lops = {
225 .flags = 0, 225 .flags = 0,
226}; 226};
227 227
228static struct ocfs2_lock_res_ops ocfs2_inode_open_lops = {
229 .get_osb = ocfs2_get_inode_osb,
230 .flags = 0,
231};
232
228static inline int ocfs2_is_inode_lock(struct ocfs2_lock_res *lockres) 233static inline int ocfs2_is_inode_lock(struct ocfs2_lock_res *lockres)
229{ 234{
230 return lockres->l_type == OCFS2_LOCK_TYPE_META || 235 return lockres->l_type == OCFS2_LOCK_TYPE_META ||
231 lockres->l_type == OCFS2_LOCK_TYPE_DATA || 236 lockres->l_type == OCFS2_LOCK_TYPE_DATA ||
232 lockres->l_type == OCFS2_LOCK_TYPE_RW; 237 lockres->l_type == OCFS2_LOCK_TYPE_RW ||
238 lockres->l_type == OCFS2_LOCK_TYPE_OPEN;
233} 239}
234 240
235static inline struct inode *ocfs2_lock_res_inode(struct ocfs2_lock_res *lockres) 241static inline struct inode *ocfs2_lock_res_inode(struct ocfs2_lock_res *lockres)
@@ -373,6 +379,9 @@ void ocfs2_inode_lock_res_init(struct ocfs2_lock_res *res,
373 case OCFS2_LOCK_TYPE_DATA: 379 case OCFS2_LOCK_TYPE_DATA:
374 ops = &ocfs2_inode_data_lops; 380 ops = &ocfs2_inode_data_lops;
375 break; 381 break;
382 case OCFS2_LOCK_TYPE_OPEN:
383 ops = &ocfs2_inode_open_lops;
384 break;
376 default: 385 default:
377 mlog_bug_on_msg(1, "type: %d\n", type); 386 mlog_bug_on_msg(1, "type: %d\n", type);
378 ops = NULL; /* thanks, gcc */ 387 ops = NULL; /* thanks, gcc */
@@ -1129,6 +1138,12 @@ int ocfs2_create_new_inode_locks(struct inode *inode)
1129 goto bail; 1138 goto bail;
1130 } 1139 }
1131 1140
1141 ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_open_lockres, 0, 0);
1142 if (ret) {
1143 mlog_errno(ret);
1144 goto bail;
1145 }
1146
1132bail: 1147bail:
1133 mlog_exit(ret); 1148 mlog_exit(ret);
1134 return ret; 1149 return ret;
@@ -1182,6 +1197,99 @@ void ocfs2_rw_unlock(struct inode *inode, int write)
1182 mlog_exit_void(); 1197 mlog_exit_void();
1183} 1198}
1184 1199
1200/*
1201 * ocfs2_open_lock always get PR mode lock.
1202 */
1203int ocfs2_open_lock(struct inode *inode)
1204{
1205 int status = 0;
1206 struct ocfs2_lock_res *lockres;
1207 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1208
1209 BUG_ON(!inode);
1210
1211 mlog_entry_void();
1212
1213 mlog(0, "inode %llu take PRMODE open lock\n",
1214 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1215
1216 if (ocfs2_mount_local(osb))
1217 goto out;
1218
1219 lockres = &OCFS2_I(inode)->ip_open_lockres;
1220
1221 status = ocfs2_cluster_lock(OCFS2_SB(inode->i_sb), lockres,
1222 LKM_PRMODE, 0, 0);
1223 if (status < 0)
1224 mlog_errno(status);
1225
1226out:
1227 mlog_exit(status);
1228 return status;
1229}
1230
1231int ocfs2_try_open_lock(struct inode *inode, int write)
1232{
1233 int status = 0, level;
1234 struct ocfs2_lock_res *lockres;
1235 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1236
1237 BUG_ON(!inode);
1238
1239 mlog_entry_void();
1240
1241 mlog(0, "inode %llu try to take %s open lock\n",
1242 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1243 write ? "EXMODE" : "PRMODE");
1244
1245 if (ocfs2_mount_local(osb))
1246 goto out;
1247
1248 lockres = &OCFS2_I(inode)->ip_open_lockres;
1249
1250 level = write ? LKM_EXMODE : LKM_PRMODE;
1251
1252 /*
1253 * The file system may already holding a PRMODE/EXMODE open lock.
1254 * Since we pass LKM_NOQUEUE, the request won't block waiting on
1255 * other nodes and the -EAGAIN will indicate to the caller that
1256 * this inode is still in use.
1257 */
1258 status = ocfs2_cluster_lock(OCFS2_SB(inode->i_sb), lockres,
1259 level, LKM_NOQUEUE, 0);
1260
1261out:
1262 mlog_exit(status);
1263 return status;
1264}
1265
1266/*
1267 * ocfs2_open_unlock unlock PR and EX mode open locks.
1268 */
1269void ocfs2_open_unlock(struct inode *inode)
1270{
1271 struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_open_lockres;
1272 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1273
1274 mlog_entry_void();
1275
1276 mlog(0, "inode %llu drop open lock\n",
1277 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1278
1279 if (ocfs2_mount_local(osb))
1280 goto out;
1281
1282 if(lockres->l_ro_holders)
1283 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres,
1284 LKM_PRMODE);
1285 if(lockres->l_ex_holders)
1286 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres,
1287 LKM_EXMODE);
1288
1289out:
1290 mlog_exit_void();
1291}
1292
1185int ocfs2_data_lock_full(struct inode *inode, 1293int ocfs2_data_lock_full(struct inode *inode,
1186 int write, 1294 int write,
1187 int arg_flags) 1295 int arg_flags)
@@ -1387,8 +1495,7 @@ static void ocfs2_refresh_inode_from_lvb(struct inode *inode)
1387 if (S_ISLNK(inode->i_mode) && !oi->ip_clusters) 1495 if (S_ISLNK(inode->i_mode) && !oi->ip_clusters)
1388 inode->i_blocks = 0; 1496 inode->i_blocks = 0;
1389 else 1497 else
1390 inode->i_blocks = 1498 inode->i_blocks = ocfs2_inode_sector_count(inode);
1391 ocfs2_align_bytes_to_sectors(i_size_read(inode));
1392 1499
1393 inode->i_uid = be32_to_cpu(lvb->lvb_iuid); 1500 inode->i_uid = be32_to_cpu(lvb->lvb_iuid);
1394 inode->i_gid = be32_to_cpu(lvb->lvb_igid); 1501 inode->i_gid = be32_to_cpu(lvb->lvb_igid);
@@ -1479,12 +1586,15 @@ static int ocfs2_meta_lock_update(struct inode *inode,
1479{ 1586{
1480 int status = 0; 1587 int status = 0;
1481 struct ocfs2_inode_info *oi = OCFS2_I(inode); 1588 struct ocfs2_inode_info *oi = OCFS2_I(inode);
1482 struct ocfs2_lock_res *lockres = NULL; 1589 struct ocfs2_lock_res *lockres = &oi->ip_meta_lockres;
1483 struct ocfs2_dinode *fe; 1590 struct ocfs2_dinode *fe;
1484 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 1591 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1485 1592
1486 mlog_entry_void(); 1593 mlog_entry_void();
1487 1594
1595 if (ocfs2_mount_local(osb))
1596 goto bail;
1597
1488 spin_lock(&oi->ip_lock); 1598 spin_lock(&oi->ip_lock);
1489 if (oi->ip_flags & OCFS2_INODE_DELETED) { 1599 if (oi->ip_flags & OCFS2_INODE_DELETED) {
1490 mlog(0, "Orphaned inode %llu was deleted while we " 1600 mlog(0, "Orphaned inode %llu was deleted while we "
@@ -1496,22 +1606,16 @@ static int ocfs2_meta_lock_update(struct inode *inode,
1496 } 1606 }
1497 spin_unlock(&oi->ip_lock); 1607 spin_unlock(&oi->ip_lock);
1498 1608
1499 if (!ocfs2_mount_local(osb)) { 1609 if (!ocfs2_should_refresh_lock_res(lockres))
1500 lockres = &oi->ip_meta_lockres; 1610 goto bail;
1501
1502 if (!ocfs2_should_refresh_lock_res(lockres))
1503 goto bail;
1504 }
1505 1611
1506 /* This will discard any caching information we might have had 1612 /* This will discard any caching information we might have had
1507 * for the inode metadata. */ 1613 * for the inode metadata. */
1508 ocfs2_metadata_cache_purge(inode); 1614 ocfs2_metadata_cache_purge(inode);
1509 1615
1510 /* will do nothing for inode types that don't use the extent
1511 * map (directories, bitmap files, etc) */
1512 ocfs2_extent_map_trunc(inode, 0); 1616 ocfs2_extent_map_trunc(inode, 0);
1513 1617
1514 if (lockres && ocfs2_meta_lvb_is_trustable(inode, lockres)) { 1618 if (ocfs2_meta_lvb_is_trustable(inode, lockres)) {
1515 mlog(0, "Trusting LVB on inode %llu\n", 1619 mlog(0, "Trusting LVB on inode %llu\n",
1516 (unsigned long long)oi->ip_blkno); 1620 (unsigned long long)oi->ip_blkno);
1517 ocfs2_refresh_inode_from_lvb(inode); 1621 ocfs2_refresh_inode_from_lvb(inode);
@@ -1558,8 +1662,7 @@ static int ocfs2_meta_lock_update(struct inode *inode,
1558 1662
1559 status = 0; 1663 status = 0;
1560bail_refresh: 1664bail_refresh:
1561 if (lockres) 1665 ocfs2_complete_lock_res_refresh(lockres, status);
1562 ocfs2_complete_lock_res_refresh(lockres, status);
1563bail: 1666bail:
1564 mlog_exit(status); 1667 mlog_exit(status);
1565 return status; 1668 return status;
@@ -1630,7 +1733,6 @@ int ocfs2_meta_lock_full(struct inode *inode,
1630 wait_event(osb->recovery_event, 1733 wait_event(osb->recovery_event,
1631 ocfs2_node_map_is_empty(osb, &osb->recovery_map)); 1734 ocfs2_node_map_is_empty(osb, &osb->recovery_map));
1632 1735
1633 acquired = 0;
1634 lockres = &OCFS2_I(inode)->ip_meta_lockres; 1736 lockres = &OCFS2_I(inode)->ip_meta_lockres;
1635 level = ex ? LKM_EXMODE : LKM_PRMODE; 1737 level = ex ? LKM_EXMODE : LKM_PRMODE;
1636 dlm_flags = 0; 1738 dlm_flags = 0;
@@ -2458,13 +2560,20 @@ int ocfs2_drop_inode_locks(struct inode *inode)
2458 * ocfs2_clear_inode has done it for us. */ 2560 * ocfs2_clear_inode has done it for us. */
2459 2561
2460 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb), 2562 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
2461 &OCFS2_I(inode)->ip_data_lockres); 2563 &OCFS2_I(inode)->ip_open_lockres);
2462 if (err < 0) 2564 if (err < 0)
2463 mlog_errno(err); 2565 mlog_errno(err);
2464 2566
2465 status = err; 2567 status = err;
2466 2568
2467 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb), 2569 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
2570 &OCFS2_I(inode)->ip_data_lockres);
2571 if (err < 0)
2572 mlog_errno(err);
2573 if (err < 0 && !status)
2574 status = err;
2575
2576 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
2468 &OCFS2_I(inode)->ip_meta_lockres); 2577 &OCFS2_I(inode)->ip_meta_lockres);
2469 if (err < 0) 2578 if (err < 0)
2470 mlog_errno(err); 2579 mlog_errno(err);
diff --git a/fs/ocfs2/dlmglue.h b/fs/ocfs2/dlmglue.h
index c343fca68cf1..59cb566e7983 100644
--- a/fs/ocfs2/dlmglue.h
+++ b/fs/ocfs2/dlmglue.h
@@ -80,6 +80,9 @@ void ocfs2_data_unlock(struct inode *inode,
80 int write); 80 int write);
81int ocfs2_rw_lock(struct inode *inode, int write); 81int ocfs2_rw_lock(struct inode *inode, int write);
82void ocfs2_rw_unlock(struct inode *inode, int write); 82void ocfs2_rw_unlock(struct inode *inode, int write);
83int ocfs2_open_lock(struct inode *inode);
84int ocfs2_try_open_lock(struct inode *inode, int write);
85void ocfs2_open_unlock(struct inode *inode);
83int ocfs2_meta_lock_atime(struct inode *inode, 86int ocfs2_meta_lock_atime(struct inode *inode,
84 struct vfsmount *vfsmnt, 87 struct vfsmount *vfsmnt,
85 int *level); 88 int *level);
diff --git a/fs/ocfs2/extent_map.c b/fs/ocfs2/extent_map.c
index 80ac69f11d9f..ba2b2ab1c6e4 100644
--- a/fs/ocfs2/extent_map.c
+++ b/fs/ocfs2/extent_map.c
@@ -3,8 +3,7 @@
3 * 3 *
4 * extent_map.c 4 * extent_map.c
5 * 5 *
6 * In-memory extent map for OCFS2. Man, this code was prettier in 6 * Block/Cluster mapping functions
7 * the library.
8 * 7 *
9 * Copyright (C) 2004 Oracle. All rights reserved. 8 * Copyright (C) 2004 Oracle. All rights reserved.
10 * 9 *
@@ -26,1016 +25,528 @@
26#include <linux/fs.h> 25#include <linux/fs.h>
27#include <linux/init.h> 26#include <linux/init.h>
28#include <linux/types.h> 27#include <linux/types.h>
29#include <linux/slab.h>
30#include <linux/rbtree.h>
31 28
32#define MLOG_MASK_PREFIX ML_EXTENT_MAP 29#define MLOG_MASK_PREFIX ML_EXTENT_MAP
33#include <cluster/masklog.h> 30#include <cluster/masklog.h>
34 31
35#include "ocfs2.h" 32#include "ocfs2.h"
36 33
34#include "alloc.h"
37#include "extent_map.h" 35#include "extent_map.h"
38#include "inode.h" 36#include "inode.h"
39#include "super.h" 37#include "super.h"
40 38
41#include "buffer_head_io.h" 39#include "buffer_head_io.h"
42 40
43
44/* 41/*
45 * SUCK SUCK SUCK 42 * The extent caching implementation is intentionally trivial.
46 * Our headers are so bad that struct ocfs2_extent_map is in ocfs.h
47 */
48
49struct ocfs2_extent_map_entry {
50 struct rb_node e_node;
51 int e_tree_depth;
52 struct ocfs2_extent_rec e_rec;
53};
54
55struct ocfs2_em_insert_context {
56 int need_left;
57 int need_right;
58 struct ocfs2_extent_map_entry *new_ent;
59 struct ocfs2_extent_map_entry *old_ent;
60 struct ocfs2_extent_map_entry *left_ent;
61 struct ocfs2_extent_map_entry *right_ent;
62};
63
64static struct kmem_cache *ocfs2_em_ent_cachep = NULL;
65
66
67static struct ocfs2_extent_map_entry *
68ocfs2_extent_map_lookup(struct ocfs2_extent_map *em,
69 u32 cpos, u32 clusters,
70 struct rb_node ***ret_p,
71 struct rb_node **ret_parent);
72static int ocfs2_extent_map_insert(struct inode *inode,
73 struct ocfs2_extent_rec *rec,
74 int tree_depth);
75static int ocfs2_extent_map_insert_entry(struct ocfs2_extent_map *em,
76 struct ocfs2_extent_map_entry *ent);
77static int ocfs2_extent_map_find_leaf(struct inode *inode,
78 u32 cpos, u32 clusters,
79 struct ocfs2_extent_list *el);
80static int ocfs2_extent_map_lookup_read(struct inode *inode,
81 u32 cpos, u32 clusters,
82 struct ocfs2_extent_map_entry **ret_ent);
83static int ocfs2_extent_map_try_insert(struct inode *inode,
84 struct ocfs2_extent_rec *rec,
85 int tree_depth,
86 struct ocfs2_em_insert_context *ctxt);
87
88/* returns 1 only if the rec contains all the given clusters -- that is that
89 * rec's cpos is <= the cluster cpos and that the rec endpoint (cpos +
90 * clusters) is >= the argument's endpoint */
91static int ocfs2_extent_rec_contains_clusters(struct ocfs2_extent_rec *rec,
92 u32 cpos, u32 clusters)
93{
94 if (le32_to_cpu(rec->e_cpos) > cpos)
95 return 0;
96 if (cpos + clusters > le32_to_cpu(rec->e_cpos) +
97 le32_to_cpu(rec->e_clusters))
98 return 0;
99 return 1;
100}
101
102
103/*
104 * Find an entry in the tree that intersects the region passed in.
105 * Note that this will find straddled intervals, it is up to the
106 * callers to enforce any boundary conditions.
107 *
108 * Callers must hold ip_lock. This lookup is not guaranteed to return
109 * a tree_depth 0 match, and as such can race inserts if the lock
110 * were not held.
111 * 43 *
112 * The rb_node garbage lets insertion share the search. Trivial 44 * We only cache a small number of extents stored directly on the
113 * callers pass NULL. 45 * inode, so linear order operations are acceptable. If we ever want
46 * to increase the size of the extent map, then these algorithms must
47 * get smarter.
114 */ 48 */
115static struct ocfs2_extent_map_entry * 49
116ocfs2_extent_map_lookup(struct ocfs2_extent_map *em, 50void ocfs2_extent_map_init(struct inode *inode)
117 u32 cpos, u32 clusters,
118 struct rb_node ***ret_p,
119 struct rb_node **ret_parent)
120{ 51{
121 struct rb_node **p = &em->em_extents.rb_node; 52 struct ocfs2_inode_info *oi = OCFS2_I(inode);
122 struct rb_node *parent = NULL;
123 struct ocfs2_extent_map_entry *ent = NULL;
124
125 while (*p)
126 {
127 parent = *p;
128 ent = rb_entry(parent, struct ocfs2_extent_map_entry,
129 e_node);
130 if ((cpos + clusters) <= le32_to_cpu(ent->e_rec.e_cpos)) {
131 p = &(*p)->rb_left;
132 ent = NULL;
133 } else if (cpos >= (le32_to_cpu(ent->e_rec.e_cpos) +
134 le32_to_cpu(ent->e_rec.e_clusters))) {
135 p = &(*p)->rb_right;
136 ent = NULL;
137 } else
138 break;
139 }
140 53
141 if (ret_p != NULL) 54 oi->ip_extent_map.em_num_items = 0;
142 *ret_p = p; 55 INIT_LIST_HEAD(&oi->ip_extent_map.em_list);
143 if (ret_parent != NULL)
144 *ret_parent = parent;
145 return ent;
146} 56}
147 57
148/* 58static void __ocfs2_extent_map_lookup(struct ocfs2_extent_map *em,
149 * Find the leaf containing the interval we want. While we're on our 59 unsigned int cpos,
150 * way down the tree, fill in every record we see at any depth, because 60 struct ocfs2_extent_map_item **ret_emi)
151 * we might want it later.
152 *
153 * Note that this code is run without ip_lock. That's because it
154 * sleeps while reading. If someone is also filling the extent list at
155 * the same time we are, we might have to restart.
156 */
157static int ocfs2_extent_map_find_leaf(struct inode *inode,
158 u32 cpos, u32 clusters,
159 struct ocfs2_extent_list *el)
160{ 61{
161 int i, ret; 62 unsigned int range;
162 struct buffer_head *eb_bh = NULL; 63 struct ocfs2_extent_map_item *emi;
163 u64 blkno;
164 u32 rec_end;
165 struct ocfs2_extent_block *eb;
166 struct ocfs2_extent_rec *rec;
167
168 /*
169 * The bh data containing the el cannot change here, because
170 * we hold alloc_sem. So we can do this without other
171 * locks.
172 */
173 while (el->l_tree_depth)
174 {
175 blkno = 0;
176 for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
177 rec = &el->l_recs[i];
178 rec_end = (le32_to_cpu(rec->e_cpos) +
179 le32_to_cpu(rec->e_clusters));
180
181 ret = -EBADR;
182 if (rec_end > OCFS2_I(inode)->ip_clusters) {
183 mlog_errno(ret);
184 ocfs2_error(inode->i_sb,
185 "Extent %d at e_blkno %llu of inode %llu goes past ip_clusters of %u\n",
186 i,
187 (unsigned long long)le64_to_cpu(rec->e_blkno),
188 (unsigned long long)OCFS2_I(inode)->ip_blkno,
189 OCFS2_I(inode)->ip_clusters);
190 goto out_free;
191 }
192
193 if (rec_end <= cpos) {
194 ret = ocfs2_extent_map_insert(inode, rec,
195 le16_to_cpu(el->l_tree_depth));
196 if (ret && (ret != -EEXIST)) {
197 mlog_errno(ret);
198 goto out_free;
199 }
200 continue;
201 }
202 if ((cpos + clusters) <= le32_to_cpu(rec->e_cpos)) {
203 ret = ocfs2_extent_map_insert(inode, rec,
204 le16_to_cpu(el->l_tree_depth));
205 if (ret && (ret != -EEXIST)) {
206 mlog_errno(ret);
207 goto out_free;
208 }
209 continue;
210 }
211 64
212 /* 65 *ret_emi = NULL;
213 * We've found a record that matches our
214 * interval. We don't insert it because we're
215 * about to traverse it.
216 */
217
218 /* Check to see if we're stradling */
219 ret = -ESRCH;
220 if (!ocfs2_extent_rec_contains_clusters(rec,
221 cpos,
222 clusters)) {
223 mlog_errno(ret);
224 goto out_free;
225 }
226 66
227 /* 67 list_for_each_entry(emi, &em->em_list, ei_list) {
228 * If we've already found a record, the el has 68 range = emi->ei_cpos + emi->ei_clusters;
229 * two records covering the same interval.
230 * EEEK!
231 */
232 ret = -EBADR;
233 if (blkno) {
234 mlog_errno(ret);
235 ocfs2_error(inode->i_sb,
236 "Multiple extents for (cpos = %u, clusters = %u) on inode %llu; e_blkno %llu and rec %d at e_blkno %llu\n",
237 cpos, clusters,
238 (unsigned long long)OCFS2_I(inode)->ip_blkno,
239 (unsigned long long)blkno, i,
240 (unsigned long long)le64_to_cpu(rec->e_blkno));
241 goto out_free;
242 }
243 69
244 blkno = le64_to_cpu(rec->e_blkno); 70 if (cpos >= emi->ei_cpos && cpos < range) {
245 } 71 list_move(&emi->ei_list, &em->em_list);
246 72
247 /* 73 *ret_emi = emi;
248 * We don't support holes, and we're still up 74 break;
249 * in the branches, so we'd better have found someone
250 */
251 ret = -EBADR;
252 if (!blkno) {
253 ocfs2_error(inode->i_sb,
254 "No record found for (cpos = %u, clusters = %u) on inode %llu\n",
255 cpos, clusters,
256 (unsigned long long)OCFS2_I(inode)->ip_blkno);
257 mlog_errno(ret);
258 goto out_free;
259 }
260
261 if (eb_bh) {
262 brelse(eb_bh);
263 eb_bh = NULL;
264 }
265 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
266 blkno, &eb_bh, OCFS2_BH_CACHED,
267 inode);
268 if (ret) {
269 mlog_errno(ret);
270 goto out_free;
271 }
272 eb = (struct ocfs2_extent_block *)eb_bh->b_data;
273 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
274 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
275 ret = -EIO;
276 goto out_free;
277 } 75 }
278 el = &eb->h_list;
279 } 76 }
77}
280 78
281 BUG_ON(el->l_tree_depth); 79static int ocfs2_extent_map_lookup(struct inode *inode, unsigned int cpos,
282 80 unsigned int *phys, unsigned int *len,
283 for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) { 81 unsigned int *flags)
284 rec = &el->l_recs[i]; 82{
285 83 unsigned int coff;
286 if ((le32_to_cpu(rec->e_cpos) + le32_to_cpu(rec->e_clusters)) > 84 struct ocfs2_inode_info *oi = OCFS2_I(inode);
287 OCFS2_I(inode)->ip_clusters) { 85 struct ocfs2_extent_map_item *emi;
288 ret = -EBADR; 86
289 mlog_errno(ret); 87 spin_lock(&oi->ip_lock);
290 ocfs2_error(inode->i_sb, 88
291 "Extent %d at e_blkno %llu of inode %llu goes past ip_clusters of %u\n", 89 __ocfs2_extent_map_lookup(&oi->ip_extent_map, cpos, &emi);
292 i, 90 if (emi) {
293 (unsigned long long)le64_to_cpu(rec->e_blkno), 91 coff = cpos - emi->ei_cpos;
294 (unsigned long long)OCFS2_I(inode)->ip_blkno, 92 *phys = emi->ei_phys + coff;
295 OCFS2_I(inode)->ip_clusters); 93 if (len)
296 return ret; 94 *len = emi->ei_clusters - coff;
297 } 95 if (flags)
298 96 *flags = emi->ei_flags;
299 ret = ocfs2_extent_map_insert(inode, rec,
300 le16_to_cpu(el->l_tree_depth));
301 if (ret && (ret != -EEXIST)) {
302 mlog_errno(ret);
303 goto out_free;
304 }
305 } 97 }
306 98
307 ret = 0; 99 spin_unlock(&oi->ip_lock);
308 100
309out_free: 101 if (emi == NULL)
310 if (eb_bh) 102 return -ENOENT;
311 brelse(eb_bh);
312 103
313 return ret; 104 return 0;
314} 105}
315 106
316/* 107/*
317 * This lookup actually will read from disk. It has one invariant: 108 * Forget about all clusters equal to or greater than cpos.
318 * It will never re-traverse blocks. This means that all inserts should
319 * be new regions or more granular regions (both allowed by insert).
320 */ 109 */
321static int ocfs2_extent_map_lookup_read(struct inode *inode, 110void ocfs2_extent_map_trunc(struct inode *inode, unsigned int cpos)
322 u32 cpos,
323 u32 clusters,
324 struct ocfs2_extent_map_entry **ret_ent)
325{ 111{
326 int ret; 112 struct list_head *p, *n;
327 u64 blkno; 113 struct ocfs2_extent_map_item *emi;
328 struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map; 114 struct ocfs2_inode_info *oi = OCFS2_I(inode);
329 struct ocfs2_extent_map_entry *ent; 115 struct ocfs2_extent_map *em = &oi->ip_extent_map;
330 struct buffer_head *bh = NULL; 116 LIST_HEAD(tmp_list);
331 struct ocfs2_extent_block *eb; 117 unsigned int range;
332 struct ocfs2_dinode *di; 118
333 struct ocfs2_extent_list *el; 119 spin_lock(&oi->ip_lock);
334 120 list_for_each_safe(p, n, &em->em_list) {
335 spin_lock(&OCFS2_I(inode)->ip_lock); 121 emi = list_entry(p, struct ocfs2_extent_map_item, ei_list);
336 ent = ocfs2_extent_map_lookup(em, cpos, clusters, NULL, NULL); 122
337 if (ent) { 123 if (emi->ei_cpos >= cpos) {
338 if (!ent->e_tree_depth) { 124 /* Full truncate of this record. */
339 spin_unlock(&OCFS2_I(inode)->ip_lock); 125 list_move(&emi->ei_list, &tmp_list);
340 *ret_ent = ent; 126 BUG_ON(em->em_num_items == 0);
341 return 0; 127 em->em_num_items--;
342 } 128 continue;
343 blkno = le64_to_cpu(ent->e_rec.e_blkno);
344 spin_unlock(&OCFS2_I(inode)->ip_lock);
345
346 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb), blkno, &bh,
347 OCFS2_BH_CACHED, inode);
348 if (ret) {
349 mlog_errno(ret);
350 if (bh)
351 brelse(bh);
352 return ret;
353 } 129 }
354 eb = (struct ocfs2_extent_block *)bh->b_data;
355 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
356 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
357 brelse(bh);
358 return -EIO;
359 }
360 el = &eb->h_list;
361 } else {
362 spin_unlock(&OCFS2_I(inode)->ip_lock);
363 130
364 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb), 131 range = emi->ei_cpos + emi->ei_clusters;
365 OCFS2_I(inode)->ip_blkno, &bh, 132 if (range > cpos) {
366 OCFS2_BH_CACHED, inode); 133 /* Partial truncate */
367 if (ret) { 134 emi->ei_clusters = cpos - emi->ei_cpos;
368 mlog_errno(ret);
369 if (bh)
370 brelse(bh);
371 return ret;
372 } 135 }
373 di = (struct ocfs2_dinode *)bh->b_data;
374 if (!OCFS2_IS_VALID_DINODE(di)) {
375 brelse(bh);
376 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, di);
377 return -EIO;
378 }
379 el = &di->id2.i_list;
380 }
381
382 ret = ocfs2_extent_map_find_leaf(inode, cpos, clusters, el);
383 brelse(bh);
384 if (ret) {
385 mlog_errno(ret);
386 return ret;
387 } 136 }
137 spin_unlock(&oi->ip_lock);
388 138
389 ent = ocfs2_extent_map_lookup(em, cpos, clusters, NULL, NULL); 139 list_for_each_safe(p, n, &tmp_list) {
390 if (!ent) { 140 emi = list_entry(p, struct ocfs2_extent_map_item, ei_list);
391 ret = -ESRCH; 141 list_del(&emi->ei_list);
392 mlog_errno(ret); 142 kfree(emi);
393 return ret;
394 } 143 }
395
396 /* FIXME: Make sure this isn't a corruption */
397 BUG_ON(ent->e_tree_depth);
398
399 *ret_ent = ent;
400
401 return 0;
402} 144}
403 145
404/* 146/*
405 * Callers must hold ip_lock. This can insert pieces of the tree, 147 * Is any part of emi2 contained within emi1
406 * thus racing lookup if the lock weren't held.
407 */ 148 */
408static int ocfs2_extent_map_insert_entry(struct ocfs2_extent_map *em, 149static int ocfs2_ei_is_contained(struct ocfs2_extent_map_item *emi1,
409 struct ocfs2_extent_map_entry *ent) 150 struct ocfs2_extent_map_item *emi2)
410{ 151{
411 struct rb_node **p, *parent; 152 unsigned int range1, range2;
412 struct ocfs2_extent_map_entry *old_ent;
413 153
414 old_ent = ocfs2_extent_map_lookup(em, le32_to_cpu(ent->e_rec.e_cpos), 154 /*
415 le32_to_cpu(ent->e_rec.e_clusters), 155 * Check if logical start of emi2 is inside emi1
416 &p, &parent); 156 */
417 if (old_ent) 157 range1 = emi1->ei_cpos + emi1->ei_clusters;
418 return -EEXIST; 158 if (emi2->ei_cpos >= emi1->ei_cpos && emi2->ei_cpos < range1)
159 return 1;
419 160
420 rb_link_node(&ent->e_node, parent, p); 161 /*
421 rb_insert_color(&ent->e_node, &em->em_extents); 162 * Check if logical end of emi2 is inside emi1
163 */
164 range2 = emi2->ei_cpos + emi2->ei_clusters;
165 if (range2 > emi1->ei_cpos && range2 <= range1)
166 return 1;
422 167
423 return 0; 168 return 0;
424} 169}
425 170
171static void ocfs2_copy_emi_fields(struct ocfs2_extent_map_item *dest,
172 struct ocfs2_extent_map_item *src)
173{
174 dest->ei_cpos = src->ei_cpos;
175 dest->ei_phys = src->ei_phys;
176 dest->ei_clusters = src->ei_clusters;
177 dest->ei_flags = src->ei_flags;
178}
426 179
427/* 180/*
428 * Simple rule: on any return code other than -EAGAIN, anything left 181 * Try to merge emi with ins. Returns 1 if merge succeeds, zero
429 * in the insert_context will be freed. 182 * otherwise.
430 *
431 * Simple rule #2: A return code of -EEXIST from this function or
432 * its calls to ocfs2_extent_map_insert_entry() signifies that another
433 * thread beat us to the insert. It is not an actual error, but it
434 * tells the caller we have no more work to do.
435 */ 183 */
436static int ocfs2_extent_map_try_insert(struct inode *inode, 184static int ocfs2_try_to_merge_extent_map(struct ocfs2_extent_map_item *emi,
437 struct ocfs2_extent_rec *rec, 185 struct ocfs2_extent_map_item *ins)
438 int tree_depth,
439 struct ocfs2_em_insert_context *ctxt)
440{ 186{
441 int ret;
442 struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map;
443 struct ocfs2_extent_map_entry *old_ent;
444
445 ctxt->need_left = 0;
446 ctxt->need_right = 0;
447 ctxt->old_ent = NULL;
448
449 spin_lock(&OCFS2_I(inode)->ip_lock);
450 ret = ocfs2_extent_map_insert_entry(em, ctxt->new_ent);
451 if (!ret) {
452 ctxt->new_ent = NULL;
453 goto out_unlock;
454 }
455
456 /* Since insert_entry failed, the map MUST have old_ent */
457 old_ent = ocfs2_extent_map_lookup(em, le32_to_cpu(rec->e_cpos),
458 le32_to_cpu(rec->e_clusters),
459 NULL, NULL);
460
461 BUG_ON(!old_ent);
462
463 if (old_ent->e_tree_depth < tree_depth) {
464 /* Another thread beat us to the lower tree_depth */
465 ret = -EEXIST;
466 goto out_unlock;
467 }
468
469 if (old_ent->e_tree_depth == tree_depth) {
470 /*
471 * Another thread beat us to this tree_depth.
472 * Let's make sure we agree with that thread (the
473 * extent_rec should be identical).
474 */
475 if (!memcmp(rec, &old_ent->e_rec,
476 sizeof(struct ocfs2_extent_rec)))
477 ret = 0;
478 else
479 /* FIXME: Should this be ESRCH/EBADR??? */
480 ret = -EEXIST;
481
482 goto out_unlock;
483 }
484
485 /* 187 /*
486 * We do it in this order specifically so that no actual tree 188 * Handle contiguousness
487 * changes occur until we have all the pieces we need. We
488 * don't want malloc failures to leave an inconsistent tree.
489 * Whenever we drop the lock, another process could be
490 * inserting. Also note that, if another process just beat us
491 * to an insert, we might not need the same pieces we needed
492 * the first go round. In the end, the pieces we need will
493 * be used, and the pieces we don't will be freed.
494 */ 189 */
495 ctxt->need_left = !!(le32_to_cpu(rec->e_cpos) > 190 if (ins->ei_phys == (emi->ei_phys + emi->ei_clusters) &&
496 le32_to_cpu(old_ent->e_rec.e_cpos)); 191 ins->ei_cpos == (emi->ei_cpos + emi->ei_clusters) &&
497 ctxt->need_right = !!((le32_to_cpu(old_ent->e_rec.e_cpos) + 192 ins->ei_flags == emi->ei_flags) {
498 le32_to_cpu(old_ent->e_rec.e_clusters)) > 193 emi->ei_clusters += ins->ei_clusters;
499 (le32_to_cpu(rec->e_cpos) + le32_to_cpu(rec->e_clusters))); 194 return 1;
500 ret = -EAGAIN; 195 } else if ((ins->ei_phys + ins->ei_clusters) == emi->ei_phys &&
501 if (ctxt->need_left) { 196 (ins->ei_cpos + ins->ei_clusters) == emi->ei_phys &&
502 if (!ctxt->left_ent) 197 ins->ei_flags == emi->ei_flags) {
503 goto out_unlock; 198 emi->ei_phys = ins->ei_phys;
504 *(ctxt->left_ent) = *old_ent; 199 emi->ei_cpos = ins->ei_cpos;
505 ctxt->left_ent->e_rec.e_clusters = 200 emi->ei_clusters += ins->ei_clusters;
506 cpu_to_le32(le32_to_cpu(rec->e_cpos) - 201 return 1;
507 le32_to_cpu(ctxt->left_ent->e_rec.e_cpos));
508 }
509 if (ctxt->need_right) {
510 if (!ctxt->right_ent)
511 goto out_unlock;
512 *(ctxt->right_ent) = *old_ent;
513 ctxt->right_ent->e_rec.e_cpos =
514 cpu_to_le32(le32_to_cpu(rec->e_cpos) +
515 le32_to_cpu(rec->e_clusters));
516 ctxt->right_ent->e_rec.e_clusters =
517 cpu_to_le32((le32_to_cpu(old_ent->e_rec.e_cpos) +
518 le32_to_cpu(old_ent->e_rec.e_clusters)) -
519 le32_to_cpu(ctxt->right_ent->e_rec.e_cpos));
520 }
521
522 rb_erase(&old_ent->e_node, &em->em_extents);
523 /* Now that he's erased, set him up for deletion */
524 ctxt->old_ent = old_ent;
525
526 if (ctxt->need_left) {
527 ret = ocfs2_extent_map_insert_entry(em,
528 ctxt->left_ent);
529 if (ret)
530 goto out_unlock;
531 ctxt->left_ent = NULL;
532 } 202 }
533 203
534 if (ctxt->need_right) { 204 /*
535 ret = ocfs2_extent_map_insert_entry(em, 205 * Overlapping extents - this shouldn't happen unless we've
536 ctxt->right_ent); 206 * split an extent to change it's flags. That is exceedingly
537 if (ret) 207 * rare, so there's no sense in trying to optimize it yet.
538 goto out_unlock; 208 */
539 ctxt->right_ent = NULL; 209 if (ocfs2_ei_is_contained(emi, ins) ||
210 ocfs2_ei_is_contained(ins, emi)) {
211 ocfs2_copy_emi_fields(emi, ins);
212 return 1;
540 } 213 }
541 214
542 ret = ocfs2_extent_map_insert_entry(em, ctxt->new_ent); 215 /* No merge was possible. */
543 216 return 0;
544 if (!ret)
545 ctxt->new_ent = NULL;
546
547out_unlock:
548 spin_unlock(&OCFS2_I(inode)->ip_lock);
549
550 return ret;
551} 217}
552 218
553 219/*
554static int ocfs2_extent_map_insert(struct inode *inode, 220 * In order to reduce complexity on the caller, this insert function
555 struct ocfs2_extent_rec *rec, 221 * is intentionally liberal in what it will accept.
556 int tree_depth) 222 *
223 * The only rule is that the truncate call *must* be used whenever
224 * records have been deleted. This avoids inserting overlapping
225 * records with different physical mappings.
226 */
227void ocfs2_extent_map_insert_rec(struct inode *inode,
228 struct ocfs2_extent_rec *rec)
557{ 229{
558 int ret; 230 struct ocfs2_inode_info *oi = OCFS2_I(inode);
559 struct ocfs2_em_insert_context ctxt = {0, }; 231 struct ocfs2_extent_map *em = &oi->ip_extent_map;
560 232 struct ocfs2_extent_map_item *emi, *new_emi = NULL;
561 if ((le32_to_cpu(rec->e_cpos) + le32_to_cpu(rec->e_clusters)) > 233 struct ocfs2_extent_map_item ins;
562 OCFS2_I(inode)->ip_map.em_clusters) { 234
563 ret = -EBADR; 235 ins.ei_cpos = le32_to_cpu(rec->e_cpos);
564 mlog_errno(ret); 236 ins.ei_phys = ocfs2_blocks_to_clusters(inode->i_sb,
565 return ret; 237 le64_to_cpu(rec->e_blkno));
238 ins.ei_clusters = le16_to_cpu(rec->e_leaf_clusters);
239 ins.ei_flags = rec->e_flags;
240
241search:
242 spin_lock(&oi->ip_lock);
243
244 list_for_each_entry(emi, &em->em_list, ei_list) {
245 if (ocfs2_try_to_merge_extent_map(emi, &ins)) {
246 list_move(&emi->ei_list, &em->em_list);
247 spin_unlock(&oi->ip_lock);
248 goto out;
249 }
566 } 250 }
567 251
568 /* Zero e_clusters means a truncated tail record. It better be EOF */ 252 /*
569 if (!rec->e_clusters) { 253 * No item could be merged.
570 if ((le32_to_cpu(rec->e_cpos) + le32_to_cpu(rec->e_clusters)) != 254 *
571 OCFS2_I(inode)->ip_map.em_clusters) { 255 * Either allocate and add a new item, or overwrite the last recently
572 ret = -EBADR; 256 * inserted.
573 mlog_errno(ret); 257 */
574 ocfs2_error(inode->i_sb,
575 "Zero e_clusters on non-tail extent record at e_blkno %llu on inode %llu\n",
576 (unsigned long long)le64_to_cpu(rec->e_blkno),
577 (unsigned long long)OCFS2_I(inode)->ip_blkno);
578 return ret;
579 }
580 258
581 /* Ignore the truncated tail */ 259 if (em->em_num_items < OCFS2_MAX_EXTENT_MAP_ITEMS) {
582 return 0; 260 if (new_emi == NULL) {
583 } 261 spin_unlock(&oi->ip_lock);
584 262
585 ret = -ENOMEM; 263 new_emi = kmalloc(sizeof(*new_emi), GFP_NOFS);
586 ctxt.new_ent = kmem_cache_alloc(ocfs2_em_ent_cachep, 264 if (new_emi == NULL)
587 GFP_NOFS); 265 goto out;
588 if (!ctxt.new_ent) {
589 mlog_errno(ret);
590 return ret;
591 }
592 266
593 ctxt.new_ent->e_rec = *rec; 267 goto search;
594 ctxt.new_ent->e_tree_depth = tree_depth;
595
596 do {
597 ret = -ENOMEM;
598 if (ctxt.need_left && !ctxt.left_ent) {
599 ctxt.left_ent =
600 kmem_cache_alloc(ocfs2_em_ent_cachep,
601 GFP_NOFS);
602 if (!ctxt.left_ent)
603 break;
604 }
605 if (ctxt.need_right && !ctxt.right_ent) {
606 ctxt.right_ent =
607 kmem_cache_alloc(ocfs2_em_ent_cachep,
608 GFP_NOFS);
609 if (!ctxt.right_ent)
610 break;
611 } 268 }
612 269
613 ret = ocfs2_extent_map_try_insert(inode, rec, 270 ocfs2_copy_emi_fields(new_emi, &ins);
614 tree_depth, &ctxt); 271 list_add(&new_emi->ei_list, &em->em_list);
615 } while (ret == -EAGAIN); 272 em->em_num_items++;
616 273 new_emi = NULL;
617 if ((ret < 0) && (ret != -EEXIST)) 274 } else {
618 mlog_errno(ret); 275 BUG_ON(list_empty(&em->em_list) || em->em_num_items == 0);
276 emi = list_entry(em->em_list.prev,
277 struct ocfs2_extent_map_item, ei_list);
278 list_move(&emi->ei_list, &em->em_list);
279 ocfs2_copy_emi_fields(emi, &ins);
280 }
619 281
620 if (ctxt.left_ent) 282 spin_unlock(&oi->ip_lock);
621 kmem_cache_free(ocfs2_em_ent_cachep, ctxt.left_ent);
622 if (ctxt.right_ent)
623 kmem_cache_free(ocfs2_em_ent_cachep, ctxt.right_ent);
624 if (ctxt.old_ent)
625 kmem_cache_free(ocfs2_em_ent_cachep, ctxt.old_ent);
626 if (ctxt.new_ent)
627 kmem_cache_free(ocfs2_em_ent_cachep, ctxt.new_ent);
628 283
629 return ret; 284out:
285 if (new_emi)
286 kfree(new_emi);
630} 287}
631 288
632/* 289/*
633 * Append this record to the tail of the extent map. It must be 290 * Return the 1st index within el which contains an extent start
634 * tree_depth 0. The record might be an extension of an existing 291 * larger than v_cluster.
635 * record, and as such that needs to be handled. eg:
636 *
637 * Existing record in the extent map:
638 *
639 * cpos = 10, len = 10
640 * |---------|
641 *
642 * New Record:
643 *
644 * cpos = 10, len = 20
645 * |------------------|
646 *
647 * The passed record is the new on-disk record. The new_clusters value
648 * is how many clusters were added to the file. If the append is a
649 * contiguous append, the new_clusters has been added to
650 * rec->e_clusters. If the append is an entirely new extent, then
651 * rec->e_clusters is == new_clusters.
652 */ 292 */
653int ocfs2_extent_map_append(struct inode *inode, 293static int ocfs2_search_for_hole_index(struct ocfs2_extent_list *el,
654 struct ocfs2_extent_rec *rec, 294 u32 v_cluster)
655 u32 new_clusters)
656{ 295{
657 int ret; 296 int i;
658 struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map; 297 struct ocfs2_extent_rec *rec;
659 struct ocfs2_extent_map_entry *ent;
660 struct ocfs2_extent_rec *old;
661
662 BUG_ON(!new_clusters);
663 BUG_ON(le32_to_cpu(rec->e_clusters) < new_clusters);
664 298
665 if (em->em_clusters < OCFS2_I(inode)->ip_clusters) { 299 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
666 /* 300 rec = &el->l_recs[i];
667 * Size changed underneath us on disk. Drop any
668 * straddling records and update our idea of
669 * i_clusters
670 */
671 ocfs2_extent_map_drop(inode, em->em_clusters - 1);
672 em->em_clusters = OCFS2_I(inode)->ip_clusters;
673 }
674 301
675 mlog_bug_on_msg((le32_to_cpu(rec->e_cpos) + 302 if (v_cluster < le32_to_cpu(rec->e_cpos))
676 le32_to_cpu(rec->e_clusters)) != 303 break;
677 (em->em_clusters + new_clusters),
678 "Inode %llu:\n"
679 "rec->e_cpos = %u + rec->e_clusters = %u = %u\n"
680 "em->em_clusters = %u + new_clusters = %u = %u\n",
681 (unsigned long long)OCFS2_I(inode)->ip_blkno,
682 le32_to_cpu(rec->e_cpos), le32_to_cpu(rec->e_clusters),
683 le32_to_cpu(rec->e_cpos) + le32_to_cpu(rec->e_clusters),
684 em->em_clusters, new_clusters,
685 em->em_clusters + new_clusters);
686
687 em->em_clusters += new_clusters;
688
689 ret = -ENOENT;
690 if (le32_to_cpu(rec->e_clusters) > new_clusters) {
691 /* This is a contiguous append */
692 ent = ocfs2_extent_map_lookup(em, le32_to_cpu(rec->e_cpos), 1,
693 NULL, NULL);
694 if (ent) {
695 old = &ent->e_rec;
696 BUG_ON((le32_to_cpu(rec->e_cpos) +
697 le32_to_cpu(rec->e_clusters)) !=
698 (le32_to_cpu(old->e_cpos) +
699 le32_to_cpu(old->e_clusters) +
700 new_clusters));
701 if (ent->e_tree_depth == 0) {
702 BUG_ON(le32_to_cpu(old->e_cpos) !=
703 le32_to_cpu(rec->e_cpos));
704 BUG_ON(le64_to_cpu(old->e_blkno) !=
705 le64_to_cpu(rec->e_blkno));
706 ret = 0;
707 }
708 /*
709 * Let non-leafs fall through as -ENOENT to
710 * force insertion of the new leaf.
711 */
712 le32_add_cpu(&old->e_clusters, new_clusters);
713 }
714 } 304 }
715 305
716 if (ret == -ENOENT) 306 return i;
717 ret = ocfs2_extent_map_insert(inode, rec, 0);
718 if (ret < 0)
719 mlog_errno(ret);
720 return ret;
721} 307}
722 308
723#if 0
724/* Code here is included but defined out as it completes the extent
725 * map api and may be used in the future. */
726
727/* 309/*
728 * Look up the record containing this cluster offset. This record is 310 * Figure out the size of a hole which starts at v_cluster within the given
729 * part of the extent map. Do not free it. Any changes you make to 311 * extent list.
730 * it will reflect in the extent map. So, if your last extent
731 * is (cpos = 10, clusters = 10) and you truncate the file by 5
732 * clusters, you can do:
733 * 312 *
734 * ret = ocfs2_extent_map_get_rec(em, orig_size - 5, &rec); 313 * If there is no more allocation past v_cluster, we return the maximum
735 * rec->e_clusters -= 5; 314 * cluster size minus v_cluster.
736 * 315 *
737 * The lookup does not read from disk. If the map isn't filled in for 316 * If we have in-inode extents, then el points to the dinode list and
738 * an entry, you won't find it. 317 * eb_bh is NULL. Otherwise, eb_bh should point to the extent block
739 * 318 * containing el.
740 * Also note that the returned record is valid until alloc_sem is
741 * dropped. After that, truncate and extend can happen. Caveat Emptor.
742 */ 319 */
743int ocfs2_extent_map_get_rec(struct inode *inode, u32 cpos, 320static int ocfs2_figure_hole_clusters(struct inode *inode,
744 struct ocfs2_extent_rec **rec, 321 struct ocfs2_extent_list *el,
745 int *tree_depth) 322 struct buffer_head *eb_bh,
323 u32 v_cluster,
324 u32 *num_clusters)
746{ 325{
747 int ret = -ENOENT; 326 int ret, i;
748 struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map; 327 struct buffer_head *next_eb_bh = NULL;
749 struct ocfs2_extent_map_entry *ent; 328 struct ocfs2_extent_block *eb, *next_eb;
750 329
751 *rec = NULL; 330 i = ocfs2_search_for_hole_index(el, v_cluster);
752 331
753 if (cpos >= OCFS2_I(inode)->ip_clusters) 332 if (i == le16_to_cpu(el->l_next_free_rec) && eb_bh) {
754 return -EINVAL; 333 eb = (struct ocfs2_extent_block *)eb_bh->b_data;
755 334
756 if (cpos >= em->em_clusters) {
757 /* 335 /*
758 * Size changed underneath us on disk. Drop any 336 * Check the next leaf for any extents.
759 * straddling records and update our idea of
760 * i_clusters
761 */ 337 */
762 ocfs2_extent_map_drop(inode, em->em_clusters - 1);
763 em->em_clusters = OCFS2_I(inode)->ip_clusters ;
764 }
765
766 ent = ocfs2_extent_map_lookup(&OCFS2_I(inode)->ip_map, cpos, 1,
767 NULL, NULL);
768 338
769 if (ent) { 339 if (le64_to_cpu(eb->h_next_leaf_blk) == 0ULL)
770 *rec = &ent->e_rec; 340 goto no_more_extents;
771 if (tree_depth)
772 *tree_depth = ent->e_tree_depth;
773 ret = 0;
774 }
775 341
776 return ret; 342 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
777} 343 le64_to_cpu(eb->h_next_leaf_blk),
344 &next_eb_bh, OCFS2_BH_CACHED, inode);
345 if (ret) {
346 mlog_errno(ret);
347 goto out;
348 }
349 next_eb = (struct ocfs2_extent_block *)next_eb_bh->b_data;
778 350
779int ocfs2_extent_map_get_clusters(struct inode *inode, 351 if (!OCFS2_IS_VALID_EXTENT_BLOCK(next_eb)) {
780 u32 v_cpos, int count, 352 ret = -EROFS;
781 u32 *p_cpos, int *ret_count) 353 OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, next_eb);
782{ 354 goto out;
783 int ret; 355 }
784 u32 coff, ccount;
785 struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map;
786 struct ocfs2_extent_map_entry *ent = NULL;
787 356
788 *p_cpos = ccount = 0; 357 el = &next_eb->h_list;
789 358
790 if ((v_cpos + count) > OCFS2_I(inode)->ip_clusters) 359 i = ocfs2_search_for_hole_index(el, v_cluster);
791 return -EINVAL; 360 }
792 361
793 if ((v_cpos + count) > em->em_clusters) { 362no_more_extents:
363 if (i == le16_to_cpu(el->l_next_free_rec)) {
794 /* 364 /*
795 * Size changed underneath us on disk. Drop any 365 * We're at the end of our existing allocation. Just
796 * straddling records and update our idea of 366 * return the maximum number of clusters we could
797 * i_clusters 367 * possibly allocate.
798 */ 368 */
799 ocfs2_extent_map_drop(inode, em->em_clusters - 1); 369 *num_clusters = UINT_MAX - v_cluster;
800 em->em_clusters = OCFS2_I(inode)->ip_clusters; 370 } else {
371 *num_clusters = le32_to_cpu(el->l_recs[i].e_cpos) - v_cluster;
801 } 372 }
802 373
374 ret = 0;
375out:
376 brelse(next_eb_bh);
377 return ret;
378}
803 379
804 ret = ocfs2_extent_map_lookup_read(inode, v_cpos, count, &ent); 380/*
805 if (ret) 381 * Return the index of the extent record which contains cluster #v_cluster.
806 return ret; 382 * -1 is returned if it was not found.
383 *
384 * Should work fine on interior and exterior nodes.
385 */
386static int ocfs2_search_extent_list(struct ocfs2_extent_list *el,
387 u32 v_cluster)
388{
389 int ret = -1;
390 int i;
391 struct ocfs2_extent_rec *rec;
392 u32 rec_end, rec_start, clusters;
807 393
808 if (ent) { 394 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
809 /* We should never find ourselves straddling an interval */ 395 rec = &el->l_recs[i];
810 if (!ocfs2_extent_rec_contains_clusters(&ent->e_rec,
811 v_cpos,
812 count))
813 return -ESRCH;
814 396
815 coff = v_cpos - le32_to_cpu(ent->e_rec.e_cpos); 397 rec_start = le32_to_cpu(rec->e_cpos);
816 *p_cpos = ocfs2_blocks_to_clusters(inode->i_sb, 398 clusters = ocfs2_rec_clusters(el, rec);
817 le64_to_cpu(ent->e_rec.e_blkno)) +
818 coff;
819 399
820 if (ret_count) 400 rec_end = rec_start + clusters;
821 *ret_count = le32_to_cpu(ent->e_rec.e_clusters) - coff;
822 401
823 return 0; 402 if (v_cluster >= rec_start && v_cluster < rec_end) {
403 ret = i;
404 break;
405 }
824 } 406 }
825 407
826 408 return ret;
827 return -ENOENT;
828} 409}
829 410
830#endif /* 0 */ 411int ocfs2_get_clusters(struct inode *inode, u32 v_cluster,
831 412 u32 *p_cluster, u32 *num_clusters,
832int ocfs2_extent_map_get_blocks(struct inode *inode, 413 unsigned int *extent_flags)
833 u64 v_blkno, int count,
834 u64 *p_blkno, int *ret_count)
835{ 414{
836 int ret; 415 int ret, i;
837 u64 boff; 416 unsigned int flags = 0;
838 u32 cpos, clusters; 417 struct buffer_head *di_bh = NULL;
839 int bpc = ocfs2_clusters_to_blocks(inode->i_sb, 1); 418 struct buffer_head *eb_bh = NULL;
840 struct ocfs2_extent_map_entry *ent = NULL; 419 struct ocfs2_dinode *di;
841 struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map; 420 struct ocfs2_extent_block *eb;
421 struct ocfs2_extent_list *el;
842 struct ocfs2_extent_rec *rec; 422 struct ocfs2_extent_rec *rec;
423 u32 coff;
843 424
844 *p_blkno = 0; 425 ret = ocfs2_extent_map_lookup(inode, v_cluster, p_cluster,
845 426 num_clusters, extent_flags);
846 cpos = ocfs2_blocks_to_clusters(inode->i_sb, v_blkno); 427 if (ret == 0)
847 clusters = ocfs2_blocks_to_clusters(inode->i_sb, 428 goto out;
848 (u64)count + bpc - 1);
849 if ((cpos + clusters) > OCFS2_I(inode)->ip_clusters) {
850 ret = -EINVAL;
851 mlog_errno(ret);
852 return ret;
853 }
854
855 if ((cpos + clusters) > em->em_clusters) {
856 /*
857 * Size changed underneath us on disk. Drop any
858 * straddling records and update our idea of
859 * i_clusters
860 */
861 ocfs2_extent_map_drop(inode, em->em_clusters - 1);
862 em->em_clusters = OCFS2_I(inode)->ip_clusters;
863 }
864 429
865 ret = ocfs2_extent_map_lookup_read(inode, cpos, clusters, &ent); 430 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb), OCFS2_I(inode)->ip_blkno,
431 &di_bh, OCFS2_BH_CACHED, inode);
866 if (ret) { 432 if (ret) {
867 mlog_errno(ret); 433 mlog_errno(ret);
868 return ret; 434 goto out;
869 } 435 }
870 436
871 if (ent) 437 di = (struct ocfs2_dinode *) di_bh->b_data;
872 { 438 el = &di->id2.i_list;
873 rec = &ent->e_rec;
874 439
875 /* We should never find ourselves straddling an interval */ 440 if (el->l_tree_depth) {
876 if (!ocfs2_extent_rec_contains_clusters(rec, cpos, clusters)) { 441 ret = ocfs2_find_leaf(inode, el, v_cluster, &eb_bh);
877 ret = -ESRCH; 442 if (ret) {
878 mlog_errno(ret); 443 mlog_errno(ret);
879 return ret; 444 goto out;
880 } 445 }
881 446
882 boff = ocfs2_clusters_to_blocks(inode->i_sb, cpos - 447 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
883 le32_to_cpu(rec->e_cpos)); 448 el = &eb->h_list;
884 boff += (v_blkno & (u64)(bpc - 1));
885 *p_blkno = le64_to_cpu(rec->e_blkno) + boff;
886 449
887 if (ret_count) { 450 if (el->l_tree_depth) {
888 *ret_count = ocfs2_clusters_to_blocks(inode->i_sb, 451 ocfs2_error(inode->i_sb,
889 le32_to_cpu(rec->e_clusters)) - boff; 452 "Inode %lu has non zero tree depth in "
453 "leaf block %llu\n", inode->i_ino,
454 (unsigned long long)eb_bh->b_blocknr);
455 ret = -EROFS;
456 goto out;
890 } 457 }
891
892 return 0;
893 } 458 }
894 459
895 return -ENOENT; 460 i = ocfs2_search_extent_list(el, v_cluster);
896} 461 if (i == -1) {
897 462 /*
898int ocfs2_extent_map_init(struct inode *inode) 463 * A hole was found. Return some canned values that
899{ 464 * callers can key on. If asked for, num_clusters will
900 struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map; 465 * be populated with the size of the hole.
901 466 */
902 em->em_extents = RB_ROOT; 467 *p_cluster = 0;
903 em->em_clusters = 0; 468 if (num_clusters) {
904 469 ret = ocfs2_figure_hole_clusters(inode, el, eb_bh,
905 return 0; 470 v_cluster,
906} 471 num_clusters);
907 472 if (ret) {
908/* Needs the lock */ 473 mlog_errno(ret);
909static void __ocfs2_extent_map_drop(struct inode *inode, 474 goto out;
910 u32 new_clusters, 475 }
911 struct rb_node **free_head, 476 }
912 struct ocfs2_extent_map_entry **tail_ent) 477 } else {
913{ 478 rec = &el->l_recs[i];
914 struct rb_node *node, *next;
915 struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map;
916 struct ocfs2_extent_map_entry *ent;
917 479
918 *free_head = NULL; 480 BUG_ON(v_cluster < le32_to_cpu(rec->e_cpos));
919 481
920 ent = NULL; 482 if (!rec->e_blkno) {
921 node = rb_last(&em->em_extents); 483 ocfs2_error(inode->i_sb, "Inode %lu has bad extent "
922 while (node) 484 "record (%u, %u, 0)", inode->i_ino,
923 { 485 le32_to_cpu(rec->e_cpos),
924 next = rb_prev(node); 486 ocfs2_rec_clusters(el, rec));
487 ret = -EROFS;
488 goto out;
489 }
925 490
926 ent = rb_entry(node, struct ocfs2_extent_map_entry, 491 coff = v_cluster - le32_to_cpu(rec->e_cpos);
927 e_node);
928 if (le32_to_cpu(ent->e_rec.e_cpos) < new_clusters)
929 break;
930 492
931 rb_erase(&ent->e_node, &em->em_extents); 493 *p_cluster = ocfs2_blocks_to_clusters(inode->i_sb,
494 le64_to_cpu(rec->e_blkno));
495 *p_cluster = *p_cluster + coff;
932 496
933 node->rb_right = *free_head; 497 if (num_clusters)
934 *free_head = node; 498 *num_clusters = ocfs2_rec_clusters(el, rec) - coff;
935 499
936 ent = NULL; 500 flags = rec->e_flags;
937 node = next;
938 }
939 501
940 /* Do we have an entry straddling new_clusters? */ 502 ocfs2_extent_map_insert_rec(inode, rec);
941 if (tail_ent) {
942 if (ent &&
943 ((le32_to_cpu(ent->e_rec.e_cpos) +
944 le32_to_cpu(ent->e_rec.e_clusters)) > new_clusters))
945 *tail_ent = ent;
946 else
947 *tail_ent = NULL;
948 } 503 }
949}
950
951static void __ocfs2_extent_map_drop_cleanup(struct rb_node *free_head)
952{
953 struct rb_node *node;
954 struct ocfs2_extent_map_entry *ent;
955 504
956 while (free_head) { 505 if (extent_flags)
957 node = free_head; 506 *extent_flags = flags;
958 free_head = node->rb_right;
959 507
960 ent = rb_entry(node, struct ocfs2_extent_map_entry, 508out:
961 e_node); 509 brelse(di_bh);
962 kmem_cache_free(ocfs2_em_ent_cachep, ent); 510 brelse(eb_bh);
963 } 511 return ret;
964} 512}
965 513
966/* 514/*
967 * Remove all entries past new_clusters, inclusive of an entry that 515 * This expects alloc_sem to be held. The allocation cannot change at
968 * contains new_clusters. This is effectively a cache forget. 516 * all while the map is in the process of being updated.
969 *
970 * If you want to also clip the last extent by some number of clusters,
971 * you need to call ocfs2_extent_map_trunc().
972 * This code does not check or modify ip_clusters.
973 */ 517 */
974int ocfs2_extent_map_drop(struct inode *inode, u32 new_clusters) 518int ocfs2_extent_map_get_blocks(struct inode *inode, u64 v_blkno, u64 *p_blkno,
519 u64 *ret_count, unsigned int *extent_flags)
975{ 520{
976 struct rb_node *free_head = NULL; 521 int ret;
977 struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map; 522 int bpc = ocfs2_clusters_to_blocks(inode->i_sb, 1);
978 struct ocfs2_extent_map_entry *ent; 523 u32 cpos, num_clusters, p_cluster;
979 524 u64 boff = 0;
980 spin_lock(&OCFS2_I(inode)->ip_lock);
981 525
982 __ocfs2_extent_map_drop(inode, new_clusters, &free_head, &ent); 526 cpos = ocfs2_blocks_to_clusters(inode->i_sb, v_blkno);
983 527
984 if (ent) { 528 ret = ocfs2_get_clusters(inode, cpos, &p_cluster, &num_clusters,
985 rb_erase(&ent->e_node, &em->em_extents); 529 extent_flags);
986 ent->e_node.rb_right = free_head; 530 if (ret) {
987 free_head = &ent->e_node; 531 mlog_errno(ret);
532 goto out;
988 } 533 }
989 534
990 spin_unlock(&OCFS2_I(inode)->ip_lock); 535 /*
991 536 * p_cluster == 0 indicates a hole.
992 if (free_head) 537 */
993 __ocfs2_extent_map_drop_cleanup(free_head); 538 if (p_cluster) {
994 539 boff = ocfs2_clusters_to_blocks(inode->i_sb, p_cluster);
995 return 0; 540 boff += (v_blkno & (u64)(bpc - 1));
996} 541 }
997
998/*
999 * Remove all entries past new_clusters and also clip any extent
1000 * straddling new_clusters, if there is one. This does not check
1001 * or modify ip_clusters
1002 */
1003int ocfs2_extent_map_trunc(struct inode *inode, u32 new_clusters)
1004{
1005 struct rb_node *free_head = NULL;
1006 struct ocfs2_extent_map_entry *ent = NULL;
1007
1008 spin_lock(&OCFS2_I(inode)->ip_lock);
1009
1010 __ocfs2_extent_map_drop(inode, new_clusters, &free_head, &ent);
1011
1012 if (ent)
1013 ent->e_rec.e_clusters = cpu_to_le32(new_clusters -
1014 le32_to_cpu(ent->e_rec.e_cpos));
1015
1016 OCFS2_I(inode)->ip_map.em_clusters = new_clusters;
1017
1018 spin_unlock(&OCFS2_I(inode)->ip_lock);
1019
1020 if (free_head)
1021 __ocfs2_extent_map_drop_cleanup(free_head);
1022
1023 return 0;
1024}
1025 542
1026int __init init_ocfs2_extent_maps(void) 543 *p_blkno = boff;
1027{
1028 ocfs2_em_ent_cachep =
1029 kmem_cache_create("ocfs2_em_ent",
1030 sizeof(struct ocfs2_extent_map_entry),
1031 0, SLAB_HWCACHE_ALIGN, NULL, NULL);
1032 if (!ocfs2_em_ent_cachep)
1033 return -ENOMEM;
1034 544
1035 return 0; 545 if (ret_count) {
1036} 546 *ret_count = ocfs2_clusters_to_blocks(inode->i_sb, num_clusters);
547 *ret_count -= v_blkno & (u64)(bpc - 1);
548 }
1037 549
1038void exit_ocfs2_extent_maps(void) 550out:
1039{ 551 return ret;
1040 kmem_cache_destroy(ocfs2_em_ent_cachep);
1041} 552}
diff --git a/fs/ocfs2/extent_map.h b/fs/ocfs2/extent_map.h
index fa3745efa886..de91e3e41a22 100644
--- a/fs/ocfs2/extent_map.h
+++ b/fs/ocfs2/extent_map.h
@@ -25,22 +25,29 @@
25#ifndef _EXTENT_MAP_H 25#ifndef _EXTENT_MAP_H
26#define _EXTENT_MAP_H 26#define _EXTENT_MAP_H
27 27
28int init_ocfs2_extent_maps(void); 28struct ocfs2_extent_map_item {
29void exit_ocfs2_extent_maps(void); 29 unsigned int ei_cpos;
30 unsigned int ei_phys;
31 unsigned int ei_clusters;
32 unsigned int ei_flags;
30 33
31/* 34 struct list_head ei_list;
32 * EVERY CALL here except _init, _trunc, and _drop expects alloc_sem 35};
33 * to be held. The allocation cannot change at all while the map is 36
34 * in the process of being updated. 37#define OCFS2_MAX_EXTENT_MAP_ITEMS 3
35 */ 38struct ocfs2_extent_map {
36int ocfs2_extent_map_init(struct inode *inode); 39 unsigned int em_num_items;
37int ocfs2_extent_map_append(struct inode *inode, 40 struct list_head em_list;
38 struct ocfs2_extent_rec *rec, 41};
39 u32 new_clusters); 42
40int ocfs2_extent_map_get_blocks(struct inode *inode, 43void ocfs2_extent_map_init(struct inode *inode);
41 u64 v_blkno, int count, 44void ocfs2_extent_map_trunc(struct inode *inode, unsigned int cluster);
42 u64 *p_blkno, int *ret_count); 45void ocfs2_extent_map_insert_rec(struct inode *inode,
43int ocfs2_extent_map_drop(struct inode *inode, u32 new_clusters); 46 struct ocfs2_extent_rec *rec);
44int ocfs2_extent_map_trunc(struct inode *inode, u32 new_clusters); 47
48int ocfs2_get_clusters(struct inode *inode, u32 v_cluster, u32 *p_cluster,
49 u32 *num_clusters, unsigned int *extent_flags);
50int ocfs2_extent_map_get_blocks(struct inode *inode, u64 v_blkno, u64 *p_blkno,
51 u64 *ret_count, unsigned int *extent_flags);
45 52
46#endif /* _EXTENT_MAP_H */ 53#endif /* _EXTENT_MAP_H */
diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c
index f2cd3bf9efb2..520a2a6d7670 100644
--- a/fs/ocfs2/file.c
+++ b/fs/ocfs2/file.c
@@ -33,6 +33,7 @@
33#include <linux/sched.h> 33#include <linux/sched.h>
34#include <linux/pipe_fs_i.h> 34#include <linux/pipe_fs_i.h>
35#include <linux/mount.h> 35#include <linux/mount.h>
36#include <linux/writeback.h>
36 37
37#define MLOG_MASK_PREFIX ML_INODE 38#define MLOG_MASK_PREFIX ML_INODE
38#include <cluster/masklog.h> 39#include <cluster/masklog.h>
@@ -215,7 +216,7 @@ int ocfs2_set_inode_size(handle_t *handle,
215 216
216 mlog_entry_void(); 217 mlog_entry_void();
217 i_size_write(inode, new_i_size); 218 i_size_write(inode, new_i_size);
218 inode->i_blocks = ocfs2_align_bytes_to_sectors(new_i_size); 219 inode->i_blocks = ocfs2_inode_sector_count(inode);
219 inode->i_ctime = inode->i_mtime = CURRENT_TIME; 220 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
220 221
221 status = ocfs2_mark_inode_dirty(handle, inode, fe_bh); 222 status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
@@ -261,6 +262,7 @@ static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
261{ 262{
262 int status; 263 int status;
263 handle_t *handle; 264 handle_t *handle;
265 struct ocfs2_dinode *di;
264 266
265 mlog_entry_void(); 267 mlog_entry_void();
266 268
@@ -274,12 +276,39 @@ static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
274 goto out; 276 goto out;
275 } 277 }
276 278
277 status = ocfs2_set_inode_size(handle, inode, fe_bh, new_i_size); 279 status = ocfs2_journal_access(handle, inode, fe_bh,
280 OCFS2_JOURNAL_ACCESS_WRITE);
281 if (status < 0) {
282 mlog_errno(status);
283 goto out_commit;
284 }
285
286 /*
287 * Do this before setting i_size.
288 */
289 status = ocfs2_zero_tail_for_truncate(inode, handle, new_i_size);
290 if (status) {
291 mlog_errno(status);
292 goto out_commit;
293 }
294
295 i_size_write(inode, new_i_size);
296 inode->i_blocks = ocfs2_align_bytes_to_sectors(new_i_size);
297 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
298
299 di = (struct ocfs2_dinode *) fe_bh->b_data;
300 di->i_size = cpu_to_le64(new_i_size);
301 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
302 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
303
304 status = ocfs2_journal_dirty(handle, fe_bh);
278 if (status < 0) 305 if (status < 0)
279 mlog_errno(status); 306 mlog_errno(status);
280 307
308out_commit:
281 ocfs2_commit_trans(osb, handle); 309 ocfs2_commit_trans(osb, handle);
282out: 310out:
311
283 mlog_exit(status); 312 mlog_exit(status);
284 return status; 313 return status;
285} 314}
@@ -342,19 +371,6 @@ static int ocfs2_truncate_file(struct inode *inode,
342 mlog_errno(status); 371 mlog_errno(status);
343 goto bail; 372 goto bail;
344 } 373 }
345 ocfs2_data_unlock(inode, 1);
346
347 if (le32_to_cpu(fe->i_clusters) ==
348 ocfs2_clusters_for_bytes(osb->sb, new_i_size)) {
349 mlog(0, "fe->i_clusters = %u, so we do a simple truncate\n",
350 fe->i_clusters);
351 /* No allocation change is required, so lets fast path
352 * this truncate. */
353 status = ocfs2_simple_size_update(inode, di_bh, new_i_size);
354 if (status < 0)
355 mlog_errno(status);
356 goto bail;
357 }
358 374
359 /* alright, we're going to need to do a full blown alloc size 375 /* alright, we're going to need to do a full blown alloc size
360 * change. Orphan the inode so that recovery can complete the 376 * change. Orphan the inode so that recovery can complete the
@@ -363,22 +379,25 @@ static int ocfs2_truncate_file(struct inode *inode,
363 status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size); 379 status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
364 if (status < 0) { 380 if (status < 0) {
365 mlog_errno(status); 381 mlog_errno(status);
366 goto bail; 382 goto bail_unlock_data;
367 } 383 }
368 384
369 status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc); 385 status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
370 if (status < 0) { 386 if (status < 0) {
371 mlog_errno(status); 387 mlog_errno(status);
372 goto bail; 388 goto bail_unlock_data;
373 } 389 }
374 390
375 status = ocfs2_commit_truncate(osb, inode, di_bh, tc); 391 status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
376 if (status < 0) { 392 if (status < 0) {
377 mlog_errno(status); 393 mlog_errno(status);
378 goto bail; 394 goto bail_unlock_data;
379 } 395 }
380 396
381 /* TODO: orphan dir cleanup here. */ 397 /* TODO: orphan dir cleanup here. */
398bail_unlock_data:
399 ocfs2_data_unlock(inode, 1);
400
382bail: 401bail:
383 402
384 mlog_exit(status); 403 mlog_exit(status);
@@ -397,6 +416,7 @@ bail:
397 */ 416 */
398int ocfs2_do_extend_allocation(struct ocfs2_super *osb, 417int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
399 struct inode *inode, 418 struct inode *inode,
419 u32 *logical_offset,
400 u32 clusters_to_add, 420 u32 clusters_to_add,
401 struct buffer_head *fe_bh, 421 struct buffer_head *fe_bh,
402 handle_t *handle, 422 handle_t *handle,
@@ -460,18 +480,14 @@ int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
460 block = ocfs2_clusters_to_blocks(osb->sb, bit_off); 480 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
461 mlog(0, "Allocating %u clusters at block %u for inode %llu\n", 481 mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
462 num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno); 482 num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
463 status = ocfs2_insert_extent(osb, handle, inode, fe_bh, block, 483 status = ocfs2_insert_extent(osb, handle, inode, fe_bh,
464 num_bits, meta_ac); 484 *logical_offset, block, num_bits,
485 meta_ac);
465 if (status < 0) { 486 if (status < 0) {
466 mlog_errno(status); 487 mlog_errno(status);
467 goto leave; 488 goto leave;
468 } 489 }
469 490
470 le32_add_cpu(&fe->i_clusters, num_bits);
471 spin_lock(&OCFS2_I(inode)->ip_lock);
472 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters);
473 spin_unlock(&OCFS2_I(inode)->ip_lock);
474
475 status = ocfs2_journal_dirty(handle, fe_bh); 491 status = ocfs2_journal_dirty(handle, fe_bh);
476 if (status < 0) { 492 if (status < 0) {
477 mlog_errno(status); 493 mlog_errno(status);
@@ -479,6 +495,7 @@ int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
479 } 495 }
480 496
481 clusters_to_add -= num_bits; 497 clusters_to_add -= num_bits;
498 *logical_offset += num_bits;
482 499
483 if (clusters_to_add) { 500 if (clusters_to_add) {
484 mlog(0, "need to alloc once more, clusters = %u, wanted = " 501 mlog(0, "need to alloc once more, clusters = %u, wanted = "
@@ -494,14 +511,87 @@ leave:
494 return status; 511 return status;
495} 512}
496 513
514/*
515 * For a given allocation, determine which allocators will need to be
516 * accessed, and lock them, reserving the appropriate number of bits.
517 *
518 * Called from ocfs2_extend_allocation() for file systems which don't
519 * support holes, and from ocfs2_write() for file systems which
520 * understand sparse inodes.
521 */
522int ocfs2_lock_allocators(struct inode *inode, struct ocfs2_dinode *di,
523 u32 clusters_to_add,
524 struct ocfs2_alloc_context **data_ac,
525 struct ocfs2_alloc_context **meta_ac)
526{
527 int ret, num_free_extents;
528 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
529
530 *meta_ac = NULL;
531 *data_ac = NULL;
532
533 mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
534 "clusters_to_add = %u\n",
535 (unsigned long long)OCFS2_I(inode)->ip_blkno, i_size_read(inode),
536 le32_to_cpu(di->i_clusters), clusters_to_add);
537
538 num_free_extents = ocfs2_num_free_extents(osb, inode, di);
539 if (num_free_extents < 0) {
540 ret = num_free_extents;
541 mlog_errno(ret);
542 goto out;
543 }
544
545 /*
546 * Sparse allocation file systems need to be more conservative
547 * with reserving room for expansion - the actual allocation
548 * happens while we've got a journal handle open so re-taking
549 * a cluster lock (because we ran out of room for another
550 * extent) will violate ordering rules.
551 *
552 * Most of the time we'll only be seeing this 1 cluster at a time
553 * anyway.
554 */
555 if (!num_free_extents ||
556 (ocfs2_sparse_alloc(osb) && num_free_extents < clusters_to_add)) {
557 ret = ocfs2_reserve_new_metadata(osb, di, meta_ac);
558 if (ret < 0) {
559 if (ret != -ENOSPC)
560 mlog_errno(ret);
561 goto out;
562 }
563 }
564
565 ret = ocfs2_reserve_clusters(osb, clusters_to_add, data_ac);
566 if (ret < 0) {
567 if (ret != -ENOSPC)
568 mlog_errno(ret);
569 goto out;
570 }
571
572out:
573 if (ret) {
574 if (*meta_ac) {
575 ocfs2_free_alloc_context(*meta_ac);
576 *meta_ac = NULL;
577 }
578
579 /*
580 * We cannot have an error and a non null *data_ac.
581 */
582 }
583
584 return ret;
585}
586
497static int ocfs2_extend_allocation(struct inode *inode, 587static int ocfs2_extend_allocation(struct inode *inode,
498 u32 clusters_to_add) 588 u32 clusters_to_add)
499{ 589{
500 int status = 0; 590 int status = 0;
501 int restart_func = 0; 591 int restart_func = 0;
502 int drop_alloc_sem = 0; 592 int drop_alloc_sem = 0;
503 int credits, num_free_extents; 593 int credits;
504 u32 prev_clusters; 594 u32 prev_clusters, logical_start;
505 struct buffer_head *bh = NULL; 595 struct buffer_head *bh = NULL;
506 struct ocfs2_dinode *fe = NULL; 596 struct ocfs2_dinode *fe = NULL;
507 handle_t *handle = NULL; 597 handle_t *handle = NULL;
@@ -512,6 +602,12 @@ static int ocfs2_extend_allocation(struct inode *inode,
512 602
513 mlog_entry("(clusters_to_add = %u)\n", clusters_to_add); 603 mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
514 604
605 /*
606 * This function only exists for file systems which don't
607 * support holes.
608 */
609 BUG_ON(ocfs2_sparse_alloc(osb));
610
515 status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh, 611 status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh,
516 OCFS2_BH_CACHED, inode); 612 OCFS2_BH_CACHED, inode);
517 if (status < 0) { 613 if (status < 0) {
@@ -526,39 +622,11 @@ static int ocfs2_extend_allocation(struct inode *inode,
526 goto leave; 622 goto leave;
527 } 623 }
528 624
625 logical_start = OCFS2_I(inode)->ip_clusters;
626
529restart_all: 627restart_all:
530 BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters); 628 BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
531 629
532 mlog(0, "extend inode %llu, i_size = %lld, fe->i_clusters = %u, "
533 "clusters_to_add = %u\n",
534 (unsigned long long)OCFS2_I(inode)->ip_blkno, i_size_read(inode),
535 fe->i_clusters, clusters_to_add);
536
537 num_free_extents = ocfs2_num_free_extents(osb,
538 inode,
539 fe);
540 if (num_free_extents < 0) {
541 status = num_free_extents;
542 mlog_errno(status);
543 goto leave;
544 }
545
546 if (!num_free_extents) {
547 status = ocfs2_reserve_new_metadata(osb, fe, &meta_ac);
548 if (status < 0) {
549 if (status != -ENOSPC)
550 mlog_errno(status);
551 goto leave;
552 }
553 }
554
555 status = ocfs2_reserve_clusters(osb, clusters_to_add, &data_ac);
556 if (status < 0) {
557 if (status != -ENOSPC)
558 mlog_errno(status);
559 goto leave;
560 }
561
562 /* blocks peope in read/write from reading our allocation 630 /* blocks peope in read/write from reading our allocation
563 * until we're done changing it. We depend on i_mutex to block 631 * until we're done changing it. We depend on i_mutex to block
564 * other extend/truncate calls while we're here. Ordering wrt 632 * other extend/truncate calls while we're here. Ordering wrt
@@ -566,6 +634,13 @@ restart_all:
566 down_write(&OCFS2_I(inode)->ip_alloc_sem); 634 down_write(&OCFS2_I(inode)->ip_alloc_sem);
567 drop_alloc_sem = 1; 635 drop_alloc_sem = 1;
568 636
637 status = ocfs2_lock_allocators(inode, fe, clusters_to_add, &data_ac,
638 &meta_ac);
639 if (status) {
640 mlog_errno(status);
641 goto leave;
642 }
643
569 credits = ocfs2_calc_extend_credits(osb->sb, fe, clusters_to_add); 644 credits = ocfs2_calc_extend_credits(osb->sb, fe, clusters_to_add);
570 handle = ocfs2_start_trans(osb, credits); 645 handle = ocfs2_start_trans(osb, credits);
571 if (IS_ERR(handle)) { 646 if (IS_ERR(handle)) {
@@ -590,6 +665,7 @@ restarted_transaction:
590 665
591 status = ocfs2_do_extend_allocation(osb, 666 status = ocfs2_do_extend_allocation(osb,
592 inode, 667 inode,
668 &logical_start,
593 clusters_to_add, 669 clusters_to_add,
594 bh, 670 bh,
595 handle, 671 handle,
@@ -778,7 +854,7 @@ static int ocfs2_extend_file(struct inode *inode,
778 size_t tail_to_skip) 854 size_t tail_to_skip)
779{ 855{
780 int ret = 0; 856 int ret = 0;
781 u32 clusters_to_add; 857 u32 clusters_to_add = 0;
782 858
783 BUG_ON(!tail_to_skip && !di_bh); 859 BUG_ON(!tail_to_skip && !di_bh);
784 860
@@ -790,6 +866,11 @@ static int ocfs2_extend_file(struct inode *inode,
790 goto out; 866 goto out;
791 BUG_ON(new_i_size < i_size_read(inode)); 867 BUG_ON(new_i_size < i_size_read(inode));
792 868
869 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
870 BUG_ON(tail_to_skip != 0);
871 goto out_update_size;
872 }
873
793 clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size) - 874 clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size) -
794 OCFS2_I(inode)->ip_clusters; 875 OCFS2_I(inode)->ip_clusters;
795 876
@@ -825,6 +906,7 @@ static int ocfs2_extend_file(struct inode *inode,
825 goto out_unlock; 906 goto out_unlock;
826 } 907 }
827 908
909out_update_size:
828 if (!tail_to_skip) { 910 if (!tail_to_skip) {
829 /* We're being called from ocfs2_setattr() which wants 911 /* We're being called from ocfs2_setattr() which wants
830 * us to update i_size */ 912 * us to update i_size */
@@ -834,7 +916,8 @@ static int ocfs2_extend_file(struct inode *inode,
834 } 916 }
835 917
836out_unlock: 918out_unlock:
837 ocfs2_data_unlock(inode, 1); 919 if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
920 ocfs2_data_unlock(inode, 1);
838 921
839out: 922out:
840 return ret; 923 return ret;
@@ -972,7 +1055,8 @@ int ocfs2_permission(struct inode *inode, int mask, struct nameidata *nd)
972 1055
973 ret = ocfs2_meta_lock(inode, NULL, 0); 1056 ret = ocfs2_meta_lock(inode, NULL, 0);
974 if (ret) { 1057 if (ret) {
975 mlog_errno(ret); 1058 if (ret != -ENOENT)
1059 mlog_errno(ret);
976 goto out; 1060 goto out;
977 } 1061 }
978 1062
@@ -1035,10 +1119,49 @@ out:
1035 return ret; 1119 return ret;
1036} 1120}
1037 1121
1122/*
1123 * Will look for holes and unwritten extents in the range starting at
1124 * pos for count bytes (inclusive).
1125 */
1126static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1127 size_t count)
1128{
1129 int ret = 0;
1130 unsigned int extent_flags;
1131 u32 cpos, clusters, extent_len, phys_cpos;
1132 struct super_block *sb = inode->i_sb;
1133
1134 cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1135 clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1136
1137 while (clusters) {
1138 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1139 &extent_flags);
1140 if (ret < 0) {
1141 mlog_errno(ret);
1142 goto out;
1143 }
1144
1145 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1146 ret = 1;
1147 break;
1148 }
1149
1150 if (extent_len > clusters)
1151 extent_len = clusters;
1152
1153 clusters -= extent_len;
1154 cpos += extent_len;
1155 }
1156out:
1157 return ret;
1158}
1159
1038static int ocfs2_prepare_inode_for_write(struct dentry *dentry, 1160static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1039 loff_t *ppos, 1161 loff_t *ppos,
1040 size_t count, 1162 size_t count,
1041 int appending) 1163 int appending,
1164 int *direct_io)
1042{ 1165{
1043 int ret = 0, meta_level = appending; 1166 int ret = 0, meta_level = appending;
1044 struct inode *inode = dentry->d_inode; 1167 struct inode *inode = dentry->d_inode;
@@ -1089,6 +1212,49 @@ static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1089 } else { 1212 } else {
1090 saved_pos = *ppos; 1213 saved_pos = *ppos;
1091 } 1214 }
1215
1216 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
1217 loff_t end = saved_pos + count;
1218
1219 /*
1220 * Skip the O_DIRECT checks if we don't need
1221 * them.
1222 */
1223 if (!direct_io || !(*direct_io))
1224 break;
1225
1226 /*
1227 * Allowing concurrent direct writes means
1228 * i_size changes wouldn't be synchronized, so
1229 * one node could wind up truncating another
1230 * nodes writes.
1231 */
1232 if (end > i_size_read(inode)) {
1233 *direct_io = 0;
1234 break;
1235 }
1236
1237 /*
1238 * We don't fill holes during direct io, so
1239 * check for them here. If any are found, the
1240 * caller will have to retake some cluster
1241 * locks and initiate the io as buffered.
1242 */
1243 ret = ocfs2_check_range_for_holes(inode, saved_pos,
1244 count);
1245 if (ret == 1) {
1246 *direct_io = 0;
1247 ret = 0;
1248 } else if (ret < 0)
1249 mlog_errno(ret);
1250 break;
1251 }
1252
1253 /*
1254 * The rest of this loop is concerned with legacy file
1255 * systems which don't support sparse files.
1256 */
1257
1092 newsize = count + saved_pos; 1258 newsize = count + saved_pos;
1093 1259
1094 mlog(0, "pos=%lld newsize=%lld cursize=%lld\n", 1260 mlog(0, "pos=%lld newsize=%lld cursize=%lld\n",
@@ -1141,55 +1307,264 @@ out:
1141 return ret; 1307 return ret;
1142} 1308}
1143 1309
1310static inline void
1311ocfs2_set_next_iovec(const struct iovec **iovp, size_t *basep, size_t bytes)
1312{
1313 const struct iovec *iov = *iovp;
1314 size_t base = *basep;
1315
1316 do {
1317 int copy = min(bytes, iov->iov_len - base);
1318
1319 bytes -= copy;
1320 base += copy;
1321 if (iov->iov_len == base) {
1322 iov++;
1323 base = 0;
1324 }
1325 } while (bytes);
1326 *iovp = iov;
1327 *basep = base;
1328}
1329
1330static struct page * ocfs2_get_write_source(struct ocfs2_buffered_write_priv *bp,
1331 const struct iovec *cur_iov,
1332 size_t iov_offset)
1333{
1334 int ret;
1335 char *buf;
1336 struct page *src_page = NULL;
1337
1338 buf = cur_iov->iov_base + iov_offset;
1339
1340 if (!segment_eq(get_fs(), KERNEL_DS)) {
1341 /*
1342 * Pull in the user page. We want to do this outside
1343 * of the meta data locks in order to preserve locking
1344 * order in case of page fault.
1345 */
1346 ret = get_user_pages(current, current->mm,
1347 (unsigned long)buf & PAGE_CACHE_MASK, 1,
1348 0, 0, &src_page, NULL);
1349 if (ret == 1)
1350 bp->b_src_buf = kmap(src_page);
1351 else
1352 src_page = ERR_PTR(-EFAULT);
1353 } else {
1354 bp->b_src_buf = buf;
1355 }
1356
1357 return src_page;
1358}
1359
1360static void ocfs2_put_write_source(struct ocfs2_buffered_write_priv *bp,
1361 struct page *page)
1362{
1363 if (page) {
1364 kunmap(page);
1365 page_cache_release(page);
1366 }
1367}
1368
1369static ssize_t ocfs2_file_buffered_write(struct file *file, loff_t *ppos,
1370 const struct iovec *iov,
1371 unsigned long nr_segs,
1372 size_t count,
1373 ssize_t o_direct_written)
1374{
1375 int ret = 0;
1376 ssize_t copied, total = 0;
1377 size_t iov_offset = 0;
1378 const struct iovec *cur_iov = iov;
1379 struct ocfs2_buffered_write_priv bp;
1380 struct page *page;
1381
1382 /*
1383 * handle partial DIO write. Adjust cur_iov if needed.
1384 */
1385 ocfs2_set_next_iovec(&cur_iov, &iov_offset, o_direct_written);
1386
1387 do {
1388 bp.b_cur_off = iov_offset;
1389 bp.b_cur_iov = cur_iov;
1390
1391 page = ocfs2_get_write_source(&bp, cur_iov, iov_offset);
1392 if (IS_ERR(page)) {
1393 ret = PTR_ERR(page);
1394 goto out;
1395 }
1396
1397 copied = ocfs2_buffered_write_cluster(file, *ppos, count,
1398 ocfs2_map_and_write_user_data,
1399 &bp);
1400
1401 ocfs2_put_write_source(&bp, page);
1402
1403 if (copied < 0) {
1404 mlog_errno(copied);
1405 ret = copied;
1406 goto out;
1407 }
1408
1409 total += copied;
1410 *ppos = *ppos + copied;
1411 count -= copied;
1412
1413 ocfs2_set_next_iovec(&cur_iov, &iov_offset, copied);
1414 } while(count);
1415
1416out:
1417 return total ? total : ret;
1418}
1419
1420static int ocfs2_check_iovec(const struct iovec *iov, size_t *counted,
1421 unsigned long *nr_segs)
1422{
1423 size_t ocount; /* original count */
1424 unsigned long seg;
1425
1426 ocount = 0;
1427 for (seg = 0; seg < *nr_segs; seg++) {
1428 const struct iovec *iv = &iov[seg];
1429
1430 /*
1431 * If any segment has a negative length, or the cumulative
1432 * length ever wraps negative then return -EINVAL.
1433 */
1434 ocount += iv->iov_len;
1435 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
1436 return -EINVAL;
1437 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
1438 continue;
1439 if (seg == 0)
1440 return -EFAULT;
1441 *nr_segs = seg;
1442 ocount -= iv->iov_len; /* This segment is no good */
1443 break;
1444 }
1445
1446 *counted = ocount;
1447 return 0;
1448}
1449
1144static ssize_t ocfs2_file_aio_write(struct kiocb *iocb, 1450static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1145 const struct iovec *iov, 1451 const struct iovec *iov,
1146 unsigned long nr_segs, 1452 unsigned long nr_segs,
1147 loff_t pos) 1453 loff_t pos)
1148{ 1454{
1149 int ret, rw_level, have_alloc_sem = 0; 1455 int ret, direct_io, appending, rw_level, have_alloc_sem = 0;
1150 struct file *filp = iocb->ki_filp; 1456 int can_do_direct, sync = 0;
1151 struct inode *inode = filp->f_path.dentry->d_inode; 1457 ssize_t written = 0;
1152 int appending = filp->f_flags & O_APPEND ? 1 : 0; 1458 size_t ocount; /* original count */
1153 1459 size_t count; /* after file limit checks */
1154 mlog_entry("(0x%p, %u, '%.*s')\n", filp, 1460 loff_t *ppos = &iocb->ki_pos;
1461 struct file *file = iocb->ki_filp;
1462 struct inode *inode = file->f_path.dentry->d_inode;
1463
1464 mlog_entry("(0x%p, %u, '%.*s')\n", file,
1155 (unsigned int)nr_segs, 1465 (unsigned int)nr_segs,
1156 filp->f_path.dentry->d_name.len, 1466 file->f_path.dentry->d_name.len,
1157 filp->f_path.dentry->d_name.name); 1467 file->f_path.dentry->d_name.name);
1158 1468
1159 /* happy write of zero bytes */
1160 if (iocb->ki_left == 0) 1469 if (iocb->ki_left == 0)
1161 return 0; 1470 return 0;
1162 1471
1472 ret = ocfs2_check_iovec(iov, &ocount, &nr_segs);
1473 if (ret)
1474 return ret;
1475
1476 count = ocount;
1477
1478 vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1479
1480 appending = file->f_flags & O_APPEND ? 1 : 0;
1481 direct_io = file->f_flags & O_DIRECT ? 1 : 0;
1482
1163 mutex_lock(&inode->i_mutex); 1483 mutex_lock(&inode->i_mutex);
1484
1485relock:
1164 /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */ 1486 /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1165 if (filp->f_flags & O_DIRECT) { 1487 if (direct_io) {
1166 have_alloc_sem = 1;
1167 down_read(&inode->i_alloc_sem); 1488 down_read(&inode->i_alloc_sem);
1489 have_alloc_sem = 1;
1168 } 1490 }
1169 1491
1170 /* concurrent O_DIRECT writes are allowed */ 1492 /* concurrent O_DIRECT writes are allowed */
1171 rw_level = (filp->f_flags & O_DIRECT) ? 0 : 1; 1493 rw_level = !direct_io;
1172 ret = ocfs2_rw_lock(inode, rw_level); 1494 ret = ocfs2_rw_lock(inode, rw_level);
1173 if (ret < 0) { 1495 if (ret < 0) {
1174 rw_level = -1;
1175 mlog_errno(ret); 1496 mlog_errno(ret);
1176 goto out; 1497 goto out_sems;
1177 } 1498 }
1178 1499
1179 ret = ocfs2_prepare_inode_for_write(filp->f_path.dentry, &iocb->ki_pos, 1500 can_do_direct = direct_io;
1180 iocb->ki_left, appending); 1501 ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
1502 iocb->ki_left, appending,
1503 &can_do_direct);
1181 if (ret < 0) { 1504 if (ret < 0) {
1182 mlog_errno(ret); 1505 mlog_errno(ret);
1183 goto out; 1506 goto out;
1184 } 1507 }
1185 1508
1186 /* communicate with ocfs2_dio_end_io */ 1509 /*
1187 ocfs2_iocb_set_rw_locked(iocb); 1510 * We can't complete the direct I/O as requested, fall back to
1511 * buffered I/O.
1512 */
1513 if (direct_io && !can_do_direct) {
1514 ocfs2_rw_unlock(inode, rw_level);
1515 up_read(&inode->i_alloc_sem);
1516
1517 have_alloc_sem = 0;
1518 rw_level = -1;
1188 1519
1189 ret = generic_file_aio_write_nolock(iocb, iov, nr_segs, iocb->ki_pos); 1520 direct_io = 0;
1521 sync = 1;
1522 goto relock;
1523 }
1524
1525 if (!sync && ((file->f_flags & O_SYNC) || IS_SYNC(inode)))
1526 sync = 1;
1527
1528 /*
1529 * XXX: Is it ok to execute these checks a second time?
1530 */
1531 ret = generic_write_checks(file, ppos, &count, S_ISBLK(inode->i_mode));
1532 if (ret)
1533 goto out;
1534
1535 /*
1536 * Set pos so that sync_page_range_nolock() below understands
1537 * where to start from. We might've moved it around via the
1538 * calls above. The range we want to actually sync starts from
1539 * *ppos here.
1540 *
1541 */
1542 pos = *ppos;
1543
1544 /* communicate with ocfs2_dio_end_io */
1545 ocfs2_iocb_set_rw_locked(iocb, rw_level);
1546
1547 if (direct_io) {
1548 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
1549 ppos, count, ocount);
1550 if (written < 0) {
1551 ret = written;
1552 goto out_dio;
1553 }
1554 } else {
1555 written = ocfs2_file_buffered_write(file, ppos, iov, nr_segs,
1556 count, written);
1557 if (written < 0) {
1558 ret = written;
1559 if (ret != -EFAULT || ret != -ENOSPC)
1560 mlog_errno(ret);
1561 goto out;
1562 }
1563 }
1190 1564
1565out_dio:
1191 /* buffered aio wouldn't have proper lock coverage today */ 1566 /* buffered aio wouldn't have proper lock coverage today */
1192 BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT)); 1567 BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
1193 1568
1194 /* 1569 /*
1195 * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io 1570 * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
@@ -1207,13 +1582,102 @@ static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1207 } 1582 }
1208 1583
1209out: 1584out:
1585 if (rw_level != -1)
1586 ocfs2_rw_unlock(inode, rw_level);
1587
1588out_sems:
1210 if (have_alloc_sem) 1589 if (have_alloc_sem)
1211 up_read(&inode->i_alloc_sem); 1590 up_read(&inode->i_alloc_sem);
1212 if (rw_level != -1) 1591
1213 ocfs2_rw_unlock(inode, rw_level); 1592 if (written > 0 && sync) {
1593 ssize_t err;
1594
1595 err = sync_page_range_nolock(inode, file->f_mapping, pos, count);
1596 if (err < 0)
1597 written = err;
1598 }
1599
1214 mutex_unlock(&inode->i_mutex); 1600 mutex_unlock(&inode->i_mutex);
1215 1601
1216 mlog_exit(ret); 1602 mlog_exit(ret);
1603 return written ? written : ret;
1604}
1605
1606static int ocfs2_splice_write_actor(struct pipe_inode_info *pipe,
1607 struct pipe_buffer *buf,
1608 struct splice_desc *sd)
1609{
1610 int ret, count, total = 0;
1611 ssize_t copied = 0;
1612 struct ocfs2_splice_write_priv sp;
1613
1614 ret = buf->ops->pin(pipe, buf);
1615 if (ret)
1616 goto out;
1617
1618 sp.s_sd = sd;
1619 sp.s_buf = buf;
1620 sp.s_pipe = pipe;
1621 sp.s_offset = sd->pos & ~PAGE_CACHE_MASK;
1622 sp.s_buf_offset = buf->offset;
1623
1624 count = sd->len;
1625 if (count + sp.s_offset > PAGE_CACHE_SIZE)
1626 count = PAGE_CACHE_SIZE - sp.s_offset;
1627
1628 do {
1629 /*
1630 * splice wants us to copy up to one page at a
1631 * time. For pagesize > cluster size, this means we
1632 * might enter ocfs2_buffered_write_cluster() more
1633 * than once, so keep track of our progress here.
1634 */
1635 copied = ocfs2_buffered_write_cluster(sd->file,
1636 (loff_t)sd->pos + total,
1637 count,
1638 ocfs2_map_and_write_splice_data,
1639 &sp);
1640 if (copied < 0) {
1641 mlog_errno(copied);
1642 ret = copied;
1643 goto out;
1644 }
1645
1646 count -= copied;
1647 sp.s_offset += copied;
1648 sp.s_buf_offset += copied;
1649 total += copied;
1650 } while (count);
1651
1652 ret = 0;
1653out:
1654
1655 return total ? total : ret;
1656}
1657
1658static ssize_t __ocfs2_file_splice_write(struct pipe_inode_info *pipe,
1659 struct file *out,
1660 loff_t *ppos,
1661 size_t len,
1662 unsigned int flags)
1663{
1664 int ret, err;
1665 struct address_space *mapping = out->f_mapping;
1666 struct inode *inode = mapping->host;
1667
1668 ret = __splice_from_pipe(pipe, out, ppos, len, flags,
1669 ocfs2_splice_write_actor);
1670 if (ret > 0) {
1671 *ppos += ret;
1672
1673 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
1674 err = generic_osync_inode(inode, mapping,
1675 OSYNC_METADATA|OSYNC_DATA);
1676 if (err)
1677 ret = err;
1678 }
1679 }
1680
1217 return ret; 1681 return ret;
1218} 1682}
1219 1683
@@ -1239,14 +1703,15 @@ static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
1239 goto out; 1703 goto out;
1240 } 1704 }
1241 1705
1242 ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, ppos, len, 0); 1706 ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, ppos, len, 0,
1707 NULL);
1243 if (ret < 0) { 1708 if (ret < 0) {
1244 mlog_errno(ret); 1709 mlog_errno(ret);
1245 goto out_unlock; 1710 goto out_unlock;
1246 } 1711 }
1247 1712
1248 /* ok, we're done with i_size and alloc work */ 1713 /* ok, we're done with i_size and alloc work */
1249 ret = generic_file_splice_write_nolock(pipe, out, ppos, len, flags); 1714 ret = __ocfs2_file_splice_write(pipe, out, ppos, len, flags);
1250 1715
1251out_unlock: 1716out_unlock:
1252 ocfs2_rw_unlock(inode, 1); 1717 ocfs2_rw_unlock(inode, 1);
@@ -1323,7 +1788,7 @@ static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
1323 } 1788 }
1324 rw_level = 0; 1789 rw_level = 0;
1325 /* communicate with ocfs2_dio_end_io */ 1790 /* communicate with ocfs2_dio_end_io */
1326 ocfs2_iocb_set_rw_locked(iocb); 1791 ocfs2_iocb_set_rw_locked(iocb, rw_level);
1327 } 1792 }
1328 1793
1329 /* 1794 /*
diff --git a/fs/ocfs2/file.h b/fs/ocfs2/file.h
index cc973f01f6ce..2c4460fced52 100644
--- a/fs/ocfs2/file.h
+++ b/fs/ocfs2/file.h
@@ -39,12 +39,17 @@ enum ocfs2_alloc_restarted {
39}; 39};
40int ocfs2_do_extend_allocation(struct ocfs2_super *osb, 40int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
41 struct inode *inode, 41 struct inode *inode,
42 u32 *cluster_start,
42 u32 clusters_to_add, 43 u32 clusters_to_add,
43 struct buffer_head *fe_bh, 44 struct buffer_head *fe_bh,
44 handle_t *handle, 45 handle_t *handle,
45 struct ocfs2_alloc_context *data_ac, 46 struct ocfs2_alloc_context *data_ac,
46 struct ocfs2_alloc_context *meta_ac, 47 struct ocfs2_alloc_context *meta_ac,
47 enum ocfs2_alloc_restarted *reason); 48 enum ocfs2_alloc_restarted *reason);
49int ocfs2_lock_allocators(struct inode *inode, struct ocfs2_dinode *di,
50 u32 clusters_to_add,
51 struct ocfs2_alloc_context **data_ac,
52 struct ocfs2_alloc_context **meta_ac);
48int ocfs2_setattr(struct dentry *dentry, struct iattr *attr); 53int ocfs2_setattr(struct dentry *dentry, struct iattr *attr);
49int ocfs2_getattr(struct vfsmount *mnt, struct dentry *dentry, 54int ocfs2_getattr(struct vfsmount *mnt, struct dentry *dentry,
50 struct kstat *stat); 55 struct kstat *stat);
diff --git a/fs/ocfs2/inode.c b/fs/ocfs2/inode.c
index 28ab56f2b98c..21a605079c62 100644
--- a/fs/ocfs2/inode.c
+++ b/fs/ocfs2/inode.c
@@ -89,24 +89,6 @@ void ocfs2_set_inode_flags(struct inode *inode)
89 inode->i_flags |= S_DIRSYNC; 89 inode->i_flags |= S_DIRSYNC;
90} 90}
91 91
92struct inode *ocfs2_ilookup_for_vote(struct ocfs2_super *osb,
93 u64 blkno,
94 int delete_vote)
95{
96 struct ocfs2_find_inode_args args;
97
98 /* ocfs2_ilookup_for_vote should *only* be called from the
99 * vote thread */
100 BUG_ON(current != osb->vote_task);
101
102 args.fi_blkno = blkno;
103 args.fi_flags = OCFS2_FI_FLAG_NOWAIT;
104 if (delete_vote)
105 args.fi_flags |= OCFS2_FI_FLAG_DELETE;
106 args.fi_ino = ino_from_blkno(osb->sb, blkno);
107 return ilookup5(osb->sb, args.fi_ino, ocfs2_find_actor, &args);
108}
109
110struct inode *ocfs2_iget(struct ocfs2_super *osb, u64 blkno, int flags) 92struct inode *ocfs2_iget(struct ocfs2_super *osb, u64 blkno, int flags)
111{ 93{
112 struct inode *inode = NULL; 94 struct inode *inode = NULL;
@@ -182,28 +164,6 @@ static int ocfs2_find_actor(struct inode *inode, void *opaque)
182 if (oi->ip_blkno != args->fi_blkno) 164 if (oi->ip_blkno != args->fi_blkno)
183 goto bail; 165 goto bail;
184 166
185 /* OCFS2_FI_FLAG_NOWAIT is *only* set from
186 * ocfs2_ilookup_for_vote which won't create an inode for one
187 * that isn't found. The vote thread which doesn't want to get
188 * an inode which is in the process of going away - otherwise
189 * the call to __wait_on_freeing_inode in find_inode_fast will
190 * cause it to deadlock on an inode which may be waiting on a
191 * vote (or lock release) in delete_inode */
192 if ((args->fi_flags & OCFS2_FI_FLAG_NOWAIT) &&
193 (inode->i_state & (I_FREEING|I_CLEAR))) {
194 /* As stated above, we're not going to return an
195 * inode. In the case of a delete vote, the voting
196 * code is going to signal the other node to go
197 * ahead. Mark that state here, so this freeing inode
198 * has the state when it gets to delete_inode. */
199 if (args->fi_flags & OCFS2_FI_FLAG_DELETE) {
200 spin_lock(&oi->ip_lock);
201 ocfs2_mark_inode_remotely_deleted(inode);
202 spin_unlock(&oi->ip_lock);
203 }
204 goto bail;
205 }
206
207 ret = 1; 167 ret = 1;
208bail: 168bail:
209 mlog_exit(ret); 169 mlog_exit(ret);
@@ -261,6 +221,9 @@ int ocfs2_populate_inode(struct inode *inode, struct ocfs2_dinode *fe,
261 goto bail; 221 goto bail;
262 } 222 }
263 223
224 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters);
225 OCFS2_I(inode)->ip_attr = le32_to_cpu(fe->i_attr);
226
264 inode->i_version = 1; 227 inode->i_version = 1;
265 inode->i_generation = le32_to_cpu(fe->i_generation); 228 inode->i_generation = le32_to_cpu(fe->i_generation);
266 inode->i_rdev = huge_decode_dev(le64_to_cpu(fe->id1.dev1.i_rdev)); 229 inode->i_rdev = huge_decode_dev(le64_to_cpu(fe->id1.dev1.i_rdev));
@@ -272,8 +235,7 @@ int ocfs2_populate_inode(struct inode *inode, struct ocfs2_dinode *fe,
272 if (S_ISLNK(inode->i_mode) && !fe->i_clusters) 235 if (S_ISLNK(inode->i_mode) && !fe->i_clusters)
273 inode->i_blocks = 0; 236 inode->i_blocks = 0;
274 else 237 else
275 inode->i_blocks = 238 inode->i_blocks = ocfs2_inode_sector_count(inode);
276 ocfs2_align_bytes_to_sectors(le64_to_cpu(fe->i_size));
277 inode->i_mapping->a_ops = &ocfs2_aops; 239 inode->i_mapping->a_ops = &ocfs2_aops;
278 inode->i_atime.tv_sec = le64_to_cpu(fe->i_atime); 240 inode->i_atime.tv_sec = le64_to_cpu(fe->i_atime);
279 inode->i_atime.tv_nsec = le32_to_cpu(fe->i_atime_nsec); 241 inode->i_atime.tv_nsec = le32_to_cpu(fe->i_atime_nsec);
@@ -288,10 +250,6 @@ int ocfs2_populate_inode(struct inode *inode, struct ocfs2_dinode *fe,
288 (unsigned long long)OCFS2_I(inode)->ip_blkno, 250 (unsigned long long)OCFS2_I(inode)->ip_blkno,
289 (unsigned long long)fe->i_blkno); 251 (unsigned long long)fe->i_blkno);
290 252
291 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters);
292 OCFS2_I(inode)->ip_orphaned_slot = OCFS2_INVALID_SLOT;
293 OCFS2_I(inode)->ip_attr = le32_to_cpu(fe->i_attr);
294
295 inode->i_nlink = le16_to_cpu(fe->i_links_count); 253 inode->i_nlink = le16_to_cpu(fe->i_links_count);
296 254
297 if (fe->i_flags & cpu_to_le32(OCFS2_SYSTEM_FL)) 255 if (fe->i_flags & cpu_to_le32(OCFS2_SYSTEM_FL))
@@ -347,6 +305,9 @@ int ocfs2_populate_inode(struct inode *inode, struct ocfs2_dinode *fe,
347 305
348 ocfs2_inode_lock_res_init(&OCFS2_I(inode)->ip_meta_lockres, 306 ocfs2_inode_lock_res_init(&OCFS2_I(inode)->ip_meta_lockres,
349 OCFS2_LOCK_TYPE_META, 0, inode); 307 OCFS2_LOCK_TYPE_META, 0, inode);
308
309 ocfs2_inode_lock_res_init(&OCFS2_I(inode)->ip_open_lockres,
310 OCFS2_LOCK_TYPE_OPEN, 0, inode);
350 } 311 }
351 312
352 ocfs2_inode_lock_res_init(&OCFS2_I(inode)->ip_rw_lockres, 313 ocfs2_inode_lock_res_init(&OCFS2_I(inode)->ip_rw_lockres,
@@ -421,7 +382,7 @@ static int ocfs2_read_locked_inode(struct inode *inode,
421 * cluster lock before trusting anything anyway. 382 * cluster lock before trusting anything anyway.
422 */ 383 */
423 can_lock = !(args->fi_flags & OCFS2_FI_FLAG_SYSFILE) 384 can_lock = !(args->fi_flags & OCFS2_FI_FLAG_SYSFILE)
424 && !(args->fi_flags & OCFS2_FI_FLAG_NOLOCK) 385 && !(args->fi_flags & OCFS2_FI_FLAG_ORPHAN_RECOVERY)
425 && !ocfs2_mount_local(osb); 386 && !ocfs2_mount_local(osb);
426 387
427 /* 388 /*
@@ -438,7 +399,17 @@ static int ocfs2_read_locked_inode(struct inode *inode,
438 OCFS2_LOCK_TYPE_META, 399 OCFS2_LOCK_TYPE_META,
439 generation, inode); 400 generation, inode);
440 401
402 ocfs2_inode_lock_res_init(&OCFS2_I(inode)->ip_open_lockres,
403 OCFS2_LOCK_TYPE_OPEN,
404 0, inode);
405
441 if (can_lock) { 406 if (can_lock) {
407 status = ocfs2_open_lock(inode);
408 if (status) {
409 make_bad_inode(inode);
410 mlog_errno(status);
411 return status;
412 }
442 status = ocfs2_meta_lock(inode, NULL, 0); 413 status = ocfs2_meta_lock(inode, NULL, 0);
443 if (status) { 414 if (status) {
444 make_bad_inode(inode); 415 make_bad_inode(inode);
@@ -447,6 +418,14 @@ static int ocfs2_read_locked_inode(struct inode *inode,
447 } 418 }
448 } 419 }
449 420
421 if (args->fi_flags & OCFS2_FI_FLAG_ORPHAN_RECOVERY) {
422 status = ocfs2_try_open_lock(inode, 0);
423 if (status) {
424 make_bad_inode(inode);
425 return status;
426 }
427 }
428
450 status = ocfs2_read_block(osb, args->fi_blkno, &bh, 0, 429 status = ocfs2_read_block(osb, args->fi_blkno, &bh, 0,
451 can_lock ? inode : NULL); 430 can_lock ? inode : NULL);
452 if (status < 0) { 431 if (status < 0) {
@@ -507,50 +486,56 @@ static int ocfs2_truncate_for_delete(struct ocfs2_super *osb,
507 struct buffer_head *fe_bh) 486 struct buffer_head *fe_bh)
508{ 487{
509 int status = 0; 488 int status = 0;
510 handle_t *handle = NULL;
511 struct ocfs2_truncate_context *tc = NULL; 489 struct ocfs2_truncate_context *tc = NULL;
512 struct ocfs2_dinode *fe; 490 struct ocfs2_dinode *fe;
491 handle_t *handle = NULL;
513 492
514 mlog_entry_void(); 493 mlog_entry_void();
515 494
516 fe = (struct ocfs2_dinode *) fe_bh->b_data; 495 fe = (struct ocfs2_dinode *) fe_bh->b_data;
517 496
518 /* zero allocation, zero truncate :) */ 497 if (fe->i_clusters) {
519 if (!fe->i_clusters) 498 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
520 goto bail; 499 if (IS_ERR(handle)) {
500 status = PTR_ERR(handle);
501 mlog_errno(status);
502 goto out;
503 }
521 504
522 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS); 505 status = ocfs2_journal_access(handle, inode, fe_bh,
523 if (IS_ERR(handle)) { 506 OCFS2_JOURNAL_ACCESS_WRITE);
524 status = PTR_ERR(handle); 507 if (status < 0) {
525 handle = NULL; 508 mlog_errno(status);
526 mlog_errno(status); 509 goto out;
527 goto bail; 510 }
528 }
529 511
530 status = ocfs2_set_inode_size(handle, inode, fe_bh, 0ULL); 512 i_size_write(inode, 0);
531 if (status < 0) {
532 mlog_errno(status);
533 goto bail;
534 }
535 513
536 ocfs2_commit_trans(osb, handle); 514 status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
537 handle = NULL; 515 if (status < 0) {
516 mlog_errno(status);
517 goto out;
518 }
538 519
539 status = ocfs2_prepare_truncate(osb, inode, fe_bh, &tc); 520 ocfs2_commit_trans(osb, handle);
540 if (status < 0) { 521 handle = NULL;
541 mlog_errno(status);
542 goto bail;
543 }
544 522
545 status = ocfs2_commit_truncate(osb, inode, fe_bh, tc); 523 status = ocfs2_prepare_truncate(osb, inode, fe_bh, &tc);
546 if (status < 0) { 524 if (status < 0) {
547 mlog_errno(status); 525 mlog_errno(status);
548 goto bail; 526 goto out;
527 }
528
529 status = ocfs2_commit_truncate(osb, inode, fe_bh, tc);
530 if (status < 0) {
531 mlog_errno(status);
532 goto out;
533 }
549 } 534 }
550bail: 535
536out:
551 if (handle) 537 if (handle)
552 ocfs2_commit_trans(osb, handle); 538 ocfs2_commit_trans(osb, handle);
553
554 mlog_exit(status); 539 mlog_exit(status);
555 return status; 540 return status;
556} 541}
@@ -678,10 +663,10 @@ static int ocfs2_wipe_inode(struct inode *inode,
678 struct inode *orphan_dir_inode = NULL; 663 struct inode *orphan_dir_inode = NULL;
679 struct buffer_head *orphan_dir_bh = NULL; 664 struct buffer_head *orphan_dir_bh = NULL;
680 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 665 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
666 struct ocfs2_dinode *di;
681 667
682 /* We've already voted on this so it should be readonly - no 668 di = (struct ocfs2_dinode *) di_bh->b_data;
683 * spinlock needed. */ 669 orphaned_slot = le16_to_cpu(di->i_orphaned_slot);
684 orphaned_slot = OCFS2_I(inode)->ip_orphaned_slot;
685 670
686 status = ocfs2_check_orphan_recovery_state(osb, orphaned_slot); 671 status = ocfs2_check_orphan_recovery_state(osb, orphaned_slot);
687 if (status) 672 if (status)
@@ -839,11 +824,20 @@ static int ocfs2_query_inode_wipe(struct inode *inode,
839 goto bail; 824 goto bail;
840 } 825 }
841 826
842 status = ocfs2_request_delete_vote(inode); 827 /*
843 /* -EBUSY means that other nodes are still using the 828 * This is how ocfs2 determines whether an inode is still live
844 * inode. We're done here though, so avoid doing anything on 829 * within the cluster. Every node takes a shared read lock on
845 * disk and let them worry about deleting it. */ 830 * the inode open lock in ocfs2_read_locked_inode(). When we
846 if (status == -EBUSY) { 831 * get to ->delete_inode(), each node tries to convert it's
832 * lock to an exclusive. Trylocks are serialized by the inode
833 * meta data lock. If the upconvert suceeds, we know the inode
834 * is no longer live and can be deleted.
835 *
836 * Though we call this with the meta data lock held, the
837 * trylock keeps us from ABBA deadlock.
838 */
839 status = ocfs2_try_open_lock(inode, 1);
840 if (status == -EAGAIN) {
847 status = 0; 841 status = 0;
848 mlog(0, "Skipping delete of %llu because it is in use on" 842 mlog(0, "Skipping delete of %llu because it is in use on"
849 "other nodes\n", (unsigned long long)oi->ip_blkno); 843 "other nodes\n", (unsigned long long)oi->ip_blkno);
@@ -854,21 +848,10 @@ static int ocfs2_query_inode_wipe(struct inode *inode,
854 goto bail; 848 goto bail;
855 } 849 }
856 850
857 spin_lock(&oi->ip_lock); 851 *wipe = 1;
858 if (oi->ip_orphaned_slot == OCFS2_INVALID_SLOT) { 852 mlog(0, "Inode %llu is ok to wipe from orphan dir %u\n",
859 /* Nobody knew which slot this inode was orphaned 853 (unsigned long long)oi->ip_blkno,
860 * into. This may happen during node death and 854 le16_to_cpu(di->i_orphaned_slot));
861 * recovery knows how to clean it up so we can safely
862 * ignore this inode for now on. */
863 mlog(0, "Nobody knew where inode %llu was orphaned!\n",
864 (unsigned long long)oi->ip_blkno);
865 } else {
866 *wipe = 1;
867
868 mlog(0, "Inode %llu is ok to wipe from orphan dir %d\n",
869 (unsigned long long)oi->ip_blkno, oi->ip_orphaned_slot);
870 }
871 spin_unlock(&oi->ip_lock);
872 855
873bail: 856bail:
874 return status; 857 return status;
@@ -1001,11 +984,16 @@ void ocfs2_clear_inode(struct inode *inode)
1001 mlog_bug_on_msg(OCFS2_SB(inode->i_sb) == NULL, 984 mlog_bug_on_msg(OCFS2_SB(inode->i_sb) == NULL,
1002 "Inode=%lu\n", inode->i_ino); 985 "Inode=%lu\n", inode->i_ino);
1003 986
987 /* For remove delete_inode vote, we hold open lock before,
988 * now it is time to unlock PR and EX open locks. */
989 ocfs2_open_unlock(inode);
990
1004 /* Do these before all the other work so that we don't bounce 991 /* Do these before all the other work so that we don't bounce
1005 * the vote thread while waiting to destroy the locks. */ 992 * the vote thread while waiting to destroy the locks. */
1006 ocfs2_mark_lockres_freeing(&oi->ip_rw_lockres); 993 ocfs2_mark_lockres_freeing(&oi->ip_rw_lockres);
1007 ocfs2_mark_lockres_freeing(&oi->ip_meta_lockres); 994 ocfs2_mark_lockres_freeing(&oi->ip_meta_lockres);
1008 ocfs2_mark_lockres_freeing(&oi->ip_data_lockres); 995 ocfs2_mark_lockres_freeing(&oi->ip_data_lockres);
996 ocfs2_mark_lockres_freeing(&oi->ip_open_lockres);
1009 997
1010 /* We very well may get a clear_inode before all an inodes 998 /* We very well may get a clear_inode before all an inodes
1011 * metadata has hit disk. Of course, we can't drop any cluster 999 * metadata has hit disk. Of course, we can't drop any cluster
@@ -1020,8 +1008,7 @@ void ocfs2_clear_inode(struct inode *inode)
1020 "Clear inode of %llu, inode has io markers\n", 1008 "Clear inode of %llu, inode has io markers\n",
1021 (unsigned long long)oi->ip_blkno); 1009 (unsigned long long)oi->ip_blkno);
1022 1010
1023 ocfs2_extent_map_drop(inode, 0); 1011 ocfs2_extent_map_trunc(inode, 0);
1024 ocfs2_extent_map_init(inode);
1025 1012
1026 status = ocfs2_drop_inode_locks(inode); 1013 status = ocfs2_drop_inode_locks(inode);
1027 if (status < 0) 1014 if (status < 0)
@@ -1030,6 +1017,7 @@ void ocfs2_clear_inode(struct inode *inode)
1030 ocfs2_lock_res_free(&oi->ip_rw_lockres); 1017 ocfs2_lock_res_free(&oi->ip_rw_lockres);
1031 ocfs2_lock_res_free(&oi->ip_meta_lockres); 1018 ocfs2_lock_res_free(&oi->ip_meta_lockres);
1032 ocfs2_lock_res_free(&oi->ip_data_lockres); 1019 ocfs2_lock_res_free(&oi->ip_data_lockres);
1020 ocfs2_lock_res_free(&oi->ip_open_lockres);
1033 1021
1034 ocfs2_metadata_cache_purge(inode); 1022 ocfs2_metadata_cache_purge(inode);
1035 1023
@@ -1086,9 +1074,6 @@ void ocfs2_drop_inode(struct inode *inode)
1086 mlog(0, "Drop inode %llu, nlink = %u, ip_flags = 0x%x\n", 1074 mlog(0, "Drop inode %llu, nlink = %u, ip_flags = 0x%x\n",
1087 (unsigned long long)oi->ip_blkno, inode->i_nlink, oi->ip_flags); 1075 (unsigned long long)oi->ip_blkno, inode->i_nlink, oi->ip_flags);
1088 1076
1089 /* Testing ip_orphaned_slot here wouldn't work because we may
1090 * not have gotten a delete_inode vote from any other nodes
1091 * yet. */
1092 if (oi->ip_flags & OCFS2_INODE_MAYBE_ORPHANED) 1077 if (oi->ip_flags & OCFS2_INODE_MAYBE_ORPHANED)
1093 generic_delete_inode(inode); 1078 generic_delete_inode(inode);
1094 else 1079 else
@@ -1121,8 +1106,8 @@ struct buffer_head *ocfs2_bread(struct inode *inode,
1121 return NULL; 1106 return NULL;
1122 } 1107 }
1123 1108
1124 tmperr = ocfs2_extent_map_get_blocks(inode, block, 1, 1109 tmperr = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
1125 &p_blkno, NULL); 1110 NULL);
1126 if (tmperr < 0) { 1111 if (tmperr < 0) {
1127 mlog_errno(tmperr); 1112 mlog_errno(tmperr);
1128 goto fail; 1113 goto fail;
@@ -1259,7 +1244,7 @@ void ocfs2_refresh_inode(struct inode *inode,
1259 if (S_ISLNK(inode->i_mode) && le32_to_cpu(fe->i_clusters) == 0) 1244 if (S_ISLNK(inode->i_mode) && le32_to_cpu(fe->i_clusters) == 0)
1260 inode->i_blocks = 0; 1245 inode->i_blocks = 0;
1261 else 1246 else
1262 inode->i_blocks = ocfs2_align_bytes_to_sectors(i_size_read(inode)); 1247 inode->i_blocks = ocfs2_inode_sector_count(inode);
1263 inode->i_atime.tv_sec = le64_to_cpu(fe->i_atime); 1248 inode->i_atime.tv_sec = le64_to_cpu(fe->i_atime);
1264 inode->i_atime.tv_nsec = le32_to_cpu(fe->i_atime_nsec); 1249 inode->i_atime.tv_nsec = le32_to_cpu(fe->i_atime_nsec);
1265 inode->i_mtime.tv_sec = le64_to_cpu(fe->i_mtime); 1250 inode->i_mtime.tv_sec = le64_to_cpu(fe->i_mtime);
diff --git a/fs/ocfs2/inode.h b/fs/ocfs2/inode.h
index 1a7dd2945b34..03ae075869ee 100644
--- a/fs/ocfs2/inode.h
+++ b/fs/ocfs2/inode.h
@@ -26,6 +26,8 @@
26#ifndef OCFS2_INODE_H 26#ifndef OCFS2_INODE_H
27#define OCFS2_INODE_H 27#define OCFS2_INODE_H
28 28
29#include "extent_map.h"
30
29/* OCFS2 Inode Private Data */ 31/* OCFS2 Inode Private Data */
30struct ocfs2_inode_info 32struct ocfs2_inode_info
31{ 33{
@@ -34,6 +36,7 @@ struct ocfs2_inode_info
34 struct ocfs2_lock_res ip_rw_lockres; 36 struct ocfs2_lock_res ip_rw_lockres;
35 struct ocfs2_lock_res ip_meta_lockres; 37 struct ocfs2_lock_res ip_meta_lockres;
36 struct ocfs2_lock_res ip_data_lockres; 38 struct ocfs2_lock_res ip_data_lockres;
39 struct ocfs2_lock_res ip_open_lockres;
37 40
38 /* protects allocation changes on this inode. */ 41 /* protects allocation changes on this inode. */
39 struct rw_semaphore ip_alloc_sem; 42 struct rw_semaphore ip_alloc_sem;
@@ -42,9 +45,7 @@ struct ocfs2_inode_info
42 spinlock_t ip_lock; 45 spinlock_t ip_lock;
43 u32 ip_open_count; 46 u32 ip_open_count;
44 u32 ip_clusters; 47 u32 ip_clusters;
45 struct ocfs2_extent_map ip_map;
46 struct list_head ip_io_markers; 48 struct list_head ip_io_markers;
47 int ip_orphaned_slot;
48 49
49 struct mutex ip_io_mutex; 50 struct mutex ip_io_mutex;
50 51
@@ -64,6 +65,8 @@ struct ocfs2_inode_info
64 65
65 struct ocfs2_caching_info ip_metadata_cache; 66 struct ocfs2_caching_info ip_metadata_cache;
66 67
68 struct ocfs2_extent_map ip_extent_map;
69
67 struct inode vfs_inode; 70 struct inode vfs_inode;
68}; 71};
69 72
@@ -117,14 +120,9 @@ void ocfs2_delete_inode(struct inode *inode);
117void ocfs2_drop_inode(struct inode *inode); 120void ocfs2_drop_inode(struct inode *inode);
118 121
119/* Flags for ocfs2_iget() */ 122/* Flags for ocfs2_iget() */
120#define OCFS2_FI_FLAG_NOWAIT 0x1 123#define OCFS2_FI_FLAG_SYSFILE 0x4
121#define OCFS2_FI_FLAG_DELETE 0x2 124#define OCFS2_FI_FLAG_ORPHAN_RECOVERY 0x8
122#define OCFS2_FI_FLAG_SYSFILE 0x4
123#define OCFS2_FI_FLAG_NOLOCK 0x8
124struct inode *ocfs2_iget(struct ocfs2_super *osb, u64 feoff, int flags); 125struct inode *ocfs2_iget(struct ocfs2_super *osb, u64 feoff, int flags);
125struct inode *ocfs2_ilookup_for_vote(struct ocfs2_super *osb,
126 u64 blkno,
127 int delete_vote);
128int ocfs2_inode_init_private(struct inode *inode); 126int ocfs2_inode_init_private(struct inode *inode);
129int ocfs2_inode_revalidate(struct dentry *dentry); 127int ocfs2_inode_revalidate(struct dentry *dentry);
130int ocfs2_populate_inode(struct inode *inode, struct ocfs2_dinode *fe, 128int ocfs2_populate_inode(struct inode *inode, struct ocfs2_dinode *fe,
@@ -144,4 +142,11 @@ int ocfs2_aio_write(struct file *file, struct kiocb *req, struct iocb *iocb);
144 142
145void ocfs2_set_inode_flags(struct inode *inode); 143void ocfs2_set_inode_flags(struct inode *inode);
146 144
145static inline blkcnt_t ocfs2_inode_sector_count(struct inode *inode)
146{
147 int c_to_s_bits = OCFS2_SB(inode->i_sb)->s_clustersize_bits - 9;
148
149 return (blkcnt_t)(OCFS2_I(inode)->ip_clusters << c_to_s_bits);
150}
151
147#endif /* OCFS2_INODE_H */ 152#endif /* OCFS2_INODE_H */
diff --git a/fs/ocfs2/journal.c b/fs/ocfs2/journal.c
index 825cb0ae1b4c..5a8a90d1c787 100644
--- a/fs/ocfs2/journal.c
+++ b/fs/ocfs2/journal.c
@@ -649,29 +649,20 @@ bail:
649static int ocfs2_force_read_journal(struct inode *inode) 649static int ocfs2_force_read_journal(struct inode *inode)
650{ 650{
651 int status = 0; 651 int status = 0;
652 int i, p_blocks; 652 int i;
653 u64 v_blkno, p_blkno; 653 u64 v_blkno, p_blkno, p_blocks, num_blocks;
654#define CONCURRENT_JOURNAL_FILL 32 654#define CONCURRENT_JOURNAL_FILL 32ULL
655 struct buffer_head *bhs[CONCURRENT_JOURNAL_FILL]; 655 struct buffer_head *bhs[CONCURRENT_JOURNAL_FILL];
656 656
657 mlog_entry_void(); 657 mlog_entry_void();
658 658
659 BUG_ON(inode->i_blocks !=
660 ocfs2_align_bytes_to_sectors(i_size_read(inode)));
661
662 memset(bhs, 0, sizeof(struct buffer_head *) * CONCURRENT_JOURNAL_FILL); 659 memset(bhs, 0, sizeof(struct buffer_head *) * CONCURRENT_JOURNAL_FILL);
663 660
664 mlog(0, "Force reading %llu blocks\n", 661 num_blocks = ocfs2_blocks_for_bytes(inode->i_sb, inode->i_size);
665 (unsigned long long)(inode->i_blocks >>
666 (inode->i_sb->s_blocksize_bits - 9)));
667
668 v_blkno = 0; 662 v_blkno = 0;
669 while (v_blkno < 663 while (v_blkno < num_blocks) {
670 (inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9))) {
671
672 status = ocfs2_extent_map_get_blocks(inode, v_blkno, 664 status = ocfs2_extent_map_get_blocks(inode, v_blkno,
673 1, &p_blkno, 665 &p_blkno, &p_blocks, NULL);
674 &p_blocks);
675 if (status < 0) { 666 if (status < 0) {
676 mlog_errno(status); 667 mlog_errno(status);
677 goto bail; 668 goto bail;
@@ -1306,7 +1297,7 @@ static int ocfs2_queue_orphans(struct ocfs2_super *osb,
1306 continue; 1297 continue;
1307 1298
1308 iter = ocfs2_iget(osb, le64_to_cpu(de->inode), 1299 iter = ocfs2_iget(osb, le64_to_cpu(de->inode),
1309 OCFS2_FI_FLAG_NOLOCK); 1300 OCFS2_FI_FLAG_ORPHAN_RECOVERY);
1310 if (IS_ERR(iter)) 1301 if (IS_ERR(iter))
1311 continue; 1302 continue;
1312 1303
@@ -1418,7 +1409,6 @@ static int ocfs2_recover_orphans(struct ocfs2_super *osb,
1418 /* Set the proper information to get us going into 1409 /* Set the proper information to get us going into
1419 * ocfs2_delete_inode. */ 1410 * ocfs2_delete_inode. */
1420 oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED; 1411 oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
1421 oi->ip_orphaned_slot = slot;
1422 spin_unlock(&oi->ip_lock); 1412 spin_unlock(&oi->ip_lock);
1423 1413
1424 iput(inode); 1414 iput(inode);
diff --git a/fs/ocfs2/journal.h b/fs/ocfs2/journal.h
index d026b4f27757..3db5de4506da 100644
--- a/fs/ocfs2/journal.h
+++ b/fs/ocfs2/journal.h
@@ -390,7 +390,7 @@ static inline int ocfs2_calc_tree_trunc_credits(struct super_block *sb,
390 /* We may be deleting metadata blocks, so metadata alloc dinode + 390 /* We may be deleting metadata blocks, so metadata alloc dinode +
391 one desc. block for each possible delete. */ 391 one desc. block for each possible delete. */
392 if (tree_depth && next_free == 1 && 392 if (tree_depth && next_free == 1 &&
393 le32_to_cpu(last_el->l_recs[i].e_clusters) == clusters_to_del) 393 ocfs2_rec_clusters(last_el, &last_el->l_recs[i]) == clusters_to_del)
394 credits += 1 + tree_depth; 394 credits += 1 + tree_depth;
395 395
396 /* update to the truncate log. */ 396 /* update to the truncate log. */
diff --git a/fs/ocfs2/mmap.c b/fs/ocfs2/mmap.c
index 51b020447683..af01158b39f5 100644
--- a/fs/ocfs2/mmap.c
+++ b/fs/ocfs2/mmap.c
@@ -85,8 +85,11 @@ int ocfs2_mmap(struct file *file, struct vm_area_struct *vma)
85 int ret = 0, lock_level = 0; 85 int ret = 0, lock_level = 0;
86 struct ocfs2_super *osb = OCFS2_SB(file->f_dentry->d_inode->i_sb); 86 struct ocfs2_super *osb = OCFS2_SB(file->f_dentry->d_inode->i_sb);
87 87
88 /* We don't want to support shared writable mappings yet. */ 88 /*
89 if (!ocfs2_mount_local(osb) && 89 * Only support shared writeable mmap for local mounts which
90 * don't know about holes.
91 */
92 if ((!ocfs2_mount_local(osb) || ocfs2_sparse_alloc(osb)) &&
90 ((vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_MAYSHARE)) && 93 ((vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_MAYSHARE)) &&
91 ((vma->vm_flags & VM_WRITE) || (vma->vm_flags & VM_MAYWRITE))) { 94 ((vma->vm_flags & VM_WRITE) || (vma->vm_flags & VM_MAYWRITE))) {
92 mlog(0, "disallow shared writable mmaps %lx\n", vma->vm_flags); 95 mlog(0, "disallow shared writable mmaps %lx\n", vma->vm_flags);
diff --git a/fs/ocfs2/namei.c b/fs/ocfs2/namei.c
index 28dd757ff67d..2bcf353fd7c5 100644
--- a/fs/ocfs2/namei.c
+++ b/fs/ocfs2/namei.c
@@ -175,8 +175,6 @@ static struct dentry *ocfs2_lookup(struct inode *dir, struct dentry *dentry,
175 175
176 inode = ocfs2_iget(OCFS2_SB(dir->i_sb), blkno, 0); 176 inode = ocfs2_iget(OCFS2_SB(dir->i_sb), blkno, 0);
177 if (IS_ERR(inode)) { 177 if (IS_ERR(inode)) {
178 mlog(ML_ERROR, "Unable to create inode %llu\n",
179 (unsigned long long)blkno);
180 ret = ERR_PTR(-EACCES); 178 ret = ERR_PTR(-EACCES);
181 goto bail_unlock; 179 goto bail_unlock;
182 } 180 }
@@ -189,7 +187,6 @@ static struct dentry *ocfs2_lookup(struct inode *dir, struct dentry *dentry,
189 * unlink. */ 187 * unlink. */
190 spin_lock(&oi->ip_lock); 188 spin_lock(&oi->ip_lock);
191 oi->ip_flags &= ~OCFS2_INODE_MAYBE_ORPHANED; 189 oi->ip_flags &= ~OCFS2_INODE_MAYBE_ORPHANED;
192 oi->ip_orphaned_slot = OCFS2_INVALID_SLOT;
193 spin_unlock(&oi->ip_lock); 190 spin_unlock(&oi->ip_lock);
194 191
195bail_add: 192bail_add:
@@ -288,7 +285,7 @@ static int ocfs2_fill_new_dir(struct ocfs2_super *osb,
288 285
289 i_size_write(inode, inode->i_sb->s_blocksize); 286 i_size_write(inode, inode->i_sb->s_blocksize);
290 inode->i_nlink = 2; 287 inode->i_nlink = 2;
291 inode->i_blocks = ocfs2_align_bytes_to_sectors(inode->i_sb->s_blocksize); 288 inode->i_blocks = ocfs2_inode_sector_count(inode);
292 status = ocfs2_mark_inode_dirty(handle, inode, fe_bh); 289 status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
293 if (status < 0) { 290 if (status < 0) {
294 mlog_errno(status); 291 mlog_errno(status);
@@ -1486,8 +1483,7 @@ static int ocfs2_create_symlink_data(struct ocfs2_super *osb,
1486 struct buffer_head **bhs = NULL; 1483 struct buffer_head **bhs = NULL;
1487 const char *c; 1484 const char *c;
1488 struct super_block *sb = osb->sb; 1485 struct super_block *sb = osb->sb;
1489 u64 p_blkno; 1486 u64 p_blkno, p_blocks;
1490 int p_blocks;
1491 int virtual, blocks, status, i, bytes_left; 1487 int virtual, blocks, status, i, bytes_left;
1492 1488
1493 bytes_left = i_size_read(inode) + 1; 1489 bytes_left = i_size_read(inode) + 1;
@@ -1514,8 +1510,8 @@ static int ocfs2_create_symlink_data(struct ocfs2_super *osb,
1514 goto bail; 1510 goto bail;
1515 } 1511 }
1516 1512
1517 status = ocfs2_extent_map_get_blocks(inode, 0, 1, &p_blkno, 1513 status = ocfs2_extent_map_get_blocks(inode, 0, &p_blkno, &p_blocks,
1518 &p_blocks); 1514 NULL);
1519 if (status < 0) { 1515 if (status < 0) {
1520 mlog_errno(status); 1516 mlog_errno(status);
1521 goto bail; 1517 goto bail;
@@ -1674,8 +1670,11 @@ static int ocfs2_symlink(struct inode *dir,
1674 inode->i_rdev = 0; 1670 inode->i_rdev = 0;
1675 newsize = l - 1; 1671 newsize = l - 1;
1676 if (l > ocfs2_fast_symlink_chars(sb)) { 1672 if (l > ocfs2_fast_symlink_chars(sb)) {
1673 u32 offset = 0;
1674
1677 inode->i_op = &ocfs2_symlink_inode_operations; 1675 inode->i_op = &ocfs2_symlink_inode_operations;
1678 status = ocfs2_do_extend_allocation(osb, inode, 1, new_fe_bh, 1676 status = ocfs2_do_extend_allocation(osb, inode, &offset, 1,
1677 new_fe_bh,
1679 handle, data_ac, NULL, 1678 handle, data_ac, NULL,
1680 NULL); 1679 NULL);
1681 if (status < 0) { 1680 if (status < 0) {
@@ -1689,7 +1688,7 @@ static int ocfs2_symlink(struct inode *dir,
1689 goto bail; 1688 goto bail;
1690 } 1689 }
1691 i_size_write(inode, newsize); 1690 i_size_write(inode, newsize);
1692 inode->i_blocks = ocfs2_align_bytes_to_sectors(newsize); 1691 inode->i_blocks = ocfs2_inode_sector_count(inode);
1693 } else { 1692 } else {
1694 inode->i_op = &ocfs2_fast_symlink_inode_operations; 1693 inode->i_op = &ocfs2_fast_symlink_inode_operations;
1695 memcpy((char *) fe->id2.i_symlink, symname, l); 1694 memcpy((char *) fe->id2.i_symlink, symname, l);
@@ -2222,9 +2221,7 @@ static int ocfs2_orphan_add(struct ocfs2_super *osb,
2222 /* Record which orphan dir our inode now resides 2221 /* Record which orphan dir our inode now resides
2223 * in. delete_inode will use this to determine which orphan 2222 * in. delete_inode will use this to determine which orphan
2224 * dir to lock. */ 2223 * dir to lock. */
2225 spin_lock(&OCFS2_I(inode)->ip_lock); 2224 fe->i_orphaned_slot = cpu_to_le16(osb->slot_num);
2226 OCFS2_I(inode)->ip_orphaned_slot = osb->slot_num;
2227 spin_unlock(&OCFS2_I(inode)->ip_lock);
2228 2225
2229 mlog(0, "Inode %llu orphaned in slot %d\n", 2226 mlog(0, "Inode %llu orphaned in slot %d\n",
2230 (unsigned long long)OCFS2_I(inode)->ip_blkno, osb->slot_num); 2227 (unsigned long long)OCFS2_I(inode)->ip_blkno, osb->slot_num);
diff --git a/fs/ocfs2/ocfs2.h b/fs/ocfs2/ocfs2.h
index db8e77cd35d3..82cc92dcf8a6 100644
--- a/fs/ocfs2/ocfs2.h
+++ b/fs/ocfs2/ocfs2.h
@@ -46,11 +46,6 @@
46#include "endian.h" 46#include "endian.h"
47#include "ocfs2_lockid.h" 47#include "ocfs2_lockid.h"
48 48
49struct ocfs2_extent_map {
50 u32 em_clusters;
51 struct rb_root em_extents;
52};
53
54/* Most user visible OCFS2 inodes will have very few pieces of 49/* Most user visible OCFS2 inodes will have very few pieces of
55 * metadata, but larger files (including bitmaps, etc) must be taken 50 * metadata, but larger files (including bitmaps, etc) must be taken
56 * into account when designing an access scheme. We allow a small 51 * into account when designing an access scheme. We allow a small
@@ -303,6 +298,13 @@ static inline int ocfs2_should_order_data(struct inode *inode)
303 return 1; 298 return 1;
304} 299}
305 300
301static inline int ocfs2_sparse_alloc(struct ocfs2_super *osb)
302{
303 if (osb->s_feature_incompat & OCFS2_FEATURE_INCOMPAT_SPARSE_ALLOC)
304 return 1;
305 return 0;
306}
307
306/* set / clear functions because cluster events can make these happen 308/* set / clear functions because cluster events can make these happen
307 * in parallel so we want the transitions to be atomic. this also 309 * in parallel so we want the transitions to be atomic. this also
308 * means that any future flags osb_flags must be protected by spinlock 310 * means that any future flags osb_flags must be protected by spinlock
@@ -461,6 +463,49 @@ static inline unsigned long ocfs2_align_bytes_to_sectors(u64 bytes)
461 return (unsigned long)((bytes + 511) >> 9); 463 return (unsigned long)((bytes + 511) >> 9);
462} 464}
463 465
466static inline unsigned int ocfs2_page_index_to_clusters(struct super_block *sb,
467 unsigned long pg_index)
468{
469 u32 clusters = pg_index;
470 unsigned int cbits = OCFS2_SB(sb)->s_clustersize_bits;
471
472 if (unlikely(PAGE_CACHE_SHIFT > cbits))
473 clusters = pg_index << (PAGE_CACHE_SHIFT - cbits);
474 else if (PAGE_CACHE_SHIFT < cbits)
475 clusters = pg_index >> (cbits - PAGE_CACHE_SHIFT);
476
477 return clusters;
478}
479
480/*
481 * Find the 1st page index which covers the given clusters.
482 */
483static inline unsigned long ocfs2_align_clusters_to_page_index(struct super_block *sb,
484 u32 clusters)
485{
486 unsigned int cbits = OCFS2_SB(sb)->s_clustersize_bits;
487 unsigned long index = clusters;
488
489 if (PAGE_CACHE_SHIFT > cbits) {
490 index = clusters >> (PAGE_CACHE_SHIFT - cbits);
491 } else if (PAGE_CACHE_SHIFT < cbits) {
492 index = clusters << (cbits - PAGE_CACHE_SHIFT);
493 }
494
495 return index;
496}
497
498static inline unsigned int ocfs2_pages_per_cluster(struct super_block *sb)
499{
500 unsigned int cbits = OCFS2_SB(sb)->s_clustersize_bits;
501 unsigned int pages_per_cluster = 1;
502
503 if (PAGE_CACHE_SHIFT < cbits)
504 pages_per_cluster = 1 << (cbits - PAGE_CACHE_SHIFT);
505
506 return pages_per_cluster;
507}
508
464#define ocfs2_set_bit ext2_set_bit 509#define ocfs2_set_bit ext2_set_bit
465#define ocfs2_clear_bit ext2_clear_bit 510#define ocfs2_clear_bit ext2_clear_bit
466#define ocfs2_test_bit ext2_test_bit 511#define ocfs2_test_bit ext2_test_bit
diff --git a/fs/ocfs2/ocfs2_fs.h b/fs/ocfs2/ocfs2_fs.h
index e61e218f5e0b..71306479c68f 100644
--- a/fs/ocfs2/ocfs2_fs.h
+++ b/fs/ocfs2/ocfs2_fs.h
@@ -86,7 +86,8 @@
86 OCFS2_SB(sb)->s_feature_incompat &= ~(mask) 86 OCFS2_SB(sb)->s_feature_incompat &= ~(mask)
87 87
88#define OCFS2_FEATURE_COMPAT_SUPP OCFS2_FEATURE_COMPAT_BACKUP_SB 88#define OCFS2_FEATURE_COMPAT_SUPP OCFS2_FEATURE_COMPAT_BACKUP_SB
89#define OCFS2_FEATURE_INCOMPAT_SUPP OCFS2_FEATURE_INCOMPAT_LOCAL_MOUNT 89#define OCFS2_FEATURE_INCOMPAT_SUPP (OCFS2_FEATURE_INCOMPAT_LOCAL_MOUNT \
90 | OCFS2_FEATURE_INCOMPAT_SPARSE_ALLOC)
90#define OCFS2_FEATURE_RO_COMPAT_SUPP 0 91#define OCFS2_FEATURE_RO_COMPAT_SUPP 0
91 92
92/* 93/*
@@ -155,6 +156,12 @@
155#define OCFS2_FL_MODIFIABLE (0x000100FF) /* User modifiable flags */ 156#define OCFS2_FL_MODIFIABLE (0x000100FF) /* User modifiable flags */
156 157
157/* 158/*
159 * Extent record flags (e_node.leaf.flags)
160 */
161#define OCFS2_EXT_UNWRITTEN (0x01) /* Extent is allocated but
162 * unwritten */
163
164/*
158 * ioctl commands 165 * ioctl commands
159 */ 166 */
160#define OCFS2_IOC_GETFLAGS _IOR('f', 1, long) 167#define OCFS2_IOC_GETFLAGS _IOR('f', 1, long)
@@ -282,10 +289,21 @@ static unsigned char ocfs2_type_by_mode[S_IFMT >> S_SHIFT] = {
282/* 289/*
283 * On disk extent record for OCFS2 290 * On disk extent record for OCFS2
284 * It describes a range of clusters on disk. 291 * It describes a range of clusters on disk.
292 *
293 * Length fields are divided into interior and leaf node versions.
294 * This leaves room for a flags field (OCFS2_EXT_*) in the leaf nodes.
285 */ 295 */
286struct ocfs2_extent_rec { 296struct ocfs2_extent_rec {
287/*00*/ __le32 e_cpos; /* Offset into the file, in clusters */ 297/*00*/ __le32 e_cpos; /* Offset into the file, in clusters */
288 __le32 e_clusters; /* Clusters covered by this extent */ 298 union {
299 __le32 e_int_clusters; /* Clusters covered by all children */
300 struct {
301 __le16 e_leaf_clusters; /* Clusters covered by this
302 extent */
303 __u8 e_reserved1;
304 __u8 e_flags; /* Extent flags */
305 };
306 };
289 __le64 e_blkno; /* Physical disk offset, in blocks */ 307 __le64 e_blkno; /* Physical disk offset, in blocks */
290/*10*/ 308/*10*/
291}; 309};
@@ -311,7 +329,10 @@ struct ocfs2_extent_list {
311/*00*/ __le16 l_tree_depth; /* Extent tree depth from this 329/*00*/ __le16 l_tree_depth; /* Extent tree depth from this
312 point. 0 means data extents 330 point. 0 means data extents
313 hang directly off this 331 hang directly off this
314 header (a leaf) */ 332 header (a leaf)
333 NOTE: The high 8 bits cannot be
334 used - tree_depth is never that big.
335 */
315 __le16 l_count; /* Number of extent records */ 336 __le16 l_count; /* Number of extent records */
316 __le16 l_next_free_rec; /* Next unused extent slot */ 337 __le16 l_next_free_rec; /* Next unused extent slot */
317 __le16 l_reserved1; 338 __le16 l_reserved1;
@@ -446,7 +467,9 @@ struct ocfs2_dinode {
446 __le32 i_ctime_nsec; 467 __le32 i_ctime_nsec;
447 __le32 i_mtime_nsec; 468 __le32 i_mtime_nsec;
448 __le32 i_attr; 469 __le32 i_attr;
449 __le32 i_reserved1; 470 __le16 i_orphaned_slot; /* Only valid when OCFS2_ORPHANED_FL
471 was set in i_flags */
472 __le16 i_reserved1;
450/*70*/ __le64 i_reserved2[8]; 473/*70*/ __le64 i_reserved2[8];
451/*B8*/ union { 474/*B8*/ union {
452 __le64 i_pad1; /* Generic way to refer to this 475 __le64 i_pad1; /* Generic way to refer to this
diff --git a/fs/ocfs2/ocfs2_lockid.h b/fs/ocfs2/ocfs2_lockid.h
index 4d5d5655c185..4ca02b1c38ac 100644
--- a/fs/ocfs2/ocfs2_lockid.h
+++ b/fs/ocfs2/ocfs2_lockid.h
@@ -44,6 +44,7 @@ enum ocfs2_lock_type {
44 OCFS2_LOCK_TYPE_RENAME, 44 OCFS2_LOCK_TYPE_RENAME,
45 OCFS2_LOCK_TYPE_RW, 45 OCFS2_LOCK_TYPE_RW,
46 OCFS2_LOCK_TYPE_DENTRY, 46 OCFS2_LOCK_TYPE_DENTRY,
47 OCFS2_LOCK_TYPE_OPEN,
47 OCFS2_NUM_LOCK_TYPES 48 OCFS2_NUM_LOCK_TYPES
48}; 49};
49 50
@@ -69,6 +70,9 @@ static inline char ocfs2_lock_type_char(enum ocfs2_lock_type type)
69 case OCFS2_LOCK_TYPE_DENTRY: 70 case OCFS2_LOCK_TYPE_DENTRY:
70 c = 'N'; 71 c = 'N';
71 break; 72 break;
73 case OCFS2_LOCK_TYPE_OPEN:
74 c = 'O';
75 break;
72 default: 76 default:
73 c = '\0'; 77 c = '\0';
74 } 78 }
@@ -85,6 +89,7 @@ static char *ocfs2_lock_type_strings[] = {
85 * important job it does, anyway. */ 89 * important job it does, anyway. */
86 [OCFS2_LOCK_TYPE_RW] = "Write/Read", 90 [OCFS2_LOCK_TYPE_RW] = "Write/Read",
87 [OCFS2_LOCK_TYPE_DENTRY] = "Dentry", 91 [OCFS2_LOCK_TYPE_DENTRY] = "Dentry",
92 [OCFS2_LOCK_TYPE_OPEN] = "Open",
88}; 93};
89 94
90static inline const char *ocfs2_lock_type_string(enum ocfs2_lock_type type) 95static inline const char *ocfs2_lock_type_string(enum ocfs2_lock_type type)
diff --git a/fs/ocfs2/slot_map.c b/fs/ocfs2/slot_map.c
index 2d3ac32cb74e..d921a28329dc 100644
--- a/fs/ocfs2/slot_map.c
+++ b/fs/ocfs2/slot_map.c
@@ -197,7 +197,7 @@ int ocfs2_init_slot_info(struct ocfs2_super *osb)
197 goto bail; 197 goto bail;
198 } 198 }
199 199
200 status = ocfs2_extent_map_get_blocks(inode, 0ULL, 1, &blkno, NULL); 200 status = ocfs2_extent_map_get_blocks(inode, 0ULL, &blkno, NULL, NULL);
201 if (status < 0) { 201 if (status < 0) {
202 mlog_errno(status); 202 mlog_errno(status);
203 goto bail; 203 goto bail;
diff --git a/fs/ocfs2/suballoc.c b/fs/ocfs2/suballoc.c
index 6dbb11762759..0da655ae5d6f 100644
--- a/fs/ocfs2/suballoc.c
+++ b/fs/ocfs2/suballoc.c
@@ -381,8 +381,7 @@ static int ocfs2_block_group_alloc(struct ocfs2_super *osb,
381 le32_to_cpu(fe->i_clusters))); 381 le32_to_cpu(fe->i_clusters)));
382 spin_unlock(&OCFS2_I(alloc_inode)->ip_lock); 382 spin_unlock(&OCFS2_I(alloc_inode)->ip_lock);
383 i_size_write(alloc_inode, le64_to_cpu(fe->i_size)); 383 i_size_write(alloc_inode, le64_to_cpu(fe->i_size));
384 alloc_inode->i_blocks = 384 alloc_inode->i_blocks = ocfs2_inode_sector_count(alloc_inode);
385 ocfs2_align_bytes_to_sectors(i_size_read(alloc_inode));
386 385
387 status = 0; 386 status = 0;
388bail: 387bail:
diff --git a/fs/ocfs2/super.c b/fs/ocfs2/super.c
index 6534f92424dd..5c9e8243691f 100644
--- a/fs/ocfs2/super.c
+++ b/fs/ocfs2/super.c
@@ -806,9 +806,6 @@ static int __init ocfs2_init(void)
806 806
807 ocfs2_print_version(); 807 ocfs2_print_version();
808 808
809 if (init_ocfs2_extent_maps())
810 return -ENOMEM;
811
812 status = init_ocfs2_uptodate_cache(); 809 status = init_ocfs2_uptodate_cache();
813 if (status < 0) { 810 if (status < 0) {
814 mlog_errno(status); 811 mlog_errno(status);
@@ -837,7 +834,6 @@ leave:
837 if (status < 0) { 834 if (status < 0) {
838 ocfs2_free_mem_caches(); 835 ocfs2_free_mem_caches();
839 exit_ocfs2_uptodate_cache(); 836 exit_ocfs2_uptodate_cache();
840 exit_ocfs2_extent_maps();
841 } 837 }
842 838
843 mlog_exit(status); 839 mlog_exit(status);
@@ -863,8 +859,6 @@ static void __exit ocfs2_exit(void)
863 859
864 unregister_filesystem(&ocfs2_fs_type); 860 unregister_filesystem(&ocfs2_fs_type);
865 861
866 exit_ocfs2_extent_maps();
867
868 exit_ocfs2_uptodate_cache(); 862 exit_ocfs2_uptodate_cache();
869 863
870 mlog_exit_void(); 864 mlog_exit_void();
@@ -963,6 +957,7 @@ static void ocfs2_inode_init_once(void *data,
963 ocfs2_lock_res_init_once(&oi->ip_rw_lockres); 957 ocfs2_lock_res_init_once(&oi->ip_rw_lockres);
964 ocfs2_lock_res_init_once(&oi->ip_meta_lockres); 958 ocfs2_lock_res_init_once(&oi->ip_meta_lockres);
965 ocfs2_lock_res_init_once(&oi->ip_data_lockres); 959 ocfs2_lock_res_init_once(&oi->ip_data_lockres);
960 ocfs2_lock_res_init_once(&oi->ip_open_lockres);
966 961
967 ocfs2_metadata_cache_init(&oi->vfs_inode); 962 ocfs2_metadata_cache_init(&oi->vfs_inode);
968 963
diff --git a/fs/ocfs2/vote.c b/fs/ocfs2/vote.c
index f30e63b9910c..4f82a2f0efef 100644
--- a/fs/ocfs2/vote.c
+++ b/fs/ocfs2/vote.c
@@ -63,17 +63,10 @@ struct ocfs2_msg_hdr
63 __be32 h_node_num; /* node sending this particular message. */ 63 __be32 h_node_num; /* node sending this particular message. */
64}; 64};
65 65
66/* OCFS2_MAX_FILENAME_LEN is 255 characters, but we want to align this
67 * for the network. */
68#define OCFS2_VOTE_FILENAME_LEN 256
69struct ocfs2_vote_msg 66struct ocfs2_vote_msg
70{ 67{
71 struct ocfs2_msg_hdr v_hdr; 68 struct ocfs2_msg_hdr v_hdr;
72 union { 69 __be32 v_reserved1;
73 __be32 v_generic1;
74 __be32 v_orphaned_slot; /* Used during delete votes */
75 __be32 v_nlink; /* Used during unlink votes */
76 } md1; /* Message type dependant 1 */
77}; 70};
78 71
79/* Responses are given these values to maintain backwards 72/* Responses are given these values to maintain backwards
@@ -86,7 +79,6 @@ struct ocfs2_response_msg
86{ 79{
87 struct ocfs2_msg_hdr r_hdr; 80 struct ocfs2_msg_hdr r_hdr;
88 __be32 r_response; 81 __be32 r_response;
89 __be32 r_orphaned_slot;
90}; 82};
91 83
92struct ocfs2_vote_work { 84struct ocfs2_vote_work {
@@ -96,7 +88,6 @@ struct ocfs2_vote_work {
96 88
97enum ocfs2_vote_request { 89enum ocfs2_vote_request {
98 OCFS2_VOTE_REQ_INVALID = 0, 90 OCFS2_VOTE_REQ_INVALID = 0,
99 OCFS2_VOTE_REQ_DELETE,
100 OCFS2_VOTE_REQ_MOUNT, 91 OCFS2_VOTE_REQ_MOUNT,
101 OCFS2_VOTE_REQ_UMOUNT, 92 OCFS2_VOTE_REQ_UMOUNT,
102 OCFS2_VOTE_REQ_LAST 93 OCFS2_VOTE_REQ_LAST
@@ -151,135 +142,23 @@ static void ocfs2_process_umount_request(struct ocfs2_super *osb,
151 ocfs2_node_map_set_bit(osb, &osb->umount_map, node_num); 142 ocfs2_node_map_set_bit(osb, &osb->umount_map, node_num);
152} 143}
153 144
154void ocfs2_mark_inode_remotely_deleted(struct inode *inode)
155{
156 struct ocfs2_inode_info *oi = OCFS2_I(inode);
157
158 assert_spin_locked(&oi->ip_lock);
159 /* We set the SKIP_DELETE flag on the inode so we don't try to
160 * delete it in delete_inode ourselves, thus avoiding
161 * unecessary lock pinging. If the other node failed to wipe
162 * the inode as a result of a crash, then recovery will pick
163 * up the slack. */
164 oi->ip_flags |= OCFS2_INODE_DELETED|OCFS2_INODE_SKIP_DELETE;
165}
166
167static int ocfs2_process_delete_request(struct inode *inode,
168 int *orphaned_slot)
169{
170 int response = OCFS2_RESPONSE_BUSY;
171
172 mlog(0, "DELETE vote on inode %lu, read lnk_cnt = %u, slot = %d\n",
173 inode->i_ino, inode->i_nlink, *orphaned_slot);
174
175 spin_lock(&OCFS2_I(inode)->ip_lock);
176
177 /* Whatever our vote response is, we want to make sure that
178 * the orphaned slot is recorded properly on this node *and*
179 * on the requesting node. Technically, if the requesting node
180 * did not know which slot the inode is orphaned in but we
181 * respond with BUSY he doesn't actually need the orphaned
182 * slot, but it doesn't hurt to do it here anyway. */
183 if ((*orphaned_slot) != OCFS2_INVALID_SLOT) {
184 mlog_bug_on_msg(OCFS2_I(inode)->ip_orphaned_slot !=
185 OCFS2_INVALID_SLOT &&
186 OCFS2_I(inode)->ip_orphaned_slot !=
187 (*orphaned_slot),
188 "Inode %llu: This node thinks it's "
189 "orphaned in slot %d, messaged it's in %d\n",
190 (unsigned long long)OCFS2_I(inode)->ip_blkno,
191 OCFS2_I(inode)->ip_orphaned_slot,
192 *orphaned_slot);
193
194 mlog(0, "Setting orphaned slot for inode %llu to %d\n",
195 (unsigned long long)OCFS2_I(inode)->ip_blkno,
196 *orphaned_slot);
197
198 OCFS2_I(inode)->ip_orphaned_slot = *orphaned_slot;
199 } else {
200 mlog(0, "Sending back orphaned slot %d for inode %llu\n",
201 OCFS2_I(inode)->ip_orphaned_slot,
202 (unsigned long long)OCFS2_I(inode)->ip_blkno);
203
204 *orphaned_slot = OCFS2_I(inode)->ip_orphaned_slot;
205 }
206
207 /* vote no if the file is still open. */
208 if (OCFS2_I(inode)->ip_open_count) {
209 mlog(0, "open count = %u\n",
210 OCFS2_I(inode)->ip_open_count);
211 spin_unlock(&OCFS2_I(inode)->ip_lock);
212 goto done;
213 }
214 spin_unlock(&OCFS2_I(inode)->ip_lock);
215
216 /* directories are a bit ugly... What if someone is sitting in
217 * it? We want to make sure the inode is removed completely as
218 * a result of the iput in process_vote. */
219 if (S_ISDIR(inode->i_mode) && (atomic_read(&inode->i_count) != 1)) {
220 mlog(0, "i_count = %u\n", atomic_read(&inode->i_count));
221 goto done;
222 }
223
224 if (filemap_fdatawrite(inode->i_mapping)) {
225 mlog(ML_ERROR, "Could not sync inode %llu for delete!\n",
226 (unsigned long long)OCFS2_I(inode)->ip_blkno);
227 goto done;
228 }
229 sync_mapping_buffers(inode->i_mapping);
230 truncate_inode_pages(inode->i_mapping, 0);
231 ocfs2_extent_map_trunc(inode, 0);
232
233 spin_lock(&OCFS2_I(inode)->ip_lock);
234 /* double check open count - someone might have raced this
235 * thread into ocfs2_file_open while we were writing out
236 * data. If we're to allow a wipe of this inode now, we *must*
237 * hold the spinlock until we've marked it. */
238 if (OCFS2_I(inode)->ip_open_count) {
239 mlog(0, "Raced to wipe! open count = %u\n",
240 OCFS2_I(inode)->ip_open_count);
241 spin_unlock(&OCFS2_I(inode)->ip_lock);
242 goto done;
243 }
244
245 /* Mark the inode as being wiped from disk. */
246 ocfs2_mark_inode_remotely_deleted(inode);
247 spin_unlock(&OCFS2_I(inode)->ip_lock);
248
249 /* Not sure this is necessary anymore. */
250 d_prune_aliases(inode);
251
252 /* If we get here, then we're voting 'yes', so commit the
253 * delete on our side. */
254 response = OCFS2_RESPONSE_OK;
255done:
256 return response;
257}
258
259static void ocfs2_process_vote(struct ocfs2_super *osb, 145static void ocfs2_process_vote(struct ocfs2_super *osb,
260 struct ocfs2_vote_msg *msg) 146 struct ocfs2_vote_msg *msg)
261{ 147{
262 int net_status, vote_response; 148 int net_status, vote_response;
263 int orphaned_slot = 0; 149 unsigned int node_num;
264 unsigned int node_num, generation;
265 u64 blkno; 150 u64 blkno;
266 enum ocfs2_vote_request request; 151 enum ocfs2_vote_request request;
267 struct inode *inode = NULL;
268 struct ocfs2_msg_hdr *hdr = &msg->v_hdr; 152 struct ocfs2_msg_hdr *hdr = &msg->v_hdr;
269 struct ocfs2_response_msg response; 153 struct ocfs2_response_msg response;
270 154
271 /* decode the network mumbo jumbo into local variables. */ 155 /* decode the network mumbo jumbo into local variables. */
272 request = be32_to_cpu(hdr->h_request); 156 request = be32_to_cpu(hdr->h_request);
273 blkno = be64_to_cpu(hdr->h_blkno); 157 blkno = be64_to_cpu(hdr->h_blkno);
274 generation = be32_to_cpu(hdr->h_generation);
275 node_num = be32_to_cpu(hdr->h_node_num); 158 node_num = be32_to_cpu(hdr->h_node_num);
276 if (request == OCFS2_VOTE_REQ_DELETE)
277 orphaned_slot = be32_to_cpu(msg->md1.v_orphaned_slot);
278 159
279 mlog(0, "processing vote: request = %u, blkno = %llu, " 160 mlog(0, "processing vote: request = %u, blkno = %llu, node_num = %u\n",
280 "generation = %u, node_num = %u, priv1 = %u\n", request, 161 request, (unsigned long long)blkno, node_num);
281 (unsigned long long)blkno, generation, node_num,
282 be32_to_cpu(msg->md1.v_generic1));
283 162
284 if (!ocfs2_is_valid_vote_request(request)) { 163 if (!ocfs2_is_valid_vote_request(request)) {
285 mlog(ML_ERROR, "Invalid vote request %d from node %u\n", 164 mlog(ML_ERROR, "Invalid vote request %d from node %u\n",
@@ -302,52 +181,6 @@ static void ocfs2_process_vote(struct ocfs2_super *osb,
302 break; 181 break;
303 } 182 }
304 183
305 /* We cannot process the remaining message types before we're
306 * fully mounted. It's perfectly safe however to send a 'yes'
307 * response as we can't possibly have any of the state they're
308 * asking us to modify yet. */
309 if (atomic_read(&osb->vol_state) == VOLUME_INIT)
310 goto respond;
311
312 /* If we get here, then the request is against an inode. */
313 inode = ocfs2_ilookup_for_vote(osb, blkno,
314 request == OCFS2_VOTE_REQ_DELETE);
315
316 /* Not finding the inode is perfectly valid - it means we're
317 * not interested in what the other node is about to do to it
318 * so in those cases we automatically respond with an
319 * affirmative. Cluster locking ensures that we won't race
320 * interest in the inode with this vote request. */
321 if (!inode)
322 goto respond;
323
324 /* Check generation values. It's possible for us to get a
325 * request against a stale inode. If so then we proceed as if
326 * we had not found an inode in the first place. */
327 if (inode->i_generation != generation) {
328 mlog(0, "generation passed %u != inode generation = %u, "
329 "ip_flags = %x, ip_blkno = %llu, msg %llu, i_count = %u, "
330 "message type = %u\n", generation, inode->i_generation,
331 OCFS2_I(inode)->ip_flags,
332 (unsigned long long)OCFS2_I(inode)->ip_blkno,
333 (unsigned long long)blkno, atomic_read(&inode->i_count),
334 request);
335 iput(inode);
336 inode = NULL;
337 goto respond;
338 }
339
340 switch (request) {
341 case OCFS2_VOTE_REQ_DELETE:
342 vote_response = ocfs2_process_delete_request(inode,
343 &orphaned_slot);
344 break;
345 default:
346 mlog(ML_ERROR, "node %u, invalid request: %u\n",
347 node_num, request);
348 vote_response = OCFS2_RESPONSE_BAD_MSG;
349 }
350
351respond: 184respond:
352 /* Response struture is small so we just put it on the stack 185 /* Response struture is small so we just put it on the stack
353 * and stuff it inline. */ 186 * and stuff it inline. */
@@ -357,7 +190,6 @@ respond:
357 response.r_hdr.h_generation = hdr->h_generation; 190 response.r_hdr.h_generation = hdr->h_generation;
358 response.r_hdr.h_node_num = cpu_to_be32(osb->node_num); 191 response.r_hdr.h_node_num = cpu_to_be32(osb->node_num);
359 response.r_response = cpu_to_be32(vote_response); 192 response.r_response = cpu_to_be32(vote_response);
360 response.r_orphaned_slot = cpu_to_be32(orphaned_slot);
361 193
362 net_status = o2net_send_message(OCFS2_MESSAGE_TYPE_RESPONSE, 194 net_status = o2net_send_message(OCFS2_MESSAGE_TYPE_RESPONSE,
363 osb->net_key, 195 osb->net_key,
@@ -373,9 +205,6 @@ respond:
373 && net_status != -ENOTCONN) 205 && net_status != -ENOTCONN)
374 mlog(ML_ERROR, "message to node %u fails with error %d!\n", 206 mlog(ML_ERROR, "message to node %u fails with error %d!\n",
375 node_num, net_status); 207 node_num, net_status);
376
377 if (inode)
378 iput(inode);
379} 208}
380 209
381static void ocfs2_vote_thread_do_work(struct ocfs2_super *osb) 210static void ocfs2_vote_thread_do_work(struct ocfs2_super *osb)
@@ -634,8 +463,7 @@ bail:
634static struct ocfs2_vote_msg * ocfs2_new_vote_request(struct ocfs2_super *osb, 463static struct ocfs2_vote_msg * ocfs2_new_vote_request(struct ocfs2_super *osb,
635 u64 blkno, 464 u64 blkno,
636 unsigned int generation, 465 unsigned int generation,
637 enum ocfs2_vote_request type, 466 enum ocfs2_vote_request type)
638 u32 priv)
639{ 467{
640 struct ocfs2_vote_msg *request; 468 struct ocfs2_vote_msg *request;
641 struct ocfs2_msg_hdr *hdr; 469 struct ocfs2_msg_hdr *hdr;
@@ -651,8 +479,6 @@ static struct ocfs2_vote_msg * ocfs2_new_vote_request(struct ocfs2_super *osb,
651 hdr->h_request = cpu_to_be32(type); 479 hdr->h_request = cpu_to_be32(type);
652 hdr->h_blkno = cpu_to_be64(blkno); 480 hdr->h_blkno = cpu_to_be64(blkno);
653 hdr->h_generation = cpu_to_be32(generation); 481 hdr->h_generation = cpu_to_be32(generation);
654
655 request->md1.v_generic1 = cpu_to_be32(priv);
656 } 482 }
657 483
658 return request; 484 return request;
@@ -664,7 +490,7 @@ static int ocfs2_do_request_vote(struct ocfs2_super *osb,
664 struct ocfs2_vote_msg *request, 490 struct ocfs2_vote_msg *request,
665 struct ocfs2_net_response_cb *callback) 491 struct ocfs2_net_response_cb *callback)
666{ 492{
667 int status, response; 493 int status, response = -EBUSY;
668 unsigned int response_id; 494 unsigned int response_id;
669 struct ocfs2_msg_hdr *hdr; 495 struct ocfs2_msg_hdr *hdr;
670 496
@@ -686,109 +512,12 @@ bail:
686 return status; 512 return status;
687} 513}
688 514
689static int ocfs2_request_vote(struct inode *inode,
690 struct ocfs2_vote_msg *request,
691 struct ocfs2_net_response_cb *callback)
692{
693 int status;
694 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
695
696 if (ocfs2_inode_is_new(inode))
697 return 0;
698
699 status = -EAGAIN;
700 while (status == -EAGAIN) {
701 if (!(osb->s_mount_opt & OCFS2_MOUNT_NOINTR) &&
702 signal_pending(current))
703 return -ERESTARTSYS;
704
705 status = ocfs2_super_lock(osb, 0);
706 if (status < 0) {
707 mlog_errno(status);
708 break;
709 }
710
711 status = 0;
712 if (!ocfs2_node_map_is_only(osb, &osb->mounted_map,
713 osb->node_num))
714 status = ocfs2_do_request_vote(osb, request, callback);
715
716 ocfs2_super_unlock(osb, 0);
717 }
718 return status;
719}
720
721static void ocfs2_delete_response_cb(void *priv,
722 struct ocfs2_response_msg *resp)
723{
724 int orphaned_slot, node;
725 struct inode *inode = priv;
726
727 orphaned_slot = be32_to_cpu(resp->r_orphaned_slot);
728 node = be32_to_cpu(resp->r_hdr.h_node_num);
729 mlog(0, "node %d tells us that inode %llu is orphaned in slot %d\n",
730 node, (unsigned long long)OCFS2_I(inode)->ip_blkno,
731 orphaned_slot);
732
733 /* The other node may not actually know which slot the inode
734 * is orphaned in. */
735 if (orphaned_slot == OCFS2_INVALID_SLOT)
736 return;
737
738 /* Ok, the responding node knows which slot this inode is
739 * orphaned in. We verify that the information is correct and
740 * then record this in the inode. ocfs2_delete_inode will use
741 * this information to determine which lock to take. */
742 spin_lock(&OCFS2_I(inode)->ip_lock);
743 mlog_bug_on_msg(OCFS2_I(inode)->ip_orphaned_slot != orphaned_slot &&
744 OCFS2_I(inode)->ip_orphaned_slot
745 != OCFS2_INVALID_SLOT, "Inode %llu: Node %d says it's "
746 "orphaned in slot %d, we think it's in %d\n",
747 (unsigned long long)OCFS2_I(inode)->ip_blkno,
748 be32_to_cpu(resp->r_hdr.h_node_num),
749 orphaned_slot, OCFS2_I(inode)->ip_orphaned_slot);
750
751 OCFS2_I(inode)->ip_orphaned_slot = orphaned_slot;
752 spin_unlock(&OCFS2_I(inode)->ip_lock);
753}
754
755int ocfs2_request_delete_vote(struct inode *inode)
756{
757 int orphaned_slot, status;
758 struct ocfs2_net_response_cb delete_cb;
759 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
760 struct ocfs2_vote_msg *request;
761
762 spin_lock(&OCFS2_I(inode)->ip_lock);
763 orphaned_slot = OCFS2_I(inode)->ip_orphaned_slot;
764 spin_unlock(&OCFS2_I(inode)->ip_lock);
765
766 delete_cb.rc_cb = ocfs2_delete_response_cb;
767 delete_cb.rc_priv = inode;
768
769 mlog(0, "Inode %llu, we start thinking orphaned slot is %d\n",
770 (unsigned long long)OCFS2_I(inode)->ip_blkno, orphaned_slot);
771
772 status = -ENOMEM;
773 request = ocfs2_new_vote_request(osb, OCFS2_I(inode)->ip_blkno,
774 inode->i_generation,
775 OCFS2_VOTE_REQ_DELETE, orphaned_slot);
776 if (request) {
777 status = ocfs2_request_vote(inode, request, &delete_cb);
778
779 kfree(request);
780 }
781
782 return status;
783}
784
785int ocfs2_request_mount_vote(struct ocfs2_super *osb) 515int ocfs2_request_mount_vote(struct ocfs2_super *osb)
786{ 516{
787 int status; 517 int status;
788 struct ocfs2_vote_msg *request = NULL; 518 struct ocfs2_vote_msg *request = NULL;
789 519
790 request = ocfs2_new_vote_request(osb, 0ULL, 0, 520 request = ocfs2_new_vote_request(osb, 0ULL, 0, OCFS2_VOTE_REQ_MOUNT);
791 OCFS2_VOTE_REQ_MOUNT, 0);
792 if (!request) { 521 if (!request) {
793 status = -ENOMEM; 522 status = -ENOMEM;
794 goto bail; 523 goto bail;
@@ -821,8 +550,7 @@ int ocfs2_request_umount_vote(struct ocfs2_super *osb)
821 int status; 550 int status;
822 struct ocfs2_vote_msg *request = NULL; 551 struct ocfs2_vote_msg *request = NULL;
823 552
824 request = ocfs2_new_vote_request(osb, 0ULL, 0, 553 request = ocfs2_new_vote_request(osb, 0ULL, 0, OCFS2_VOTE_REQ_UMOUNT);
825 OCFS2_VOTE_REQ_UMOUNT, 0);
826 if (!request) { 554 if (!request) {
827 status = -ENOMEM; 555 status = -ENOMEM;
828 goto bail; 556 goto bail;
@@ -969,7 +697,6 @@ static int ocfs2_handle_vote_message(struct o2net_msg *msg,
969 be32_to_cpu(work->w_msg.v_hdr.h_generation)); 697 be32_to_cpu(work->w_msg.v_hdr.h_generation));
970 mlog(0, "h_node_num = %u\n", 698 mlog(0, "h_node_num = %u\n",
971 be32_to_cpu(work->w_msg.v_hdr.h_node_num)); 699 be32_to_cpu(work->w_msg.v_hdr.h_node_num));
972 mlog(0, "v_generic1 = %u\n", be32_to_cpu(work->w_msg.md1.v_generic1));
973 700
974 spin_lock(&osb->vote_task_lock); 701 spin_lock(&osb->vote_task_lock);
975 list_add_tail(&work->w_list, &osb->vote_list); 702 list_add_tail(&work->w_list, &osb->vote_list);
diff --git a/fs/ocfs2/vote.h b/fs/ocfs2/vote.h
index 53ebc1c69e56..9ea46f62de31 100644
--- a/fs/ocfs2/vote.h
+++ b/fs/ocfs2/vote.h
@@ -38,14 +38,11 @@ static inline void ocfs2_kick_vote_thread(struct ocfs2_super *osb)
38 wake_up(&osb->vote_event); 38 wake_up(&osb->vote_event);
39} 39}
40 40
41int ocfs2_request_delete_vote(struct inode *inode);
42int ocfs2_request_mount_vote(struct ocfs2_super *osb); 41int ocfs2_request_mount_vote(struct ocfs2_super *osb);
43int ocfs2_request_umount_vote(struct ocfs2_super *osb); 42int ocfs2_request_umount_vote(struct ocfs2_super *osb);
44int ocfs2_register_net_handlers(struct ocfs2_super *osb); 43int ocfs2_register_net_handlers(struct ocfs2_super *osb);
45void ocfs2_unregister_net_handlers(struct ocfs2_super *osb); 44void ocfs2_unregister_net_handlers(struct ocfs2_super *osb);
46 45
47void ocfs2_mark_inode_remotely_deleted(struct inode *inode);
48
49void ocfs2_remove_node_from_vote_queues(struct ocfs2_super *osb, 46void ocfs2_remove_node_from_vote_queues(struct ocfs2_super *osb,
50 int node_num); 47 int node_num);
51#endif 48#endif
diff --git a/fs/sync.c b/fs/sync.c
index d0feff61e6aa..5cb9e7e43383 100644
--- a/fs/sync.c
+++ b/fs/sync.c
@@ -239,13 +239,11 @@ out:
239/* 239/*
240 * `endbyte' is inclusive 240 * `endbyte' is inclusive
241 */ 241 */
242int do_sync_file_range(struct file *file, loff_t offset, loff_t endbyte, 242int do_sync_mapping_range(struct address_space *mapping, loff_t offset,
243 unsigned int flags) 243 loff_t endbyte, unsigned int flags)
244{ 244{
245 int ret; 245 int ret;
246 struct address_space *mapping;
247 246
248 mapping = file->f_mapping;
249 if (!mapping) { 247 if (!mapping) {
250 ret = -EINVAL; 248 ret = -EINVAL;
251 goto out; 249 goto out;
@@ -275,4 +273,4 @@ int do_sync_file_range(struct file *file, loff_t offset, loff_t endbyte,
275out: 273out:
276 return ret; 274 return ret;
277} 275}
278EXPORT_SYMBOL_GPL(do_sync_file_range); 276EXPORT_SYMBOL_GPL(do_sync_mapping_range);
diff --git a/include/linux/fs.h b/include/linux/fs.h
index 86ec3f4a7da6..095a9c9a64fb 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -843,8 +843,13 @@ extern int fcntl_setlease(unsigned int fd, struct file *filp, long arg);
843extern int fcntl_getlease(struct file *filp); 843extern int fcntl_getlease(struct file *filp);
844 844
845/* fs/sync.c */ 845/* fs/sync.c */
846extern int do_sync_file_range(struct file *file, loff_t offset, loff_t endbyte, 846extern int do_sync_mapping_range(struct address_space *mapping, loff_t offset,
847 unsigned int flags); 847 loff_t endbyte, unsigned int flags);
848static inline int do_sync_file_range(struct file *file, loff_t offset,
849 loff_t endbyte, unsigned int flags)
850{
851 return do_sync_mapping_range(file->f_mapping, offset, endbyte, flags);
852}
848 853
849/* fs/locks.c */ 854/* fs/locks.c */
850extern void locks_init_lock(struct file_lock *); 855extern void locks_init_lock(struct file_lock *);
generated by cgit v1.2.2 (git 2.25.0) at 2025-11-16 17:16:16 -0500