Loading...
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * alloc.c
5 *
6 * Extent allocs and frees
7 *
8 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 */
25
26#include <linux/fs.h>
27#include <linux/types.h>
28#include <linux/slab.h>
29#include <linux/highmem.h>
30#include <linux/swap.h>
31#include <linux/quotaops.h>
32#include <linux/blkdev.h>
33
34#include <cluster/masklog.h>
35
36#include "ocfs2.h"
37
38#include "alloc.h"
39#include "aops.h"
40#include "blockcheck.h"
41#include "dlmglue.h"
42#include "extent_map.h"
43#include "inode.h"
44#include "journal.h"
45#include "localalloc.h"
46#include "suballoc.h"
47#include "sysfile.h"
48#include "file.h"
49#include "super.h"
50#include "uptodate.h"
51#include "xattr.h"
52#include "refcounttree.h"
53#include "ocfs2_trace.h"
54
55#include "buffer_head_io.h"
56
57enum ocfs2_contig_type {
58 CONTIG_NONE = 0,
59 CONTIG_LEFT,
60 CONTIG_RIGHT,
61 CONTIG_LEFTRIGHT,
62};
63
64static enum ocfs2_contig_type
65 ocfs2_extent_rec_contig(struct super_block *sb,
66 struct ocfs2_extent_rec *ext,
67 struct ocfs2_extent_rec *insert_rec);
68/*
69 * Operations for a specific extent tree type.
70 *
71 * To implement an on-disk btree (extent tree) type in ocfs2, add
72 * an ocfs2_extent_tree_operations structure and the matching
73 * ocfs2_init_<thingy>_extent_tree() function. That's pretty much it
74 * for the allocation portion of the extent tree.
75 */
76struct ocfs2_extent_tree_operations {
77 /*
78 * last_eb_blk is the block number of the right most leaf extent
79 * block. Most on-disk structures containing an extent tree store
80 * this value for fast access. The ->eo_set_last_eb_blk() and
81 * ->eo_get_last_eb_blk() operations access this value. They are
82 * both required.
83 */
84 void (*eo_set_last_eb_blk)(struct ocfs2_extent_tree *et,
85 u64 blkno);
86 u64 (*eo_get_last_eb_blk)(struct ocfs2_extent_tree *et);
87
88 /*
89 * The on-disk structure usually keeps track of how many total
90 * clusters are stored in this extent tree. This function updates
91 * that value. new_clusters is the delta, and must be
92 * added to the total. Required.
93 */
94 void (*eo_update_clusters)(struct ocfs2_extent_tree *et,
95 u32 new_clusters);
96
97 /*
98 * If this extent tree is supported by an extent map, insert
99 * a record into the map.
100 */
101 void (*eo_extent_map_insert)(struct ocfs2_extent_tree *et,
102 struct ocfs2_extent_rec *rec);
103
104 /*
105 * If this extent tree is supported by an extent map, truncate the
106 * map to clusters,
107 */
108 void (*eo_extent_map_truncate)(struct ocfs2_extent_tree *et,
109 u32 clusters);
110
111 /*
112 * If ->eo_insert_check() exists, it is called before rec is
113 * inserted into the extent tree. It is optional.
114 */
115 int (*eo_insert_check)(struct ocfs2_extent_tree *et,
116 struct ocfs2_extent_rec *rec);
117 int (*eo_sanity_check)(struct ocfs2_extent_tree *et);
118
119 /*
120 * --------------------------------------------------------------
121 * The remaining are internal to ocfs2_extent_tree and don't have
122 * accessor functions
123 */
124
125 /*
126 * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el.
127 * It is required.
128 */
129 void (*eo_fill_root_el)(struct ocfs2_extent_tree *et);
130
131 /*
132 * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if
133 * it exists. If it does not, et->et_max_leaf_clusters is set
134 * to 0 (unlimited). Optional.
135 */
136 void (*eo_fill_max_leaf_clusters)(struct ocfs2_extent_tree *et);
137
138 /*
139 * ->eo_extent_contig test whether the 2 ocfs2_extent_rec
140 * are contiguous or not. Optional. Don't need to set it if use
141 * ocfs2_extent_rec as the tree leaf.
142 */
143 enum ocfs2_contig_type
144 (*eo_extent_contig)(struct ocfs2_extent_tree *et,
145 struct ocfs2_extent_rec *ext,
146 struct ocfs2_extent_rec *insert_rec);
147};
148
149
150/*
151 * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check
152 * in the methods.
153 */
154static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et);
155static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
156 u64 blkno);
157static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
158 u32 clusters);
159static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
160 struct ocfs2_extent_rec *rec);
161static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
162 u32 clusters);
163static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
164 struct ocfs2_extent_rec *rec);
165static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et);
166static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et);
167static const struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops = {
168 .eo_set_last_eb_blk = ocfs2_dinode_set_last_eb_blk,
169 .eo_get_last_eb_blk = ocfs2_dinode_get_last_eb_blk,
170 .eo_update_clusters = ocfs2_dinode_update_clusters,
171 .eo_extent_map_insert = ocfs2_dinode_extent_map_insert,
172 .eo_extent_map_truncate = ocfs2_dinode_extent_map_truncate,
173 .eo_insert_check = ocfs2_dinode_insert_check,
174 .eo_sanity_check = ocfs2_dinode_sanity_check,
175 .eo_fill_root_el = ocfs2_dinode_fill_root_el,
176};
177
178static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
179 u64 blkno)
180{
181 struct ocfs2_dinode *di = et->et_object;
182
183 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
184 di->i_last_eb_blk = cpu_to_le64(blkno);
185}
186
187static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et)
188{
189 struct ocfs2_dinode *di = et->et_object;
190
191 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
192 return le64_to_cpu(di->i_last_eb_blk);
193}
194
195static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
196 u32 clusters)
197{
198 struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
199 struct ocfs2_dinode *di = et->et_object;
200
201 le32_add_cpu(&di->i_clusters, clusters);
202 spin_lock(&oi->ip_lock);
203 oi->ip_clusters = le32_to_cpu(di->i_clusters);
204 spin_unlock(&oi->ip_lock);
205}
206
207static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
208 struct ocfs2_extent_rec *rec)
209{
210 struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
211
212 ocfs2_extent_map_insert_rec(inode, rec);
213}
214
215static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
216 u32 clusters)
217{
218 struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
219
220 ocfs2_extent_map_trunc(inode, clusters);
221}
222
223static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
224 struct ocfs2_extent_rec *rec)
225{
226 struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
227 struct ocfs2_super *osb = OCFS2_SB(oi->vfs_inode.i_sb);
228
229 BUG_ON(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL);
230 mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) &&
231 (oi->ip_clusters != le32_to_cpu(rec->e_cpos)),
232 "Device %s, asking for sparse allocation: inode %llu, "
233 "cpos %u, clusters %u\n",
234 osb->dev_str,
235 (unsigned long long)oi->ip_blkno,
236 rec->e_cpos, oi->ip_clusters);
237
238 return 0;
239}
240
241static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et)
242{
243 struct ocfs2_dinode *di = et->et_object;
244
245 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
246 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
247
248 return 0;
249}
250
251static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et)
252{
253 struct ocfs2_dinode *di = et->et_object;
254
255 et->et_root_el = &di->id2.i_list;
256}
257
258
259static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree *et)
260{
261 struct ocfs2_xattr_value_buf *vb = et->et_object;
262
263 et->et_root_el = &vb->vb_xv->xr_list;
264}
265
266static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree *et,
267 u64 blkno)
268{
269 struct ocfs2_xattr_value_buf *vb = et->et_object;
270
271 vb->vb_xv->xr_last_eb_blk = cpu_to_le64(blkno);
272}
273
274static u64 ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree *et)
275{
276 struct ocfs2_xattr_value_buf *vb = et->et_object;
277
278 return le64_to_cpu(vb->vb_xv->xr_last_eb_blk);
279}
280
281static void ocfs2_xattr_value_update_clusters(struct ocfs2_extent_tree *et,
282 u32 clusters)
283{
284 struct ocfs2_xattr_value_buf *vb = et->et_object;
285
286 le32_add_cpu(&vb->vb_xv->xr_clusters, clusters);
287}
288
289static const struct ocfs2_extent_tree_operations ocfs2_xattr_value_et_ops = {
290 .eo_set_last_eb_blk = ocfs2_xattr_value_set_last_eb_blk,
291 .eo_get_last_eb_blk = ocfs2_xattr_value_get_last_eb_blk,
292 .eo_update_clusters = ocfs2_xattr_value_update_clusters,
293 .eo_fill_root_el = ocfs2_xattr_value_fill_root_el,
294};
295
296static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree *et)
297{
298 struct ocfs2_xattr_block *xb = et->et_object;
299
300 et->et_root_el = &xb->xb_attrs.xb_root.xt_list;
301}
302
303static void ocfs2_xattr_tree_fill_max_leaf_clusters(struct ocfs2_extent_tree *et)
304{
305 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
306 et->et_max_leaf_clusters =
307 ocfs2_clusters_for_bytes(sb, OCFS2_MAX_XATTR_TREE_LEAF_SIZE);
308}
309
310static void ocfs2_xattr_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
311 u64 blkno)
312{
313 struct ocfs2_xattr_block *xb = et->et_object;
314 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
315
316 xt->xt_last_eb_blk = cpu_to_le64(blkno);
317}
318
319static u64 ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
320{
321 struct ocfs2_xattr_block *xb = et->et_object;
322 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
323
324 return le64_to_cpu(xt->xt_last_eb_blk);
325}
326
327static void ocfs2_xattr_tree_update_clusters(struct ocfs2_extent_tree *et,
328 u32 clusters)
329{
330 struct ocfs2_xattr_block *xb = et->et_object;
331
332 le32_add_cpu(&xb->xb_attrs.xb_root.xt_clusters, clusters);
333}
334
335static const struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops = {
336 .eo_set_last_eb_blk = ocfs2_xattr_tree_set_last_eb_blk,
337 .eo_get_last_eb_blk = ocfs2_xattr_tree_get_last_eb_blk,
338 .eo_update_clusters = ocfs2_xattr_tree_update_clusters,
339 .eo_fill_root_el = ocfs2_xattr_tree_fill_root_el,
340 .eo_fill_max_leaf_clusters = ocfs2_xattr_tree_fill_max_leaf_clusters,
341};
342
343static void ocfs2_dx_root_set_last_eb_blk(struct ocfs2_extent_tree *et,
344 u64 blkno)
345{
346 struct ocfs2_dx_root_block *dx_root = et->et_object;
347
348 dx_root->dr_last_eb_blk = cpu_to_le64(blkno);
349}
350
351static u64 ocfs2_dx_root_get_last_eb_blk(struct ocfs2_extent_tree *et)
352{
353 struct ocfs2_dx_root_block *dx_root = et->et_object;
354
355 return le64_to_cpu(dx_root->dr_last_eb_blk);
356}
357
358static void ocfs2_dx_root_update_clusters(struct ocfs2_extent_tree *et,
359 u32 clusters)
360{
361 struct ocfs2_dx_root_block *dx_root = et->et_object;
362
363 le32_add_cpu(&dx_root->dr_clusters, clusters);
364}
365
366static int ocfs2_dx_root_sanity_check(struct ocfs2_extent_tree *et)
367{
368 struct ocfs2_dx_root_block *dx_root = et->et_object;
369
370 BUG_ON(!OCFS2_IS_VALID_DX_ROOT(dx_root));
371
372 return 0;
373}
374
375static void ocfs2_dx_root_fill_root_el(struct ocfs2_extent_tree *et)
376{
377 struct ocfs2_dx_root_block *dx_root = et->et_object;
378
379 et->et_root_el = &dx_root->dr_list;
380}
381
382static const struct ocfs2_extent_tree_operations ocfs2_dx_root_et_ops = {
383 .eo_set_last_eb_blk = ocfs2_dx_root_set_last_eb_blk,
384 .eo_get_last_eb_blk = ocfs2_dx_root_get_last_eb_blk,
385 .eo_update_clusters = ocfs2_dx_root_update_clusters,
386 .eo_sanity_check = ocfs2_dx_root_sanity_check,
387 .eo_fill_root_el = ocfs2_dx_root_fill_root_el,
388};
389
390static void ocfs2_refcount_tree_fill_root_el(struct ocfs2_extent_tree *et)
391{
392 struct ocfs2_refcount_block *rb = et->et_object;
393
394 et->et_root_el = &rb->rf_list;
395}
396
397static void ocfs2_refcount_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
398 u64 blkno)
399{
400 struct ocfs2_refcount_block *rb = et->et_object;
401
402 rb->rf_last_eb_blk = cpu_to_le64(blkno);
403}
404
405static u64 ocfs2_refcount_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
406{
407 struct ocfs2_refcount_block *rb = et->et_object;
408
409 return le64_to_cpu(rb->rf_last_eb_blk);
410}
411
412static void ocfs2_refcount_tree_update_clusters(struct ocfs2_extent_tree *et,
413 u32 clusters)
414{
415 struct ocfs2_refcount_block *rb = et->et_object;
416
417 le32_add_cpu(&rb->rf_clusters, clusters);
418}
419
420static enum ocfs2_contig_type
421ocfs2_refcount_tree_extent_contig(struct ocfs2_extent_tree *et,
422 struct ocfs2_extent_rec *ext,
423 struct ocfs2_extent_rec *insert_rec)
424{
425 return CONTIG_NONE;
426}
427
428static const struct ocfs2_extent_tree_operations ocfs2_refcount_tree_et_ops = {
429 .eo_set_last_eb_blk = ocfs2_refcount_tree_set_last_eb_blk,
430 .eo_get_last_eb_blk = ocfs2_refcount_tree_get_last_eb_blk,
431 .eo_update_clusters = ocfs2_refcount_tree_update_clusters,
432 .eo_fill_root_el = ocfs2_refcount_tree_fill_root_el,
433 .eo_extent_contig = ocfs2_refcount_tree_extent_contig,
434};
435
436static void __ocfs2_init_extent_tree(struct ocfs2_extent_tree *et,
437 struct ocfs2_caching_info *ci,
438 struct buffer_head *bh,
439 ocfs2_journal_access_func access,
440 void *obj,
441 const struct ocfs2_extent_tree_operations *ops)
442{
443 et->et_ops = ops;
444 et->et_root_bh = bh;
445 et->et_ci = ci;
446 et->et_root_journal_access = access;
447 if (!obj)
448 obj = (void *)bh->b_data;
449 et->et_object = obj;
450
451 et->et_ops->eo_fill_root_el(et);
452 if (!et->et_ops->eo_fill_max_leaf_clusters)
453 et->et_max_leaf_clusters = 0;
454 else
455 et->et_ops->eo_fill_max_leaf_clusters(et);
456}
457
458void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree *et,
459 struct ocfs2_caching_info *ci,
460 struct buffer_head *bh)
461{
462 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_di,
463 NULL, &ocfs2_dinode_et_ops);
464}
465
466void ocfs2_init_xattr_tree_extent_tree(struct ocfs2_extent_tree *et,
467 struct ocfs2_caching_info *ci,
468 struct buffer_head *bh)
469{
470 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_xb,
471 NULL, &ocfs2_xattr_tree_et_ops);
472}
473
474void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree *et,
475 struct ocfs2_caching_info *ci,
476 struct ocfs2_xattr_value_buf *vb)
477{
478 __ocfs2_init_extent_tree(et, ci, vb->vb_bh, vb->vb_access, vb,
479 &ocfs2_xattr_value_et_ops);
480}
481
482void ocfs2_init_dx_root_extent_tree(struct ocfs2_extent_tree *et,
483 struct ocfs2_caching_info *ci,
484 struct buffer_head *bh)
485{
486 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_dr,
487 NULL, &ocfs2_dx_root_et_ops);
488}
489
490void ocfs2_init_refcount_extent_tree(struct ocfs2_extent_tree *et,
491 struct ocfs2_caching_info *ci,
492 struct buffer_head *bh)
493{
494 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_rb,
495 NULL, &ocfs2_refcount_tree_et_ops);
496}
497
498static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree *et,
499 u64 new_last_eb_blk)
500{
501 et->et_ops->eo_set_last_eb_blk(et, new_last_eb_blk);
502}
503
504static inline u64 ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree *et)
505{
506 return et->et_ops->eo_get_last_eb_blk(et);
507}
508
509static inline void ocfs2_et_update_clusters(struct ocfs2_extent_tree *et,
510 u32 clusters)
511{
512 et->et_ops->eo_update_clusters(et, clusters);
513}
514
515static inline void ocfs2_et_extent_map_insert(struct ocfs2_extent_tree *et,
516 struct ocfs2_extent_rec *rec)
517{
518 if (et->et_ops->eo_extent_map_insert)
519 et->et_ops->eo_extent_map_insert(et, rec);
520}
521
522static inline void ocfs2_et_extent_map_truncate(struct ocfs2_extent_tree *et,
523 u32 clusters)
524{
525 if (et->et_ops->eo_extent_map_truncate)
526 et->et_ops->eo_extent_map_truncate(et, clusters);
527}
528
529static inline int ocfs2_et_root_journal_access(handle_t *handle,
530 struct ocfs2_extent_tree *et,
531 int type)
532{
533 return et->et_root_journal_access(handle, et->et_ci, et->et_root_bh,
534 type);
535}
536
537static inline enum ocfs2_contig_type
538 ocfs2_et_extent_contig(struct ocfs2_extent_tree *et,
539 struct ocfs2_extent_rec *rec,
540 struct ocfs2_extent_rec *insert_rec)
541{
542 if (et->et_ops->eo_extent_contig)
543 return et->et_ops->eo_extent_contig(et, rec, insert_rec);
544
545 return ocfs2_extent_rec_contig(
546 ocfs2_metadata_cache_get_super(et->et_ci),
547 rec, insert_rec);
548}
549
550static inline int ocfs2_et_insert_check(struct ocfs2_extent_tree *et,
551 struct ocfs2_extent_rec *rec)
552{
553 int ret = 0;
554
555 if (et->et_ops->eo_insert_check)
556 ret = et->et_ops->eo_insert_check(et, rec);
557 return ret;
558}
559
560static inline int ocfs2_et_sanity_check(struct ocfs2_extent_tree *et)
561{
562 int ret = 0;
563
564 if (et->et_ops->eo_sanity_check)
565 ret = et->et_ops->eo_sanity_check(et);
566 return ret;
567}
568
569static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
570 struct ocfs2_extent_block *eb);
571static void ocfs2_adjust_rightmost_records(handle_t *handle,
572 struct ocfs2_extent_tree *et,
573 struct ocfs2_path *path,
574 struct ocfs2_extent_rec *insert_rec);
575/*
576 * Reset the actual path elements so that we can re-use the structure
577 * to build another path. Generally, this involves freeing the buffer
578 * heads.
579 */
580void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root)
581{
582 int i, start = 0, depth = 0;
583 struct ocfs2_path_item *node;
584
585 if (keep_root)
586 start = 1;
587
588 for(i = start; i < path_num_items(path); i++) {
589 node = &path->p_node[i];
590
591 brelse(node->bh);
592 node->bh = NULL;
593 node->el = NULL;
594 }
595
596 /*
597 * Tree depth may change during truncate, or insert. If we're
598 * keeping the root extent list, then make sure that our path
599 * structure reflects the proper depth.
600 */
601 if (keep_root)
602 depth = le16_to_cpu(path_root_el(path)->l_tree_depth);
603 else
604 path_root_access(path) = NULL;
605
606 path->p_tree_depth = depth;
607}
608
609void ocfs2_free_path(struct ocfs2_path *path)
610{
611 if (path) {
612 ocfs2_reinit_path(path, 0);
613 kfree(path);
614 }
615}
616
617/*
618 * All the elements of src into dest. After this call, src could be freed
619 * without affecting dest.
620 *
621 * Both paths should have the same root. Any non-root elements of dest
622 * will be freed.
623 */
624static void ocfs2_cp_path(struct ocfs2_path *dest, struct ocfs2_path *src)
625{
626 int i;
627
628 BUG_ON(path_root_bh(dest) != path_root_bh(src));
629 BUG_ON(path_root_el(dest) != path_root_el(src));
630 BUG_ON(path_root_access(dest) != path_root_access(src));
631
632 ocfs2_reinit_path(dest, 1);
633
634 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
635 dest->p_node[i].bh = src->p_node[i].bh;
636 dest->p_node[i].el = src->p_node[i].el;
637
638 if (dest->p_node[i].bh)
639 get_bh(dest->p_node[i].bh);
640 }
641}
642
643/*
644 * Make the *dest path the same as src and re-initialize src path to
645 * have a root only.
646 */
647static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src)
648{
649 int i;
650
651 BUG_ON(path_root_bh(dest) != path_root_bh(src));
652 BUG_ON(path_root_access(dest) != path_root_access(src));
653
654 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
655 brelse(dest->p_node[i].bh);
656
657 dest->p_node[i].bh = src->p_node[i].bh;
658 dest->p_node[i].el = src->p_node[i].el;
659
660 src->p_node[i].bh = NULL;
661 src->p_node[i].el = NULL;
662 }
663}
664
665/*
666 * Insert an extent block at given index.
667 *
668 * This will not take an additional reference on eb_bh.
669 */
670static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index,
671 struct buffer_head *eb_bh)
672{
673 struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data;
674
675 /*
676 * Right now, no root bh is an extent block, so this helps
677 * catch code errors with dinode trees. The assertion can be
678 * safely removed if we ever need to insert extent block
679 * structures at the root.
680 */
681 BUG_ON(index == 0);
682
683 path->p_node[index].bh = eb_bh;
684 path->p_node[index].el = &eb->h_list;
685}
686
687static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh,
688 struct ocfs2_extent_list *root_el,
689 ocfs2_journal_access_func access)
690{
691 struct ocfs2_path *path;
692
693 BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH);
694
695 path = kzalloc(sizeof(*path), GFP_NOFS);
696 if (path) {
697 path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth);
698 get_bh(root_bh);
699 path_root_bh(path) = root_bh;
700 path_root_el(path) = root_el;
701 path_root_access(path) = access;
702 }
703
704 return path;
705}
706
707struct ocfs2_path *ocfs2_new_path_from_path(struct ocfs2_path *path)
708{
709 return ocfs2_new_path(path_root_bh(path), path_root_el(path),
710 path_root_access(path));
711}
712
713struct ocfs2_path *ocfs2_new_path_from_et(struct ocfs2_extent_tree *et)
714{
715 return ocfs2_new_path(et->et_root_bh, et->et_root_el,
716 et->et_root_journal_access);
717}
718
719/*
720 * Journal the buffer at depth idx. All idx>0 are extent_blocks,
721 * otherwise it's the root_access function.
722 *
723 * I don't like the way this function's name looks next to
724 * ocfs2_journal_access_path(), but I don't have a better one.
725 */
726int ocfs2_path_bh_journal_access(handle_t *handle,
727 struct ocfs2_caching_info *ci,
728 struct ocfs2_path *path,
729 int idx)
730{
731 ocfs2_journal_access_func access = path_root_access(path);
732
733 if (!access)
734 access = ocfs2_journal_access;
735
736 if (idx)
737 access = ocfs2_journal_access_eb;
738
739 return access(handle, ci, path->p_node[idx].bh,
740 OCFS2_JOURNAL_ACCESS_WRITE);
741}
742
743/*
744 * Convenience function to journal all components in a path.
745 */
746int ocfs2_journal_access_path(struct ocfs2_caching_info *ci,
747 handle_t *handle,
748 struct ocfs2_path *path)
749{
750 int i, ret = 0;
751
752 if (!path)
753 goto out;
754
755 for(i = 0; i < path_num_items(path); i++) {
756 ret = ocfs2_path_bh_journal_access(handle, ci, path, i);
757 if (ret < 0) {
758 mlog_errno(ret);
759 goto out;
760 }
761 }
762
763out:
764 return ret;
765}
766
767/*
768 * Return the index of the extent record which contains cluster #v_cluster.
769 * -1 is returned if it was not found.
770 *
771 * Should work fine on interior and exterior nodes.
772 */
773int ocfs2_search_extent_list(struct ocfs2_extent_list *el, u32 v_cluster)
774{
775 int ret = -1;
776 int i;
777 struct ocfs2_extent_rec *rec;
778 u32 rec_end, rec_start, clusters;
779
780 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
781 rec = &el->l_recs[i];
782
783 rec_start = le32_to_cpu(rec->e_cpos);
784 clusters = ocfs2_rec_clusters(el, rec);
785
786 rec_end = rec_start + clusters;
787
788 if (v_cluster >= rec_start && v_cluster < rec_end) {
789 ret = i;
790 break;
791 }
792 }
793
794 return ret;
795}
796
797/*
798 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
799 * ocfs2_extent_rec_contig only work properly against leaf nodes!
800 */
801static int ocfs2_block_extent_contig(struct super_block *sb,
802 struct ocfs2_extent_rec *ext,
803 u64 blkno)
804{
805 u64 blk_end = le64_to_cpu(ext->e_blkno);
806
807 blk_end += ocfs2_clusters_to_blocks(sb,
808 le16_to_cpu(ext->e_leaf_clusters));
809
810 return blkno == blk_end;
811}
812
813static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
814 struct ocfs2_extent_rec *right)
815{
816 u32 left_range;
817
818 left_range = le32_to_cpu(left->e_cpos) +
819 le16_to_cpu(left->e_leaf_clusters);
820
821 return (left_range == le32_to_cpu(right->e_cpos));
822}
823
824static enum ocfs2_contig_type
825 ocfs2_extent_rec_contig(struct super_block *sb,
826 struct ocfs2_extent_rec *ext,
827 struct ocfs2_extent_rec *insert_rec)
828{
829 u64 blkno = le64_to_cpu(insert_rec->e_blkno);
830
831 /*
832 * Refuse to coalesce extent records with different flag
833 * fields - we don't want to mix unwritten extents with user
834 * data.
835 */
836 if (ext->e_flags != insert_rec->e_flags)
837 return CONTIG_NONE;
838
839 if (ocfs2_extents_adjacent(ext, insert_rec) &&
840 ocfs2_block_extent_contig(sb, ext, blkno))
841 return CONTIG_RIGHT;
842
843 blkno = le64_to_cpu(ext->e_blkno);
844 if (ocfs2_extents_adjacent(insert_rec, ext) &&
845 ocfs2_block_extent_contig(sb, insert_rec, blkno))
846 return CONTIG_LEFT;
847
848 return CONTIG_NONE;
849}
850
851/*
852 * NOTE: We can have pretty much any combination of contiguousness and
853 * appending.
854 *
855 * The usefulness of APPEND_TAIL is more in that it lets us know that
856 * we'll have to update the path to that leaf.
857 */
858enum ocfs2_append_type {
859 APPEND_NONE = 0,
860 APPEND_TAIL,
861};
862
863enum ocfs2_split_type {
864 SPLIT_NONE = 0,
865 SPLIT_LEFT,
866 SPLIT_RIGHT,
867};
868
869struct ocfs2_insert_type {
870 enum ocfs2_split_type ins_split;
871 enum ocfs2_append_type ins_appending;
872 enum ocfs2_contig_type ins_contig;
873 int ins_contig_index;
874 int ins_tree_depth;
875};
876
877struct ocfs2_merge_ctxt {
878 enum ocfs2_contig_type c_contig_type;
879 int c_has_empty_extent;
880 int c_split_covers_rec;
881};
882
883static int ocfs2_validate_extent_block(struct super_block *sb,
884 struct buffer_head *bh)
885{
886 int rc;
887 struct ocfs2_extent_block *eb =
888 (struct ocfs2_extent_block *)bh->b_data;
889
890 trace_ocfs2_validate_extent_block((unsigned long long)bh->b_blocknr);
891
892 BUG_ON(!buffer_uptodate(bh));
893
894 /*
895 * If the ecc fails, we return the error but otherwise
896 * leave the filesystem running. We know any error is
897 * local to this block.
898 */
899 rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &eb->h_check);
900 if (rc) {
901 mlog(ML_ERROR, "Checksum failed for extent block %llu\n",
902 (unsigned long long)bh->b_blocknr);
903 return rc;
904 }
905
906 /*
907 * Errors after here are fatal.
908 */
909
910 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
911 rc = ocfs2_error(sb,
912 "Extent block #%llu has bad signature %.*s\n",
913 (unsigned long long)bh->b_blocknr, 7,
914 eb->h_signature);
915 goto bail;
916 }
917
918 if (le64_to_cpu(eb->h_blkno) != bh->b_blocknr) {
919 rc = ocfs2_error(sb,
920 "Extent block #%llu has an invalid h_blkno of %llu\n",
921 (unsigned long long)bh->b_blocknr,
922 (unsigned long long)le64_to_cpu(eb->h_blkno));
923 goto bail;
924 }
925
926 if (le32_to_cpu(eb->h_fs_generation) != OCFS2_SB(sb)->fs_generation) {
927 rc = ocfs2_error(sb,
928 "Extent block #%llu has an invalid h_fs_generation of #%u\n",
929 (unsigned long long)bh->b_blocknr,
930 le32_to_cpu(eb->h_fs_generation));
931 goto bail;
932 }
933bail:
934 return rc;
935}
936
937int ocfs2_read_extent_block(struct ocfs2_caching_info *ci, u64 eb_blkno,
938 struct buffer_head **bh)
939{
940 int rc;
941 struct buffer_head *tmp = *bh;
942
943 rc = ocfs2_read_block(ci, eb_blkno, &tmp,
944 ocfs2_validate_extent_block);
945
946 /* If ocfs2_read_block() got us a new bh, pass it up. */
947 if (!rc && !*bh)
948 *bh = tmp;
949
950 return rc;
951}
952
953
954/*
955 * How many free extents have we got before we need more meta data?
956 */
957int ocfs2_num_free_extents(struct ocfs2_super *osb,
958 struct ocfs2_extent_tree *et)
959{
960 int retval;
961 struct ocfs2_extent_list *el = NULL;
962 struct ocfs2_extent_block *eb;
963 struct buffer_head *eb_bh = NULL;
964 u64 last_eb_blk = 0;
965
966 el = et->et_root_el;
967 last_eb_blk = ocfs2_et_get_last_eb_blk(et);
968
969 if (last_eb_blk) {
970 retval = ocfs2_read_extent_block(et->et_ci, last_eb_blk,
971 &eb_bh);
972 if (retval < 0) {
973 mlog_errno(retval);
974 goto bail;
975 }
976 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
977 el = &eb->h_list;
978 }
979
980 BUG_ON(el->l_tree_depth != 0);
981
982 retval = le16_to_cpu(el->l_count) - le16_to_cpu(el->l_next_free_rec);
983bail:
984 brelse(eb_bh);
985
986 trace_ocfs2_num_free_extents(retval);
987 return retval;
988}
989
990/* expects array to already be allocated
991 *
992 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
993 * l_count for you
994 */
995static int ocfs2_create_new_meta_bhs(handle_t *handle,
996 struct ocfs2_extent_tree *et,
997 int wanted,
998 struct ocfs2_alloc_context *meta_ac,
999 struct buffer_head *bhs[])
1000{
1001 int count, status, i;
1002 u16 suballoc_bit_start;
1003 u32 num_got;
1004 u64 suballoc_loc, first_blkno;
1005 struct ocfs2_super *osb =
1006 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
1007 struct ocfs2_extent_block *eb;
1008
1009 count = 0;
1010 while (count < wanted) {
1011 status = ocfs2_claim_metadata(handle,
1012 meta_ac,
1013 wanted - count,
1014 &suballoc_loc,
1015 &suballoc_bit_start,
1016 &num_got,
1017 &first_blkno);
1018 if (status < 0) {
1019 mlog_errno(status);
1020 goto bail;
1021 }
1022
1023 for(i = count; i < (num_got + count); i++) {
1024 bhs[i] = sb_getblk(osb->sb, first_blkno);
1025 if (bhs[i] == NULL) {
1026 status = -ENOMEM;
1027 mlog_errno(status);
1028 goto bail;
1029 }
1030 ocfs2_set_new_buffer_uptodate(et->et_ci, bhs[i]);
1031
1032 status = ocfs2_journal_access_eb(handle, et->et_ci,
1033 bhs[i],
1034 OCFS2_JOURNAL_ACCESS_CREATE);
1035 if (status < 0) {
1036 mlog_errno(status);
1037 goto bail;
1038 }
1039
1040 memset(bhs[i]->b_data, 0, osb->sb->s_blocksize);
1041 eb = (struct ocfs2_extent_block *) bhs[i]->b_data;
1042 /* Ok, setup the minimal stuff here. */
1043 strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
1044 eb->h_blkno = cpu_to_le64(first_blkno);
1045 eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
1046 eb->h_suballoc_slot =
1047 cpu_to_le16(meta_ac->ac_alloc_slot);
1048 eb->h_suballoc_loc = cpu_to_le64(suballoc_loc);
1049 eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1050 eb->h_list.l_count =
1051 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
1052
1053 suballoc_bit_start++;
1054 first_blkno++;
1055
1056 /* We'll also be dirtied by the caller, so
1057 * this isn't absolutely necessary. */
1058 ocfs2_journal_dirty(handle, bhs[i]);
1059 }
1060
1061 count += num_got;
1062 }
1063
1064 status = 0;
1065bail:
1066 if (status < 0) {
1067 for(i = 0; i < wanted; i++) {
1068 brelse(bhs[i]);
1069 bhs[i] = NULL;
1070 }
1071 mlog_errno(status);
1072 }
1073 return status;
1074}
1075
1076/*
1077 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
1078 *
1079 * Returns the sum of the rightmost extent rec logical offset and
1080 * cluster count.
1081 *
1082 * ocfs2_add_branch() uses this to determine what logical cluster
1083 * value should be populated into the leftmost new branch records.
1084 *
1085 * ocfs2_shift_tree_depth() uses this to determine the # clusters
1086 * value for the new topmost tree record.
1087 */
1088static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list *el)
1089{
1090 int i;
1091
1092 i = le16_to_cpu(el->l_next_free_rec) - 1;
1093
1094 return le32_to_cpu(el->l_recs[i].e_cpos) +
1095 ocfs2_rec_clusters(el, &el->l_recs[i]);
1096}
1097
1098/*
1099 * Change range of the branches in the right most path according to the leaf
1100 * extent block's rightmost record.
1101 */
1102static int ocfs2_adjust_rightmost_branch(handle_t *handle,
1103 struct ocfs2_extent_tree *et)
1104{
1105 int status;
1106 struct ocfs2_path *path = NULL;
1107 struct ocfs2_extent_list *el;
1108 struct ocfs2_extent_rec *rec;
1109
1110 path = ocfs2_new_path_from_et(et);
1111 if (!path) {
1112 status = -ENOMEM;
1113 return status;
1114 }
1115
1116 status = ocfs2_find_path(et->et_ci, path, UINT_MAX);
1117 if (status < 0) {
1118 mlog_errno(status);
1119 goto out;
1120 }
1121
1122 status = ocfs2_extend_trans(handle, path_num_items(path));
1123 if (status < 0) {
1124 mlog_errno(status);
1125 goto out;
1126 }
1127
1128 status = ocfs2_journal_access_path(et->et_ci, handle, path);
1129 if (status < 0) {
1130 mlog_errno(status);
1131 goto out;
1132 }
1133
1134 el = path_leaf_el(path);
1135 rec = &el->l_recs[le16_to_cpu(el->l_next_free_rec) - 1];
1136
1137 ocfs2_adjust_rightmost_records(handle, et, path, rec);
1138
1139out:
1140 ocfs2_free_path(path);
1141 return status;
1142}
1143
1144/*
1145 * Add an entire tree branch to our inode. eb_bh is the extent block
1146 * to start at, if we don't want to start the branch at the root
1147 * structure.
1148 *
1149 * last_eb_bh is required as we have to update it's next_leaf pointer
1150 * for the new last extent block.
1151 *
1152 * the new branch will be 'empty' in the sense that every block will
1153 * contain a single record with cluster count == 0.
1154 */
1155static int ocfs2_add_branch(handle_t *handle,
1156 struct ocfs2_extent_tree *et,
1157 struct buffer_head *eb_bh,
1158 struct buffer_head **last_eb_bh,
1159 struct ocfs2_alloc_context *meta_ac)
1160{
1161 int status, new_blocks, i;
1162 u64 next_blkno, new_last_eb_blk;
1163 struct buffer_head *bh;
1164 struct buffer_head **new_eb_bhs = NULL;
1165 struct ocfs2_extent_block *eb;
1166 struct ocfs2_extent_list *eb_el;
1167 struct ocfs2_extent_list *el;
1168 u32 new_cpos, root_end;
1169
1170 BUG_ON(!last_eb_bh || !*last_eb_bh);
1171
1172 if (eb_bh) {
1173 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1174 el = &eb->h_list;
1175 } else
1176 el = et->et_root_el;
1177
1178 /* we never add a branch to a leaf. */
1179 BUG_ON(!el->l_tree_depth);
1180
1181 new_blocks = le16_to_cpu(el->l_tree_depth);
1182
1183 eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
1184 new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
1185 root_end = ocfs2_sum_rightmost_rec(et->et_root_el);
1186
1187 /*
1188 * If there is a gap before the root end and the real end
1189 * of the righmost leaf block, we need to remove the gap
1190 * between new_cpos and root_end first so that the tree
1191 * is consistent after we add a new branch(it will start
1192 * from new_cpos).
1193 */
1194 if (root_end > new_cpos) {
1195 trace_ocfs2_adjust_rightmost_branch(
1196 (unsigned long long)
1197 ocfs2_metadata_cache_owner(et->et_ci),
1198 root_end, new_cpos);
1199
1200 status = ocfs2_adjust_rightmost_branch(handle, et);
1201 if (status) {
1202 mlog_errno(status);
1203 goto bail;
1204 }
1205 }
1206
1207 /* allocate the number of new eb blocks we need */
1208 new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
1209 GFP_KERNEL);
1210 if (!new_eb_bhs) {
1211 status = -ENOMEM;
1212 mlog_errno(status);
1213 goto bail;
1214 }
1215
1216 status = ocfs2_create_new_meta_bhs(handle, et, new_blocks,
1217 meta_ac, new_eb_bhs);
1218 if (status < 0) {
1219 mlog_errno(status);
1220 goto bail;
1221 }
1222
1223 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
1224 * linked with the rest of the tree.
1225 * conversly, new_eb_bhs[0] is the new bottommost leaf.
1226 *
1227 * when we leave the loop, new_last_eb_blk will point to the
1228 * newest leaf, and next_blkno will point to the topmost extent
1229 * block. */
1230 next_blkno = new_last_eb_blk = 0;
1231 for(i = 0; i < new_blocks; i++) {
1232 bh = new_eb_bhs[i];
1233 eb = (struct ocfs2_extent_block *) bh->b_data;
1234 /* ocfs2_create_new_meta_bhs() should create it right! */
1235 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1236 eb_el = &eb->h_list;
1237
1238 status = ocfs2_journal_access_eb(handle, et->et_ci, bh,
1239 OCFS2_JOURNAL_ACCESS_CREATE);
1240 if (status < 0) {
1241 mlog_errno(status);
1242 goto bail;
1243 }
1244
1245 eb->h_next_leaf_blk = 0;
1246 eb_el->l_tree_depth = cpu_to_le16(i);
1247 eb_el->l_next_free_rec = cpu_to_le16(1);
1248 /*
1249 * This actually counts as an empty extent as
1250 * c_clusters == 0
1251 */
1252 eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
1253 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
1254 /*
1255 * eb_el isn't always an interior node, but even leaf
1256 * nodes want a zero'd flags and reserved field so
1257 * this gets the whole 32 bits regardless of use.
1258 */
1259 eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
1260 if (!eb_el->l_tree_depth)
1261 new_last_eb_blk = le64_to_cpu(eb->h_blkno);
1262
1263 ocfs2_journal_dirty(handle, bh);
1264 next_blkno = le64_to_cpu(eb->h_blkno);
1265 }
1266
1267 /* This is a bit hairy. We want to update up to three blocks
1268 * here without leaving any of them in an inconsistent state
1269 * in case of error. We don't have to worry about
1270 * journal_dirty erroring as it won't unless we've aborted the
1271 * handle (in which case we would never be here) so reserving
1272 * the write with journal_access is all we need to do. */
1273 status = ocfs2_journal_access_eb(handle, et->et_ci, *last_eb_bh,
1274 OCFS2_JOURNAL_ACCESS_WRITE);
1275 if (status < 0) {
1276 mlog_errno(status);
1277 goto bail;
1278 }
1279 status = ocfs2_et_root_journal_access(handle, et,
1280 OCFS2_JOURNAL_ACCESS_WRITE);
1281 if (status < 0) {
1282 mlog_errno(status);
1283 goto bail;
1284 }
1285 if (eb_bh) {
1286 status = ocfs2_journal_access_eb(handle, et->et_ci, eb_bh,
1287 OCFS2_JOURNAL_ACCESS_WRITE);
1288 if (status < 0) {
1289 mlog_errno(status);
1290 goto bail;
1291 }
1292 }
1293
1294 /* Link the new branch into the rest of the tree (el will
1295 * either be on the root_bh, or the extent block passed in. */
1296 i = le16_to_cpu(el->l_next_free_rec);
1297 el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
1298 el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1299 el->l_recs[i].e_int_clusters = 0;
1300 le16_add_cpu(&el->l_next_free_rec, 1);
1301
1302 /* fe needs a new last extent block pointer, as does the
1303 * next_leaf on the previously last-extent-block. */
1304 ocfs2_et_set_last_eb_blk(et, new_last_eb_blk);
1305
1306 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
1307 eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);
1308
1309 ocfs2_journal_dirty(handle, *last_eb_bh);
1310 ocfs2_journal_dirty(handle, et->et_root_bh);
1311 if (eb_bh)
1312 ocfs2_journal_dirty(handle, eb_bh);
1313
1314 /*
1315 * Some callers want to track the rightmost leaf so pass it
1316 * back here.
1317 */
1318 brelse(*last_eb_bh);
1319 get_bh(new_eb_bhs[0]);
1320 *last_eb_bh = new_eb_bhs[0];
1321
1322 status = 0;
1323bail:
1324 if (new_eb_bhs) {
1325 for (i = 0; i < new_blocks; i++)
1326 brelse(new_eb_bhs[i]);
1327 kfree(new_eb_bhs);
1328 }
1329
1330 return status;
1331}
1332
1333/*
1334 * adds another level to the allocation tree.
1335 * returns back the new extent block so you can add a branch to it
1336 * after this call.
1337 */
1338static int ocfs2_shift_tree_depth(handle_t *handle,
1339 struct ocfs2_extent_tree *et,
1340 struct ocfs2_alloc_context *meta_ac,
1341 struct buffer_head **ret_new_eb_bh)
1342{
1343 int status, i;
1344 u32 new_clusters;
1345 struct buffer_head *new_eb_bh = NULL;
1346 struct ocfs2_extent_block *eb;
1347 struct ocfs2_extent_list *root_el;
1348 struct ocfs2_extent_list *eb_el;
1349
1350 status = ocfs2_create_new_meta_bhs(handle, et, 1, meta_ac,
1351 &new_eb_bh);
1352 if (status < 0) {
1353 mlog_errno(status);
1354 goto bail;
1355 }
1356
1357 eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
1358 /* ocfs2_create_new_meta_bhs() should create it right! */
1359 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1360
1361 eb_el = &eb->h_list;
1362 root_el = et->et_root_el;
1363
1364 status = ocfs2_journal_access_eb(handle, et->et_ci, new_eb_bh,
1365 OCFS2_JOURNAL_ACCESS_CREATE);
1366 if (status < 0) {
1367 mlog_errno(status);
1368 goto bail;
1369 }
1370
1371 /* copy the root extent list data into the new extent block */
1372 eb_el->l_tree_depth = root_el->l_tree_depth;
1373 eb_el->l_next_free_rec = root_el->l_next_free_rec;
1374 for (i = 0; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1375 eb_el->l_recs[i] = root_el->l_recs[i];
1376
1377 ocfs2_journal_dirty(handle, new_eb_bh);
1378
1379 status = ocfs2_et_root_journal_access(handle, et,
1380 OCFS2_JOURNAL_ACCESS_WRITE);
1381 if (status < 0) {
1382 mlog_errno(status);
1383 goto bail;
1384 }
1385
1386 new_clusters = ocfs2_sum_rightmost_rec(eb_el);
1387
1388 /* update root_bh now */
1389 le16_add_cpu(&root_el->l_tree_depth, 1);
1390 root_el->l_recs[0].e_cpos = 0;
1391 root_el->l_recs[0].e_blkno = eb->h_blkno;
1392 root_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
1393 for (i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1394 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
1395 root_el->l_next_free_rec = cpu_to_le16(1);
1396
1397 /* If this is our 1st tree depth shift, then last_eb_blk
1398 * becomes the allocated extent block */
1399 if (root_el->l_tree_depth == cpu_to_le16(1))
1400 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
1401
1402 ocfs2_journal_dirty(handle, et->et_root_bh);
1403
1404 *ret_new_eb_bh = new_eb_bh;
1405 new_eb_bh = NULL;
1406 status = 0;
1407bail:
1408 brelse(new_eb_bh);
1409
1410 return status;
1411}
1412
1413/*
1414 * Should only be called when there is no space left in any of the
1415 * leaf nodes. What we want to do is find the lowest tree depth
1416 * non-leaf extent block with room for new records. There are three
1417 * valid results of this search:
1418 *
1419 * 1) a lowest extent block is found, then we pass it back in
1420 * *lowest_eb_bh and return '0'
1421 *
1422 * 2) the search fails to find anything, but the root_el has room. We
1423 * pass NULL back in *lowest_eb_bh, but still return '0'
1424 *
1425 * 3) the search fails to find anything AND the root_el is full, in
1426 * which case we return > 0
1427 *
1428 * return status < 0 indicates an error.
1429 */
1430static int ocfs2_find_branch_target(struct ocfs2_extent_tree *et,
1431 struct buffer_head **target_bh)
1432{
1433 int status = 0, i;
1434 u64 blkno;
1435 struct ocfs2_extent_block *eb;
1436 struct ocfs2_extent_list *el;
1437 struct buffer_head *bh = NULL;
1438 struct buffer_head *lowest_bh = NULL;
1439
1440 *target_bh = NULL;
1441
1442 el = et->et_root_el;
1443
1444 while(le16_to_cpu(el->l_tree_depth) > 1) {
1445 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1446 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1447 "Owner %llu has empty extent list (next_free_rec == 0)\n",
1448 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
1449 status = -EIO;
1450 goto bail;
1451 }
1452 i = le16_to_cpu(el->l_next_free_rec) - 1;
1453 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1454 if (!blkno) {
1455 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1456 "Owner %llu has extent list where extent # %d has no physical block start\n",
1457 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), i);
1458 status = -EIO;
1459 goto bail;
1460 }
1461
1462 brelse(bh);
1463 bh = NULL;
1464
1465 status = ocfs2_read_extent_block(et->et_ci, blkno, &bh);
1466 if (status < 0) {
1467 mlog_errno(status);
1468 goto bail;
1469 }
1470
1471 eb = (struct ocfs2_extent_block *) bh->b_data;
1472 el = &eb->h_list;
1473
1474 if (le16_to_cpu(el->l_next_free_rec) <
1475 le16_to_cpu(el->l_count)) {
1476 brelse(lowest_bh);
1477 lowest_bh = bh;
1478 get_bh(lowest_bh);
1479 }
1480 }
1481
1482 /* If we didn't find one and the fe doesn't have any room,
1483 * then return '1' */
1484 el = et->et_root_el;
1485 if (!lowest_bh && (el->l_next_free_rec == el->l_count))
1486 status = 1;
1487
1488 *target_bh = lowest_bh;
1489bail:
1490 brelse(bh);
1491
1492 return status;
1493}
1494
1495/*
1496 * Grow a b-tree so that it has more records.
1497 *
1498 * We might shift the tree depth in which case existing paths should
1499 * be considered invalid.
1500 *
1501 * Tree depth after the grow is returned via *final_depth.
1502 *
1503 * *last_eb_bh will be updated by ocfs2_add_branch().
1504 */
1505static int ocfs2_grow_tree(handle_t *handle, struct ocfs2_extent_tree *et,
1506 int *final_depth, struct buffer_head **last_eb_bh,
1507 struct ocfs2_alloc_context *meta_ac)
1508{
1509 int ret, shift;
1510 struct ocfs2_extent_list *el = et->et_root_el;
1511 int depth = le16_to_cpu(el->l_tree_depth);
1512 struct buffer_head *bh = NULL;
1513
1514 BUG_ON(meta_ac == NULL);
1515
1516 shift = ocfs2_find_branch_target(et, &bh);
1517 if (shift < 0) {
1518 ret = shift;
1519 mlog_errno(ret);
1520 goto out;
1521 }
1522
1523 /* We traveled all the way to the bottom of the allocation tree
1524 * and didn't find room for any more extents - we need to add
1525 * another tree level */
1526 if (shift) {
1527 BUG_ON(bh);
1528 trace_ocfs2_grow_tree(
1529 (unsigned long long)
1530 ocfs2_metadata_cache_owner(et->et_ci),
1531 depth);
1532
1533 /* ocfs2_shift_tree_depth will return us a buffer with
1534 * the new extent block (so we can pass that to
1535 * ocfs2_add_branch). */
1536 ret = ocfs2_shift_tree_depth(handle, et, meta_ac, &bh);
1537 if (ret < 0) {
1538 mlog_errno(ret);
1539 goto out;
1540 }
1541 depth++;
1542 if (depth == 1) {
1543 /*
1544 * Special case: we have room now if we shifted from
1545 * tree_depth 0, so no more work needs to be done.
1546 *
1547 * We won't be calling add_branch, so pass
1548 * back *last_eb_bh as the new leaf. At depth
1549 * zero, it should always be null so there's
1550 * no reason to brelse.
1551 */
1552 BUG_ON(*last_eb_bh);
1553 get_bh(bh);
1554 *last_eb_bh = bh;
1555 goto out;
1556 }
1557 }
1558
1559 /* call ocfs2_add_branch to add the final part of the tree with
1560 * the new data. */
1561 ret = ocfs2_add_branch(handle, et, bh, last_eb_bh,
1562 meta_ac);
1563 if (ret < 0) {
1564 mlog_errno(ret);
1565 goto out;
1566 }
1567
1568out:
1569 if (final_depth)
1570 *final_depth = depth;
1571 brelse(bh);
1572 return ret;
1573}
1574
1575/*
1576 * This function will discard the rightmost extent record.
1577 */
1578static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
1579{
1580 int next_free = le16_to_cpu(el->l_next_free_rec);
1581 int count = le16_to_cpu(el->l_count);
1582 unsigned int num_bytes;
1583
1584 BUG_ON(!next_free);
1585 /* This will cause us to go off the end of our extent list. */
1586 BUG_ON(next_free >= count);
1587
1588 num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
1589
1590 memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
1591}
1592
1593static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
1594 struct ocfs2_extent_rec *insert_rec)
1595{
1596 int i, insert_index, next_free, has_empty, num_bytes;
1597 u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
1598 struct ocfs2_extent_rec *rec;
1599
1600 next_free = le16_to_cpu(el->l_next_free_rec);
1601 has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
1602
1603 BUG_ON(!next_free);
1604
1605 /* The tree code before us didn't allow enough room in the leaf. */
1606 BUG_ON(el->l_next_free_rec == el->l_count && !has_empty);
1607
1608 /*
1609 * The easiest way to approach this is to just remove the
1610 * empty extent and temporarily decrement next_free.
1611 */
1612 if (has_empty) {
1613 /*
1614 * If next_free was 1 (only an empty extent), this
1615 * loop won't execute, which is fine. We still want
1616 * the decrement above to happen.
1617 */
1618 for(i = 0; i < (next_free - 1); i++)
1619 el->l_recs[i] = el->l_recs[i+1];
1620
1621 next_free--;
1622 }
1623
1624 /*
1625 * Figure out what the new record index should be.
1626 */
1627 for(i = 0; i < next_free; i++) {
1628 rec = &el->l_recs[i];
1629
1630 if (insert_cpos < le32_to_cpu(rec->e_cpos))
1631 break;
1632 }
1633 insert_index = i;
1634
1635 trace_ocfs2_rotate_leaf(insert_cpos, insert_index,
1636 has_empty, next_free,
1637 le16_to_cpu(el->l_count));
1638
1639 BUG_ON(insert_index < 0);
1640 BUG_ON(insert_index >= le16_to_cpu(el->l_count));
1641 BUG_ON(insert_index > next_free);
1642
1643 /*
1644 * No need to memmove if we're just adding to the tail.
1645 */
1646 if (insert_index != next_free) {
1647 BUG_ON(next_free >= le16_to_cpu(el->l_count));
1648
1649 num_bytes = next_free - insert_index;
1650 num_bytes *= sizeof(struct ocfs2_extent_rec);
1651 memmove(&el->l_recs[insert_index + 1],
1652 &el->l_recs[insert_index],
1653 num_bytes);
1654 }
1655
1656 /*
1657 * Either we had an empty extent, and need to re-increment or
1658 * there was no empty extent on a non full rightmost leaf node,
1659 * in which case we still need to increment.
1660 */
1661 next_free++;
1662 el->l_next_free_rec = cpu_to_le16(next_free);
1663 /*
1664 * Make sure none of the math above just messed up our tree.
1665 */
1666 BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
1667
1668 el->l_recs[insert_index] = *insert_rec;
1669
1670}
1671
1672static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
1673{
1674 int size, num_recs = le16_to_cpu(el->l_next_free_rec);
1675
1676 BUG_ON(num_recs == 0);
1677
1678 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
1679 num_recs--;
1680 size = num_recs * sizeof(struct ocfs2_extent_rec);
1681 memmove(&el->l_recs[0], &el->l_recs[1], size);
1682 memset(&el->l_recs[num_recs], 0,
1683 sizeof(struct ocfs2_extent_rec));
1684 el->l_next_free_rec = cpu_to_le16(num_recs);
1685 }
1686}
1687
1688/*
1689 * Create an empty extent record .
1690 *
1691 * l_next_free_rec may be updated.
1692 *
1693 * If an empty extent already exists do nothing.
1694 */
1695static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
1696{
1697 int next_free = le16_to_cpu(el->l_next_free_rec);
1698
1699 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1700
1701 if (next_free == 0)
1702 goto set_and_inc;
1703
1704 if (ocfs2_is_empty_extent(&el->l_recs[0]))
1705 return;
1706
1707 mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
1708 "Asked to create an empty extent in a full list:\n"
1709 "count = %u, tree depth = %u",
1710 le16_to_cpu(el->l_count),
1711 le16_to_cpu(el->l_tree_depth));
1712
1713 ocfs2_shift_records_right(el);
1714
1715set_and_inc:
1716 le16_add_cpu(&el->l_next_free_rec, 1);
1717 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1718}
1719
1720/*
1721 * For a rotation which involves two leaf nodes, the "root node" is
1722 * the lowest level tree node which contains a path to both leafs. This
1723 * resulting set of information can be used to form a complete "subtree"
1724 *
1725 * This function is passed two full paths from the dinode down to a
1726 * pair of adjacent leaves. It's task is to figure out which path
1727 * index contains the subtree root - this can be the root index itself
1728 * in a worst-case rotation.
1729 *
1730 * The array index of the subtree root is passed back.
1731 */
1732int ocfs2_find_subtree_root(struct ocfs2_extent_tree *et,
1733 struct ocfs2_path *left,
1734 struct ocfs2_path *right)
1735{
1736 int i = 0;
1737
1738 /*
1739 * Check that the caller passed in two paths from the same tree.
1740 */
1741 BUG_ON(path_root_bh(left) != path_root_bh(right));
1742
1743 do {
1744 i++;
1745
1746 /*
1747 * The caller didn't pass two adjacent paths.
1748 */
1749 mlog_bug_on_msg(i > left->p_tree_depth,
1750 "Owner %llu, left depth %u, right depth %u\n"
1751 "left leaf blk %llu, right leaf blk %llu\n",
1752 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
1753 left->p_tree_depth, right->p_tree_depth,
1754 (unsigned long long)path_leaf_bh(left)->b_blocknr,
1755 (unsigned long long)path_leaf_bh(right)->b_blocknr);
1756 } while (left->p_node[i].bh->b_blocknr ==
1757 right->p_node[i].bh->b_blocknr);
1758
1759 return i - 1;
1760}
1761
1762typedef void (path_insert_t)(void *, struct buffer_head *);
1763
1764/*
1765 * Traverse a btree path in search of cpos, starting at root_el.
1766 *
1767 * This code can be called with a cpos larger than the tree, in which
1768 * case it will return the rightmost path.
1769 */
1770static int __ocfs2_find_path(struct ocfs2_caching_info *ci,
1771 struct ocfs2_extent_list *root_el, u32 cpos,
1772 path_insert_t *func, void *data)
1773{
1774 int i, ret = 0;
1775 u32 range;
1776 u64 blkno;
1777 struct buffer_head *bh = NULL;
1778 struct ocfs2_extent_block *eb;
1779 struct ocfs2_extent_list *el;
1780 struct ocfs2_extent_rec *rec;
1781
1782 el = root_el;
1783 while (el->l_tree_depth) {
1784 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1785 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1786 "Owner %llu has empty extent list at depth %u\n",
1787 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1788 le16_to_cpu(el->l_tree_depth));
1789 ret = -EROFS;
1790 goto out;
1791
1792 }
1793
1794 for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
1795 rec = &el->l_recs[i];
1796
1797 /*
1798 * In the case that cpos is off the allocation
1799 * tree, this should just wind up returning the
1800 * rightmost record.
1801 */
1802 range = le32_to_cpu(rec->e_cpos) +
1803 ocfs2_rec_clusters(el, rec);
1804 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1805 break;
1806 }
1807
1808 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1809 if (blkno == 0) {
1810 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1811 "Owner %llu has bad blkno in extent list at depth %u (index %d)\n",
1812 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1813 le16_to_cpu(el->l_tree_depth), i);
1814 ret = -EROFS;
1815 goto out;
1816 }
1817
1818 brelse(bh);
1819 bh = NULL;
1820 ret = ocfs2_read_extent_block(ci, blkno, &bh);
1821 if (ret) {
1822 mlog_errno(ret);
1823 goto out;
1824 }
1825
1826 eb = (struct ocfs2_extent_block *) bh->b_data;
1827 el = &eb->h_list;
1828
1829 if (le16_to_cpu(el->l_next_free_rec) >
1830 le16_to_cpu(el->l_count)) {
1831 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1832 "Owner %llu has bad count in extent list at block %llu (next free=%u, count=%u)\n",
1833 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1834 (unsigned long long)bh->b_blocknr,
1835 le16_to_cpu(el->l_next_free_rec),
1836 le16_to_cpu(el->l_count));
1837 ret = -EROFS;
1838 goto out;
1839 }
1840
1841 if (func)
1842 func(data, bh);
1843 }
1844
1845out:
1846 /*
1847 * Catch any trailing bh that the loop didn't handle.
1848 */
1849 brelse(bh);
1850
1851 return ret;
1852}
1853
1854/*
1855 * Given an initialized path (that is, it has a valid root extent
1856 * list), this function will traverse the btree in search of the path
1857 * which would contain cpos.
1858 *
1859 * The path traveled is recorded in the path structure.
1860 *
1861 * Note that this will not do any comparisons on leaf node extent
1862 * records, so it will work fine in the case that we just added a tree
1863 * branch.
1864 */
1865struct find_path_data {
1866 int index;
1867 struct ocfs2_path *path;
1868};
1869static void find_path_ins(void *data, struct buffer_head *bh)
1870{
1871 struct find_path_data *fp = data;
1872
1873 get_bh(bh);
1874 ocfs2_path_insert_eb(fp->path, fp->index, bh);
1875 fp->index++;
1876}
1877int ocfs2_find_path(struct ocfs2_caching_info *ci,
1878 struct ocfs2_path *path, u32 cpos)
1879{
1880 struct find_path_data data;
1881
1882 data.index = 1;
1883 data.path = path;
1884 return __ocfs2_find_path(ci, path_root_el(path), cpos,
1885 find_path_ins, &data);
1886}
1887
1888static void find_leaf_ins(void *data, struct buffer_head *bh)
1889{
1890 struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1891 struct ocfs2_extent_list *el = &eb->h_list;
1892 struct buffer_head **ret = data;
1893
1894 /* We want to retain only the leaf block. */
1895 if (le16_to_cpu(el->l_tree_depth) == 0) {
1896 get_bh(bh);
1897 *ret = bh;
1898 }
1899}
1900/*
1901 * Find the leaf block in the tree which would contain cpos. No
1902 * checking of the actual leaf is done.
1903 *
1904 * Some paths want to call this instead of allocating a path structure
1905 * and calling ocfs2_find_path().
1906 *
1907 * This function doesn't handle non btree extent lists.
1908 */
1909int ocfs2_find_leaf(struct ocfs2_caching_info *ci,
1910 struct ocfs2_extent_list *root_el, u32 cpos,
1911 struct buffer_head **leaf_bh)
1912{
1913 int ret;
1914 struct buffer_head *bh = NULL;
1915
1916 ret = __ocfs2_find_path(ci, root_el, cpos, find_leaf_ins, &bh);
1917 if (ret) {
1918 mlog_errno(ret);
1919 goto out;
1920 }
1921
1922 *leaf_bh = bh;
1923out:
1924 return ret;
1925}
1926
1927/*
1928 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1929 *
1930 * Basically, we've moved stuff around at the bottom of the tree and
1931 * we need to fix up the extent records above the changes to reflect
1932 * the new changes.
1933 *
1934 * left_rec: the record on the left.
1935 * left_child_el: is the child list pointed to by left_rec
1936 * right_rec: the record to the right of left_rec
1937 * right_child_el: is the child list pointed to by right_rec
1938 *
1939 * By definition, this only works on interior nodes.
1940 */
1941static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1942 struct ocfs2_extent_list *left_child_el,
1943 struct ocfs2_extent_rec *right_rec,
1944 struct ocfs2_extent_list *right_child_el)
1945{
1946 u32 left_clusters, right_end;
1947
1948 /*
1949 * Interior nodes never have holes. Their cpos is the cpos of
1950 * the leftmost record in their child list. Their cluster
1951 * count covers the full theoretical range of their child list
1952 * - the range between their cpos and the cpos of the record
1953 * immediately to their right.
1954 */
1955 left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1956 if (!ocfs2_rec_clusters(right_child_el, &right_child_el->l_recs[0])) {
1957 BUG_ON(right_child_el->l_tree_depth);
1958 BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
1959 left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
1960 }
1961 left_clusters -= le32_to_cpu(left_rec->e_cpos);
1962 left_rec->e_int_clusters = cpu_to_le32(left_clusters);
1963
1964 /*
1965 * Calculate the rightmost cluster count boundary before
1966 * moving cpos - we will need to adjust clusters after
1967 * updating e_cpos to keep the same highest cluster count.
1968 */
1969 right_end = le32_to_cpu(right_rec->e_cpos);
1970 right_end += le32_to_cpu(right_rec->e_int_clusters);
1971
1972 right_rec->e_cpos = left_rec->e_cpos;
1973 le32_add_cpu(&right_rec->e_cpos, left_clusters);
1974
1975 right_end -= le32_to_cpu(right_rec->e_cpos);
1976 right_rec->e_int_clusters = cpu_to_le32(right_end);
1977}
1978
1979/*
1980 * Adjust the adjacent root node records involved in a
1981 * rotation. left_el_blkno is passed in as a key so that we can easily
1982 * find it's index in the root list.
1983 */
1984static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
1985 struct ocfs2_extent_list *left_el,
1986 struct ocfs2_extent_list *right_el,
1987 u64 left_el_blkno)
1988{
1989 int i;
1990
1991 BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
1992 le16_to_cpu(left_el->l_tree_depth));
1993
1994 for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
1995 if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
1996 break;
1997 }
1998
1999 /*
2000 * The path walking code should have never returned a root and
2001 * two paths which are not adjacent.
2002 */
2003 BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
2004
2005 ocfs2_adjust_adjacent_records(&root_el->l_recs[i], left_el,
2006 &root_el->l_recs[i + 1], right_el);
2007}
2008
2009/*
2010 * We've changed a leaf block (in right_path) and need to reflect that
2011 * change back up the subtree.
2012 *
2013 * This happens in multiple places:
2014 * - When we've moved an extent record from the left path leaf to the right
2015 * path leaf to make room for an empty extent in the left path leaf.
2016 * - When our insert into the right path leaf is at the leftmost edge
2017 * and requires an update of the path immediately to it's left. This
2018 * can occur at the end of some types of rotation and appending inserts.
2019 * - When we've adjusted the last extent record in the left path leaf and the
2020 * 1st extent record in the right path leaf during cross extent block merge.
2021 */
2022static void ocfs2_complete_edge_insert(handle_t *handle,
2023 struct ocfs2_path *left_path,
2024 struct ocfs2_path *right_path,
2025 int subtree_index)
2026{
2027 int i, idx;
2028 struct ocfs2_extent_list *el, *left_el, *right_el;
2029 struct ocfs2_extent_rec *left_rec, *right_rec;
2030 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2031
2032 /*
2033 * Update the counts and position values within all the
2034 * interior nodes to reflect the leaf rotation we just did.
2035 *
2036 * The root node is handled below the loop.
2037 *
2038 * We begin the loop with right_el and left_el pointing to the
2039 * leaf lists and work our way up.
2040 *
2041 * NOTE: within this loop, left_el and right_el always refer
2042 * to the *child* lists.
2043 */
2044 left_el = path_leaf_el(left_path);
2045 right_el = path_leaf_el(right_path);
2046 for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
2047 trace_ocfs2_complete_edge_insert(i);
2048
2049 /*
2050 * One nice property of knowing that all of these
2051 * nodes are below the root is that we only deal with
2052 * the leftmost right node record and the rightmost
2053 * left node record.
2054 */
2055 el = left_path->p_node[i].el;
2056 idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
2057 left_rec = &el->l_recs[idx];
2058
2059 el = right_path->p_node[i].el;
2060 right_rec = &el->l_recs[0];
2061
2062 ocfs2_adjust_adjacent_records(left_rec, left_el, right_rec,
2063 right_el);
2064
2065 ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
2066 ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
2067
2068 /*
2069 * Setup our list pointers now so that the current
2070 * parents become children in the next iteration.
2071 */
2072 left_el = left_path->p_node[i].el;
2073 right_el = right_path->p_node[i].el;
2074 }
2075
2076 /*
2077 * At the root node, adjust the two adjacent records which
2078 * begin our path to the leaves.
2079 */
2080
2081 el = left_path->p_node[subtree_index].el;
2082 left_el = left_path->p_node[subtree_index + 1].el;
2083 right_el = right_path->p_node[subtree_index + 1].el;
2084
2085 ocfs2_adjust_root_records(el, left_el, right_el,
2086 left_path->p_node[subtree_index + 1].bh->b_blocknr);
2087
2088 root_bh = left_path->p_node[subtree_index].bh;
2089
2090 ocfs2_journal_dirty(handle, root_bh);
2091}
2092
2093static int ocfs2_rotate_subtree_right(handle_t *handle,
2094 struct ocfs2_extent_tree *et,
2095 struct ocfs2_path *left_path,
2096 struct ocfs2_path *right_path,
2097 int subtree_index)
2098{
2099 int ret, i;
2100 struct buffer_head *right_leaf_bh;
2101 struct buffer_head *left_leaf_bh = NULL;
2102 struct buffer_head *root_bh;
2103 struct ocfs2_extent_list *right_el, *left_el;
2104 struct ocfs2_extent_rec move_rec;
2105
2106 left_leaf_bh = path_leaf_bh(left_path);
2107 left_el = path_leaf_el(left_path);
2108
2109 if (left_el->l_next_free_rec != left_el->l_count) {
2110 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
2111 "Inode %llu has non-full interior leaf node %llu (next free = %u)\n",
2112 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2113 (unsigned long long)left_leaf_bh->b_blocknr,
2114 le16_to_cpu(left_el->l_next_free_rec));
2115 return -EROFS;
2116 }
2117
2118 /*
2119 * This extent block may already have an empty record, so we
2120 * return early if so.
2121 */
2122 if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
2123 return 0;
2124
2125 root_bh = left_path->p_node[subtree_index].bh;
2126 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2127
2128 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2129 subtree_index);
2130 if (ret) {
2131 mlog_errno(ret);
2132 goto out;
2133 }
2134
2135 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2136 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2137 right_path, i);
2138 if (ret) {
2139 mlog_errno(ret);
2140 goto out;
2141 }
2142
2143 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2144 left_path, i);
2145 if (ret) {
2146 mlog_errno(ret);
2147 goto out;
2148 }
2149 }
2150
2151 right_leaf_bh = path_leaf_bh(right_path);
2152 right_el = path_leaf_el(right_path);
2153
2154 /* This is a code error, not a disk corruption. */
2155 mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
2156 "because rightmost leaf block %llu is empty\n",
2157 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2158 (unsigned long long)right_leaf_bh->b_blocknr);
2159
2160 ocfs2_create_empty_extent(right_el);
2161
2162 ocfs2_journal_dirty(handle, right_leaf_bh);
2163
2164 /* Do the copy now. */
2165 i = le16_to_cpu(left_el->l_next_free_rec) - 1;
2166 move_rec = left_el->l_recs[i];
2167 right_el->l_recs[0] = move_rec;
2168
2169 /*
2170 * Clear out the record we just copied and shift everything
2171 * over, leaving an empty extent in the left leaf.
2172 *
2173 * We temporarily subtract from next_free_rec so that the
2174 * shift will lose the tail record (which is now defunct).
2175 */
2176 le16_add_cpu(&left_el->l_next_free_rec, -1);
2177 ocfs2_shift_records_right(left_el);
2178 memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2179 le16_add_cpu(&left_el->l_next_free_rec, 1);
2180
2181 ocfs2_journal_dirty(handle, left_leaf_bh);
2182
2183 ocfs2_complete_edge_insert(handle, left_path, right_path,
2184 subtree_index);
2185
2186out:
2187 return ret;
2188}
2189
2190/*
2191 * Given a full path, determine what cpos value would return us a path
2192 * containing the leaf immediately to the left of the current one.
2193 *
2194 * Will return zero if the path passed in is already the leftmost path.
2195 */
2196int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
2197 struct ocfs2_path *path, u32 *cpos)
2198{
2199 int i, j, ret = 0;
2200 u64 blkno;
2201 struct ocfs2_extent_list *el;
2202
2203 BUG_ON(path->p_tree_depth == 0);
2204
2205 *cpos = 0;
2206
2207 blkno = path_leaf_bh(path)->b_blocknr;
2208
2209 /* Start at the tree node just above the leaf and work our way up. */
2210 i = path->p_tree_depth - 1;
2211 while (i >= 0) {
2212 el = path->p_node[i].el;
2213
2214 /*
2215 * Find the extent record just before the one in our
2216 * path.
2217 */
2218 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2219 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2220 if (j == 0) {
2221 if (i == 0) {
2222 /*
2223 * We've determined that the
2224 * path specified is already
2225 * the leftmost one - return a
2226 * cpos of zero.
2227 */
2228 goto out;
2229 }
2230 /*
2231 * The leftmost record points to our
2232 * leaf - we need to travel up the
2233 * tree one level.
2234 */
2235 goto next_node;
2236 }
2237
2238 *cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
2239 *cpos = *cpos + ocfs2_rec_clusters(el,
2240 &el->l_recs[j - 1]);
2241 *cpos = *cpos - 1;
2242 goto out;
2243 }
2244 }
2245
2246 /*
2247 * If we got here, we never found a valid node where
2248 * the tree indicated one should be.
2249 */
2250 ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
2251 (unsigned long long)blkno);
2252 ret = -EROFS;
2253 goto out;
2254
2255next_node:
2256 blkno = path->p_node[i].bh->b_blocknr;
2257 i--;
2258 }
2259
2260out:
2261 return ret;
2262}
2263
2264/*
2265 * Extend the transaction by enough credits to complete the rotation,
2266 * and still leave at least the original number of credits allocated
2267 * to this transaction.
2268 */
2269static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
2270 int op_credits,
2271 struct ocfs2_path *path)
2272{
2273 int ret = 0;
2274 int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;
2275
2276 if (handle->h_buffer_credits < credits)
2277 ret = ocfs2_extend_trans(handle,
2278 credits - handle->h_buffer_credits);
2279
2280 return ret;
2281}
2282
2283/*
2284 * Trap the case where we're inserting into the theoretical range past
2285 * the _actual_ left leaf range. Otherwise, we'll rotate a record
2286 * whose cpos is less than ours into the right leaf.
2287 *
2288 * It's only necessary to look at the rightmost record of the left
2289 * leaf because the logic that calls us should ensure that the
2290 * theoretical ranges in the path components above the leaves are
2291 * correct.
2292 */
2293static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
2294 u32 insert_cpos)
2295{
2296 struct ocfs2_extent_list *left_el;
2297 struct ocfs2_extent_rec *rec;
2298 int next_free;
2299
2300 left_el = path_leaf_el(left_path);
2301 next_free = le16_to_cpu(left_el->l_next_free_rec);
2302 rec = &left_el->l_recs[next_free - 1];
2303
2304 if (insert_cpos > le32_to_cpu(rec->e_cpos))
2305 return 1;
2306 return 0;
2307}
2308
2309static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
2310{
2311 int next_free = le16_to_cpu(el->l_next_free_rec);
2312 unsigned int range;
2313 struct ocfs2_extent_rec *rec;
2314
2315 if (next_free == 0)
2316 return 0;
2317
2318 rec = &el->l_recs[0];
2319 if (ocfs2_is_empty_extent(rec)) {
2320 /* Empty list. */
2321 if (next_free == 1)
2322 return 0;
2323 rec = &el->l_recs[1];
2324 }
2325
2326 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2327 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
2328 return 1;
2329 return 0;
2330}
2331
2332/*
2333 * Rotate all the records in a btree right one record, starting at insert_cpos.
2334 *
2335 * The path to the rightmost leaf should be passed in.
2336 *
2337 * The array is assumed to be large enough to hold an entire path (tree depth).
2338 *
2339 * Upon successful return from this function:
2340 *
2341 * - The 'right_path' array will contain a path to the leaf block
2342 * whose range contains e_cpos.
2343 * - That leaf block will have a single empty extent in list index 0.
2344 * - In the case that the rotation requires a post-insert update,
2345 * *ret_left_path will contain a valid path which can be passed to
2346 * ocfs2_insert_path().
2347 */
2348static int ocfs2_rotate_tree_right(handle_t *handle,
2349 struct ocfs2_extent_tree *et,
2350 enum ocfs2_split_type split,
2351 u32 insert_cpos,
2352 struct ocfs2_path *right_path,
2353 struct ocfs2_path **ret_left_path)
2354{
2355 int ret, start, orig_credits = handle->h_buffer_credits;
2356 u32 cpos;
2357 struct ocfs2_path *left_path = NULL;
2358 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2359
2360 *ret_left_path = NULL;
2361
2362 left_path = ocfs2_new_path_from_path(right_path);
2363 if (!left_path) {
2364 ret = -ENOMEM;
2365 mlog_errno(ret);
2366 goto out;
2367 }
2368
2369 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2370 if (ret) {
2371 mlog_errno(ret);
2372 goto out;
2373 }
2374
2375 trace_ocfs2_rotate_tree_right(
2376 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2377 insert_cpos, cpos);
2378
2379 /*
2380 * What we want to do here is:
2381 *
2382 * 1) Start with the rightmost path.
2383 *
2384 * 2) Determine a path to the leaf block directly to the left
2385 * of that leaf.
2386 *
2387 * 3) Determine the 'subtree root' - the lowest level tree node
2388 * which contains a path to both leaves.
2389 *
2390 * 4) Rotate the subtree.
2391 *
2392 * 5) Find the next subtree by considering the left path to be
2393 * the new right path.
2394 *
2395 * The check at the top of this while loop also accepts
2396 * insert_cpos == cpos because cpos is only a _theoretical_
2397 * value to get us the left path - insert_cpos might very well
2398 * be filling that hole.
2399 *
2400 * Stop at a cpos of '0' because we either started at the
2401 * leftmost branch (i.e., a tree with one branch and a
2402 * rotation inside of it), or we've gone as far as we can in
2403 * rotating subtrees.
2404 */
2405 while (cpos && insert_cpos <= cpos) {
2406 trace_ocfs2_rotate_tree_right(
2407 (unsigned long long)
2408 ocfs2_metadata_cache_owner(et->et_ci),
2409 insert_cpos, cpos);
2410
2411 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
2412 if (ret) {
2413 mlog_errno(ret);
2414 goto out;
2415 }
2416
2417 mlog_bug_on_msg(path_leaf_bh(left_path) ==
2418 path_leaf_bh(right_path),
2419 "Owner %llu: error during insert of %u "
2420 "(left path cpos %u) results in two identical "
2421 "paths ending at %llu\n",
2422 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2423 insert_cpos, cpos,
2424 (unsigned long long)
2425 path_leaf_bh(left_path)->b_blocknr);
2426
2427 if (split == SPLIT_NONE &&
2428 ocfs2_rotate_requires_path_adjustment(left_path,
2429 insert_cpos)) {
2430
2431 /*
2432 * We've rotated the tree as much as we
2433 * should. The rest is up to
2434 * ocfs2_insert_path() to complete, after the
2435 * record insertion. We indicate this
2436 * situation by returning the left path.
2437 *
2438 * The reason we don't adjust the records here
2439 * before the record insert is that an error
2440 * later might break the rule where a parent
2441 * record e_cpos will reflect the actual
2442 * e_cpos of the 1st nonempty record of the
2443 * child list.
2444 */
2445 *ret_left_path = left_path;
2446 goto out_ret_path;
2447 }
2448
2449 start = ocfs2_find_subtree_root(et, left_path, right_path);
2450
2451 trace_ocfs2_rotate_subtree(start,
2452 (unsigned long long)
2453 right_path->p_node[start].bh->b_blocknr,
2454 right_path->p_tree_depth);
2455
2456 ret = ocfs2_extend_rotate_transaction(handle, start,
2457 orig_credits, right_path);
2458 if (ret) {
2459 mlog_errno(ret);
2460 goto out;
2461 }
2462
2463 ret = ocfs2_rotate_subtree_right(handle, et, left_path,
2464 right_path, start);
2465 if (ret) {
2466 mlog_errno(ret);
2467 goto out;
2468 }
2469
2470 if (split != SPLIT_NONE &&
2471 ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
2472 insert_cpos)) {
2473 /*
2474 * A rotate moves the rightmost left leaf
2475 * record over to the leftmost right leaf
2476 * slot. If we're doing an extent split
2477 * instead of a real insert, then we have to
2478 * check that the extent to be split wasn't
2479 * just moved over. If it was, then we can
2480 * exit here, passing left_path back -
2481 * ocfs2_split_extent() is smart enough to
2482 * search both leaves.
2483 */
2484 *ret_left_path = left_path;
2485 goto out_ret_path;
2486 }
2487
2488 /*
2489 * There is no need to re-read the next right path
2490 * as we know that it'll be our current left
2491 * path. Optimize by copying values instead.
2492 */
2493 ocfs2_mv_path(right_path, left_path);
2494
2495 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2496 if (ret) {
2497 mlog_errno(ret);
2498 goto out;
2499 }
2500 }
2501
2502out:
2503 ocfs2_free_path(left_path);
2504
2505out_ret_path:
2506 return ret;
2507}
2508
2509static int ocfs2_update_edge_lengths(handle_t *handle,
2510 struct ocfs2_extent_tree *et,
2511 int subtree_index, struct ocfs2_path *path)
2512{
2513 int i, idx, ret;
2514 struct ocfs2_extent_rec *rec;
2515 struct ocfs2_extent_list *el;
2516 struct ocfs2_extent_block *eb;
2517 u32 range;
2518
2519 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
2520 if (ret) {
2521 mlog_errno(ret);
2522 goto out;
2523 }
2524
2525 /* Path should always be rightmost. */
2526 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2527 BUG_ON(eb->h_next_leaf_blk != 0ULL);
2528
2529 el = &eb->h_list;
2530 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
2531 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2532 rec = &el->l_recs[idx];
2533 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2534
2535 for (i = 0; i < path->p_tree_depth; i++) {
2536 el = path->p_node[i].el;
2537 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2538 rec = &el->l_recs[idx];
2539
2540 rec->e_int_clusters = cpu_to_le32(range);
2541 le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));
2542
2543 ocfs2_journal_dirty(handle, path->p_node[i].bh);
2544 }
2545out:
2546 return ret;
2547}
2548
2549static void ocfs2_unlink_path(handle_t *handle,
2550 struct ocfs2_extent_tree *et,
2551 struct ocfs2_cached_dealloc_ctxt *dealloc,
2552 struct ocfs2_path *path, int unlink_start)
2553{
2554 int ret, i;
2555 struct ocfs2_extent_block *eb;
2556 struct ocfs2_extent_list *el;
2557 struct buffer_head *bh;
2558
2559 for(i = unlink_start; i < path_num_items(path); i++) {
2560 bh = path->p_node[i].bh;
2561
2562 eb = (struct ocfs2_extent_block *)bh->b_data;
2563 /*
2564 * Not all nodes might have had their final count
2565 * decremented by the caller - handle this here.
2566 */
2567 el = &eb->h_list;
2568 if (le16_to_cpu(el->l_next_free_rec) > 1) {
2569 mlog(ML_ERROR,
2570 "Inode %llu, attempted to remove extent block "
2571 "%llu with %u records\n",
2572 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2573 (unsigned long long)le64_to_cpu(eb->h_blkno),
2574 le16_to_cpu(el->l_next_free_rec));
2575
2576 ocfs2_journal_dirty(handle, bh);
2577 ocfs2_remove_from_cache(et->et_ci, bh);
2578 continue;
2579 }
2580
2581 el->l_next_free_rec = 0;
2582 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2583
2584 ocfs2_journal_dirty(handle, bh);
2585
2586 ret = ocfs2_cache_extent_block_free(dealloc, eb);
2587 if (ret)
2588 mlog_errno(ret);
2589
2590 ocfs2_remove_from_cache(et->et_ci, bh);
2591 }
2592}
2593
2594static void ocfs2_unlink_subtree(handle_t *handle,
2595 struct ocfs2_extent_tree *et,
2596 struct ocfs2_path *left_path,
2597 struct ocfs2_path *right_path,
2598 int subtree_index,
2599 struct ocfs2_cached_dealloc_ctxt *dealloc)
2600{
2601 int i;
2602 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2603 struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2604 struct ocfs2_extent_list *el;
2605 struct ocfs2_extent_block *eb;
2606
2607 el = path_leaf_el(left_path);
2608
2609 eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2610
2611 for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2612 if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2613 break;
2614
2615 BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2616
2617 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2618 le16_add_cpu(&root_el->l_next_free_rec, -1);
2619
2620 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2621 eb->h_next_leaf_blk = 0;
2622
2623 ocfs2_journal_dirty(handle, root_bh);
2624 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2625
2626 ocfs2_unlink_path(handle, et, dealloc, right_path,
2627 subtree_index + 1);
2628}
2629
2630static int ocfs2_rotate_subtree_left(handle_t *handle,
2631 struct ocfs2_extent_tree *et,
2632 struct ocfs2_path *left_path,
2633 struct ocfs2_path *right_path,
2634 int subtree_index,
2635 struct ocfs2_cached_dealloc_ctxt *dealloc,
2636 int *deleted)
2637{
2638 int ret, i, del_right_subtree = 0, right_has_empty = 0;
2639 struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path);
2640 struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2641 struct ocfs2_extent_block *eb;
2642
2643 *deleted = 0;
2644
2645 right_leaf_el = path_leaf_el(right_path);
2646 left_leaf_el = path_leaf_el(left_path);
2647 root_bh = left_path->p_node[subtree_index].bh;
2648 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2649
2650 if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2651 return 0;
2652
2653 eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2654 if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2655 /*
2656 * It's legal for us to proceed if the right leaf is
2657 * the rightmost one and it has an empty extent. There
2658 * are two cases to handle - whether the leaf will be
2659 * empty after removal or not. If the leaf isn't empty
2660 * then just remove the empty extent up front. The
2661 * next block will handle empty leaves by flagging
2662 * them for unlink.
2663 *
2664 * Non rightmost leaves will throw -EAGAIN and the
2665 * caller can manually move the subtree and retry.
2666 */
2667
2668 if (eb->h_next_leaf_blk != 0ULL)
2669 return -EAGAIN;
2670
2671 if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2672 ret = ocfs2_journal_access_eb(handle, et->et_ci,
2673 path_leaf_bh(right_path),
2674 OCFS2_JOURNAL_ACCESS_WRITE);
2675 if (ret) {
2676 mlog_errno(ret);
2677 goto out;
2678 }
2679
2680 ocfs2_remove_empty_extent(right_leaf_el);
2681 } else
2682 right_has_empty = 1;
2683 }
2684
2685 if (eb->h_next_leaf_blk == 0ULL &&
2686 le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2687 /*
2688 * We have to update i_last_eb_blk during the meta
2689 * data delete.
2690 */
2691 ret = ocfs2_et_root_journal_access(handle, et,
2692 OCFS2_JOURNAL_ACCESS_WRITE);
2693 if (ret) {
2694 mlog_errno(ret);
2695 goto out;
2696 }
2697
2698 del_right_subtree = 1;
2699 }
2700
2701 /*
2702 * Getting here with an empty extent in the right path implies
2703 * that it's the rightmost path and will be deleted.
2704 */
2705 BUG_ON(right_has_empty && !del_right_subtree);
2706
2707 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2708 subtree_index);
2709 if (ret) {
2710 mlog_errno(ret);
2711 goto out;
2712 }
2713
2714 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2715 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2716 right_path, i);
2717 if (ret) {
2718 mlog_errno(ret);
2719 goto out;
2720 }
2721
2722 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2723 left_path, i);
2724 if (ret) {
2725 mlog_errno(ret);
2726 goto out;
2727 }
2728 }
2729
2730 if (!right_has_empty) {
2731 /*
2732 * Only do this if we're moving a real
2733 * record. Otherwise, the action is delayed until
2734 * after removal of the right path in which case we
2735 * can do a simple shift to remove the empty extent.
2736 */
2737 ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2738 memset(&right_leaf_el->l_recs[0], 0,
2739 sizeof(struct ocfs2_extent_rec));
2740 }
2741 if (eb->h_next_leaf_blk == 0ULL) {
2742 /*
2743 * Move recs over to get rid of empty extent, decrease
2744 * next_free. This is allowed to remove the last
2745 * extent in our leaf (setting l_next_free_rec to
2746 * zero) - the delete code below won't care.
2747 */
2748 ocfs2_remove_empty_extent(right_leaf_el);
2749 }
2750
2751 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2752 ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2753
2754 if (del_right_subtree) {
2755 ocfs2_unlink_subtree(handle, et, left_path, right_path,
2756 subtree_index, dealloc);
2757 ret = ocfs2_update_edge_lengths(handle, et, subtree_index,
2758 left_path);
2759 if (ret) {
2760 mlog_errno(ret);
2761 goto out;
2762 }
2763
2764 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2765 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2766
2767 /*
2768 * Removal of the extent in the left leaf was skipped
2769 * above so we could delete the right path
2770 * 1st.
2771 */
2772 if (right_has_empty)
2773 ocfs2_remove_empty_extent(left_leaf_el);
2774
2775 ocfs2_journal_dirty(handle, et_root_bh);
2776
2777 *deleted = 1;
2778 } else
2779 ocfs2_complete_edge_insert(handle, left_path, right_path,
2780 subtree_index);
2781
2782out:
2783 return ret;
2784}
2785
2786/*
2787 * Given a full path, determine what cpos value would return us a path
2788 * containing the leaf immediately to the right of the current one.
2789 *
2790 * Will return zero if the path passed in is already the rightmost path.
2791 *
2792 * This looks similar, but is subtly different to
2793 * ocfs2_find_cpos_for_left_leaf().
2794 */
2795int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
2796 struct ocfs2_path *path, u32 *cpos)
2797{
2798 int i, j, ret = 0;
2799 u64 blkno;
2800 struct ocfs2_extent_list *el;
2801
2802 *cpos = 0;
2803
2804 if (path->p_tree_depth == 0)
2805 return 0;
2806
2807 blkno = path_leaf_bh(path)->b_blocknr;
2808
2809 /* Start at the tree node just above the leaf and work our way up. */
2810 i = path->p_tree_depth - 1;
2811 while (i >= 0) {
2812 int next_free;
2813
2814 el = path->p_node[i].el;
2815
2816 /*
2817 * Find the extent record just after the one in our
2818 * path.
2819 */
2820 next_free = le16_to_cpu(el->l_next_free_rec);
2821 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2822 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2823 if (j == (next_free - 1)) {
2824 if (i == 0) {
2825 /*
2826 * We've determined that the
2827 * path specified is already
2828 * the rightmost one - return a
2829 * cpos of zero.
2830 */
2831 goto out;
2832 }
2833 /*
2834 * The rightmost record points to our
2835 * leaf - we need to travel up the
2836 * tree one level.
2837 */
2838 goto next_node;
2839 }
2840
2841 *cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
2842 goto out;
2843 }
2844 }
2845
2846 /*
2847 * If we got here, we never found a valid node where
2848 * the tree indicated one should be.
2849 */
2850 ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
2851 (unsigned long long)blkno);
2852 ret = -EROFS;
2853 goto out;
2854
2855next_node:
2856 blkno = path->p_node[i].bh->b_blocknr;
2857 i--;
2858 }
2859
2860out:
2861 return ret;
2862}
2863
2864static int ocfs2_rotate_rightmost_leaf_left(handle_t *handle,
2865 struct ocfs2_extent_tree *et,
2866 struct ocfs2_path *path)
2867{
2868 int ret;
2869 struct buffer_head *bh = path_leaf_bh(path);
2870 struct ocfs2_extent_list *el = path_leaf_el(path);
2871
2872 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2873 return 0;
2874
2875 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
2876 path_num_items(path) - 1);
2877 if (ret) {
2878 mlog_errno(ret);
2879 goto out;
2880 }
2881
2882 ocfs2_remove_empty_extent(el);
2883 ocfs2_journal_dirty(handle, bh);
2884
2885out:
2886 return ret;
2887}
2888
2889static int __ocfs2_rotate_tree_left(handle_t *handle,
2890 struct ocfs2_extent_tree *et,
2891 int orig_credits,
2892 struct ocfs2_path *path,
2893 struct ocfs2_cached_dealloc_ctxt *dealloc,
2894 struct ocfs2_path **empty_extent_path)
2895{
2896 int ret, subtree_root, deleted;
2897 u32 right_cpos;
2898 struct ocfs2_path *left_path = NULL;
2899 struct ocfs2_path *right_path = NULL;
2900 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2901
2902 if (!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])))
2903 return 0;
2904
2905 *empty_extent_path = NULL;
2906
2907 ret = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
2908 if (ret) {
2909 mlog_errno(ret);
2910 goto out;
2911 }
2912
2913 left_path = ocfs2_new_path_from_path(path);
2914 if (!left_path) {
2915 ret = -ENOMEM;
2916 mlog_errno(ret);
2917 goto out;
2918 }
2919
2920 ocfs2_cp_path(left_path, path);
2921
2922 right_path = ocfs2_new_path_from_path(path);
2923 if (!right_path) {
2924 ret = -ENOMEM;
2925 mlog_errno(ret);
2926 goto out;
2927 }
2928
2929 while (right_cpos) {
2930 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
2931 if (ret) {
2932 mlog_errno(ret);
2933 goto out;
2934 }
2935
2936 subtree_root = ocfs2_find_subtree_root(et, left_path,
2937 right_path);
2938
2939 trace_ocfs2_rotate_subtree(subtree_root,
2940 (unsigned long long)
2941 right_path->p_node[subtree_root].bh->b_blocknr,
2942 right_path->p_tree_depth);
2943
2944 ret = ocfs2_extend_rotate_transaction(handle, 0,
2945 orig_credits, left_path);
2946 if (ret) {
2947 mlog_errno(ret);
2948 goto out;
2949 }
2950
2951 /*
2952 * Caller might still want to make changes to the
2953 * tree root, so re-add it to the journal here.
2954 */
2955 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2956 left_path, 0);
2957 if (ret) {
2958 mlog_errno(ret);
2959 goto out;
2960 }
2961
2962 ret = ocfs2_rotate_subtree_left(handle, et, left_path,
2963 right_path, subtree_root,
2964 dealloc, &deleted);
2965 if (ret == -EAGAIN) {
2966 /*
2967 * The rotation has to temporarily stop due to
2968 * the right subtree having an empty
2969 * extent. Pass it back to the caller for a
2970 * fixup.
2971 */
2972 *empty_extent_path = right_path;
2973 right_path = NULL;
2974 goto out;
2975 }
2976 if (ret) {
2977 mlog_errno(ret);
2978 goto out;
2979 }
2980
2981 /*
2982 * The subtree rotate might have removed records on
2983 * the rightmost edge. If so, then rotation is
2984 * complete.
2985 */
2986 if (deleted)
2987 break;
2988
2989 ocfs2_mv_path(left_path, right_path);
2990
2991 ret = ocfs2_find_cpos_for_right_leaf(sb, left_path,
2992 &right_cpos);
2993 if (ret) {
2994 mlog_errno(ret);
2995 goto out;
2996 }
2997 }
2998
2999out:
3000 ocfs2_free_path(right_path);
3001 ocfs2_free_path(left_path);
3002
3003 return ret;
3004}
3005
3006static int ocfs2_remove_rightmost_path(handle_t *handle,
3007 struct ocfs2_extent_tree *et,
3008 struct ocfs2_path *path,
3009 struct ocfs2_cached_dealloc_ctxt *dealloc)
3010{
3011 int ret, subtree_index;
3012 u32 cpos;
3013 struct ocfs2_path *left_path = NULL;
3014 struct ocfs2_extent_block *eb;
3015 struct ocfs2_extent_list *el;
3016
3017 ret = ocfs2_et_sanity_check(et);
3018 if (ret)
3019 goto out;
3020
3021 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
3022 if (ret) {
3023 mlog_errno(ret);
3024 goto out;
3025 }
3026
3027 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3028 path, &cpos);
3029 if (ret) {
3030 mlog_errno(ret);
3031 goto out;
3032 }
3033
3034 if (cpos) {
3035 /*
3036 * We have a path to the left of this one - it needs
3037 * an update too.
3038 */
3039 left_path = ocfs2_new_path_from_path(path);
3040 if (!left_path) {
3041 ret = -ENOMEM;
3042 mlog_errno(ret);
3043 goto out;
3044 }
3045
3046 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
3047 if (ret) {
3048 mlog_errno(ret);
3049 goto out;
3050 }
3051
3052 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
3053 if (ret) {
3054 mlog_errno(ret);
3055 goto out;
3056 }
3057
3058 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
3059
3060 ocfs2_unlink_subtree(handle, et, left_path, path,
3061 subtree_index, dealloc);
3062 ret = ocfs2_update_edge_lengths(handle, et, subtree_index,
3063 left_path);
3064 if (ret) {
3065 mlog_errno(ret);
3066 goto out;
3067 }
3068
3069 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
3070 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
3071 } else {
3072 /*
3073 * 'path' is also the leftmost path which
3074 * means it must be the only one. This gets
3075 * handled differently because we want to
3076 * revert the root back to having extents
3077 * in-line.
3078 */
3079 ocfs2_unlink_path(handle, et, dealloc, path, 1);
3080
3081 el = et->et_root_el;
3082 el->l_tree_depth = 0;
3083 el->l_next_free_rec = 0;
3084 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3085
3086 ocfs2_et_set_last_eb_blk(et, 0);
3087 }
3088
3089 ocfs2_journal_dirty(handle, path_root_bh(path));
3090
3091out:
3092 ocfs2_free_path(left_path);
3093 return ret;
3094}
3095
3096static int ocfs2_remove_rightmost_empty_extent(struct ocfs2_super *osb,
3097 struct ocfs2_extent_tree *et,
3098 struct ocfs2_path *path,
3099 struct ocfs2_cached_dealloc_ctxt *dealloc)
3100{
3101 handle_t *handle;
3102 int ret;
3103 int credits = path->p_tree_depth * 2 + 1;
3104
3105 handle = ocfs2_start_trans(osb, credits);
3106 if (IS_ERR(handle)) {
3107 ret = PTR_ERR(handle);
3108 mlog_errno(ret);
3109 return ret;
3110 }
3111
3112 ret = ocfs2_remove_rightmost_path(handle, et, path, dealloc);
3113 if (ret)
3114 mlog_errno(ret);
3115
3116 ocfs2_commit_trans(osb, handle);
3117 return ret;
3118}
3119
3120/*
3121 * Left rotation of btree records.
3122 *
3123 * In many ways, this is (unsurprisingly) the opposite of right
3124 * rotation. We start at some non-rightmost path containing an empty
3125 * extent in the leaf block. The code works its way to the rightmost
3126 * path by rotating records to the left in every subtree.
3127 *
3128 * This is used by any code which reduces the number of extent records
3129 * in a leaf. After removal, an empty record should be placed in the
3130 * leftmost list position.
3131 *
3132 * This won't handle a length update of the rightmost path records if
3133 * the rightmost tree leaf record is removed so the caller is
3134 * responsible for detecting and correcting that.
3135 */
3136static int ocfs2_rotate_tree_left(handle_t *handle,
3137 struct ocfs2_extent_tree *et,
3138 struct ocfs2_path *path,
3139 struct ocfs2_cached_dealloc_ctxt *dealloc)
3140{
3141 int ret, orig_credits = handle->h_buffer_credits;
3142 struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
3143 struct ocfs2_extent_block *eb;
3144 struct ocfs2_extent_list *el;
3145
3146 el = path_leaf_el(path);
3147 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
3148 return 0;
3149
3150 if (path->p_tree_depth == 0) {
3151rightmost_no_delete:
3152 /*
3153 * Inline extents. This is trivially handled, so do
3154 * it up front.
3155 */
3156 ret = ocfs2_rotate_rightmost_leaf_left(handle, et, path);
3157 if (ret)
3158 mlog_errno(ret);
3159 goto out;
3160 }
3161
3162 /*
3163 * Handle rightmost branch now. There's several cases:
3164 * 1) simple rotation leaving records in there. That's trivial.
3165 * 2) rotation requiring a branch delete - there's no more
3166 * records left. Two cases of this:
3167 * a) There are branches to the left.
3168 * b) This is also the leftmost (the only) branch.
3169 *
3170 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
3171 * 2a) we need the left branch so that we can update it with the unlink
3172 * 2b) we need to bring the root back to inline extents.
3173 */
3174
3175 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
3176 el = &eb->h_list;
3177 if (eb->h_next_leaf_blk == 0) {
3178 /*
3179 * This gets a bit tricky if we're going to delete the
3180 * rightmost path. Get the other cases out of the way
3181 * 1st.
3182 */
3183 if (le16_to_cpu(el->l_next_free_rec) > 1)
3184 goto rightmost_no_delete;
3185
3186 if (le16_to_cpu(el->l_next_free_rec) == 0) {
3187 ret = -EIO;
3188 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3189 "Owner %llu has empty extent block at %llu\n",
3190 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
3191 (unsigned long long)le64_to_cpu(eb->h_blkno));
3192 goto out;
3193 }
3194
3195 /*
3196 * XXX: The caller can not trust "path" any more after
3197 * this as it will have been deleted. What do we do?
3198 *
3199 * In theory the rotate-for-merge code will never get
3200 * here because it'll always ask for a rotate in a
3201 * nonempty list.
3202 */
3203
3204 ret = ocfs2_remove_rightmost_path(handle, et, path,
3205 dealloc);
3206 if (ret)
3207 mlog_errno(ret);
3208 goto out;
3209 }
3210
3211 /*
3212 * Now we can loop, remembering the path we get from -EAGAIN
3213 * and restarting from there.
3214 */
3215try_rotate:
3216 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits, path,
3217 dealloc, &restart_path);
3218 if (ret && ret != -EAGAIN) {
3219 mlog_errno(ret);
3220 goto out;
3221 }
3222
3223 while (ret == -EAGAIN) {
3224 tmp_path = restart_path;
3225 restart_path = NULL;
3226
3227 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits,
3228 tmp_path, dealloc,
3229 &restart_path);
3230 if (ret && ret != -EAGAIN) {
3231 mlog_errno(ret);
3232 goto out;
3233 }
3234
3235 ocfs2_free_path(tmp_path);
3236 tmp_path = NULL;
3237
3238 if (ret == 0)
3239 goto try_rotate;
3240 }
3241
3242out:
3243 ocfs2_free_path(tmp_path);
3244 ocfs2_free_path(restart_path);
3245 return ret;
3246}
3247
3248static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
3249 int index)
3250{
3251 struct ocfs2_extent_rec *rec = &el->l_recs[index];
3252 unsigned int size;
3253
3254 if (rec->e_leaf_clusters == 0) {
3255 /*
3256 * We consumed all of the merged-from record. An empty
3257 * extent cannot exist anywhere but the 1st array
3258 * position, so move things over if the merged-from
3259 * record doesn't occupy that position.
3260 *
3261 * This creates a new empty extent so the caller
3262 * should be smart enough to have removed any existing
3263 * ones.
3264 */
3265 if (index > 0) {
3266 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3267 size = index * sizeof(struct ocfs2_extent_rec);
3268 memmove(&el->l_recs[1], &el->l_recs[0], size);
3269 }
3270
3271 /*
3272 * Always memset - the caller doesn't check whether it
3273 * created an empty extent, so there could be junk in
3274 * the other fields.
3275 */
3276 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3277 }
3278}
3279
3280static int ocfs2_get_right_path(struct ocfs2_extent_tree *et,
3281 struct ocfs2_path *left_path,
3282 struct ocfs2_path **ret_right_path)
3283{
3284 int ret;
3285 u32 right_cpos;
3286 struct ocfs2_path *right_path = NULL;
3287 struct ocfs2_extent_list *left_el;
3288
3289 *ret_right_path = NULL;
3290
3291 /* This function shouldn't be called for non-trees. */
3292 BUG_ON(left_path->p_tree_depth == 0);
3293
3294 left_el = path_leaf_el(left_path);
3295 BUG_ON(left_el->l_next_free_rec != left_el->l_count);
3296
3297 ret = ocfs2_find_cpos_for_right_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3298 left_path, &right_cpos);
3299 if (ret) {
3300 mlog_errno(ret);
3301 goto out;
3302 }
3303
3304 /* This function shouldn't be called for the rightmost leaf. */
3305 BUG_ON(right_cpos == 0);
3306
3307 right_path = ocfs2_new_path_from_path(left_path);
3308 if (!right_path) {
3309 ret = -ENOMEM;
3310 mlog_errno(ret);
3311 goto out;
3312 }
3313
3314 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
3315 if (ret) {
3316 mlog_errno(ret);
3317 goto out;
3318 }
3319
3320 *ret_right_path = right_path;
3321out:
3322 if (ret)
3323 ocfs2_free_path(right_path);
3324 return ret;
3325}
3326
3327/*
3328 * Remove split_rec clusters from the record at index and merge them
3329 * onto the beginning of the record "next" to it.
3330 * For index < l_count - 1, the next means the extent rec at index + 1.
3331 * For index == l_count - 1, the "next" means the 1st extent rec of the
3332 * next extent block.
3333 */
3334static int ocfs2_merge_rec_right(struct ocfs2_path *left_path,
3335 handle_t *handle,
3336 struct ocfs2_extent_tree *et,
3337 struct ocfs2_extent_rec *split_rec,
3338 int index)
3339{
3340 int ret, next_free, i;
3341 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3342 struct ocfs2_extent_rec *left_rec;
3343 struct ocfs2_extent_rec *right_rec;
3344 struct ocfs2_extent_list *right_el;
3345 struct ocfs2_path *right_path = NULL;
3346 int subtree_index = 0;
3347 struct ocfs2_extent_list *el = path_leaf_el(left_path);
3348 struct buffer_head *bh = path_leaf_bh(left_path);
3349 struct buffer_head *root_bh = NULL;
3350
3351 BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
3352 left_rec = &el->l_recs[index];
3353
3354 if (index == le16_to_cpu(el->l_next_free_rec) - 1 &&
3355 le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count)) {
3356 /* we meet with a cross extent block merge. */
3357 ret = ocfs2_get_right_path(et, left_path, &right_path);
3358 if (ret) {
3359 mlog_errno(ret);
3360 return ret;
3361 }
3362
3363 right_el = path_leaf_el(right_path);
3364 next_free = le16_to_cpu(right_el->l_next_free_rec);
3365 BUG_ON(next_free <= 0);
3366 right_rec = &right_el->l_recs[0];
3367 if (ocfs2_is_empty_extent(right_rec)) {
3368 BUG_ON(next_free <= 1);
3369 right_rec = &right_el->l_recs[1];
3370 }
3371
3372 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3373 le16_to_cpu(left_rec->e_leaf_clusters) !=
3374 le32_to_cpu(right_rec->e_cpos));
3375
3376 subtree_index = ocfs2_find_subtree_root(et, left_path,
3377 right_path);
3378
3379 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3380 handle->h_buffer_credits,
3381 right_path);
3382 if (ret) {
3383 mlog_errno(ret);
3384 goto out;
3385 }
3386
3387 root_bh = left_path->p_node[subtree_index].bh;
3388 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3389
3390 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3391 subtree_index);
3392 if (ret) {
3393 mlog_errno(ret);
3394 goto out;
3395 }
3396
3397 for (i = subtree_index + 1;
3398 i < path_num_items(right_path); i++) {
3399 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3400 right_path, i);
3401 if (ret) {
3402 mlog_errno(ret);
3403 goto out;
3404 }
3405
3406 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3407 left_path, i);
3408 if (ret) {
3409 mlog_errno(ret);
3410 goto out;
3411 }
3412 }
3413
3414 } else {
3415 BUG_ON(index == le16_to_cpu(el->l_next_free_rec) - 1);
3416 right_rec = &el->l_recs[index + 1];
3417 }
3418
3419 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, left_path,
3420 path_num_items(left_path) - 1);
3421 if (ret) {
3422 mlog_errno(ret);
3423 goto out;
3424 }
3425
3426 le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
3427
3428 le32_add_cpu(&right_rec->e_cpos, -split_clusters);
3429 le64_add_cpu(&right_rec->e_blkno,
3430 -ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3431 split_clusters));
3432 le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
3433
3434 ocfs2_cleanup_merge(el, index);
3435
3436 ocfs2_journal_dirty(handle, bh);
3437 if (right_path) {
3438 ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
3439 ocfs2_complete_edge_insert(handle, left_path, right_path,
3440 subtree_index);
3441 }
3442out:
3443 ocfs2_free_path(right_path);
3444 return ret;
3445}
3446
3447static int ocfs2_get_left_path(struct ocfs2_extent_tree *et,
3448 struct ocfs2_path *right_path,
3449 struct ocfs2_path **ret_left_path)
3450{
3451 int ret;
3452 u32 left_cpos;
3453 struct ocfs2_path *left_path = NULL;
3454
3455 *ret_left_path = NULL;
3456
3457 /* This function shouldn't be called for non-trees. */
3458 BUG_ON(right_path->p_tree_depth == 0);
3459
3460 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3461 right_path, &left_cpos);
3462 if (ret) {
3463 mlog_errno(ret);
3464 goto out;
3465 }
3466
3467 /* This function shouldn't be called for the leftmost leaf. */
3468 BUG_ON(left_cpos == 0);
3469
3470 left_path = ocfs2_new_path_from_path(right_path);
3471 if (!left_path) {
3472 ret = -ENOMEM;
3473 mlog_errno(ret);
3474 goto out;
3475 }
3476
3477 ret = ocfs2_find_path(et->et_ci, left_path, left_cpos);
3478 if (ret) {
3479 mlog_errno(ret);
3480 goto out;
3481 }
3482
3483 *ret_left_path = left_path;
3484out:
3485 if (ret)
3486 ocfs2_free_path(left_path);
3487 return ret;
3488}
3489
3490/*
3491 * Remove split_rec clusters from the record at index and merge them
3492 * onto the tail of the record "before" it.
3493 * For index > 0, the "before" means the extent rec at index - 1.
3494 *
3495 * For index == 0, the "before" means the last record of the previous
3496 * extent block. And there is also a situation that we may need to
3497 * remove the rightmost leaf extent block in the right_path and change
3498 * the right path to indicate the new rightmost path.
3499 */
3500static int ocfs2_merge_rec_left(struct ocfs2_path *right_path,
3501 handle_t *handle,
3502 struct ocfs2_extent_tree *et,
3503 struct ocfs2_extent_rec *split_rec,
3504 struct ocfs2_cached_dealloc_ctxt *dealloc,
3505 int index)
3506{
3507 int ret, i, subtree_index = 0, has_empty_extent = 0;
3508 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3509 struct ocfs2_extent_rec *left_rec;
3510 struct ocfs2_extent_rec *right_rec;
3511 struct ocfs2_extent_list *el = path_leaf_el(right_path);
3512 struct buffer_head *bh = path_leaf_bh(right_path);
3513 struct buffer_head *root_bh = NULL;
3514 struct ocfs2_path *left_path = NULL;
3515 struct ocfs2_extent_list *left_el;
3516
3517 BUG_ON(index < 0);
3518
3519 right_rec = &el->l_recs[index];
3520 if (index == 0) {
3521 /* we meet with a cross extent block merge. */
3522 ret = ocfs2_get_left_path(et, right_path, &left_path);
3523 if (ret) {
3524 mlog_errno(ret);
3525 return ret;
3526 }
3527
3528 left_el = path_leaf_el(left_path);
3529 BUG_ON(le16_to_cpu(left_el->l_next_free_rec) !=
3530 le16_to_cpu(left_el->l_count));
3531
3532 left_rec = &left_el->l_recs[
3533 le16_to_cpu(left_el->l_next_free_rec) - 1];
3534 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3535 le16_to_cpu(left_rec->e_leaf_clusters) !=
3536 le32_to_cpu(split_rec->e_cpos));
3537
3538 subtree_index = ocfs2_find_subtree_root(et, left_path,
3539 right_path);
3540
3541 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3542 handle->h_buffer_credits,
3543 left_path);
3544 if (ret) {
3545 mlog_errno(ret);
3546 goto out;
3547 }
3548
3549 root_bh = left_path->p_node[subtree_index].bh;
3550 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3551
3552 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3553 subtree_index);
3554 if (ret) {
3555 mlog_errno(ret);
3556 goto out;
3557 }
3558
3559 for (i = subtree_index + 1;
3560 i < path_num_items(right_path); i++) {
3561 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3562 right_path, i);
3563 if (ret) {
3564 mlog_errno(ret);
3565 goto out;
3566 }
3567
3568 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3569 left_path, i);
3570 if (ret) {
3571 mlog_errno(ret);
3572 goto out;
3573 }
3574 }
3575 } else {
3576 left_rec = &el->l_recs[index - 1];
3577 if (ocfs2_is_empty_extent(&el->l_recs[0]))
3578 has_empty_extent = 1;
3579 }
3580
3581 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3582 path_num_items(right_path) - 1);
3583 if (ret) {
3584 mlog_errno(ret);
3585 goto out;
3586 }
3587
3588 if (has_empty_extent && index == 1) {
3589 /*
3590 * The easy case - we can just plop the record right in.
3591 */
3592 *left_rec = *split_rec;
3593
3594 has_empty_extent = 0;
3595 } else
3596 le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
3597
3598 le32_add_cpu(&right_rec->e_cpos, split_clusters);
3599 le64_add_cpu(&right_rec->e_blkno,
3600 ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3601 split_clusters));
3602 le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
3603
3604 ocfs2_cleanup_merge(el, index);
3605
3606 ocfs2_journal_dirty(handle, bh);
3607 if (left_path) {
3608 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
3609
3610 /*
3611 * In the situation that the right_rec is empty and the extent
3612 * block is empty also, ocfs2_complete_edge_insert can't handle
3613 * it and we need to delete the right extent block.
3614 */
3615 if (le16_to_cpu(right_rec->e_leaf_clusters) == 0 &&
3616 le16_to_cpu(el->l_next_free_rec) == 1) {
3617 /* extend credit for ocfs2_remove_rightmost_path */
3618 ret = ocfs2_extend_rotate_transaction(handle, 0,
3619 handle->h_buffer_credits,
3620 right_path);
3621 if (ret) {
3622 mlog_errno(ret);
3623 goto out;
3624 }
3625
3626 ret = ocfs2_remove_rightmost_path(handle, et,
3627 right_path,
3628 dealloc);
3629 if (ret) {
3630 mlog_errno(ret);
3631 goto out;
3632 }
3633
3634 /* Now the rightmost extent block has been deleted.
3635 * So we use the new rightmost path.
3636 */
3637 ocfs2_mv_path(right_path, left_path);
3638 left_path = NULL;
3639 } else
3640 ocfs2_complete_edge_insert(handle, left_path,
3641 right_path, subtree_index);
3642 }
3643out:
3644 ocfs2_free_path(left_path);
3645 return ret;
3646}
3647
3648static int ocfs2_try_to_merge_extent(handle_t *handle,
3649 struct ocfs2_extent_tree *et,
3650 struct ocfs2_path *path,
3651 int split_index,
3652 struct ocfs2_extent_rec *split_rec,
3653 struct ocfs2_cached_dealloc_ctxt *dealloc,
3654 struct ocfs2_merge_ctxt *ctxt)
3655{
3656 int ret = 0;
3657 struct ocfs2_extent_list *el = path_leaf_el(path);
3658 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3659
3660 BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
3661
3662 if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) {
3663 /* extend credit for ocfs2_remove_rightmost_path */
3664 ret = ocfs2_extend_rotate_transaction(handle, 0,
3665 handle->h_buffer_credits,
3666 path);
3667 if (ret) {
3668 mlog_errno(ret);
3669 goto out;
3670 }
3671 /*
3672 * The merge code will need to create an empty
3673 * extent to take the place of the newly
3674 * emptied slot. Remove any pre-existing empty
3675 * extents - having more than one in a leaf is
3676 * illegal.
3677 */
3678 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3679 if (ret) {
3680 mlog_errno(ret);
3681 goto out;
3682 }
3683 split_index--;
3684 rec = &el->l_recs[split_index];
3685 }
3686
3687 if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
3688 /*
3689 * Left-right contig implies this.
3690 */
3691 BUG_ON(!ctxt->c_split_covers_rec);
3692
3693 /*
3694 * Since the leftright insert always covers the entire
3695 * extent, this call will delete the insert record
3696 * entirely, resulting in an empty extent record added to
3697 * the extent block.
3698 *
3699 * Since the adding of an empty extent shifts
3700 * everything back to the right, there's no need to
3701 * update split_index here.
3702 *
3703 * When the split_index is zero, we need to merge it to the
3704 * prevoius extent block. It is more efficient and easier
3705 * if we do merge_right first and merge_left later.
3706 */
3707 ret = ocfs2_merge_rec_right(path, handle, et, split_rec,
3708 split_index);
3709 if (ret) {
3710 mlog_errno(ret);
3711 goto out;
3712 }
3713
3714 /*
3715 * We can only get this from logic error above.
3716 */
3717 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
3718
3719 /* extend credit for ocfs2_remove_rightmost_path */
3720 ret = ocfs2_extend_rotate_transaction(handle, 0,
3721 handle->h_buffer_credits,
3722 path);
3723 if (ret) {
3724 mlog_errno(ret);
3725 goto out;
3726 }
3727
3728 /* The merge left us with an empty extent, remove it. */
3729 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3730 if (ret) {
3731 mlog_errno(ret);
3732 goto out;
3733 }
3734
3735 rec = &el->l_recs[split_index];
3736
3737 /*
3738 * Note that we don't pass split_rec here on purpose -
3739 * we've merged it into the rec already.
3740 */
3741 ret = ocfs2_merge_rec_left(path, handle, et, rec,
3742 dealloc, split_index);
3743
3744 if (ret) {
3745 mlog_errno(ret);
3746 goto out;
3747 }
3748
3749 /* extend credit for ocfs2_remove_rightmost_path */
3750 ret = ocfs2_extend_rotate_transaction(handle, 0,
3751 handle->h_buffer_credits,
3752 path);
3753 if (ret) {
3754 mlog_errno(ret);
3755 goto out;
3756 }
3757
3758 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3759 /*
3760 * Error from this last rotate is not critical, so
3761 * print but don't bubble it up.
3762 */
3763 if (ret)
3764 mlog_errno(ret);
3765 ret = 0;
3766 } else {
3767 /*
3768 * Merge a record to the left or right.
3769 *
3770 * 'contig_type' is relative to the existing record,
3771 * so for example, if we're "right contig", it's to
3772 * the record on the left (hence the left merge).
3773 */
3774 if (ctxt->c_contig_type == CONTIG_RIGHT) {
3775 ret = ocfs2_merge_rec_left(path, handle, et,
3776 split_rec, dealloc,
3777 split_index);
3778 if (ret) {
3779 mlog_errno(ret);
3780 goto out;
3781 }
3782 } else {
3783 ret = ocfs2_merge_rec_right(path, handle,
3784 et, split_rec,
3785 split_index);
3786 if (ret) {
3787 mlog_errno(ret);
3788 goto out;
3789 }
3790 }
3791
3792 if (ctxt->c_split_covers_rec) {
3793 /* extend credit for ocfs2_remove_rightmost_path */
3794 ret = ocfs2_extend_rotate_transaction(handle, 0,
3795 handle->h_buffer_credits,
3796 path);
3797 if (ret) {
3798 mlog_errno(ret);
3799 ret = 0;
3800 goto out;
3801 }
3802
3803 /*
3804 * The merge may have left an empty extent in
3805 * our leaf. Try to rotate it away.
3806 */
3807 ret = ocfs2_rotate_tree_left(handle, et, path,
3808 dealloc);
3809 if (ret)
3810 mlog_errno(ret);
3811 ret = 0;
3812 }
3813 }
3814
3815out:
3816 return ret;
3817}
3818
3819static void ocfs2_subtract_from_rec(struct super_block *sb,
3820 enum ocfs2_split_type split,
3821 struct ocfs2_extent_rec *rec,
3822 struct ocfs2_extent_rec *split_rec)
3823{
3824 u64 len_blocks;
3825
3826 len_blocks = ocfs2_clusters_to_blocks(sb,
3827 le16_to_cpu(split_rec->e_leaf_clusters));
3828
3829 if (split == SPLIT_LEFT) {
3830 /*
3831 * Region is on the left edge of the existing
3832 * record.
3833 */
3834 le32_add_cpu(&rec->e_cpos,
3835 le16_to_cpu(split_rec->e_leaf_clusters));
3836 le64_add_cpu(&rec->e_blkno, len_blocks);
3837 le16_add_cpu(&rec->e_leaf_clusters,
3838 -le16_to_cpu(split_rec->e_leaf_clusters));
3839 } else {
3840 /*
3841 * Region is on the right edge of the existing
3842 * record.
3843 */
3844 le16_add_cpu(&rec->e_leaf_clusters,
3845 -le16_to_cpu(split_rec->e_leaf_clusters));
3846 }
3847}
3848
3849/*
3850 * Do the final bits of extent record insertion at the target leaf
3851 * list. If this leaf is part of an allocation tree, it is assumed
3852 * that the tree above has been prepared.
3853 */
3854static void ocfs2_insert_at_leaf(struct ocfs2_extent_tree *et,
3855 struct ocfs2_extent_rec *insert_rec,
3856 struct ocfs2_extent_list *el,
3857 struct ocfs2_insert_type *insert)
3858{
3859 int i = insert->ins_contig_index;
3860 unsigned int range;
3861 struct ocfs2_extent_rec *rec;
3862
3863 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3864
3865 if (insert->ins_split != SPLIT_NONE) {
3866 i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
3867 BUG_ON(i == -1);
3868 rec = &el->l_recs[i];
3869 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
3870 insert->ins_split, rec,
3871 insert_rec);
3872 goto rotate;
3873 }
3874
3875 /*
3876 * Contiguous insert - either left or right.
3877 */
3878 if (insert->ins_contig != CONTIG_NONE) {
3879 rec = &el->l_recs[i];
3880 if (insert->ins_contig == CONTIG_LEFT) {
3881 rec->e_blkno = insert_rec->e_blkno;
3882 rec->e_cpos = insert_rec->e_cpos;
3883 }
3884 le16_add_cpu(&rec->e_leaf_clusters,
3885 le16_to_cpu(insert_rec->e_leaf_clusters));
3886 return;
3887 }
3888
3889 /*
3890 * Handle insert into an empty leaf.
3891 */
3892 if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3893 ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3894 ocfs2_is_empty_extent(&el->l_recs[0]))) {
3895 el->l_recs[0] = *insert_rec;
3896 el->l_next_free_rec = cpu_to_le16(1);
3897 return;
3898 }
3899
3900 /*
3901 * Appending insert.
3902 */
3903 if (insert->ins_appending == APPEND_TAIL) {
3904 i = le16_to_cpu(el->l_next_free_rec) - 1;
3905 rec = &el->l_recs[i];
3906 range = le32_to_cpu(rec->e_cpos)
3907 + le16_to_cpu(rec->e_leaf_clusters);
3908 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3909
3910 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3911 le16_to_cpu(el->l_count),
3912 "owner %llu, depth %u, count %u, next free %u, "
3913 "rec.cpos %u, rec.clusters %u, "
3914 "insert.cpos %u, insert.clusters %u\n",
3915 ocfs2_metadata_cache_owner(et->et_ci),
3916 le16_to_cpu(el->l_tree_depth),
3917 le16_to_cpu(el->l_count),
3918 le16_to_cpu(el->l_next_free_rec),
3919 le32_to_cpu(el->l_recs[i].e_cpos),
3920 le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3921 le32_to_cpu(insert_rec->e_cpos),
3922 le16_to_cpu(insert_rec->e_leaf_clusters));
3923 i++;
3924 el->l_recs[i] = *insert_rec;
3925 le16_add_cpu(&el->l_next_free_rec, 1);
3926 return;
3927 }
3928
3929rotate:
3930 /*
3931 * Ok, we have to rotate.
3932 *
3933 * At this point, it is safe to assume that inserting into an
3934 * empty leaf and appending to a leaf have both been handled
3935 * above.
3936 *
3937 * This leaf needs to have space, either by the empty 1st
3938 * extent record, or by virtue of an l_next_rec < l_count.
3939 */
3940 ocfs2_rotate_leaf(el, insert_rec);
3941}
3942
3943static void ocfs2_adjust_rightmost_records(handle_t *handle,
3944 struct ocfs2_extent_tree *et,
3945 struct ocfs2_path *path,
3946 struct ocfs2_extent_rec *insert_rec)
3947{
3948 int ret, i, next_free;
3949 struct buffer_head *bh;
3950 struct ocfs2_extent_list *el;
3951 struct ocfs2_extent_rec *rec;
3952
3953 /*
3954 * Update everything except the leaf block.
3955 */
3956 for (i = 0; i < path->p_tree_depth; i++) {
3957 bh = path->p_node[i].bh;
3958 el = path->p_node[i].el;
3959
3960 next_free = le16_to_cpu(el->l_next_free_rec);
3961 if (next_free == 0) {
3962 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3963 "Owner %llu has a bad extent list\n",
3964 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
3965 ret = -EIO;
3966 return;
3967 }
3968
3969 rec = &el->l_recs[next_free - 1];
3970
3971 rec->e_int_clusters = insert_rec->e_cpos;
3972 le32_add_cpu(&rec->e_int_clusters,
3973 le16_to_cpu(insert_rec->e_leaf_clusters));
3974 le32_add_cpu(&rec->e_int_clusters,
3975 -le32_to_cpu(rec->e_cpos));
3976
3977 ocfs2_journal_dirty(handle, bh);
3978 }
3979}
3980
3981static int ocfs2_append_rec_to_path(handle_t *handle,
3982 struct ocfs2_extent_tree *et,
3983 struct ocfs2_extent_rec *insert_rec,
3984 struct ocfs2_path *right_path,
3985 struct ocfs2_path **ret_left_path)
3986{
3987 int ret, next_free;
3988 struct ocfs2_extent_list *el;
3989 struct ocfs2_path *left_path = NULL;
3990
3991 *ret_left_path = NULL;
3992
3993 /*
3994 * This shouldn't happen for non-trees. The extent rec cluster
3995 * count manipulation below only works for interior nodes.
3996 */
3997 BUG_ON(right_path->p_tree_depth == 0);
3998
3999 /*
4000 * If our appending insert is at the leftmost edge of a leaf,
4001 * then we might need to update the rightmost records of the
4002 * neighboring path.
4003 */
4004 el = path_leaf_el(right_path);
4005 next_free = le16_to_cpu(el->l_next_free_rec);
4006 if (next_free == 0 ||
4007 (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
4008 u32 left_cpos;
4009
4010 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
4011 right_path, &left_cpos);
4012 if (ret) {
4013 mlog_errno(ret);
4014 goto out;
4015 }
4016
4017 trace_ocfs2_append_rec_to_path(
4018 (unsigned long long)
4019 ocfs2_metadata_cache_owner(et->et_ci),
4020 le32_to_cpu(insert_rec->e_cpos),
4021 left_cpos);
4022
4023 /*
4024 * No need to worry if the append is already in the
4025 * leftmost leaf.
4026 */
4027 if (left_cpos) {
4028 left_path = ocfs2_new_path_from_path(right_path);
4029 if (!left_path) {
4030 ret = -ENOMEM;
4031 mlog_errno(ret);
4032 goto out;
4033 }
4034
4035 ret = ocfs2_find_path(et->et_ci, left_path,
4036 left_cpos);
4037 if (ret) {
4038 mlog_errno(ret);
4039 goto out;
4040 }
4041
4042 /*
4043 * ocfs2_insert_path() will pass the left_path to the
4044 * journal for us.
4045 */
4046 }
4047 }
4048
4049 ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4050 if (ret) {
4051 mlog_errno(ret);
4052 goto out;
4053 }
4054
4055 ocfs2_adjust_rightmost_records(handle, et, right_path, insert_rec);
4056
4057 *ret_left_path = left_path;
4058 ret = 0;
4059out:
4060 if (ret != 0)
4061 ocfs2_free_path(left_path);
4062
4063 return ret;
4064}
4065
4066static void ocfs2_split_record(struct ocfs2_extent_tree *et,
4067 struct ocfs2_path *left_path,
4068 struct ocfs2_path *right_path,
4069 struct ocfs2_extent_rec *split_rec,
4070 enum ocfs2_split_type split)
4071{
4072 int index;
4073 u32 cpos = le32_to_cpu(split_rec->e_cpos);
4074 struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
4075 struct ocfs2_extent_rec *rec, *tmprec;
4076
4077 right_el = path_leaf_el(right_path);
4078 if (left_path)
4079 left_el = path_leaf_el(left_path);
4080
4081 el = right_el;
4082 insert_el = right_el;
4083 index = ocfs2_search_extent_list(el, cpos);
4084 if (index != -1) {
4085 if (index == 0 && left_path) {
4086 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
4087
4088 /*
4089 * This typically means that the record
4090 * started in the left path but moved to the
4091 * right as a result of rotation. We either
4092 * move the existing record to the left, or we
4093 * do the later insert there.
4094 *
4095 * In this case, the left path should always
4096 * exist as the rotate code will have passed
4097 * it back for a post-insert update.
4098 */
4099
4100 if (split == SPLIT_LEFT) {
4101 /*
4102 * It's a left split. Since we know
4103 * that the rotate code gave us an
4104 * empty extent in the left path, we
4105 * can just do the insert there.
4106 */
4107 insert_el = left_el;
4108 } else {
4109 /*
4110 * Right split - we have to move the
4111 * existing record over to the left
4112 * leaf. The insert will be into the
4113 * newly created empty extent in the
4114 * right leaf.
4115 */
4116 tmprec = &right_el->l_recs[index];
4117 ocfs2_rotate_leaf(left_el, tmprec);
4118 el = left_el;
4119
4120 memset(tmprec, 0, sizeof(*tmprec));
4121 index = ocfs2_search_extent_list(left_el, cpos);
4122 BUG_ON(index == -1);
4123 }
4124 }
4125 } else {
4126 BUG_ON(!left_path);
4127 BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
4128 /*
4129 * Left path is easy - we can just allow the insert to
4130 * happen.
4131 */
4132 el = left_el;
4133 insert_el = left_el;
4134 index = ocfs2_search_extent_list(el, cpos);
4135 BUG_ON(index == -1);
4136 }
4137
4138 rec = &el->l_recs[index];
4139 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4140 split, rec, split_rec);
4141 ocfs2_rotate_leaf(insert_el, split_rec);
4142}
4143
4144/*
4145 * This function only does inserts on an allocation b-tree. For tree
4146 * depth = 0, ocfs2_insert_at_leaf() is called directly.
4147 *
4148 * right_path is the path we want to do the actual insert
4149 * in. left_path should only be passed in if we need to update that
4150 * portion of the tree after an edge insert.
4151 */
4152static int ocfs2_insert_path(handle_t *handle,
4153 struct ocfs2_extent_tree *et,
4154 struct ocfs2_path *left_path,
4155 struct ocfs2_path *right_path,
4156 struct ocfs2_extent_rec *insert_rec,
4157 struct ocfs2_insert_type *insert)
4158{
4159 int ret, subtree_index;
4160 struct buffer_head *leaf_bh = path_leaf_bh(right_path);
4161
4162 if (left_path) {
4163 /*
4164 * There's a chance that left_path got passed back to
4165 * us without being accounted for in the
4166 * journal. Extend our transaction here to be sure we
4167 * can change those blocks.
4168 */
4169 ret = ocfs2_extend_trans(handle, left_path->p_tree_depth);
4170 if (ret < 0) {
4171 mlog_errno(ret);
4172 goto out;
4173 }
4174
4175 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
4176 if (ret < 0) {
4177 mlog_errno(ret);
4178 goto out;
4179 }
4180 }
4181
4182 /*
4183 * Pass both paths to the journal. The majority of inserts
4184 * will be touching all components anyway.
4185 */
4186 ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4187 if (ret < 0) {
4188 mlog_errno(ret);
4189 goto out;
4190 }
4191
4192 if (insert->ins_split != SPLIT_NONE) {
4193 /*
4194 * We could call ocfs2_insert_at_leaf() for some types
4195 * of splits, but it's easier to just let one separate
4196 * function sort it all out.
4197 */
4198 ocfs2_split_record(et, left_path, right_path,
4199 insert_rec, insert->ins_split);
4200
4201 /*
4202 * Split might have modified either leaf and we don't
4203 * have a guarantee that the later edge insert will
4204 * dirty this for us.
4205 */
4206 if (left_path)
4207 ocfs2_journal_dirty(handle,
4208 path_leaf_bh(left_path));
4209 } else
4210 ocfs2_insert_at_leaf(et, insert_rec, path_leaf_el(right_path),
4211 insert);
4212
4213 ocfs2_journal_dirty(handle, leaf_bh);
4214
4215 if (left_path) {
4216 /*
4217 * The rotate code has indicated that we need to fix
4218 * up portions of the tree after the insert.
4219 *
4220 * XXX: Should we extend the transaction here?
4221 */
4222 subtree_index = ocfs2_find_subtree_root(et, left_path,
4223 right_path);
4224 ocfs2_complete_edge_insert(handle, left_path, right_path,
4225 subtree_index);
4226 }
4227
4228 ret = 0;
4229out:
4230 return ret;
4231}
4232
4233static int ocfs2_do_insert_extent(handle_t *handle,
4234 struct ocfs2_extent_tree *et,
4235 struct ocfs2_extent_rec *insert_rec,
4236 struct ocfs2_insert_type *type)
4237{
4238 int ret, rotate = 0;
4239 u32 cpos;
4240 struct ocfs2_path *right_path = NULL;
4241 struct ocfs2_path *left_path = NULL;
4242 struct ocfs2_extent_list *el;
4243
4244 el = et->et_root_el;
4245
4246 ret = ocfs2_et_root_journal_access(handle, et,
4247 OCFS2_JOURNAL_ACCESS_WRITE);
4248 if (ret) {
4249 mlog_errno(ret);
4250 goto out;
4251 }
4252
4253 if (le16_to_cpu(el->l_tree_depth) == 0) {
4254 ocfs2_insert_at_leaf(et, insert_rec, el, type);
4255 goto out_update_clusters;
4256 }
4257
4258 right_path = ocfs2_new_path_from_et(et);
4259 if (!right_path) {
4260 ret = -ENOMEM;
4261 mlog_errno(ret);
4262 goto out;
4263 }
4264
4265 /*
4266 * Determine the path to start with. Rotations need the
4267 * rightmost path, everything else can go directly to the
4268 * target leaf.
4269 */
4270 cpos = le32_to_cpu(insert_rec->e_cpos);
4271 if (type->ins_appending == APPEND_NONE &&
4272 type->ins_contig == CONTIG_NONE) {
4273 rotate = 1;
4274 cpos = UINT_MAX;
4275 }
4276
4277 ret = ocfs2_find_path(et->et_ci, right_path, cpos);
4278 if (ret) {
4279 mlog_errno(ret);
4280 goto out;
4281 }
4282
4283 /*
4284 * Rotations and appends need special treatment - they modify
4285 * parts of the tree's above them.
4286 *
4287 * Both might pass back a path immediate to the left of the
4288 * one being inserted to. This will be cause
4289 * ocfs2_insert_path() to modify the rightmost records of
4290 * left_path to account for an edge insert.
4291 *
4292 * XXX: When modifying this code, keep in mind that an insert
4293 * can wind up skipping both of these two special cases...
4294 */
4295 if (rotate) {
4296 ret = ocfs2_rotate_tree_right(handle, et, type->ins_split,
4297 le32_to_cpu(insert_rec->e_cpos),
4298 right_path, &left_path);
4299 if (ret) {
4300 mlog_errno(ret);
4301 goto out;
4302 }
4303
4304 /*
4305 * ocfs2_rotate_tree_right() might have extended the
4306 * transaction without re-journaling our tree root.
4307 */
4308 ret = ocfs2_et_root_journal_access(handle, et,
4309 OCFS2_JOURNAL_ACCESS_WRITE);
4310 if (ret) {
4311 mlog_errno(ret);
4312 goto out;
4313 }
4314 } else if (type->ins_appending == APPEND_TAIL
4315 && type->ins_contig != CONTIG_LEFT) {
4316 ret = ocfs2_append_rec_to_path(handle, et, insert_rec,
4317 right_path, &left_path);
4318 if (ret) {
4319 mlog_errno(ret);
4320 goto out;
4321 }
4322 }
4323
4324 ret = ocfs2_insert_path(handle, et, left_path, right_path,
4325 insert_rec, type);
4326 if (ret) {
4327 mlog_errno(ret);
4328 goto out;
4329 }
4330
4331out_update_clusters:
4332 if (type->ins_split == SPLIT_NONE)
4333 ocfs2_et_update_clusters(et,
4334 le16_to_cpu(insert_rec->e_leaf_clusters));
4335
4336 ocfs2_journal_dirty(handle, et->et_root_bh);
4337
4338out:
4339 ocfs2_free_path(left_path);
4340 ocfs2_free_path(right_path);
4341
4342 return ret;
4343}
4344
4345static int ocfs2_figure_merge_contig_type(struct ocfs2_extent_tree *et,
4346 struct ocfs2_path *path,
4347 struct ocfs2_extent_list *el, int index,
4348 struct ocfs2_extent_rec *split_rec,
4349 struct ocfs2_merge_ctxt *ctxt)
4350{
4351 int status = 0;
4352 enum ocfs2_contig_type ret = CONTIG_NONE;
4353 u32 left_cpos, right_cpos;
4354 struct ocfs2_extent_rec *rec = NULL;
4355 struct ocfs2_extent_list *new_el;
4356 struct ocfs2_path *left_path = NULL, *right_path = NULL;
4357 struct buffer_head *bh;
4358 struct ocfs2_extent_block *eb;
4359 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
4360
4361 if (index > 0) {
4362 rec = &el->l_recs[index - 1];
4363 } else if (path->p_tree_depth > 0) {
4364 status = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
4365 if (status)
4366 goto exit;
4367
4368 if (left_cpos != 0) {
4369 left_path = ocfs2_new_path_from_path(path);
4370 if (!left_path) {
4371 status = -ENOMEM;
4372 mlog_errno(status);
4373 goto exit;
4374 }
4375
4376 status = ocfs2_find_path(et->et_ci, left_path,
4377 left_cpos);
4378 if (status)
4379 goto free_left_path;
4380
4381 new_el = path_leaf_el(left_path);
4382
4383 if (le16_to_cpu(new_el->l_next_free_rec) !=
4384 le16_to_cpu(new_el->l_count)) {
4385 bh = path_leaf_bh(left_path);
4386 eb = (struct ocfs2_extent_block *)bh->b_data;
4387 ocfs2_error(sb,
4388 "Extent block #%llu has an invalid l_next_free_rec of %d. It should have matched the l_count of %d\n",
4389 (unsigned long long)le64_to_cpu(eb->h_blkno),
4390 le16_to_cpu(new_el->l_next_free_rec),
4391 le16_to_cpu(new_el->l_count));
4392 status = -EINVAL;
4393 goto free_left_path;
4394 }
4395 rec = &new_el->l_recs[
4396 le16_to_cpu(new_el->l_next_free_rec) - 1];
4397 }
4398 }
4399
4400 /*
4401 * We're careful to check for an empty extent record here -
4402 * the merge code will know what to do if it sees one.
4403 */
4404 if (rec) {
4405 if (index == 1 && ocfs2_is_empty_extent(rec)) {
4406 if (split_rec->e_cpos == el->l_recs[index].e_cpos)
4407 ret = CONTIG_RIGHT;
4408 } else {
4409 ret = ocfs2_et_extent_contig(et, rec, split_rec);
4410 }
4411 }
4412
4413 rec = NULL;
4414 if (index < (le16_to_cpu(el->l_next_free_rec) - 1))
4415 rec = &el->l_recs[index + 1];
4416 else if (le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count) &&
4417 path->p_tree_depth > 0) {
4418 status = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
4419 if (status)
4420 goto free_left_path;
4421
4422 if (right_cpos == 0)
4423 goto free_left_path;
4424
4425 right_path = ocfs2_new_path_from_path(path);
4426 if (!right_path) {
4427 status = -ENOMEM;
4428 mlog_errno(status);
4429 goto free_left_path;
4430 }
4431
4432 status = ocfs2_find_path(et->et_ci, right_path, right_cpos);
4433 if (status)
4434 goto free_right_path;
4435
4436 new_el = path_leaf_el(right_path);
4437 rec = &new_el->l_recs[0];
4438 if (ocfs2_is_empty_extent(rec)) {
4439 if (le16_to_cpu(new_el->l_next_free_rec) <= 1) {
4440 bh = path_leaf_bh(right_path);
4441 eb = (struct ocfs2_extent_block *)bh->b_data;
4442 ocfs2_error(sb,
4443 "Extent block #%llu has an invalid l_next_free_rec of %d\n",
4444 (unsigned long long)le64_to_cpu(eb->h_blkno),
4445 le16_to_cpu(new_el->l_next_free_rec));
4446 status = -EINVAL;
4447 goto free_right_path;
4448 }
4449 rec = &new_el->l_recs[1];
4450 }
4451 }
4452
4453 if (rec) {
4454 enum ocfs2_contig_type contig_type;
4455
4456 contig_type = ocfs2_et_extent_contig(et, rec, split_rec);
4457
4458 if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
4459 ret = CONTIG_LEFTRIGHT;
4460 else if (ret == CONTIG_NONE)
4461 ret = contig_type;
4462 }
4463
4464free_right_path:
4465 ocfs2_free_path(right_path);
4466free_left_path:
4467 ocfs2_free_path(left_path);
4468exit:
4469 if (status == 0)
4470 ctxt->c_contig_type = ret;
4471
4472 return status;
4473}
4474
4475static void ocfs2_figure_contig_type(struct ocfs2_extent_tree *et,
4476 struct ocfs2_insert_type *insert,
4477 struct ocfs2_extent_list *el,
4478 struct ocfs2_extent_rec *insert_rec)
4479{
4480 int i;
4481 enum ocfs2_contig_type contig_type = CONTIG_NONE;
4482
4483 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4484
4485 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
4486 contig_type = ocfs2_et_extent_contig(et, &el->l_recs[i],
4487 insert_rec);
4488 if (contig_type != CONTIG_NONE) {
4489 insert->ins_contig_index = i;
4490 break;
4491 }
4492 }
4493 insert->ins_contig = contig_type;
4494
4495 if (insert->ins_contig != CONTIG_NONE) {
4496 struct ocfs2_extent_rec *rec =
4497 &el->l_recs[insert->ins_contig_index];
4498 unsigned int len = le16_to_cpu(rec->e_leaf_clusters) +
4499 le16_to_cpu(insert_rec->e_leaf_clusters);
4500
4501 /*
4502 * Caller might want us to limit the size of extents, don't
4503 * calculate contiguousness if we might exceed that limit.
4504 */
4505 if (et->et_max_leaf_clusters &&
4506 (len > et->et_max_leaf_clusters))
4507 insert->ins_contig = CONTIG_NONE;
4508 }
4509}
4510
4511/*
4512 * This should only be called against the righmost leaf extent list.
4513 *
4514 * ocfs2_figure_appending_type() will figure out whether we'll have to
4515 * insert at the tail of the rightmost leaf.
4516 *
4517 * This should also work against the root extent list for tree's with 0
4518 * depth. If we consider the root extent list to be the rightmost leaf node
4519 * then the logic here makes sense.
4520 */
4521static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
4522 struct ocfs2_extent_list *el,
4523 struct ocfs2_extent_rec *insert_rec)
4524{
4525 int i;
4526 u32 cpos = le32_to_cpu(insert_rec->e_cpos);
4527 struct ocfs2_extent_rec *rec;
4528
4529 insert->ins_appending = APPEND_NONE;
4530
4531 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4532
4533 if (!el->l_next_free_rec)
4534 goto set_tail_append;
4535
4536 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
4537 /* Were all records empty? */
4538 if (le16_to_cpu(el->l_next_free_rec) == 1)
4539 goto set_tail_append;
4540 }
4541
4542 i = le16_to_cpu(el->l_next_free_rec) - 1;
4543 rec = &el->l_recs[i];
4544
4545 if (cpos >=
4546 (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
4547 goto set_tail_append;
4548
4549 return;
4550
4551set_tail_append:
4552 insert->ins_appending = APPEND_TAIL;
4553}
4554
4555/*
4556 * Helper function called at the beginning of an insert.
4557 *
4558 * This computes a few things that are commonly used in the process of
4559 * inserting into the btree:
4560 * - Whether the new extent is contiguous with an existing one.
4561 * - The current tree depth.
4562 * - Whether the insert is an appending one.
4563 * - The total # of free records in the tree.
4564 *
4565 * All of the information is stored on the ocfs2_insert_type
4566 * structure.
4567 */
4568static int ocfs2_figure_insert_type(struct ocfs2_extent_tree *et,
4569 struct buffer_head **last_eb_bh,
4570 struct ocfs2_extent_rec *insert_rec,
4571 int *free_records,
4572 struct ocfs2_insert_type *insert)
4573{
4574 int ret;
4575 struct ocfs2_extent_block *eb;
4576 struct ocfs2_extent_list *el;
4577 struct ocfs2_path *path = NULL;
4578 struct buffer_head *bh = NULL;
4579
4580 insert->ins_split = SPLIT_NONE;
4581
4582 el = et->et_root_el;
4583 insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
4584
4585 if (el->l_tree_depth) {
4586 /*
4587 * If we have tree depth, we read in the
4588 * rightmost extent block ahead of time as
4589 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4590 * may want it later.
4591 */
4592 ret = ocfs2_read_extent_block(et->et_ci,
4593 ocfs2_et_get_last_eb_blk(et),
4594 &bh);
4595 if (ret) {
4596 mlog_errno(ret);
4597 goto out;
4598 }
4599 eb = (struct ocfs2_extent_block *) bh->b_data;
4600 el = &eb->h_list;
4601 }
4602
4603 /*
4604 * Unless we have a contiguous insert, we'll need to know if
4605 * there is room left in our allocation tree for another
4606 * extent record.
4607 *
4608 * XXX: This test is simplistic, we can search for empty
4609 * extent records too.
4610 */
4611 *free_records = le16_to_cpu(el->l_count) -
4612 le16_to_cpu(el->l_next_free_rec);
4613
4614 if (!insert->ins_tree_depth) {
4615 ocfs2_figure_contig_type(et, insert, el, insert_rec);
4616 ocfs2_figure_appending_type(insert, el, insert_rec);
4617 return 0;
4618 }
4619
4620 path = ocfs2_new_path_from_et(et);
4621 if (!path) {
4622 ret = -ENOMEM;
4623 mlog_errno(ret);
4624 goto out;
4625 }
4626
4627 /*
4628 * In the case that we're inserting past what the tree
4629 * currently accounts for, ocfs2_find_path() will return for
4630 * us the rightmost tree path. This is accounted for below in
4631 * the appending code.
4632 */
4633 ret = ocfs2_find_path(et->et_ci, path, le32_to_cpu(insert_rec->e_cpos));
4634 if (ret) {
4635 mlog_errno(ret);
4636 goto out;
4637 }
4638
4639 el = path_leaf_el(path);
4640
4641 /*
4642 * Now that we have the path, there's two things we want to determine:
4643 * 1) Contiguousness (also set contig_index if this is so)
4644 *
4645 * 2) Are we doing an append? We can trivially break this up
4646 * into two types of appends: simple record append, or a
4647 * rotate inside the tail leaf.
4648 */
4649 ocfs2_figure_contig_type(et, insert, el, insert_rec);
4650
4651 /*
4652 * The insert code isn't quite ready to deal with all cases of
4653 * left contiguousness. Specifically, if it's an insert into
4654 * the 1st record in a leaf, it will require the adjustment of
4655 * cluster count on the last record of the path directly to it's
4656 * left. For now, just catch that case and fool the layers
4657 * above us. This works just fine for tree_depth == 0, which
4658 * is why we allow that above.
4659 */
4660 if (insert->ins_contig == CONTIG_LEFT &&
4661 insert->ins_contig_index == 0)
4662 insert->ins_contig = CONTIG_NONE;
4663
4664 /*
4665 * Ok, so we can simply compare against last_eb to figure out
4666 * whether the path doesn't exist. This will only happen in
4667 * the case that we're doing a tail append, so maybe we can
4668 * take advantage of that information somehow.
4669 */
4670 if (ocfs2_et_get_last_eb_blk(et) ==
4671 path_leaf_bh(path)->b_blocknr) {
4672 /*
4673 * Ok, ocfs2_find_path() returned us the rightmost
4674 * tree path. This might be an appending insert. There are
4675 * two cases:
4676 * 1) We're doing a true append at the tail:
4677 * -This might even be off the end of the leaf
4678 * 2) We're "appending" by rotating in the tail
4679 */
4680 ocfs2_figure_appending_type(insert, el, insert_rec);
4681 }
4682
4683out:
4684 ocfs2_free_path(path);
4685
4686 if (ret == 0)
4687 *last_eb_bh = bh;
4688 else
4689 brelse(bh);
4690 return ret;
4691}
4692
4693/*
4694 * Insert an extent into a btree.
4695 *
4696 * The caller needs to update the owning btree's cluster count.
4697 */
4698int ocfs2_insert_extent(handle_t *handle,
4699 struct ocfs2_extent_tree *et,
4700 u32 cpos,
4701 u64 start_blk,
4702 u32 new_clusters,
4703 u8 flags,
4704 struct ocfs2_alloc_context *meta_ac)
4705{
4706 int status;
4707 int uninitialized_var(free_records);
4708 struct buffer_head *last_eb_bh = NULL;
4709 struct ocfs2_insert_type insert = {0, };
4710 struct ocfs2_extent_rec rec;
4711
4712 trace_ocfs2_insert_extent_start(
4713 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4714 cpos, new_clusters);
4715
4716 memset(&rec, 0, sizeof(rec));
4717 rec.e_cpos = cpu_to_le32(cpos);
4718 rec.e_blkno = cpu_to_le64(start_blk);
4719 rec.e_leaf_clusters = cpu_to_le16(new_clusters);
4720 rec.e_flags = flags;
4721 status = ocfs2_et_insert_check(et, &rec);
4722 if (status) {
4723 mlog_errno(status);
4724 goto bail;
4725 }
4726
4727 status = ocfs2_figure_insert_type(et, &last_eb_bh, &rec,
4728 &free_records, &insert);
4729 if (status < 0) {
4730 mlog_errno(status);
4731 goto bail;
4732 }
4733
4734 trace_ocfs2_insert_extent(insert.ins_appending, insert.ins_contig,
4735 insert.ins_contig_index, free_records,
4736 insert.ins_tree_depth);
4737
4738 if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
4739 status = ocfs2_grow_tree(handle, et,
4740 &insert.ins_tree_depth, &last_eb_bh,
4741 meta_ac);
4742 if (status) {
4743 mlog_errno(status);
4744 goto bail;
4745 }
4746 }
4747
4748 /* Finally, we can add clusters. This might rotate the tree for us. */
4749 status = ocfs2_do_insert_extent(handle, et, &rec, &insert);
4750 if (status < 0)
4751 mlog_errno(status);
4752 else
4753 ocfs2_et_extent_map_insert(et, &rec);
4754
4755bail:
4756 brelse(last_eb_bh);
4757
4758 return status;
4759}
4760
4761/*
4762 * Allcate and add clusters into the extent b-tree.
4763 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4764 * The extent b-tree's root is specified by et, and
4765 * it is not limited to the file storage. Any extent tree can use this
4766 * function if it implements the proper ocfs2_extent_tree.
4767 */
4768int ocfs2_add_clusters_in_btree(handle_t *handle,
4769 struct ocfs2_extent_tree *et,
4770 u32 *logical_offset,
4771 u32 clusters_to_add,
4772 int mark_unwritten,
4773 struct ocfs2_alloc_context *data_ac,
4774 struct ocfs2_alloc_context *meta_ac,
4775 enum ocfs2_alloc_restarted *reason_ret)
4776{
4777 int status = 0, err = 0;
4778 int need_free = 0;
4779 int free_extents;
4780 enum ocfs2_alloc_restarted reason = RESTART_NONE;
4781 u32 bit_off, num_bits;
4782 u64 block;
4783 u8 flags = 0;
4784 struct ocfs2_super *osb =
4785 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
4786
4787 BUG_ON(!clusters_to_add);
4788
4789 if (mark_unwritten)
4790 flags = OCFS2_EXT_UNWRITTEN;
4791
4792 free_extents = ocfs2_num_free_extents(osb, et);
4793 if (free_extents < 0) {
4794 status = free_extents;
4795 mlog_errno(status);
4796 goto leave;
4797 }
4798
4799 /* there are two cases which could cause us to EAGAIN in the
4800 * we-need-more-metadata case:
4801 * 1) we haven't reserved *any*
4802 * 2) we are so fragmented, we've needed to add metadata too
4803 * many times. */
4804 if (!free_extents && !meta_ac) {
4805 err = -1;
4806 status = -EAGAIN;
4807 reason = RESTART_META;
4808 goto leave;
4809 } else if ((!free_extents)
4810 && (ocfs2_alloc_context_bits_left(meta_ac)
4811 < ocfs2_extend_meta_needed(et->et_root_el))) {
4812 err = -2;
4813 status = -EAGAIN;
4814 reason = RESTART_META;
4815 goto leave;
4816 }
4817
4818 status = __ocfs2_claim_clusters(handle, data_ac, 1,
4819 clusters_to_add, &bit_off, &num_bits);
4820 if (status < 0) {
4821 if (status != -ENOSPC)
4822 mlog_errno(status);
4823 goto leave;
4824 }
4825
4826 BUG_ON(num_bits > clusters_to_add);
4827
4828 /* reserve our write early -- insert_extent may update the tree root */
4829 status = ocfs2_et_root_journal_access(handle, et,
4830 OCFS2_JOURNAL_ACCESS_WRITE);
4831 if (status < 0) {
4832 mlog_errno(status);
4833 need_free = 1;
4834 goto bail;
4835 }
4836
4837 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
4838 trace_ocfs2_add_clusters_in_btree(
4839 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4840 bit_off, num_bits);
4841 status = ocfs2_insert_extent(handle, et, *logical_offset, block,
4842 num_bits, flags, meta_ac);
4843 if (status < 0) {
4844 mlog_errno(status);
4845 need_free = 1;
4846 goto bail;
4847 }
4848
4849 ocfs2_journal_dirty(handle, et->et_root_bh);
4850
4851 clusters_to_add -= num_bits;
4852 *logical_offset += num_bits;
4853
4854 if (clusters_to_add) {
4855 err = clusters_to_add;
4856 status = -EAGAIN;
4857 reason = RESTART_TRANS;
4858 }
4859
4860bail:
4861 if (need_free) {
4862 if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
4863 ocfs2_free_local_alloc_bits(osb, handle, data_ac,
4864 bit_off, num_bits);
4865 else
4866 ocfs2_free_clusters(handle,
4867 data_ac->ac_inode,
4868 data_ac->ac_bh,
4869 ocfs2_clusters_to_blocks(osb->sb, bit_off),
4870 num_bits);
4871 }
4872
4873leave:
4874 if (reason_ret)
4875 *reason_ret = reason;
4876 trace_ocfs2_add_clusters_in_btree_ret(status, reason, err);
4877 return status;
4878}
4879
4880static void ocfs2_make_right_split_rec(struct super_block *sb,
4881 struct ocfs2_extent_rec *split_rec,
4882 u32 cpos,
4883 struct ocfs2_extent_rec *rec)
4884{
4885 u32 rec_cpos = le32_to_cpu(rec->e_cpos);
4886 u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
4887
4888 memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
4889
4890 split_rec->e_cpos = cpu_to_le32(cpos);
4891 split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
4892
4893 split_rec->e_blkno = rec->e_blkno;
4894 le64_add_cpu(&split_rec->e_blkno,
4895 ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
4896
4897 split_rec->e_flags = rec->e_flags;
4898}
4899
4900static int ocfs2_split_and_insert(handle_t *handle,
4901 struct ocfs2_extent_tree *et,
4902 struct ocfs2_path *path,
4903 struct buffer_head **last_eb_bh,
4904 int split_index,
4905 struct ocfs2_extent_rec *orig_split_rec,
4906 struct ocfs2_alloc_context *meta_ac)
4907{
4908 int ret = 0, depth;
4909 unsigned int insert_range, rec_range, do_leftright = 0;
4910 struct ocfs2_extent_rec tmprec;
4911 struct ocfs2_extent_list *rightmost_el;
4912 struct ocfs2_extent_rec rec;
4913 struct ocfs2_extent_rec split_rec = *orig_split_rec;
4914 struct ocfs2_insert_type insert;
4915 struct ocfs2_extent_block *eb;
4916
4917leftright:
4918 /*
4919 * Store a copy of the record on the stack - it might move
4920 * around as the tree is manipulated below.
4921 */
4922 rec = path_leaf_el(path)->l_recs[split_index];
4923
4924 rightmost_el = et->et_root_el;
4925
4926 depth = le16_to_cpu(rightmost_el->l_tree_depth);
4927 if (depth) {
4928 BUG_ON(!(*last_eb_bh));
4929 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
4930 rightmost_el = &eb->h_list;
4931 }
4932
4933 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4934 le16_to_cpu(rightmost_el->l_count)) {
4935 ret = ocfs2_grow_tree(handle, et,
4936 &depth, last_eb_bh, meta_ac);
4937 if (ret) {
4938 mlog_errno(ret);
4939 goto out;
4940 }
4941 }
4942
4943 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4944 insert.ins_appending = APPEND_NONE;
4945 insert.ins_contig = CONTIG_NONE;
4946 insert.ins_tree_depth = depth;
4947
4948 insert_range = le32_to_cpu(split_rec.e_cpos) +
4949 le16_to_cpu(split_rec.e_leaf_clusters);
4950 rec_range = le32_to_cpu(rec.e_cpos) +
4951 le16_to_cpu(rec.e_leaf_clusters);
4952
4953 if (split_rec.e_cpos == rec.e_cpos) {
4954 insert.ins_split = SPLIT_LEFT;
4955 } else if (insert_range == rec_range) {
4956 insert.ins_split = SPLIT_RIGHT;
4957 } else {
4958 /*
4959 * Left/right split. We fake this as a right split
4960 * first and then make a second pass as a left split.
4961 */
4962 insert.ins_split = SPLIT_RIGHT;
4963
4964 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4965 &tmprec, insert_range, &rec);
4966
4967 split_rec = tmprec;
4968
4969 BUG_ON(do_leftright);
4970 do_leftright = 1;
4971 }
4972
4973 ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
4974 if (ret) {
4975 mlog_errno(ret);
4976 goto out;
4977 }
4978
4979 if (do_leftright == 1) {
4980 u32 cpos;
4981 struct ocfs2_extent_list *el;
4982
4983 do_leftright++;
4984 split_rec = *orig_split_rec;
4985
4986 ocfs2_reinit_path(path, 1);
4987
4988 cpos = le32_to_cpu(split_rec.e_cpos);
4989 ret = ocfs2_find_path(et->et_ci, path, cpos);
4990 if (ret) {
4991 mlog_errno(ret);
4992 goto out;
4993 }
4994
4995 el = path_leaf_el(path);
4996 split_index = ocfs2_search_extent_list(el, cpos);
4997 if (split_index == -1) {
4998 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
4999 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5000 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5001 cpos);
5002 ret = -EROFS;
5003 goto out;
5004 }
5005 goto leftright;
5006 }
5007out:
5008
5009 return ret;
5010}
5011
5012static int ocfs2_replace_extent_rec(handle_t *handle,
5013 struct ocfs2_extent_tree *et,
5014 struct ocfs2_path *path,
5015 struct ocfs2_extent_list *el,
5016 int split_index,
5017 struct ocfs2_extent_rec *split_rec)
5018{
5019 int ret;
5020
5021 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
5022 path_num_items(path) - 1);
5023 if (ret) {
5024 mlog_errno(ret);
5025 goto out;
5026 }
5027
5028 el->l_recs[split_index] = *split_rec;
5029
5030 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5031out:
5032 return ret;
5033}
5034
5035/*
5036 * Split part or all of the extent record at split_index in the leaf
5037 * pointed to by path. Merge with the contiguous extent record if needed.
5038 *
5039 * Care is taken to handle contiguousness so as to not grow the tree.
5040 *
5041 * meta_ac is not strictly necessary - we only truly need it if growth
5042 * of the tree is required. All other cases will degrade into a less
5043 * optimal tree layout.
5044 *
5045 * last_eb_bh should be the rightmost leaf block for any extent
5046 * btree. Since a split may grow the tree or a merge might shrink it,
5047 * the caller cannot trust the contents of that buffer after this call.
5048 *
5049 * This code is optimized for readability - several passes might be
5050 * made over certain portions of the tree. All of those blocks will
5051 * have been brought into cache (and pinned via the journal), so the
5052 * extra overhead is not expressed in terms of disk reads.
5053 */
5054int ocfs2_split_extent(handle_t *handle,
5055 struct ocfs2_extent_tree *et,
5056 struct ocfs2_path *path,
5057 int split_index,
5058 struct ocfs2_extent_rec *split_rec,
5059 struct ocfs2_alloc_context *meta_ac,
5060 struct ocfs2_cached_dealloc_ctxt *dealloc)
5061{
5062 int ret = 0;
5063 struct ocfs2_extent_list *el = path_leaf_el(path);
5064 struct buffer_head *last_eb_bh = NULL;
5065 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
5066 struct ocfs2_merge_ctxt ctxt;
5067 struct ocfs2_extent_list *rightmost_el;
5068
5069 if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) ||
5070 ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) <
5071 (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) {
5072 ret = -EIO;
5073 mlog_errno(ret);
5074 goto out;
5075 }
5076
5077 ret = ocfs2_figure_merge_contig_type(et, path, el,
5078 split_index,
5079 split_rec,
5080 &ctxt);
5081 if (ret) {
5082 mlog_errno(ret);
5083 goto out;
5084 }
5085
5086 /*
5087 * The core merge / split code wants to know how much room is
5088 * left in this allocation tree, so we pass the
5089 * rightmost extent list.
5090 */
5091 if (path->p_tree_depth) {
5092 struct ocfs2_extent_block *eb;
5093
5094 ret = ocfs2_read_extent_block(et->et_ci,
5095 ocfs2_et_get_last_eb_blk(et),
5096 &last_eb_bh);
5097 if (ret) {
5098 mlog_errno(ret);
5099 goto out;
5100 }
5101
5102 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5103 rightmost_el = &eb->h_list;
5104 } else
5105 rightmost_el = path_root_el(path);
5106
5107 if (rec->e_cpos == split_rec->e_cpos &&
5108 rec->e_leaf_clusters == split_rec->e_leaf_clusters)
5109 ctxt.c_split_covers_rec = 1;
5110 else
5111 ctxt.c_split_covers_rec = 0;
5112
5113 ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
5114
5115 trace_ocfs2_split_extent(split_index, ctxt.c_contig_type,
5116 ctxt.c_has_empty_extent,
5117 ctxt.c_split_covers_rec);
5118
5119 if (ctxt.c_contig_type == CONTIG_NONE) {
5120 if (ctxt.c_split_covers_rec)
5121 ret = ocfs2_replace_extent_rec(handle, et, path, el,
5122 split_index, split_rec);
5123 else
5124 ret = ocfs2_split_and_insert(handle, et, path,
5125 &last_eb_bh, split_index,
5126 split_rec, meta_ac);
5127 if (ret)
5128 mlog_errno(ret);
5129 } else {
5130 ret = ocfs2_try_to_merge_extent(handle, et, path,
5131 split_index, split_rec,
5132 dealloc, &ctxt);
5133 if (ret)
5134 mlog_errno(ret);
5135 }
5136
5137out:
5138 brelse(last_eb_bh);
5139 return ret;
5140}
5141
5142/*
5143 * Change the flags of the already-existing extent at cpos for len clusters.
5144 *
5145 * new_flags: the flags we want to set.
5146 * clear_flags: the flags we want to clear.
5147 * phys: the new physical offset we want this new extent starts from.
5148 *
5149 * If the existing extent is larger than the request, initiate a
5150 * split. An attempt will be made at merging with adjacent extents.
5151 *
5152 * The caller is responsible for passing down meta_ac if we'll need it.
5153 */
5154int ocfs2_change_extent_flag(handle_t *handle,
5155 struct ocfs2_extent_tree *et,
5156 u32 cpos, u32 len, u32 phys,
5157 struct ocfs2_alloc_context *meta_ac,
5158 struct ocfs2_cached_dealloc_ctxt *dealloc,
5159 int new_flags, int clear_flags)
5160{
5161 int ret, index;
5162 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5163 u64 start_blkno = ocfs2_clusters_to_blocks(sb, phys);
5164 struct ocfs2_extent_rec split_rec;
5165 struct ocfs2_path *left_path = NULL;
5166 struct ocfs2_extent_list *el;
5167 struct ocfs2_extent_rec *rec;
5168
5169 left_path = ocfs2_new_path_from_et(et);
5170 if (!left_path) {
5171 ret = -ENOMEM;
5172 mlog_errno(ret);
5173 goto out;
5174 }
5175
5176 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
5177 if (ret) {
5178 mlog_errno(ret);
5179 goto out;
5180 }
5181 el = path_leaf_el(left_path);
5182
5183 index = ocfs2_search_extent_list(el, cpos);
5184 if (index == -1) {
5185 ocfs2_error(sb,
5186 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5187 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5188 cpos);
5189 ret = -EROFS;
5190 goto out;
5191 }
5192
5193 ret = -EIO;
5194 rec = &el->l_recs[index];
5195 if (new_flags && (rec->e_flags & new_flags)) {
5196 mlog(ML_ERROR, "Owner %llu tried to set %d flags on an "
5197 "extent that already had them\n",
5198 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5199 new_flags);
5200 goto out;
5201 }
5202
5203 if (clear_flags && !(rec->e_flags & clear_flags)) {
5204 mlog(ML_ERROR, "Owner %llu tried to clear %d flags on an "
5205 "extent that didn't have them\n",
5206 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5207 clear_flags);
5208 goto out;
5209 }
5210
5211 memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
5212 split_rec.e_cpos = cpu_to_le32(cpos);
5213 split_rec.e_leaf_clusters = cpu_to_le16(len);
5214 split_rec.e_blkno = cpu_to_le64(start_blkno);
5215 split_rec.e_flags = rec->e_flags;
5216 if (new_flags)
5217 split_rec.e_flags |= new_flags;
5218 if (clear_flags)
5219 split_rec.e_flags &= ~clear_flags;
5220
5221 ret = ocfs2_split_extent(handle, et, left_path,
5222 index, &split_rec, meta_ac,
5223 dealloc);
5224 if (ret)
5225 mlog_errno(ret);
5226
5227out:
5228 ocfs2_free_path(left_path);
5229 return ret;
5230
5231}
5232
5233/*
5234 * Mark the already-existing extent at cpos as written for len clusters.
5235 * This removes the unwritten extent flag.
5236 *
5237 * If the existing extent is larger than the request, initiate a
5238 * split. An attempt will be made at merging with adjacent extents.
5239 *
5240 * The caller is responsible for passing down meta_ac if we'll need it.
5241 */
5242int ocfs2_mark_extent_written(struct inode *inode,
5243 struct ocfs2_extent_tree *et,
5244 handle_t *handle, u32 cpos, u32 len, u32 phys,
5245 struct ocfs2_alloc_context *meta_ac,
5246 struct ocfs2_cached_dealloc_ctxt *dealloc)
5247{
5248 int ret;
5249
5250 trace_ocfs2_mark_extent_written(
5251 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5252 cpos, len, phys);
5253
5254 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
5255 ocfs2_error(inode->i_sb, "Inode %llu has unwritten extents that are being written to, but the feature bit is not set in the super block\n",
5256 (unsigned long long)OCFS2_I(inode)->ip_blkno);
5257 ret = -EROFS;
5258 goto out;
5259 }
5260
5261 /*
5262 * XXX: This should be fixed up so that we just re-insert the
5263 * next extent records.
5264 */
5265 ocfs2_et_extent_map_truncate(et, 0);
5266
5267 ret = ocfs2_change_extent_flag(handle, et, cpos,
5268 len, phys, meta_ac, dealloc,
5269 0, OCFS2_EXT_UNWRITTEN);
5270 if (ret)
5271 mlog_errno(ret);
5272
5273out:
5274 return ret;
5275}
5276
5277static int ocfs2_split_tree(handle_t *handle, struct ocfs2_extent_tree *et,
5278 struct ocfs2_path *path,
5279 int index, u32 new_range,
5280 struct ocfs2_alloc_context *meta_ac)
5281{
5282 int ret, depth, credits;
5283 struct buffer_head *last_eb_bh = NULL;
5284 struct ocfs2_extent_block *eb;
5285 struct ocfs2_extent_list *rightmost_el, *el;
5286 struct ocfs2_extent_rec split_rec;
5287 struct ocfs2_extent_rec *rec;
5288 struct ocfs2_insert_type insert;
5289
5290 /*
5291 * Setup the record to split before we grow the tree.
5292 */
5293 el = path_leaf_el(path);
5294 rec = &el->l_recs[index];
5295 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
5296 &split_rec, new_range, rec);
5297
5298 depth = path->p_tree_depth;
5299 if (depth > 0) {
5300 ret = ocfs2_read_extent_block(et->et_ci,
5301 ocfs2_et_get_last_eb_blk(et),
5302 &last_eb_bh);
5303 if (ret < 0) {
5304 mlog_errno(ret);
5305 goto out;
5306 }
5307
5308 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5309 rightmost_el = &eb->h_list;
5310 } else
5311 rightmost_el = path_leaf_el(path);
5312
5313 credits = path->p_tree_depth +
5314 ocfs2_extend_meta_needed(et->et_root_el);
5315 ret = ocfs2_extend_trans(handle, credits);
5316 if (ret) {
5317 mlog_errno(ret);
5318 goto out;
5319 }
5320
5321 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
5322 le16_to_cpu(rightmost_el->l_count)) {
5323 ret = ocfs2_grow_tree(handle, et, &depth, &last_eb_bh,
5324 meta_ac);
5325 if (ret) {
5326 mlog_errno(ret);
5327 goto out;
5328 }
5329 }
5330
5331 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
5332 insert.ins_appending = APPEND_NONE;
5333 insert.ins_contig = CONTIG_NONE;
5334 insert.ins_split = SPLIT_RIGHT;
5335 insert.ins_tree_depth = depth;
5336
5337 ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
5338 if (ret)
5339 mlog_errno(ret);
5340
5341out:
5342 brelse(last_eb_bh);
5343 return ret;
5344}
5345
5346static int ocfs2_truncate_rec(handle_t *handle,
5347 struct ocfs2_extent_tree *et,
5348 struct ocfs2_path *path, int index,
5349 struct ocfs2_cached_dealloc_ctxt *dealloc,
5350 u32 cpos, u32 len)
5351{
5352 int ret;
5353 u32 left_cpos, rec_range, trunc_range;
5354 int is_rightmost_tree_rec = 0;
5355 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5356 struct ocfs2_path *left_path = NULL;
5357 struct ocfs2_extent_list *el = path_leaf_el(path);
5358 struct ocfs2_extent_rec *rec;
5359 struct ocfs2_extent_block *eb;
5360
5361 if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
5362 /* extend credit for ocfs2_remove_rightmost_path */
5363 ret = ocfs2_extend_rotate_transaction(handle, 0,
5364 handle->h_buffer_credits,
5365 path);
5366 if (ret) {
5367 mlog_errno(ret);
5368 goto out;
5369 }
5370
5371 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5372 if (ret) {
5373 mlog_errno(ret);
5374 goto out;
5375 }
5376
5377 index--;
5378 }
5379
5380 if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
5381 path->p_tree_depth) {
5382 /*
5383 * Check whether this is the rightmost tree record. If
5384 * we remove all of this record or part of its right
5385 * edge then an update of the record lengths above it
5386 * will be required.
5387 */
5388 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
5389 if (eb->h_next_leaf_blk == 0)
5390 is_rightmost_tree_rec = 1;
5391 }
5392
5393 rec = &el->l_recs[index];
5394 if (index == 0 && path->p_tree_depth &&
5395 le32_to_cpu(rec->e_cpos) == cpos) {
5396 /*
5397 * Changing the leftmost offset (via partial or whole
5398 * record truncate) of an interior (or rightmost) path
5399 * means we have to update the subtree that is formed
5400 * by this leaf and the one to it's left.
5401 *
5402 * There are two cases we can skip:
5403 * 1) Path is the leftmost one in our btree.
5404 * 2) The leaf is rightmost and will be empty after
5405 * we remove the extent record - the rotate code
5406 * knows how to update the newly formed edge.
5407 */
5408
5409 ret = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
5410 if (ret) {
5411 mlog_errno(ret);
5412 goto out;
5413 }
5414
5415 if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
5416 left_path = ocfs2_new_path_from_path(path);
5417 if (!left_path) {
5418 ret = -ENOMEM;
5419 mlog_errno(ret);
5420 goto out;
5421 }
5422
5423 ret = ocfs2_find_path(et->et_ci, left_path,
5424 left_cpos);
5425 if (ret) {
5426 mlog_errno(ret);
5427 goto out;
5428 }
5429 }
5430 }
5431
5432 ret = ocfs2_extend_rotate_transaction(handle, 0,
5433 handle->h_buffer_credits,
5434 path);
5435 if (ret) {
5436 mlog_errno(ret);
5437 goto out;
5438 }
5439
5440 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
5441 if (ret) {
5442 mlog_errno(ret);
5443 goto out;
5444 }
5445
5446 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
5447 if (ret) {
5448 mlog_errno(ret);
5449 goto out;
5450 }
5451
5452 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5453 trunc_range = cpos + len;
5454
5455 if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
5456 int next_free;
5457
5458 memset(rec, 0, sizeof(*rec));
5459 ocfs2_cleanup_merge(el, index);
5460
5461 next_free = le16_to_cpu(el->l_next_free_rec);
5462 if (is_rightmost_tree_rec && next_free > 1) {
5463 /*
5464 * We skip the edge update if this path will
5465 * be deleted by the rotate code.
5466 */
5467 rec = &el->l_recs[next_free - 1];
5468 ocfs2_adjust_rightmost_records(handle, et, path,
5469 rec);
5470 }
5471 } else if (le32_to_cpu(rec->e_cpos) == cpos) {
5472 /* Remove leftmost portion of the record. */
5473 le32_add_cpu(&rec->e_cpos, len);
5474 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
5475 le16_add_cpu(&rec->e_leaf_clusters, -len);
5476 } else if (rec_range == trunc_range) {
5477 /* Remove rightmost portion of the record */
5478 le16_add_cpu(&rec->e_leaf_clusters, -len);
5479 if (is_rightmost_tree_rec)
5480 ocfs2_adjust_rightmost_records(handle, et, path, rec);
5481 } else {
5482 /* Caller should have trapped this. */
5483 mlog(ML_ERROR, "Owner %llu: Invalid record truncate: (%u, %u) "
5484 "(%u, %u)\n",
5485 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5486 le32_to_cpu(rec->e_cpos),
5487 le16_to_cpu(rec->e_leaf_clusters), cpos, len);
5488 BUG();
5489 }
5490
5491 if (left_path) {
5492 int subtree_index;
5493
5494 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
5495 ocfs2_complete_edge_insert(handle, left_path, path,
5496 subtree_index);
5497 }
5498
5499 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5500
5501 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5502 if (ret) {
5503 mlog_errno(ret);
5504 goto out;
5505 }
5506
5507out:
5508 ocfs2_free_path(left_path);
5509 return ret;
5510}
5511
5512int ocfs2_remove_extent(handle_t *handle,
5513 struct ocfs2_extent_tree *et,
5514 u32 cpos, u32 len,
5515 struct ocfs2_alloc_context *meta_ac,
5516 struct ocfs2_cached_dealloc_ctxt *dealloc)
5517{
5518 int ret, index;
5519 u32 rec_range, trunc_range;
5520 struct ocfs2_extent_rec *rec;
5521 struct ocfs2_extent_list *el;
5522 struct ocfs2_path *path = NULL;
5523
5524 /*
5525 * XXX: Why are we truncating to 0 instead of wherever this
5526 * affects us?
5527 */
5528 ocfs2_et_extent_map_truncate(et, 0);
5529
5530 path = ocfs2_new_path_from_et(et);
5531 if (!path) {
5532 ret = -ENOMEM;
5533 mlog_errno(ret);
5534 goto out;
5535 }
5536
5537 ret = ocfs2_find_path(et->et_ci, path, cpos);
5538 if (ret) {
5539 mlog_errno(ret);
5540 goto out;
5541 }
5542
5543 el = path_leaf_el(path);
5544 index = ocfs2_search_extent_list(el, cpos);
5545 if (index == -1) {
5546 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5547 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5548 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5549 cpos);
5550 ret = -EROFS;
5551 goto out;
5552 }
5553
5554 /*
5555 * We have 3 cases of extent removal:
5556 * 1) Range covers the entire extent rec
5557 * 2) Range begins or ends on one edge of the extent rec
5558 * 3) Range is in the middle of the extent rec (no shared edges)
5559 *
5560 * For case 1 we remove the extent rec and left rotate to
5561 * fill the hole.
5562 *
5563 * For case 2 we just shrink the existing extent rec, with a
5564 * tree update if the shrinking edge is also the edge of an
5565 * extent block.
5566 *
5567 * For case 3 we do a right split to turn the extent rec into
5568 * something case 2 can handle.
5569 */
5570 rec = &el->l_recs[index];
5571 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5572 trunc_range = cpos + len;
5573
5574 BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
5575
5576 trace_ocfs2_remove_extent(
5577 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5578 cpos, len, index, le32_to_cpu(rec->e_cpos),
5579 ocfs2_rec_clusters(el, rec));
5580
5581 if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
5582 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5583 cpos, len);
5584 if (ret) {
5585 mlog_errno(ret);
5586 goto out;
5587 }
5588 } else {
5589 ret = ocfs2_split_tree(handle, et, path, index,
5590 trunc_range, meta_ac);
5591 if (ret) {
5592 mlog_errno(ret);
5593 goto out;
5594 }
5595
5596 /*
5597 * The split could have manipulated the tree enough to
5598 * move the record location, so we have to look for it again.
5599 */
5600 ocfs2_reinit_path(path, 1);
5601
5602 ret = ocfs2_find_path(et->et_ci, path, cpos);
5603 if (ret) {
5604 mlog_errno(ret);
5605 goto out;
5606 }
5607
5608 el = path_leaf_el(path);
5609 index = ocfs2_search_extent_list(el, cpos);
5610 if (index == -1) {
5611 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5612 "Owner %llu: split at cpos %u lost record\n",
5613 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5614 cpos);
5615 ret = -EROFS;
5616 goto out;
5617 }
5618
5619 /*
5620 * Double check our values here. If anything is fishy,
5621 * it's easier to catch it at the top level.
5622 */
5623 rec = &el->l_recs[index];
5624 rec_range = le32_to_cpu(rec->e_cpos) +
5625 ocfs2_rec_clusters(el, rec);
5626 if (rec_range != trunc_range) {
5627 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5628 "Owner %llu: error after split at cpos %u trunc len %u, existing record is (%u,%u)\n",
5629 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5630 cpos, len, le32_to_cpu(rec->e_cpos),
5631 ocfs2_rec_clusters(el, rec));
5632 ret = -EROFS;
5633 goto out;
5634 }
5635
5636 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5637 cpos, len);
5638 if (ret) {
5639 mlog_errno(ret);
5640 goto out;
5641 }
5642 }
5643
5644out:
5645 ocfs2_free_path(path);
5646 return ret;
5647}
5648
5649/*
5650 * ocfs2_reserve_blocks_for_rec_trunc() would look basically the
5651 * same as ocfs2_lock_alloctors(), except for it accepts a blocks
5652 * number to reserve some extra blocks, and it only handles meta
5653 * data allocations.
5654 *
5655 * Currently, only ocfs2_remove_btree_range() uses it for truncating
5656 * and punching holes.
5657 */
5658static int ocfs2_reserve_blocks_for_rec_trunc(struct inode *inode,
5659 struct ocfs2_extent_tree *et,
5660 u32 extents_to_split,
5661 struct ocfs2_alloc_context **ac,
5662 int extra_blocks)
5663{
5664 int ret = 0, num_free_extents;
5665 unsigned int max_recs_needed = 2 * extents_to_split;
5666 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5667
5668 *ac = NULL;
5669
5670 num_free_extents = ocfs2_num_free_extents(osb, et);
5671 if (num_free_extents < 0) {
5672 ret = num_free_extents;
5673 mlog_errno(ret);
5674 goto out;
5675 }
5676
5677 if (!num_free_extents ||
5678 (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
5679 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
5680
5681 if (extra_blocks) {
5682 ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, ac);
5683 if (ret < 0) {
5684 if (ret != -ENOSPC)
5685 mlog_errno(ret);
5686 goto out;
5687 }
5688 }
5689
5690out:
5691 if (ret) {
5692 if (*ac) {
5693 ocfs2_free_alloc_context(*ac);
5694 *ac = NULL;
5695 }
5696 }
5697
5698 return ret;
5699}
5700
5701int ocfs2_remove_btree_range(struct inode *inode,
5702 struct ocfs2_extent_tree *et,
5703 u32 cpos, u32 phys_cpos, u32 len, int flags,
5704 struct ocfs2_cached_dealloc_ctxt *dealloc,
5705 u64 refcount_loc, bool refcount_tree_locked)
5706{
5707 int ret, credits = 0, extra_blocks = 0;
5708 u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
5709 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5710 struct inode *tl_inode = osb->osb_tl_inode;
5711 handle_t *handle;
5712 struct ocfs2_alloc_context *meta_ac = NULL;
5713 struct ocfs2_refcount_tree *ref_tree = NULL;
5714
5715 if ((flags & OCFS2_EXT_REFCOUNTED) && len) {
5716 BUG_ON(!ocfs2_is_refcount_inode(inode));
5717
5718 if (!refcount_tree_locked) {
5719 ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
5720 &ref_tree, NULL);
5721 if (ret) {
5722 mlog_errno(ret);
5723 goto bail;
5724 }
5725 }
5726
5727 ret = ocfs2_prepare_refcount_change_for_del(inode,
5728 refcount_loc,
5729 phys_blkno,
5730 len,
5731 &credits,
5732 &extra_blocks);
5733 if (ret < 0) {
5734 mlog_errno(ret);
5735 goto bail;
5736 }
5737 }
5738
5739 ret = ocfs2_reserve_blocks_for_rec_trunc(inode, et, 1, &meta_ac,
5740 extra_blocks);
5741 if (ret) {
5742 mlog_errno(ret);
5743 goto bail;
5744 }
5745
5746 inode_lock(tl_inode);
5747
5748 if (ocfs2_truncate_log_needs_flush(osb)) {
5749 ret = __ocfs2_flush_truncate_log(osb);
5750 if (ret < 0) {
5751 mlog_errno(ret);
5752 goto out;
5753 }
5754 }
5755
5756 handle = ocfs2_start_trans(osb,
5757 ocfs2_remove_extent_credits(osb->sb) + credits);
5758 if (IS_ERR(handle)) {
5759 ret = PTR_ERR(handle);
5760 mlog_errno(ret);
5761 goto out;
5762 }
5763
5764 ret = ocfs2_et_root_journal_access(handle, et,
5765 OCFS2_JOURNAL_ACCESS_WRITE);
5766 if (ret) {
5767 mlog_errno(ret);
5768 goto out_commit;
5769 }
5770
5771 dquot_free_space_nodirty(inode,
5772 ocfs2_clusters_to_bytes(inode->i_sb, len));
5773
5774 ret = ocfs2_remove_extent(handle, et, cpos, len, meta_ac, dealloc);
5775 if (ret) {
5776 mlog_errno(ret);
5777 goto out_commit;
5778 }
5779
5780 ocfs2_et_update_clusters(et, -len);
5781 ocfs2_update_inode_fsync_trans(handle, inode, 1);
5782
5783 ocfs2_journal_dirty(handle, et->et_root_bh);
5784
5785 if (phys_blkno) {
5786 if (flags & OCFS2_EXT_REFCOUNTED)
5787 ret = ocfs2_decrease_refcount(inode, handle,
5788 ocfs2_blocks_to_clusters(osb->sb,
5789 phys_blkno),
5790 len, meta_ac,
5791 dealloc, 1);
5792 else
5793 ret = ocfs2_truncate_log_append(osb, handle,
5794 phys_blkno, len);
5795 if (ret)
5796 mlog_errno(ret);
5797
5798 }
5799
5800out_commit:
5801 ocfs2_commit_trans(osb, handle);
5802out:
5803 inode_unlock(tl_inode);
5804bail:
5805 if (meta_ac)
5806 ocfs2_free_alloc_context(meta_ac);
5807
5808 if (ref_tree)
5809 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
5810
5811 return ret;
5812}
5813
5814int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
5815{
5816 struct buffer_head *tl_bh = osb->osb_tl_bh;
5817 struct ocfs2_dinode *di;
5818 struct ocfs2_truncate_log *tl;
5819
5820 di = (struct ocfs2_dinode *) tl_bh->b_data;
5821 tl = &di->id2.i_dealloc;
5822
5823 mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
5824 "slot %d, invalid truncate log parameters: used = "
5825 "%u, count = %u\n", osb->slot_num,
5826 le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
5827 return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
5828}
5829
5830static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
5831 unsigned int new_start)
5832{
5833 unsigned int tail_index;
5834 unsigned int current_tail;
5835
5836 /* No records, nothing to coalesce */
5837 if (!le16_to_cpu(tl->tl_used))
5838 return 0;
5839
5840 tail_index = le16_to_cpu(tl->tl_used) - 1;
5841 current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
5842 current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
5843
5844 return current_tail == new_start;
5845}
5846
5847int ocfs2_truncate_log_append(struct ocfs2_super *osb,
5848 handle_t *handle,
5849 u64 start_blk,
5850 unsigned int num_clusters)
5851{
5852 int status, index;
5853 unsigned int start_cluster, tl_count;
5854 struct inode *tl_inode = osb->osb_tl_inode;
5855 struct buffer_head *tl_bh = osb->osb_tl_bh;
5856 struct ocfs2_dinode *di;
5857 struct ocfs2_truncate_log *tl;
5858
5859 BUG_ON(inode_trylock(tl_inode));
5860
5861 start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
5862
5863 di = (struct ocfs2_dinode *) tl_bh->b_data;
5864
5865 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5866 * by the underlying call to ocfs2_read_inode_block(), so any
5867 * corruption is a code bug */
5868 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5869
5870 tl = &di->id2.i_dealloc;
5871 tl_count = le16_to_cpu(tl->tl_count);
5872 mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
5873 tl_count == 0,
5874 "Truncate record count on #%llu invalid "
5875 "wanted %u, actual %u\n",
5876 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5877 ocfs2_truncate_recs_per_inode(osb->sb),
5878 le16_to_cpu(tl->tl_count));
5879
5880 /* Caller should have known to flush before calling us. */
5881 index = le16_to_cpu(tl->tl_used);
5882 if (index >= tl_count) {
5883 status = -ENOSPC;
5884 mlog_errno(status);
5885 goto bail;
5886 }
5887
5888 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5889 OCFS2_JOURNAL_ACCESS_WRITE);
5890 if (status < 0) {
5891 mlog_errno(status);
5892 goto bail;
5893 }
5894
5895 trace_ocfs2_truncate_log_append(
5896 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index,
5897 start_cluster, num_clusters);
5898 if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
5899 /*
5900 * Move index back to the record we are coalescing with.
5901 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5902 */
5903 index--;
5904
5905 num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
5906 trace_ocfs2_truncate_log_append(
5907 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5908 index, le32_to_cpu(tl->tl_recs[index].t_start),
5909 num_clusters);
5910 } else {
5911 tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
5912 tl->tl_used = cpu_to_le16(index + 1);
5913 }
5914 tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
5915
5916 ocfs2_journal_dirty(handle, tl_bh);
5917
5918 osb->truncated_clusters += num_clusters;
5919bail:
5920 return status;
5921}
5922
5923static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
5924 struct inode *data_alloc_inode,
5925 struct buffer_head *data_alloc_bh)
5926{
5927 int status = 0;
5928 int i;
5929 unsigned int num_clusters;
5930 u64 start_blk;
5931 struct ocfs2_truncate_rec rec;
5932 struct ocfs2_dinode *di;
5933 struct ocfs2_truncate_log *tl;
5934 struct inode *tl_inode = osb->osb_tl_inode;
5935 struct buffer_head *tl_bh = osb->osb_tl_bh;
5936 handle_t *handle;
5937
5938 di = (struct ocfs2_dinode *) tl_bh->b_data;
5939 tl = &di->id2.i_dealloc;
5940 i = le16_to_cpu(tl->tl_used) - 1;
5941 while (i >= 0) {
5942 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
5943 if (IS_ERR(handle)) {
5944 status = PTR_ERR(handle);
5945 mlog_errno(status);
5946 goto bail;
5947 }
5948
5949 /* Caller has given us at least enough credits to
5950 * update the truncate log dinode */
5951 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5952 OCFS2_JOURNAL_ACCESS_WRITE);
5953 if (status < 0) {
5954 mlog_errno(status);
5955 goto bail;
5956 }
5957
5958 tl->tl_used = cpu_to_le16(i);
5959
5960 ocfs2_journal_dirty(handle, tl_bh);
5961
5962 rec = tl->tl_recs[i];
5963 start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
5964 le32_to_cpu(rec.t_start));
5965 num_clusters = le32_to_cpu(rec.t_clusters);
5966
5967 /* if start_blk is not set, we ignore the record as
5968 * invalid. */
5969 if (start_blk) {
5970 trace_ocfs2_replay_truncate_records(
5971 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5972 i, le32_to_cpu(rec.t_start), num_clusters);
5973
5974 status = ocfs2_free_clusters(handle, data_alloc_inode,
5975 data_alloc_bh, start_blk,
5976 num_clusters);
5977 if (status < 0) {
5978 mlog_errno(status);
5979 goto bail;
5980 }
5981 }
5982
5983 ocfs2_commit_trans(osb, handle);
5984 i--;
5985 }
5986
5987 osb->truncated_clusters = 0;
5988
5989bail:
5990 return status;
5991}
5992
5993/* Expects you to already be holding tl_inode->i_mutex */
5994int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5995{
5996 int status;
5997 unsigned int num_to_flush;
5998 struct inode *tl_inode = osb->osb_tl_inode;
5999 struct inode *data_alloc_inode = NULL;
6000 struct buffer_head *tl_bh = osb->osb_tl_bh;
6001 struct buffer_head *data_alloc_bh = NULL;
6002 struct ocfs2_dinode *di;
6003 struct ocfs2_truncate_log *tl;
6004
6005 BUG_ON(inode_trylock(tl_inode));
6006
6007 di = (struct ocfs2_dinode *) tl_bh->b_data;
6008
6009 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
6010 * by the underlying call to ocfs2_read_inode_block(), so any
6011 * corruption is a code bug */
6012 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6013
6014 tl = &di->id2.i_dealloc;
6015 num_to_flush = le16_to_cpu(tl->tl_used);
6016 trace_ocfs2_flush_truncate_log(
6017 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
6018 num_to_flush);
6019 if (!num_to_flush) {
6020 status = 0;
6021 goto out;
6022 }
6023
6024 data_alloc_inode = ocfs2_get_system_file_inode(osb,
6025 GLOBAL_BITMAP_SYSTEM_INODE,
6026 OCFS2_INVALID_SLOT);
6027 if (!data_alloc_inode) {
6028 status = -EINVAL;
6029 mlog(ML_ERROR, "Could not get bitmap inode!\n");
6030 goto out;
6031 }
6032
6033 inode_lock(data_alloc_inode);
6034
6035 status = ocfs2_inode_lock(data_alloc_inode, &data_alloc_bh, 1);
6036 if (status < 0) {
6037 mlog_errno(status);
6038 goto out_mutex;
6039 }
6040
6041 status = ocfs2_replay_truncate_records(osb, data_alloc_inode,
6042 data_alloc_bh);
6043 if (status < 0)
6044 mlog_errno(status);
6045
6046 brelse(data_alloc_bh);
6047 ocfs2_inode_unlock(data_alloc_inode, 1);
6048
6049out_mutex:
6050 inode_unlock(data_alloc_inode);
6051 iput(data_alloc_inode);
6052
6053out:
6054 return status;
6055}
6056
6057int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
6058{
6059 int status;
6060 struct inode *tl_inode = osb->osb_tl_inode;
6061
6062 inode_lock(tl_inode);
6063 status = __ocfs2_flush_truncate_log(osb);
6064 inode_unlock(tl_inode);
6065
6066 return status;
6067}
6068
6069static void ocfs2_truncate_log_worker(struct work_struct *work)
6070{
6071 int status;
6072 struct ocfs2_super *osb =
6073 container_of(work, struct ocfs2_super,
6074 osb_truncate_log_wq.work);
6075
6076 status = ocfs2_flush_truncate_log(osb);
6077 if (status < 0)
6078 mlog_errno(status);
6079 else
6080 ocfs2_init_steal_slots(osb);
6081}
6082
6083#define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
6084void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
6085 int cancel)
6086{
6087 if (osb->osb_tl_inode &&
6088 atomic_read(&osb->osb_tl_disable) == 0) {
6089 /* We want to push off log flushes while truncates are
6090 * still running. */
6091 if (cancel)
6092 cancel_delayed_work(&osb->osb_truncate_log_wq);
6093
6094 queue_delayed_work(osb->ocfs2_wq, &osb->osb_truncate_log_wq,
6095 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
6096 }
6097}
6098
6099/*
6100 * Try to flush truncate logs if we can free enough clusters from it.
6101 * As for return value, "< 0" means error, "0" no space and "1" means
6102 * we have freed enough spaces and let the caller try to allocate again.
6103 */
6104int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
6105 unsigned int needed)
6106{
6107 tid_t target;
6108 int ret = 0;
6109 unsigned int truncated_clusters;
6110
6111 inode_lock(osb->osb_tl_inode);
6112 truncated_clusters = osb->truncated_clusters;
6113 inode_unlock(osb->osb_tl_inode);
6114
6115 /*
6116 * Check whether we can succeed in allocating if we free
6117 * the truncate log.
6118 */
6119 if (truncated_clusters < needed)
6120 goto out;
6121
6122 ret = ocfs2_flush_truncate_log(osb);
6123 if (ret) {
6124 mlog_errno(ret);
6125 goto out;
6126 }
6127
6128 if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
6129 jbd2_log_wait_commit(osb->journal->j_journal, target);
6130 ret = 1;
6131 }
6132out:
6133 return ret;
6134}
6135
6136static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
6137 int slot_num,
6138 struct inode **tl_inode,
6139 struct buffer_head **tl_bh)
6140{
6141 int status;
6142 struct inode *inode = NULL;
6143 struct buffer_head *bh = NULL;
6144
6145 inode = ocfs2_get_system_file_inode(osb,
6146 TRUNCATE_LOG_SYSTEM_INODE,
6147 slot_num);
6148 if (!inode) {
6149 status = -EINVAL;
6150 mlog(ML_ERROR, "Could not get load truncate log inode!\n");
6151 goto bail;
6152 }
6153
6154 status = ocfs2_read_inode_block(inode, &bh);
6155 if (status < 0) {
6156 iput(inode);
6157 mlog_errno(status);
6158 goto bail;
6159 }
6160
6161 *tl_inode = inode;
6162 *tl_bh = bh;
6163bail:
6164 return status;
6165}
6166
6167/* called during the 1st stage of node recovery. we stamp a clean
6168 * truncate log and pass back a copy for processing later. if the
6169 * truncate log does not require processing, a *tl_copy is set to
6170 * NULL. */
6171int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
6172 int slot_num,
6173 struct ocfs2_dinode **tl_copy)
6174{
6175 int status;
6176 struct inode *tl_inode = NULL;
6177 struct buffer_head *tl_bh = NULL;
6178 struct ocfs2_dinode *di;
6179 struct ocfs2_truncate_log *tl;
6180
6181 *tl_copy = NULL;
6182
6183 trace_ocfs2_begin_truncate_log_recovery(slot_num);
6184
6185 status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
6186 if (status < 0) {
6187 mlog_errno(status);
6188 goto bail;
6189 }
6190
6191 di = (struct ocfs2_dinode *) tl_bh->b_data;
6192
6193 /* tl_bh is loaded from ocfs2_get_truncate_log_info(). It's
6194 * validated by the underlying call to ocfs2_read_inode_block(),
6195 * so any corruption is a code bug */
6196 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6197
6198 tl = &di->id2.i_dealloc;
6199 if (le16_to_cpu(tl->tl_used)) {
6200 trace_ocfs2_truncate_log_recovery_num(le16_to_cpu(tl->tl_used));
6201
6202 *tl_copy = kmalloc(tl_bh->b_size, GFP_KERNEL);
6203 if (!(*tl_copy)) {
6204 status = -ENOMEM;
6205 mlog_errno(status);
6206 goto bail;
6207 }
6208
6209 /* Assuming the write-out below goes well, this copy
6210 * will be passed back to recovery for processing. */
6211 memcpy(*tl_copy, tl_bh->b_data, tl_bh->b_size);
6212
6213 /* All we need to do to clear the truncate log is set
6214 * tl_used. */
6215 tl->tl_used = 0;
6216
6217 ocfs2_compute_meta_ecc(osb->sb, tl_bh->b_data, &di->i_check);
6218 status = ocfs2_write_block(osb, tl_bh, INODE_CACHE(tl_inode));
6219 if (status < 0) {
6220 mlog_errno(status);
6221 goto bail;
6222 }
6223 }
6224
6225bail:
6226 iput(tl_inode);
6227 brelse(tl_bh);
6228
6229 if (status < 0) {
6230 kfree(*tl_copy);
6231 *tl_copy = NULL;
6232 mlog_errno(status);
6233 }
6234
6235 return status;
6236}
6237
6238int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
6239 struct ocfs2_dinode *tl_copy)
6240{
6241 int status = 0;
6242 int i;
6243 unsigned int clusters, num_recs, start_cluster;
6244 u64 start_blk;
6245 handle_t *handle;
6246 struct inode *tl_inode = osb->osb_tl_inode;
6247 struct ocfs2_truncate_log *tl;
6248
6249 if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
6250 mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
6251 return -EINVAL;
6252 }
6253
6254 tl = &tl_copy->id2.i_dealloc;
6255 num_recs = le16_to_cpu(tl->tl_used);
6256 trace_ocfs2_complete_truncate_log_recovery(
6257 (unsigned long long)le64_to_cpu(tl_copy->i_blkno),
6258 num_recs);
6259
6260 inode_lock(tl_inode);
6261 for(i = 0; i < num_recs; i++) {
6262 if (ocfs2_truncate_log_needs_flush(osb)) {
6263 status = __ocfs2_flush_truncate_log(osb);
6264 if (status < 0) {
6265 mlog_errno(status);
6266 goto bail_up;
6267 }
6268 }
6269
6270 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6271 if (IS_ERR(handle)) {
6272 status = PTR_ERR(handle);
6273 mlog_errno(status);
6274 goto bail_up;
6275 }
6276
6277 clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
6278 start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
6279 start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
6280
6281 status = ocfs2_truncate_log_append(osb, handle,
6282 start_blk, clusters);
6283 ocfs2_commit_trans(osb, handle);
6284 if (status < 0) {
6285 mlog_errno(status);
6286 goto bail_up;
6287 }
6288 }
6289
6290bail_up:
6291 inode_unlock(tl_inode);
6292
6293 return status;
6294}
6295
6296void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
6297{
6298 int status;
6299 struct inode *tl_inode = osb->osb_tl_inode;
6300
6301 atomic_set(&osb->osb_tl_disable, 1);
6302
6303 if (tl_inode) {
6304 cancel_delayed_work(&osb->osb_truncate_log_wq);
6305 flush_workqueue(osb->ocfs2_wq);
6306
6307 status = ocfs2_flush_truncate_log(osb);
6308 if (status < 0)
6309 mlog_errno(status);
6310
6311 brelse(osb->osb_tl_bh);
6312 iput(osb->osb_tl_inode);
6313 }
6314}
6315
6316int ocfs2_truncate_log_init(struct ocfs2_super *osb)
6317{
6318 int status;
6319 struct inode *tl_inode = NULL;
6320 struct buffer_head *tl_bh = NULL;
6321
6322 status = ocfs2_get_truncate_log_info(osb,
6323 osb->slot_num,
6324 &tl_inode,
6325 &tl_bh);
6326 if (status < 0)
6327 mlog_errno(status);
6328
6329 /* ocfs2_truncate_log_shutdown keys on the existence of
6330 * osb->osb_tl_inode so we don't set any of the osb variables
6331 * until we're sure all is well. */
6332 INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
6333 ocfs2_truncate_log_worker);
6334 atomic_set(&osb->osb_tl_disable, 0);
6335 osb->osb_tl_bh = tl_bh;
6336 osb->osb_tl_inode = tl_inode;
6337
6338 return status;
6339}
6340
6341/*
6342 * Delayed de-allocation of suballocator blocks.
6343 *
6344 * Some sets of block de-allocations might involve multiple suballocator inodes.
6345 *
6346 * The locking for this can get extremely complicated, especially when
6347 * the suballocator inodes to delete from aren't known until deep
6348 * within an unrelated codepath.
6349 *
6350 * ocfs2_extent_block structures are a good example of this - an inode
6351 * btree could have been grown by any number of nodes each allocating
6352 * out of their own suballoc inode.
6353 *
6354 * These structures allow the delay of block de-allocation until a
6355 * later time, when locking of multiple cluster inodes won't cause
6356 * deadlock.
6357 */
6358
6359/*
6360 * Describe a single bit freed from a suballocator. For the block
6361 * suballocators, it represents one block. For the global cluster
6362 * allocator, it represents some clusters and free_bit indicates
6363 * clusters number.
6364 */
6365struct ocfs2_cached_block_free {
6366 struct ocfs2_cached_block_free *free_next;
6367 u64 free_bg;
6368 u64 free_blk;
6369 unsigned int free_bit;
6370};
6371
6372struct ocfs2_per_slot_free_list {
6373 struct ocfs2_per_slot_free_list *f_next_suballocator;
6374 int f_inode_type;
6375 int f_slot;
6376 struct ocfs2_cached_block_free *f_first;
6377};
6378
6379static int ocfs2_free_cached_blocks(struct ocfs2_super *osb,
6380 int sysfile_type,
6381 int slot,
6382 struct ocfs2_cached_block_free *head)
6383{
6384 int ret;
6385 u64 bg_blkno;
6386 handle_t *handle;
6387 struct inode *inode;
6388 struct buffer_head *di_bh = NULL;
6389 struct ocfs2_cached_block_free *tmp;
6390
6391 inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
6392 if (!inode) {
6393 ret = -EINVAL;
6394 mlog_errno(ret);
6395 goto out;
6396 }
6397
6398 inode_lock(inode);
6399
6400 ret = ocfs2_inode_lock(inode, &di_bh, 1);
6401 if (ret) {
6402 mlog_errno(ret);
6403 goto out_mutex;
6404 }
6405
6406 while (head) {
6407 if (head->free_bg)
6408 bg_blkno = head->free_bg;
6409 else
6410 bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
6411 head->free_bit);
6412 handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
6413 if (IS_ERR(handle)) {
6414 ret = PTR_ERR(handle);
6415 mlog_errno(ret);
6416 goto out_unlock;
6417 }
6418
6419 trace_ocfs2_free_cached_blocks(
6420 (unsigned long long)head->free_blk, head->free_bit);
6421
6422 ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
6423 head->free_bit, bg_blkno, 1);
6424 if (ret)
6425 mlog_errno(ret);
6426
6427 ocfs2_commit_trans(osb, handle);
6428
6429 tmp = head;
6430 head = head->free_next;
6431 kfree(tmp);
6432 }
6433
6434out_unlock:
6435 ocfs2_inode_unlock(inode, 1);
6436 brelse(di_bh);
6437out_mutex:
6438 inode_unlock(inode);
6439 iput(inode);
6440out:
6441 while(head) {
6442 /* Premature exit may have left some dangling items. */
6443 tmp = head;
6444 head = head->free_next;
6445 kfree(tmp);
6446 }
6447
6448 return ret;
6449}
6450
6451int ocfs2_cache_cluster_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6452 u64 blkno, unsigned int bit)
6453{
6454 int ret = 0;
6455 struct ocfs2_cached_block_free *item;
6456
6457 item = kzalloc(sizeof(*item), GFP_NOFS);
6458 if (item == NULL) {
6459 ret = -ENOMEM;
6460 mlog_errno(ret);
6461 return ret;
6462 }
6463
6464 trace_ocfs2_cache_cluster_dealloc((unsigned long long)blkno, bit);
6465
6466 item->free_blk = blkno;
6467 item->free_bit = bit;
6468 item->free_next = ctxt->c_global_allocator;
6469
6470 ctxt->c_global_allocator = item;
6471 return ret;
6472}
6473
6474static int ocfs2_free_cached_clusters(struct ocfs2_super *osb,
6475 struct ocfs2_cached_block_free *head)
6476{
6477 struct ocfs2_cached_block_free *tmp;
6478 struct inode *tl_inode = osb->osb_tl_inode;
6479 handle_t *handle;
6480 int ret = 0;
6481
6482 inode_lock(tl_inode);
6483
6484 while (head) {
6485 if (ocfs2_truncate_log_needs_flush(osb)) {
6486 ret = __ocfs2_flush_truncate_log(osb);
6487 if (ret < 0) {
6488 mlog_errno(ret);
6489 break;
6490 }
6491 }
6492
6493 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6494 if (IS_ERR(handle)) {
6495 ret = PTR_ERR(handle);
6496 mlog_errno(ret);
6497 break;
6498 }
6499
6500 ret = ocfs2_truncate_log_append(osb, handle, head->free_blk,
6501 head->free_bit);
6502
6503 ocfs2_commit_trans(osb, handle);
6504 tmp = head;
6505 head = head->free_next;
6506 kfree(tmp);
6507
6508 if (ret < 0) {
6509 mlog_errno(ret);
6510 break;
6511 }
6512 }
6513
6514 inode_unlock(tl_inode);
6515
6516 while (head) {
6517 /* Premature exit may have left some dangling items. */
6518 tmp = head;
6519 head = head->free_next;
6520 kfree(tmp);
6521 }
6522
6523 return ret;
6524}
6525
6526int ocfs2_run_deallocs(struct ocfs2_super *osb,
6527 struct ocfs2_cached_dealloc_ctxt *ctxt)
6528{
6529 int ret = 0, ret2;
6530 struct ocfs2_per_slot_free_list *fl;
6531
6532 if (!ctxt)
6533 return 0;
6534
6535 while (ctxt->c_first_suballocator) {
6536 fl = ctxt->c_first_suballocator;
6537
6538 if (fl->f_first) {
6539 trace_ocfs2_run_deallocs(fl->f_inode_type,
6540 fl->f_slot);
6541 ret2 = ocfs2_free_cached_blocks(osb,
6542 fl->f_inode_type,
6543 fl->f_slot,
6544 fl->f_first);
6545 if (ret2)
6546 mlog_errno(ret2);
6547 if (!ret)
6548 ret = ret2;
6549 }
6550
6551 ctxt->c_first_suballocator = fl->f_next_suballocator;
6552 kfree(fl);
6553 }
6554
6555 if (ctxt->c_global_allocator) {
6556 ret2 = ocfs2_free_cached_clusters(osb,
6557 ctxt->c_global_allocator);
6558 if (ret2)
6559 mlog_errno(ret2);
6560 if (!ret)
6561 ret = ret2;
6562
6563 ctxt->c_global_allocator = NULL;
6564 }
6565
6566 return ret;
6567}
6568
6569static struct ocfs2_per_slot_free_list *
6570ocfs2_find_per_slot_free_list(int type,
6571 int slot,
6572 struct ocfs2_cached_dealloc_ctxt *ctxt)
6573{
6574 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6575
6576 while (fl) {
6577 if (fl->f_inode_type == type && fl->f_slot == slot)
6578 return fl;
6579
6580 fl = fl->f_next_suballocator;
6581 }
6582
6583 fl = kmalloc(sizeof(*fl), GFP_NOFS);
6584 if (fl) {
6585 fl->f_inode_type = type;
6586 fl->f_slot = slot;
6587 fl->f_first = NULL;
6588 fl->f_next_suballocator = ctxt->c_first_suballocator;
6589
6590 ctxt->c_first_suballocator = fl;
6591 }
6592 return fl;
6593}
6594
6595int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6596 int type, int slot, u64 suballoc,
6597 u64 blkno, unsigned int bit)
6598{
6599 int ret;
6600 struct ocfs2_per_slot_free_list *fl;
6601 struct ocfs2_cached_block_free *item;
6602
6603 fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
6604 if (fl == NULL) {
6605 ret = -ENOMEM;
6606 mlog_errno(ret);
6607 goto out;
6608 }
6609
6610 item = kzalloc(sizeof(*item), GFP_NOFS);
6611 if (item == NULL) {
6612 ret = -ENOMEM;
6613 mlog_errno(ret);
6614 goto out;
6615 }
6616
6617 trace_ocfs2_cache_block_dealloc(type, slot,
6618 (unsigned long long)suballoc,
6619 (unsigned long long)blkno, bit);
6620
6621 item->free_bg = suballoc;
6622 item->free_blk = blkno;
6623 item->free_bit = bit;
6624 item->free_next = fl->f_first;
6625
6626 fl->f_first = item;
6627
6628 ret = 0;
6629out:
6630 return ret;
6631}
6632
6633static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
6634 struct ocfs2_extent_block *eb)
6635{
6636 return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
6637 le16_to_cpu(eb->h_suballoc_slot),
6638 le64_to_cpu(eb->h_suballoc_loc),
6639 le64_to_cpu(eb->h_blkno),
6640 le16_to_cpu(eb->h_suballoc_bit));
6641}
6642
6643static int ocfs2_zero_func(handle_t *handle, struct buffer_head *bh)
6644{
6645 set_buffer_uptodate(bh);
6646 mark_buffer_dirty(bh);
6647 return 0;
6648}
6649
6650void ocfs2_map_and_dirty_page(struct inode *inode, handle_t *handle,
6651 unsigned int from, unsigned int to,
6652 struct page *page, int zero, u64 *phys)
6653{
6654 int ret, partial = 0;
6655
6656 ret = ocfs2_map_page_blocks(page, phys, inode, from, to, 0);
6657 if (ret)
6658 mlog_errno(ret);
6659
6660 if (zero)
6661 zero_user_segment(page, from, to);
6662
6663 /*
6664 * Need to set the buffers we zero'd into uptodate
6665 * here if they aren't - ocfs2_map_page_blocks()
6666 * might've skipped some
6667 */
6668 ret = walk_page_buffers(handle, page_buffers(page),
6669 from, to, &partial,
6670 ocfs2_zero_func);
6671 if (ret < 0)
6672 mlog_errno(ret);
6673 else if (ocfs2_should_order_data(inode)) {
6674 ret = ocfs2_jbd2_file_inode(handle, inode);
6675 if (ret < 0)
6676 mlog_errno(ret);
6677 }
6678
6679 if (!partial)
6680 SetPageUptodate(page);
6681
6682 flush_dcache_page(page);
6683}
6684
6685static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start,
6686 loff_t end, struct page **pages,
6687 int numpages, u64 phys, handle_t *handle)
6688{
6689 int i;
6690 struct page *page;
6691 unsigned int from, to = PAGE_SIZE;
6692 struct super_block *sb = inode->i_sb;
6693
6694 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
6695
6696 if (numpages == 0)
6697 goto out;
6698
6699 to = PAGE_SIZE;
6700 for(i = 0; i < numpages; i++) {
6701 page = pages[i];
6702
6703 from = start & (PAGE_SIZE - 1);
6704 if ((end >> PAGE_SHIFT) == page->index)
6705 to = end & (PAGE_SIZE - 1);
6706
6707 BUG_ON(from > PAGE_SIZE);
6708 BUG_ON(to > PAGE_SIZE);
6709
6710 ocfs2_map_and_dirty_page(inode, handle, from, to, page, 1,
6711 &phys);
6712
6713 start = (page->index + 1) << PAGE_SHIFT;
6714 }
6715out:
6716 if (pages)
6717 ocfs2_unlock_and_free_pages(pages, numpages);
6718}
6719
6720int ocfs2_grab_pages(struct inode *inode, loff_t start, loff_t end,
6721 struct page **pages, int *num)
6722{
6723 int numpages, ret = 0;
6724 struct address_space *mapping = inode->i_mapping;
6725 unsigned long index;
6726 loff_t last_page_bytes;
6727
6728 BUG_ON(start > end);
6729
6730 numpages = 0;
6731 last_page_bytes = PAGE_ALIGN(end);
6732 index = start >> PAGE_SHIFT;
6733 do {
6734 pages[numpages] = find_or_create_page(mapping, index, GFP_NOFS);
6735 if (!pages[numpages]) {
6736 ret = -ENOMEM;
6737 mlog_errno(ret);
6738 goto out;
6739 }
6740
6741 numpages++;
6742 index++;
6743 } while (index < (last_page_bytes >> PAGE_SHIFT));
6744
6745out:
6746 if (ret != 0) {
6747 if (pages)
6748 ocfs2_unlock_and_free_pages(pages, numpages);
6749 numpages = 0;
6750 }
6751
6752 *num = numpages;
6753
6754 return ret;
6755}
6756
6757static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end,
6758 struct page **pages, int *num)
6759{
6760 struct super_block *sb = inode->i_sb;
6761
6762 BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
6763 (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
6764
6765 return ocfs2_grab_pages(inode, start, end, pages, num);
6766}
6767
6768/*
6769 * Zero the area past i_size but still within an allocated
6770 * cluster. This avoids exposing nonzero data on subsequent file
6771 * extends.
6772 *
6773 * We need to call this before i_size is updated on the inode because
6774 * otherwise block_write_full_page() will skip writeout of pages past
6775 * i_size. The new_i_size parameter is passed for this reason.
6776 */
6777int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
6778 u64 range_start, u64 range_end)
6779{
6780 int ret = 0, numpages;
6781 struct page **pages = NULL;
6782 u64 phys;
6783 unsigned int ext_flags;
6784 struct super_block *sb = inode->i_sb;
6785
6786 /*
6787 * File systems which don't support sparse files zero on every
6788 * extend.
6789 */
6790 if (!ocfs2_sparse_alloc(OCFS2_SB(sb)))
6791 return 0;
6792
6793 pages = kcalloc(ocfs2_pages_per_cluster(sb),
6794 sizeof(struct page *), GFP_NOFS);
6795 if (pages == NULL) {
6796 ret = -ENOMEM;
6797 mlog_errno(ret);
6798 goto out;
6799 }
6800
6801 if (range_start == range_end)
6802 goto out;
6803
6804 ret = ocfs2_extent_map_get_blocks(inode,
6805 range_start >> sb->s_blocksize_bits,
6806 &phys, NULL, &ext_flags);
6807 if (ret) {
6808 mlog_errno(ret);
6809 goto out;
6810 }
6811
6812 /*
6813 * Tail is a hole, or is marked unwritten. In either case, we
6814 * can count on read and write to return/push zero's.
6815 */
6816 if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN)
6817 goto out;
6818
6819 ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages,
6820 &numpages);
6821 if (ret) {
6822 mlog_errno(ret);
6823 goto out;
6824 }
6825
6826 ocfs2_zero_cluster_pages(inode, range_start, range_end, pages,
6827 numpages, phys, handle);
6828
6829 /*
6830 * Initiate writeout of the pages we zero'd here. We don't
6831 * wait on them - the truncate_inode_pages() call later will
6832 * do that for us.
6833 */
6834 ret = filemap_fdatawrite_range(inode->i_mapping, range_start,
6835 range_end - 1);
6836 if (ret)
6837 mlog_errno(ret);
6838
6839out:
6840 kfree(pages);
6841
6842 return ret;
6843}
6844
6845static void ocfs2_zero_dinode_id2_with_xattr(struct inode *inode,
6846 struct ocfs2_dinode *di)
6847{
6848 unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits;
6849 unsigned int xattrsize = le16_to_cpu(di->i_xattr_inline_size);
6850
6851 if (le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_XATTR_FL)
6852 memset(&di->id2, 0, blocksize -
6853 offsetof(struct ocfs2_dinode, id2) -
6854 xattrsize);
6855 else
6856 memset(&di->id2, 0, blocksize -
6857 offsetof(struct ocfs2_dinode, id2));
6858}
6859
6860void ocfs2_dinode_new_extent_list(struct inode *inode,
6861 struct ocfs2_dinode *di)
6862{
6863 ocfs2_zero_dinode_id2_with_xattr(inode, di);
6864 di->id2.i_list.l_tree_depth = 0;
6865 di->id2.i_list.l_next_free_rec = 0;
6866 di->id2.i_list.l_count = cpu_to_le16(
6867 ocfs2_extent_recs_per_inode_with_xattr(inode->i_sb, di));
6868}
6869
6870void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di)
6871{
6872 struct ocfs2_inode_info *oi = OCFS2_I(inode);
6873 struct ocfs2_inline_data *idata = &di->id2.i_data;
6874
6875 spin_lock(&oi->ip_lock);
6876 oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
6877 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
6878 spin_unlock(&oi->ip_lock);
6879
6880 /*
6881 * We clear the entire i_data structure here so that all
6882 * fields can be properly initialized.
6883 */
6884 ocfs2_zero_dinode_id2_with_xattr(inode, di);
6885
6886 idata->id_count = cpu_to_le16(
6887 ocfs2_max_inline_data_with_xattr(inode->i_sb, di));
6888}
6889
6890int ocfs2_convert_inline_data_to_extents(struct inode *inode,
6891 struct buffer_head *di_bh)
6892{
6893 int ret, i, has_data, num_pages = 0;
6894 int need_free = 0;
6895 u32 bit_off, num;
6896 handle_t *handle;
6897 u64 uninitialized_var(block);
6898 struct ocfs2_inode_info *oi = OCFS2_I(inode);
6899 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
6900 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
6901 struct ocfs2_alloc_context *data_ac = NULL;
6902 struct page **pages = NULL;
6903 loff_t end = osb->s_clustersize;
6904 struct ocfs2_extent_tree et;
6905 int did_quota = 0;
6906
6907 has_data = i_size_read(inode) ? 1 : 0;
6908
6909 if (has_data) {
6910 pages = kcalloc(ocfs2_pages_per_cluster(osb->sb),
6911 sizeof(struct page *), GFP_NOFS);
6912 if (pages == NULL) {
6913 ret = -ENOMEM;
6914 mlog_errno(ret);
6915 return ret;
6916 }
6917
6918 ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
6919 if (ret) {
6920 mlog_errno(ret);
6921 goto free_pages;
6922 }
6923 }
6924
6925 handle = ocfs2_start_trans(osb,
6926 ocfs2_inline_to_extents_credits(osb->sb));
6927 if (IS_ERR(handle)) {
6928 ret = PTR_ERR(handle);
6929 mlog_errno(ret);
6930 goto out;
6931 }
6932
6933 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
6934 OCFS2_JOURNAL_ACCESS_WRITE);
6935 if (ret) {
6936 mlog_errno(ret);
6937 goto out_commit;
6938 }
6939
6940 if (has_data) {
6941 unsigned int page_end;
6942 u64 phys;
6943
6944 ret = dquot_alloc_space_nodirty(inode,
6945 ocfs2_clusters_to_bytes(osb->sb, 1));
6946 if (ret)
6947 goto out_commit;
6948 did_quota = 1;
6949
6950 data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
6951
6952 ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off,
6953 &num);
6954 if (ret) {
6955 mlog_errno(ret);
6956 goto out_commit;
6957 }
6958
6959 /*
6960 * Save two copies, one for insert, and one that can
6961 * be changed by ocfs2_map_and_dirty_page() below.
6962 */
6963 block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off);
6964
6965 /*
6966 * Non sparse file systems zero on extend, so no need
6967 * to do that now.
6968 */
6969 if (!ocfs2_sparse_alloc(osb) &&
6970 PAGE_SIZE < osb->s_clustersize)
6971 end = PAGE_SIZE;
6972
6973 ret = ocfs2_grab_eof_pages(inode, 0, end, pages, &num_pages);
6974 if (ret) {
6975 mlog_errno(ret);
6976 need_free = 1;
6977 goto out_commit;
6978 }
6979
6980 /*
6981 * This should populate the 1st page for us and mark
6982 * it up to date.
6983 */
6984 ret = ocfs2_read_inline_data(inode, pages[0], di_bh);
6985 if (ret) {
6986 mlog_errno(ret);
6987 need_free = 1;
6988 goto out_unlock;
6989 }
6990
6991 page_end = PAGE_SIZE;
6992 if (PAGE_SIZE > osb->s_clustersize)
6993 page_end = osb->s_clustersize;
6994
6995 for (i = 0; i < num_pages; i++)
6996 ocfs2_map_and_dirty_page(inode, handle, 0, page_end,
6997 pages[i], i > 0, &phys);
6998 }
6999
7000 spin_lock(&oi->ip_lock);
7001 oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
7002 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7003 spin_unlock(&oi->ip_lock);
7004
7005 ocfs2_update_inode_fsync_trans(handle, inode, 1);
7006 ocfs2_dinode_new_extent_list(inode, di);
7007
7008 ocfs2_journal_dirty(handle, di_bh);
7009
7010 if (has_data) {
7011 /*
7012 * An error at this point should be extremely rare. If
7013 * this proves to be false, we could always re-build
7014 * the in-inode data from our pages.
7015 */
7016 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7017 ret = ocfs2_insert_extent(handle, &et, 0, block, 1, 0, NULL);
7018 if (ret) {
7019 mlog_errno(ret);
7020 need_free = 1;
7021 goto out_unlock;
7022 }
7023
7024 inode->i_blocks = ocfs2_inode_sector_count(inode);
7025 }
7026
7027out_unlock:
7028 if (pages)
7029 ocfs2_unlock_and_free_pages(pages, num_pages);
7030
7031out_commit:
7032 if (ret < 0 && did_quota)
7033 dquot_free_space_nodirty(inode,
7034 ocfs2_clusters_to_bytes(osb->sb, 1));
7035
7036 if (need_free) {
7037 if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
7038 ocfs2_free_local_alloc_bits(osb, handle, data_ac,
7039 bit_off, num);
7040 else
7041 ocfs2_free_clusters(handle,
7042 data_ac->ac_inode,
7043 data_ac->ac_bh,
7044 ocfs2_clusters_to_blocks(osb->sb, bit_off),
7045 num);
7046 }
7047
7048 ocfs2_commit_trans(osb, handle);
7049
7050out:
7051 if (data_ac)
7052 ocfs2_free_alloc_context(data_ac);
7053free_pages:
7054 kfree(pages);
7055 return ret;
7056}
7057
7058/*
7059 * It is expected, that by the time you call this function,
7060 * inode->i_size and fe->i_size have been adjusted.
7061 *
7062 * WARNING: This will kfree the truncate context
7063 */
7064int ocfs2_commit_truncate(struct ocfs2_super *osb,
7065 struct inode *inode,
7066 struct buffer_head *di_bh)
7067{
7068 int status = 0, i, flags = 0;
7069 u32 new_highest_cpos, range, trunc_cpos, trunc_len, phys_cpos, coff;
7070 u64 blkno = 0;
7071 struct ocfs2_extent_list *el;
7072 struct ocfs2_extent_rec *rec;
7073 struct ocfs2_path *path = NULL;
7074 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7075 struct ocfs2_extent_list *root_el = &(di->id2.i_list);
7076 u64 refcount_loc = le64_to_cpu(di->i_refcount_loc);
7077 struct ocfs2_extent_tree et;
7078 struct ocfs2_cached_dealloc_ctxt dealloc;
7079 struct ocfs2_refcount_tree *ref_tree = NULL;
7080
7081 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7082 ocfs2_init_dealloc_ctxt(&dealloc);
7083
7084 new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
7085 i_size_read(inode));
7086
7087 path = ocfs2_new_path(di_bh, &di->id2.i_list,
7088 ocfs2_journal_access_di);
7089 if (!path) {
7090 status = -ENOMEM;
7091 mlog_errno(status);
7092 goto bail;
7093 }
7094
7095 ocfs2_extent_map_trunc(inode, new_highest_cpos);
7096
7097start:
7098 /*
7099 * Check that we still have allocation to delete.
7100 */
7101 if (OCFS2_I(inode)->ip_clusters == 0) {
7102 status = 0;
7103 goto bail;
7104 }
7105
7106 /*
7107 * Truncate always works against the rightmost tree branch.
7108 */
7109 status = ocfs2_find_path(INODE_CACHE(inode), path, UINT_MAX);
7110 if (status) {
7111 mlog_errno(status);
7112 goto bail;
7113 }
7114
7115 trace_ocfs2_commit_truncate(
7116 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7117 new_highest_cpos,
7118 OCFS2_I(inode)->ip_clusters,
7119 path->p_tree_depth);
7120
7121 /*
7122 * By now, el will point to the extent list on the bottom most
7123 * portion of this tree. Only the tail record is considered in
7124 * each pass.
7125 *
7126 * We handle the following cases, in order:
7127 * - empty extent: delete the remaining branch
7128 * - remove the entire record
7129 * - remove a partial record
7130 * - no record needs to be removed (truncate has completed)
7131 */
7132 el = path_leaf_el(path);
7133 if (le16_to_cpu(el->l_next_free_rec) == 0) {
7134 ocfs2_error(inode->i_sb,
7135 "Inode %llu has empty extent block at %llu\n",
7136 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7137 (unsigned long long)path_leaf_bh(path)->b_blocknr);
7138 status = -EROFS;
7139 goto bail;
7140 }
7141
7142 i = le16_to_cpu(el->l_next_free_rec) - 1;
7143 rec = &el->l_recs[i];
7144 flags = rec->e_flags;
7145 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
7146
7147 if (i == 0 && ocfs2_is_empty_extent(rec)) {
7148 /*
7149 * Lower levels depend on this never happening, but it's best
7150 * to check it up here before changing the tree.
7151 */
7152 if (root_el->l_tree_depth && rec->e_int_clusters == 0) {
7153 mlog(ML_ERROR, "Inode %lu has an empty "
7154 "extent record, depth %u\n", inode->i_ino,
7155 le16_to_cpu(root_el->l_tree_depth));
7156 status = ocfs2_remove_rightmost_empty_extent(osb,
7157 &et, path, &dealloc);
7158 if (status) {
7159 mlog_errno(status);
7160 goto bail;
7161 }
7162
7163 ocfs2_reinit_path(path, 1);
7164 goto start;
7165 } else {
7166 trunc_cpos = le32_to_cpu(rec->e_cpos);
7167 trunc_len = 0;
7168 blkno = 0;
7169 }
7170 } else if (le32_to_cpu(rec->e_cpos) >= new_highest_cpos) {
7171 /*
7172 * Truncate entire record.
7173 */
7174 trunc_cpos = le32_to_cpu(rec->e_cpos);
7175 trunc_len = ocfs2_rec_clusters(el, rec);
7176 blkno = le64_to_cpu(rec->e_blkno);
7177 } else if (range > new_highest_cpos) {
7178 /*
7179 * Partial truncate. it also should be
7180 * the last truncate we're doing.
7181 */
7182 trunc_cpos = new_highest_cpos;
7183 trunc_len = range - new_highest_cpos;
7184 coff = new_highest_cpos - le32_to_cpu(rec->e_cpos);
7185 blkno = le64_to_cpu(rec->e_blkno) +
7186 ocfs2_clusters_to_blocks(inode->i_sb, coff);
7187 } else {
7188 /*
7189 * Truncate completed, leave happily.
7190 */
7191 status = 0;
7192 goto bail;
7193 }
7194
7195 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
7196
7197 if ((flags & OCFS2_EXT_REFCOUNTED) && trunc_len && !ref_tree) {
7198 status = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
7199 &ref_tree, NULL);
7200 if (status) {
7201 mlog_errno(status);
7202 goto bail;
7203 }
7204 }
7205
7206 status = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
7207 phys_cpos, trunc_len, flags, &dealloc,
7208 refcount_loc, true);
7209 if (status < 0) {
7210 mlog_errno(status);
7211 goto bail;
7212 }
7213
7214 ocfs2_reinit_path(path, 1);
7215
7216 /*
7217 * The check above will catch the case where we've truncated
7218 * away all allocation.
7219 */
7220 goto start;
7221
7222bail:
7223 if (ref_tree)
7224 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
7225
7226 ocfs2_schedule_truncate_log_flush(osb, 1);
7227
7228 ocfs2_run_deallocs(osb, &dealloc);
7229
7230 ocfs2_free_path(path);
7231
7232 return status;
7233}
7234
7235/*
7236 * 'start' is inclusive, 'end' is not.
7237 */
7238int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh,
7239 unsigned int start, unsigned int end, int trunc)
7240{
7241 int ret;
7242 unsigned int numbytes;
7243 handle_t *handle;
7244 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7245 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7246 struct ocfs2_inline_data *idata = &di->id2.i_data;
7247
7248 if (end > i_size_read(inode))
7249 end = i_size_read(inode);
7250
7251 BUG_ON(start > end);
7252
7253 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
7254 !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) ||
7255 !ocfs2_supports_inline_data(osb)) {
7256 ocfs2_error(inode->i_sb,
7257 "Inline data flags for inode %llu don't agree! Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7258 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7259 le16_to_cpu(di->i_dyn_features),
7260 OCFS2_I(inode)->ip_dyn_features,
7261 osb->s_feature_incompat);
7262 ret = -EROFS;
7263 goto out;
7264 }
7265
7266 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
7267 if (IS_ERR(handle)) {
7268 ret = PTR_ERR(handle);
7269 mlog_errno(ret);
7270 goto out;
7271 }
7272
7273 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7274 OCFS2_JOURNAL_ACCESS_WRITE);
7275 if (ret) {
7276 mlog_errno(ret);
7277 goto out_commit;
7278 }
7279
7280 numbytes = end - start;
7281 memset(idata->id_data + start, 0, numbytes);
7282
7283 /*
7284 * No need to worry about the data page here - it's been
7285 * truncated already and inline data doesn't need it for
7286 * pushing zero's to disk, so we'll let readpage pick it up
7287 * later.
7288 */
7289 if (trunc) {
7290 i_size_write(inode, start);
7291 di->i_size = cpu_to_le64(start);
7292 }
7293
7294 inode->i_blocks = ocfs2_inode_sector_count(inode);
7295 inode->i_ctime = inode->i_mtime = current_time(inode);
7296
7297 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
7298 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
7299
7300 ocfs2_update_inode_fsync_trans(handle, inode, 1);
7301 ocfs2_journal_dirty(handle, di_bh);
7302
7303out_commit:
7304 ocfs2_commit_trans(osb, handle);
7305
7306out:
7307 return ret;
7308}
7309
7310static int ocfs2_trim_extent(struct super_block *sb,
7311 struct ocfs2_group_desc *gd,
7312 u32 start, u32 count)
7313{
7314 u64 discard, bcount;
7315
7316 bcount = ocfs2_clusters_to_blocks(sb, count);
7317 discard = le64_to_cpu(gd->bg_blkno) +
7318 ocfs2_clusters_to_blocks(sb, start);
7319
7320 trace_ocfs2_trim_extent(sb, (unsigned long long)discard, bcount);
7321
7322 return sb_issue_discard(sb, discard, bcount, GFP_NOFS, 0);
7323}
7324
7325static int ocfs2_trim_group(struct super_block *sb,
7326 struct ocfs2_group_desc *gd,
7327 u32 start, u32 max, u32 minbits)
7328{
7329 int ret = 0, count = 0, next;
7330 void *bitmap = gd->bg_bitmap;
7331
7332 if (le16_to_cpu(gd->bg_free_bits_count) < minbits)
7333 return 0;
7334
7335 trace_ocfs2_trim_group((unsigned long long)le64_to_cpu(gd->bg_blkno),
7336 start, max, minbits);
7337
7338 while (start < max) {
7339 start = ocfs2_find_next_zero_bit(bitmap, max, start);
7340 if (start >= max)
7341 break;
7342 next = ocfs2_find_next_bit(bitmap, max, start);
7343
7344 if ((next - start) >= minbits) {
7345 ret = ocfs2_trim_extent(sb, gd,
7346 start, next - start);
7347 if (ret < 0) {
7348 mlog_errno(ret);
7349 break;
7350 }
7351 count += next - start;
7352 }
7353 start = next + 1;
7354
7355 if (fatal_signal_pending(current)) {
7356 count = -ERESTARTSYS;
7357 break;
7358 }
7359
7360 if ((le16_to_cpu(gd->bg_free_bits_count) - count) < minbits)
7361 break;
7362 }
7363
7364 if (ret < 0)
7365 count = ret;
7366
7367 return count;
7368}
7369
7370int ocfs2_trim_fs(struct super_block *sb, struct fstrim_range *range)
7371{
7372 struct ocfs2_super *osb = OCFS2_SB(sb);
7373 u64 start, len, trimmed, first_group, last_group, group;
7374 int ret, cnt;
7375 u32 first_bit, last_bit, minlen;
7376 struct buffer_head *main_bm_bh = NULL;
7377 struct inode *main_bm_inode = NULL;
7378 struct buffer_head *gd_bh = NULL;
7379 struct ocfs2_dinode *main_bm;
7380 struct ocfs2_group_desc *gd = NULL;
7381
7382 start = range->start >> osb->s_clustersize_bits;
7383 len = range->len >> osb->s_clustersize_bits;
7384 minlen = range->minlen >> osb->s_clustersize_bits;
7385
7386 if (minlen >= osb->bitmap_cpg || range->len < sb->s_blocksize)
7387 return -EINVAL;
7388
7389 main_bm_inode = ocfs2_get_system_file_inode(osb,
7390 GLOBAL_BITMAP_SYSTEM_INODE,
7391 OCFS2_INVALID_SLOT);
7392 if (!main_bm_inode) {
7393 ret = -EIO;
7394 mlog_errno(ret);
7395 goto out;
7396 }
7397
7398 inode_lock(main_bm_inode);
7399
7400 ret = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 0);
7401 if (ret < 0) {
7402 mlog_errno(ret);
7403 goto out_mutex;
7404 }
7405 main_bm = (struct ocfs2_dinode *)main_bm_bh->b_data;
7406
7407 if (start >= le32_to_cpu(main_bm->i_clusters)) {
7408 ret = -EINVAL;
7409 goto out_unlock;
7410 }
7411
7412 len = range->len >> osb->s_clustersize_bits;
7413 if (start + len > le32_to_cpu(main_bm->i_clusters))
7414 len = le32_to_cpu(main_bm->i_clusters) - start;
7415
7416 trace_ocfs2_trim_fs(start, len, minlen);
7417
7418 /* Determine first and last group to examine based on start and len */
7419 first_group = ocfs2_which_cluster_group(main_bm_inode, start);
7420 if (first_group == osb->first_cluster_group_blkno)
7421 first_bit = start;
7422 else
7423 first_bit = start - ocfs2_blocks_to_clusters(sb, first_group);
7424 last_group = ocfs2_which_cluster_group(main_bm_inode, start + len - 1);
7425 last_bit = osb->bitmap_cpg;
7426
7427 trimmed = 0;
7428 for (group = first_group; group <= last_group;) {
7429 if (first_bit + len >= osb->bitmap_cpg)
7430 last_bit = osb->bitmap_cpg;
7431 else
7432 last_bit = first_bit + len;
7433
7434 ret = ocfs2_read_group_descriptor(main_bm_inode,
7435 main_bm, group,
7436 &gd_bh);
7437 if (ret < 0) {
7438 mlog_errno(ret);
7439 break;
7440 }
7441
7442 gd = (struct ocfs2_group_desc *)gd_bh->b_data;
7443 cnt = ocfs2_trim_group(sb, gd, first_bit, last_bit, minlen);
7444 brelse(gd_bh);
7445 gd_bh = NULL;
7446 if (cnt < 0) {
7447 ret = cnt;
7448 mlog_errno(ret);
7449 break;
7450 }
7451
7452 trimmed += cnt;
7453 len -= osb->bitmap_cpg - first_bit;
7454 first_bit = 0;
7455 if (group == osb->first_cluster_group_blkno)
7456 group = ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7457 else
7458 group += ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7459 }
7460 range->len = trimmed * sb->s_blocksize;
7461out_unlock:
7462 ocfs2_inode_unlock(main_bm_inode, 0);
7463 brelse(main_bm_bh);
7464out_mutex:
7465 inode_unlock(main_bm_inode);
7466 iput(main_bm_inode);
7467out:
7468 return ret;
7469}
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * alloc.c
5 *
6 * Extent allocs and frees
7 *
8 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 */
25
26#include <linux/fs.h>
27#include <linux/types.h>
28#include <linux/slab.h>
29#include <linux/highmem.h>
30#include <linux/swap.h>
31#include <linux/quotaops.h>
32#include <linux/blkdev.h>
33
34#include <cluster/masklog.h>
35
36#include "ocfs2.h"
37
38#include "alloc.h"
39#include "aops.h"
40#include "blockcheck.h"
41#include "dlmglue.h"
42#include "extent_map.h"
43#include "inode.h"
44#include "journal.h"
45#include "localalloc.h"
46#include "suballoc.h"
47#include "sysfile.h"
48#include "file.h"
49#include "super.h"
50#include "uptodate.h"
51#include "xattr.h"
52#include "refcounttree.h"
53#include "ocfs2_trace.h"
54
55#include "buffer_head_io.h"
56
57enum ocfs2_contig_type {
58 CONTIG_NONE = 0,
59 CONTIG_LEFT,
60 CONTIG_RIGHT,
61 CONTIG_LEFTRIGHT,
62};
63
64static enum ocfs2_contig_type
65 ocfs2_extent_rec_contig(struct super_block *sb,
66 struct ocfs2_extent_rec *ext,
67 struct ocfs2_extent_rec *insert_rec);
68/*
69 * Operations for a specific extent tree type.
70 *
71 * To implement an on-disk btree (extent tree) type in ocfs2, add
72 * an ocfs2_extent_tree_operations structure and the matching
73 * ocfs2_init_<thingy>_extent_tree() function. That's pretty much it
74 * for the allocation portion of the extent tree.
75 */
76struct ocfs2_extent_tree_operations {
77 /*
78 * last_eb_blk is the block number of the right most leaf extent
79 * block. Most on-disk structures containing an extent tree store
80 * this value for fast access. The ->eo_set_last_eb_blk() and
81 * ->eo_get_last_eb_blk() operations access this value. They are
82 * both required.
83 */
84 void (*eo_set_last_eb_blk)(struct ocfs2_extent_tree *et,
85 u64 blkno);
86 u64 (*eo_get_last_eb_blk)(struct ocfs2_extent_tree *et);
87
88 /*
89 * The on-disk structure usually keeps track of how many total
90 * clusters are stored in this extent tree. This function updates
91 * that value. new_clusters is the delta, and must be
92 * added to the total. Required.
93 */
94 void (*eo_update_clusters)(struct ocfs2_extent_tree *et,
95 u32 new_clusters);
96
97 /*
98 * If this extent tree is supported by an extent map, insert
99 * a record into the map.
100 */
101 void (*eo_extent_map_insert)(struct ocfs2_extent_tree *et,
102 struct ocfs2_extent_rec *rec);
103
104 /*
105 * If this extent tree is supported by an extent map, truncate the
106 * map to clusters,
107 */
108 void (*eo_extent_map_truncate)(struct ocfs2_extent_tree *et,
109 u32 clusters);
110
111 /*
112 * If ->eo_insert_check() exists, it is called before rec is
113 * inserted into the extent tree. It is optional.
114 */
115 int (*eo_insert_check)(struct ocfs2_extent_tree *et,
116 struct ocfs2_extent_rec *rec);
117 int (*eo_sanity_check)(struct ocfs2_extent_tree *et);
118
119 /*
120 * --------------------------------------------------------------
121 * The remaining are internal to ocfs2_extent_tree and don't have
122 * accessor functions
123 */
124
125 /*
126 * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el.
127 * It is required.
128 */
129 void (*eo_fill_root_el)(struct ocfs2_extent_tree *et);
130
131 /*
132 * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if
133 * it exists. If it does not, et->et_max_leaf_clusters is set
134 * to 0 (unlimited). Optional.
135 */
136 void (*eo_fill_max_leaf_clusters)(struct ocfs2_extent_tree *et);
137
138 /*
139 * ->eo_extent_contig test whether the 2 ocfs2_extent_rec
140 * are contiguous or not. Optional. Don't need to set it if use
141 * ocfs2_extent_rec as the tree leaf.
142 */
143 enum ocfs2_contig_type
144 (*eo_extent_contig)(struct ocfs2_extent_tree *et,
145 struct ocfs2_extent_rec *ext,
146 struct ocfs2_extent_rec *insert_rec);
147};
148
149
150/*
151 * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check
152 * in the methods.
153 */
154static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et);
155static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
156 u64 blkno);
157static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
158 u32 clusters);
159static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
160 struct ocfs2_extent_rec *rec);
161static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
162 u32 clusters);
163static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
164 struct ocfs2_extent_rec *rec);
165static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et);
166static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et);
167static struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops = {
168 .eo_set_last_eb_blk = ocfs2_dinode_set_last_eb_blk,
169 .eo_get_last_eb_blk = ocfs2_dinode_get_last_eb_blk,
170 .eo_update_clusters = ocfs2_dinode_update_clusters,
171 .eo_extent_map_insert = ocfs2_dinode_extent_map_insert,
172 .eo_extent_map_truncate = ocfs2_dinode_extent_map_truncate,
173 .eo_insert_check = ocfs2_dinode_insert_check,
174 .eo_sanity_check = ocfs2_dinode_sanity_check,
175 .eo_fill_root_el = ocfs2_dinode_fill_root_el,
176};
177
178static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
179 u64 blkno)
180{
181 struct ocfs2_dinode *di = et->et_object;
182
183 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
184 di->i_last_eb_blk = cpu_to_le64(blkno);
185}
186
187static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et)
188{
189 struct ocfs2_dinode *di = et->et_object;
190
191 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
192 return le64_to_cpu(di->i_last_eb_blk);
193}
194
195static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
196 u32 clusters)
197{
198 struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
199 struct ocfs2_dinode *di = et->et_object;
200
201 le32_add_cpu(&di->i_clusters, clusters);
202 spin_lock(&oi->ip_lock);
203 oi->ip_clusters = le32_to_cpu(di->i_clusters);
204 spin_unlock(&oi->ip_lock);
205}
206
207static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
208 struct ocfs2_extent_rec *rec)
209{
210 struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
211
212 ocfs2_extent_map_insert_rec(inode, rec);
213}
214
215static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
216 u32 clusters)
217{
218 struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
219
220 ocfs2_extent_map_trunc(inode, clusters);
221}
222
223static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
224 struct ocfs2_extent_rec *rec)
225{
226 struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
227 struct ocfs2_super *osb = OCFS2_SB(oi->vfs_inode.i_sb);
228
229 BUG_ON(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL);
230 mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) &&
231 (oi->ip_clusters != le32_to_cpu(rec->e_cpos)),
232 "Device %s, asking for sparse allocation: inode %llu, "
233 "cpos %u, clusters %u\n",
234 osb->dev_str,
235 (unsigned long long)oi->ip_blkno,
236 rec->e_cpos, oi->ip_clusters);
237
238 return 0;
239}
240
241static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et)
242{
243 struct ocfs2_dinode *di = et->et_object;
244
245 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
246 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
247
248 return 0;
249}
250
251static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et)
252{
253 struct ocfs2_dinode *di = et->et_object;
254
255 et->et_root_el = &di->id2.i_list;
256}
257
258
259static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree *et)
260{
261 struct ocfs2_xattr_value_buf *vb = et->et_object;
262
263 et->et_root_el = &vb->vb_xv->xr_list;
264}
265
266static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree *et,
267 u64 blkno)
268{
269 struct ocfs2_xattr_value_buf *vb = et->et_object;
270
271 vb->vb_xv->xr_last_eb_blk = cpu_to_le64(blkno);
272}
273
274static u64 ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree *et)
275{
276 struct ocfs2_xattr_value_buf *vb = et->et_object;
277
278 return le64_to_cpu(vb->vb_xv->xr_last_eb_blk);
279}
280
281static void ocfs2_xattr_value_update_clusters(struct ocfs2_extent_tree *et,
282 u32 clusters)
283{
284 struct ocfs2_xattr_value_buf *vb = et->et_object;
285
286 le32_add_cpu(&vb->vb_xv->xr_clusters, clusters);
287}
288
289static struct ocfs2_extent_tree_operations ocfs2_xattr_value_et_ops = {
290 .eo_set_last_eb_blk = ocfs2_xattr_value_set_last_eb_blk,
291 .eo_get_last_eb_blk = ocfs2_xattr_value_get_last_eb_blk,
292 .eo_update_clusters = ocfs2_xattr_value_update_clusters,
293 .eo_fill_root_el = ocfs2_xattr_value_fill_root_el,
294};
295
296static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree *et)
297{
298 struct ocfs2_xattr_block *xb = et->et_object;
299
300 et->et_root_el = &xb->xb_attrs.xb_root.xt_list;
301}
302
303static void ocfs2_xattr_tree_fill_max_leaf_clusters(struct ocfs2_extent_tree *et)
304{
305 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
306 et->et_max_leaf_clusters =
307 ocfs2_clusters_for_bytes(sb, OCFS2_MAX_XATTR_TREE_LEAF_SIZE);
308}
309
310static void ocfs2_xattr_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
311 u64 blkno)
312{
313 struct ocfs2_xattr_block *xb = et->et_object;
314 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
315
316 xt->xt_last_eb_blk = cpu_to_le64(blkno);
317}
318
319static u64 ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
320{
321 struct ocfs2_xattr_block *xb = et->et_object;
322 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
323
324 return le64_to_cpu(xt->xt_last_eb_blk);
325}
326
327static void ocfs2_xattr_tree_update_clusters(struct ocfs2_extent_tree *et,
328 u32 clusters)
329{
330 struct ocfs2_xattr_block *xb = et->et_object;
331
332 le32_add_cpu(&xb->xb_attrs.xb_root.xt_clusters, clusters);
333}
334
335static struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops = {
336 .eo_set_last_eb_blk = ocfs2_xattr_tree_set_last_eb_blk,
337 .eo_get_last_eb_blk = ocfs2_xattr_tree_get_last_eb_blk,
338 .eo_update_clusters = ocfs2_xattr_tree_update_clusters,
339 .eo_fill_root_el = ocfs2_xattr_tree_fill_root_el,
340 .eo_fill_max_leaf_clusters = ocfs2_xattr_tree_fill_max_leaf_clusters,
341};
342
343static void ocfs2_dx_root_set_last_eb_blk(struct ocfs2_extent_tree *et,
344 u64 blkno)
345{
346 struct ocfs2_dx_root_block *dx_root = et->et_object;
347
348 dx_root->dr_last_eb_blk = cpu_to_le64(blkno);
349}
350
351static u64 ocfs2_dx_root_get_last_eb_blk(struct ocfs2_extent_tree *et)
352{
353 struct ocfs2_dx_root_block *dx_root = et->et_object;
354
355 return le64_to_cpu(dx_root->dr_last_eb_blk);
356}
357
358static void ocfs2_dx_root_update_clusters(struct ocfs2_extent_tree *et,
359 u32 clusters)
360{
361 struct ocfs2_dx_root_block *dx_root = et->et_object;
362
363 le32_add_cpu(&dx_root->dr_clusters, clusters);
364}
365
366static int ocfs2_dx_root_sanity_check(struct ocfs2_extent_tree *et)
367{
368 struct ocfs2_dx_root_block *dx_root = et->et_object;
369
370 BUG_ON(!OCFS2_IS_VALID_DX_ROOT(dx_root));
371
372 return 0;
373}
374
375static void ocfs2_dx_root_fill_root_el(struct ocfs2_extent_tree *et)
376{
377 struct ocfs2_dx_root_block *dx_root = et->et_object;
378
379 et->et_root_el = &dx_root->dr_list;
380}
381
382static struct ocfs2_extent_tree_operations ocfs2_dx_root_et_ops = {
383 .eo_set_last_eb_blk = ocfs2_dx_root_set_last_eb_blk,
384 .eo_get_last_eb_blk = ocfs2_dx_root_get_last_eb_blk,
385 .eo_update_clusters = ocfs2_dx_root_update_clusters,
386 .eo_sanity_check = ocfs2_dx_root_sanity_check,
387 .eo_fill_root_el = ocfs2_dx_root_fill_root_el,
388};
389
390static void ocfs2_refcount_tree_fill_root_el(struct ocfs2_extent_tree *et)
391{
392 struct ocfs2_refcount_block *rb = et->et_object;
393
394 et->et_root_el = &rb->rf_list;
395}
396
397static void ocfs2_refcount_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
398 u64 blkno)
399{
400 struct ocfs2_refcount_block *rb = et->et_object;
401
402 rb->rf_last_eb_blk = cpu_to_le64(blkno);
403}
404
405static u64 ocfs2_refcount_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
406{
407 struct ocfs2_refcount_block *rb = et->et_object;
408
409 return le64_to_cpu(rb->rf_last_eb_blk);
410}
411
412static void ocfs2_refcount_tree_update_clusters(struct ocfs2_extent_tree *et,
413 u32 clusters)
414{
415 struct ocfs2_refcount_block *rb = et->et_object;
416
417 le32_add_cpu(&rb->rf_clusters, clusters);
418}
419
420static enum ocfs2_contig_type
421ocfs2_refcount_tree_extent_contig(struct ocfs2_extent_tree *et,
422 struct ocfs2_extent_rec *ext,
423 struct ocfs2_extent_rec *insert_rec)
424{
425 return CONTIG_NONE;
426}
427
428static struct ocfs2_extent_tree_operations ocfs2_refcount_tree_et_ops = {
429 .eo_set_last_eb_blk = ocfs2_refcount_tree_set_last_eb_blk,
430 .eo_get_last_eb_blk = ocfs2_refcount_tree_get_last_eb_blk,
431 .eo_update_clusters = ocfs2_refcount_tree_update_clusters,
432 .eo_fill_root_el = ocfs2_refcount_tree_fill_root_el,
433 .eo_extent_contig = ocfs2_refcount_tree_extent_contig,
434};
435
436static void __ocfs2_init_extent_tree(struct ocfs2_extent_tree *et,
437 struct ocfs2_caching_info *ci,
438 struct buffer_head *bh,
439 ocfs2_journal_access_func access,
440 void *obj,
441 struct ocfs2_extent_tree_operations *ops)
442{
443 et->et_ops = ops;
444 et->et_root_bh = bh;
445 et->et_ci = ci;
446 et->et_root_journal_access = access;
447 if (!obj)
448 obj = (void *)bh->b_data;
449 et->et_object = obj;
450
451 et->et_ops->eo_fill_root_el(et);
452 if (!et->et_ops->eo_fill_max_leaf_clusters)
453 et->et_max_leaf_clusters = 0;
454 else
455 et->et_ops->eo_fill_max_leaf_clusters(et);
456}
457
458void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree *et,
459 struct ocfs2_caching_info *ci,
460 struct buffer_head *bh)
461{
462 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_di,
463 NULL, &ocfs2_dinode_et_ops);
464}
465
466void ocfs2_init_xattr_tree_extent_tree(struct ocfs2_extent_tree *et,
467 struct ocfs2_caching_info *ci,
468 struct buffer_head *bh)
469{
470 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_xb,
471 NULL, &ocfs2_xattr_tree_et_ops);
472}
473
474void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree *et,
475 struct ocfs2_caching_info *ci,
476 struct ocfs2_xattr_value_buf *vb)
477{
478 __ocfs2_init_extent_tree(et, ci, vb->vb_bh, vb->vb_access, vb,
479 &ocfs2_xattr_value_et_ops);
480}
481
482void ocfs2_init_dx_root_extent_tree(struct ocfs2_extent_tree *et,
483 struct ocfs2_caching_info *ci,
484 struct buffer_head *bh)
485{
486 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_dr,
487 NULL, &ocfs2_dx_root_et_ops);
488}
489
490void ocfs2_init_refcount_extent_tree(struct ocfs2_extent_tree *et,
491 struct ocfs2_caching_info *ci,
492 struct buffer_head *bh)
493{
494 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_rb,
495 NULL, &ocfs2_refcount_tree_et_ops);
496}
497
498static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree *et,
499 u64 new_last_eb_blk)
500{
501 et->et_ops->eo_set_last_eb_blk(et, new_last_eb_blk);
502}
503
504static inline u64 ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree *et)
505{
506 return et->et_ops->eo_get_last_eb_blk(et);
507}
508
509static inline void ocfs2_et_update_clusters(struct ocfs2_extent_tree *et,
510 u32 clusters)
511{
512 et->et_ops->eo_update_clusters(et, clusters);
513}
514
515static inline void ocfs2_et_extent_map_insert(struct ocfs2_extent_tree *et,
516 struct ocfs2_extent_rec *rec)
517{
518 if (et->et_ops->eo_extent_map_insert)
519 et->et_ops->eo_extent_map_insert(et, rec);
520}
521
522static inline void ocfs2_et_extent_map_truncate(struct ocfs2_extent_tree *et,
523 u32 clusters)
524{
525 if (et->et_ops->eo_extent_map_truncate)
526 et->et_ops->eo_extent_map_truncate(et, clusters);
527}
528
529static inline int ocfs2_et_root_journal_access(handle_t *handle,
530 struct ocfs2_extent_tree *et,
531 int type)
532{
533 return et->et_root_journal_access(handle, et->et_ci, et->et_root_bh,
534 type);
535}
536
537static inline enum ocfs2_contig_type
538 ocfs2_et_extent_contig(struct ocfs2_extent_tree *et,
539 struct ocfs2_extent_rec *rec,
540 struct ocfs2_extent_rec *insert_rec)
541{
542 if (et->et_ops->eo_extent_contig)
543 return et->et_ops->eo_extent_contig(et, rec, insert_rec);
544
545 return ocfs2_extent_rec_contig(
546 ocfs2_metadata_cache_get_super(et->et_ci),
547 rec, insert_rec);
548}
549
550static inline int ocfs2_et_insert_check(struct ocfs2_extent_tree *et,
551 struct ocfs2_extent_rec *rec)
552{
553 int ret = 0;
554
555 if (et->et_ops->eo_insert_check)
556 ret = et->et_ops->eo_insert_check(et, rec);
557 return ret;
558}
559
560static inline int ocfs2_et_sanity_check(struct ocfs2_extent_tree *et)
561{
562 int ret = 0;
563
564 if (et->et_ops->eo_sanity_check)
565 ret = et->et_ops->eo_sanity_check(et);
566 return ret;
567}
568
569static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
570 struct ocfs2_extent_block *eb);
571static void ocfs2_adjust_rightmost_records(handle_t *handle,
572 struct ocfs2_extent_tree *et,
573 struct ocfs2_path *path,
574 struct ocfs2_extent_rec *insert_rec);
575/*
576 * Reset the actual path elements so that we can re-use the structure
577 * to build another path. Generally, this involves freeing the buffer
578 * heads.
579 */
580void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root)
581{
582 int i, start = 0, depth = 0;
583 struct ocfs2_path_item *node;
584
585 if (keep_root)
586 start = 1;
587
588 for(i = start; i < path_num_items(path); i++) {
589 node = &path->p_node[i];
590
591 brelse(node->bh);
592 node->bh = NULL;
593 node->el = NULL;
594 }
595
596 /*
597 * Tree depth may change during truncate, or insert. If we're
598 * keeping the root extent list, then make sure that our path
599 * structure reflects the proper depth.
600 */
601 if (keep_root)
602 depth = le16_to_cpu(path_root_el(path)->l_tree_depth);
603 else
604 path_root_access(path) = NULL;
605
606 path->p_tree_depth = depth;
607}
608
609void ocfs2_free_path(struct ocfs2_path *path)
610{
611 if (path) {
612 ocfs2_reinit_path(path, 0);
613 kfree(path);
614 }
615}
616
617/*
618 * All the elements of src into dest. After this call, src could be freed
619 * without affecting dest.
620 *
621 * Both paths should have the same root. Any non-root elements of dest
622 * will be freed.
623 */
624static void ocfs2_cp_path(struct ocfs2_path *dest, struct ocfs2_path *src)
625{
626 int i;
627
628 BUG_ON(path_root_bh(dest) != path_root_bh(src));
629 BUG_ON(path_root_el(dest) != path_root_el(src));
630 BUG_ON(path_root_access(dest) != path_root_access(src));
631
632 ocfs2_reinit_path(dest, 1);
633
634 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
635 dest->p_node[i].bh = src->p_node[i].bh;
636 dest->p_node[i].el = src->p_node[i].el;
637
638 if (dest->p_node[i].bh)
639 get_bh(dest->p_node[i].bh);
640 }
641}
642
643/*
644 * Make the *dest path the same as src and re-initialize src path to
645 * have a root only.
646 */
647static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src)
648{
649 int i;
650
651 BUG_ON(path_root_bh(dest) != path_root_bh(src));
652 BUG_ON(path_root_access(dest) != path_root_access(src));
653
654 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
655 brelse(dest->p_node[i].bh);
656
657 dest->p_node[i].bh = src->p_node[i].bh;
658 dest->p_node[i].el = src->p_node[i].el;
659
660 src->p_node[i].bh = NULL;
661 src->p_node[i].el = NULL;
662 }
663}
664
665/*
666 * Insert an extent block at given index.
667 *
668 * This will not take an additional reference on eb_bh.
669 */
670static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index,
671 struct buffer_head *eb_bh)
672{
673 struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data;
674
675 /*
676 * Right now, no root bh is an extent block, so this helps
677 * catch code errors with dinode trees. The assertion can be
678 * safely removed if we ever need to insert extent block
679 * structures at the root.
680 */
681 BUG_ON(index == 0);
682
683 path->p_node[index].bh = eb_bh;
684 path->p_node[index].el = &eb->h_list;
685}
686
687static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh,
688 struct ocfs2_extent_list *root_el,
689 ocfs2_journal_access_func access)
690{
691 struct ocfs2_path *path;
692
693 BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH);
694
695 path = kzalloc(sizeof(*path), GFP_NOFS);
696 if (path) {
697 path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth);
698 get_bh(root_bh);
699 path_root_bh(path) = root_bh;
700 path_root_el(path) = root_el;
701 path_root_access(path) = access;
702 }
703
704 return path;
705}
706
707struct ocfs2_path *ocfs2_new_path_from_path(struct ocfs2_path *path)
708{
709 return ocfs2_new_path(path_root_bh(path), path_root_el(path),
710 path_root_access(path));
711}
712
713struct ocfs2_path *ocfs2_new_path_from_et(struct ocfs2_extent_tree *et)
714{
715 return ocfs2_new_path(et->et_root_bh, et->et_root_el,
716 et->et_root_journal_access);
717}
718
719/*
720 * Journal the buffer at depth idx. All idx>0 are extent_blocks,
721 * otherwise it's the root_access function.
722 *
723 * I don't like the way this function's name looks next to
724 * ocfs2_journal_access_path(), but I don't have a better one.
725 */
726int ocfs2_path_bh_journal_access(handle_t *handle,
727 struct ocfs2_caching_info *ci,
728 struct ocfs2_path *path,
729 int idx)
730{
731 ocfs2_journal_access_func access = path_root_access(path);
732
733 if (!access)
734 access = ocfs2_journal_access;
735
736 if (idx)
737 access = ocfs2_journal_access_eb;
738
739 return access(handle, ci, path->p_node[idx].bh,
740 OCFS2_JOURNAL_ACCESS_WRITE);
741}
742
743/*
744 * Convenience function to journal all components in a path.
745 */
746int ocfs2_journal_access_path(struct ocfs2_caching_info *ci,
747 handle_t *handle,
748 struct ocfs2_path *path)
749{
750 int i, ret = 0;
751
752 if (!path)
753 goto out;
754
755 for(i = 0; i < path_num_items(path); i++) {
756 ret = ocfs2_path_bh_journal_access(handle, ci, path, i);
757 if (ret < 0) {
758 mlog_errno(ret);
759 goto out;
760 }
761 }
762
763out:
764 return ret;
765}
766
767/*
768 * Return the index of the extent record which contains cluster #v_cluster.
769 * -1 is returned if it was not found.
770 *
771 * Should work fine on interior and exterior nodes.
772 */
773int ocfs2_search_extent_list(struct ocfs2_extent_list *el, u32 v_cluster)
774{
775 int ret = -1;
776 int i;
777 struct ocfs2_extent_rec *rec;
778 u32 rec_end, rec_start, clusters;
779
780 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
781 rec = &el->l_recs[i];
782
783 rec_start = le32_to_cpu(rec->e_cpos);
784 clusters = ocfs2_rec_clusters(el, rec);
785
786 rec_end = rec_start + clusters;
787
788 if (v_cluster >= rec_start && v_cluster < rec_end) {
789 ret = i;
790 break;
791 }
792 }
793
794 return ret;
795}
796
797/*
798 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
799 * ocfs2_extent_rec_contig only work properly against leaf nodes!
800 */
801static int ocfs2_block_extent_contig(struct super_block *sb,
802 struct ocfs2_extent_rec *ext,
803 u64 blkno)
804{
805 u64 blk_end = le64_to_cpu(ext->e_blkno);
806
807 blk_end += ocfs2_clusters_to_blocks(sb,
808 le16_to_cpu(ext->e_leaf_clusters));
809
810 return blkno == blk_end;
811}
812
813static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
814 struct ocfs2_extent_rec *right)
815{
816 u32 left_range;
817
818 left_range = le32_to_cpu(left->e_cpos) +
819 le16_to_cpu(left->e_leaf_clusters);
820
821 return (left_range == le32_to_cpu(right->e_cpos));
822}
823
824static enum ocfs2_contig_type
825 ocfs2_extent_rec_contig(struct super_block *sb,
826 struct ocfs2_extent_rec *ext,
827 struct ocfs2_extent_rec *insert_rec)
828{
829 u64 blkno = le64_to_cpu(insert_rec->e_blkno);
830
831 /*
832 * Refuse to coalesce extent records with different flag
833 * fields - we don't want to mix unwritten extents with user
834 * data.
835 */
836 if (ext->e_flags != insert_rec->e_flags)
837 return CONTIG_NONE;
838
839 if (ocfs2_extents_adjacent(ext, insert_rec) &&
840 ocfs2_block_extent_contig(sb, ext, blkno))
841 return CONTIG_RIGHT;
842
843 blkno = le64_to_cpu(ext->e_blkno);
844 if (ocfs2_extents_adjacent(insert_rec, ext) &&
845 ocfs2_block_extent_contig(sb, insert_rec, blkno))
846 return CONTIG_LEFT;
847
848 return CONTIG_NONE;
849}
850
851/*
852 * NOTE: We can have pretty much any combination of contiguousness and
853 * appending.
854 *
855 * The usefulness of APPEND_TAIL is more in that it lets us know that
856 * we'll have to update the path to that leaf.
857 */
858enum ocfs2_append_type {
859 APPEND_NONE = 0,
860 APPEND_TAIL,
861};
862
863enum ocfs2_split_type {
864 SPLIT_NONE = 0,
865 SPLIT_LEFT,
866 SPLIT_RIGHT,
867};
868
869struct ocfs2_insert_type {
870 enum ocfs2_split_type ins_split;
871 enum ocfs2_append_type ins_appending;
872 enum ocfs2_contig_type ins_contig;
873 int ins_contig_index;
874 int ins_tree_depth;
875};
876
877struct ocfs2_merge_ctxt {
878 enum ocfs2_contig_type c_contig_type;
879 int c_has_empty_extent;
880 int c_split_covers_rec;
881};
882
883static int ocfs2_validate_extent_block(struct super_block *sb,
884 struct buffer_head *bh)
885{
886 int rc;
887 struct ocfs2_extent_block *eb =
888 (struct ocfs2_extent_block *)bh->b_data;
889
890 trace_ocfs2_validate_extent_block((unsigned long long)bh->b_blocknr);
891
892 BUG_ON(!buffer_uptodate(bh));
893
894 /*
895 * If the ecc fails, we return the error but otherwise
896 * leave the filesystem running. We know any error is
897 * local to this block.
898 */
899 rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &eb->h_check);
900 if (rc) {
901 mlog(ML_ERROR, "Checksum failed for extent block %llu\n",
902 (unsigned long long)bh->b_blocknr);
903 return rc;
904 }
905
906 /*
907 * Errors after here are fatal.
908 */
909
910 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
911 ocfs2_error(sb,
912 "Extent block #%llu has bad signature %.*s",
913 (unsigned long long)bh->b_blocknr, 7,
914 eb->h_signature);
915 return -EINVAL;
916 }
917
918 if (le64_to_cpu(eb->h_blkno) != bh->b_blocknr) {
919 ocfs2_error(sb,
920 "Extent block #%llu has an invalid h_blkno "
921 "of %llu",
922 (unsigned long long)bh->b_blocknr,
923 (unsigned long long)le64_to_cpu(eb->h_blkno));
924 return -EINVAL;
925 }
926
927 if (le32_to_cpu(eb->h_fs_generation) != OCFS2_SB(sb)->fs_generation) {
928 ocfs2_error(sb,
929 "Extent block #%llu has an invalid "
930 "h_fs_generation of #%u",
931 (unsigned long long)bh->b_blocknr,
932 le32_to_cpu(eb->h_fs_generation));
933 return -EINVAL;
934 }
935
936 return 0;
937}
938
939int ocfs2_read_extent_block(struct ocfs2_caching_info *ci, u64 eb_blkno,
940 struct buffer_head **bh)
941{
942 int rc;
943 struct buffer_head *tmp = *bh;
944
945 rc = ocfs2_read_block(ci, eb_blkno, &tmp,
946 ocfs2_validate_extent_block);
947
948 /* If ocfs2_read_block() got us a new bh, pass it up. */
949 if (!rc && !*bh)
950 *bh = tmp;
951
952 return rc;
953}
954
955
956/*
957 * How many free extents have we got before we need more meta data?
958 */
959int ocfs2_num_free_extents(struct ocfs2_super *osb,
960 struct ocfs2_extent_tree *et)
961{
962 int retval;
963 struct ocfs2_extent_list *el = NULL;
964 struct ocfs2_extent_block *eb;
965 struct buffer_head *eb_bh = NULL;
966 u64 last_eb_blk = 0;
967
968 el = et->et_root_el;
969 last_eb_blk = ocfs2_et_get_last_eb_blk(et);
970
971 if (last_eb_blk) {
972 retval = ocfs2_read_extent_block(et->et_ci, last_eb_blk,
973 &eb_bh);
974 if (retval < 0) {
975 mlog_errno(retval);
976 goto bail;
977 }
978 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
979 el = &eb->h_list;
980 }
981
982 BUG_ON(el->l_tree_depth != 0);
983
984 retval = le16_to_cpu(el->l_count) - le16_to_cpu(el->l_next_free_rec);
985bail:
986 brelse(eb_bh);
987
988 trace_ocfs2_num_free_extents(retval);
989 return retval;
990}
991
992/* expects array to already be allocated
993 *
994 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
995 * l_count for you
996 */
997static int ocfs2_create_new_meta_bhs(handle_t *handle,
998 struct ocfs2_extent_tree *et,
999 int wanted,
1000 struct ocfs2_alloc_context *meta_ac,
1001 struct buffer_head *bhs[])
1002{
1003 int count, status, i;
1004 u16 suballoc_bit_start;
1005 u32 num_got;
1006 u64 suballoc_loc, first_blkno;
1007 struct ocfs2_super *osb =
1008 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
1009 struct ocfs2_extent_block *eb;
1010
1011 count = 0;
1012 while (count < wanted) {
1013 status = ocfs2_claim_metadata(handle,
1014 meta_ac,
1015 wanted - count,
1016 &suballoc_loc,
1017 &suballoc_bit_start,
1018 &num_got,
1019 &first_blkno);
1020 if (status < 0) {
1021 mlog_errno(status);
1022 goto bail;
1023 }
1024
1025 for(i = count; i < (num_got + count); i++) {
1026 bhs[i] = sb_getblk(osb->sb, first_blkno);
1027 if (bhs[i] == NULL) {
1028 status = -EIO;
1029 mlog_errno(status);
1030 goto bail;
1031 }
1032 ocfs2_set_new_buffer_uptodate(et->et_ci, bhs[i]);
1033
1034 status = ocfs2_journal_access_eb(handle, et->et_ci,
1035 bhs[i],
1036 OCFS2_JOURNAL_ACCESS_CREATE);
1037 if (status < 0) {
1038 mlog_errno(status);
1039 goto bail;
1040 }
1041
1042 memset(bhs[i]->b_data, 0, osb->sb->s_blocksize);
1043 eb = (struct ocfs2_extent_block *) bhs[i]->b_data;
1044 /* Ok, setup the minimal stuff here. */
1045 strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
1046 eb->h_blkno = cpu_to_le64(first_blkno);
1047 eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
1048 eb->h_suballoc_slot =
1049 cpu_to_le16(meta_ac->ac_alloc_slot);
1050 eb->h_suballoc_loc = cpu_to_le64(suballoc_loc);
1051 eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1052 eb->h_list.l_count =
1053 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
1054
1055 suballoc_bit_start++;
1056 first_blkno++;
1057
1058 /* We'll also be dirtied by the caller, so
1059 * this isn't absolutely necessary. */
1060 ocfs2_journal_dirty(handle, bhs[i]);
1061 }
1062
1063 count += num_got;
1064 }
1065
1066 status = 0;
1067bail:
1068 if (status < 0) {
1069 for(i = 0; i < wanted; i++) {
1070 brelse(bhs[i]);
1071 bhs[i] = NULL;
1072 }
1073 mlog_errno(status);
1074 }
1075 return status;
1076}
1077
1078/*
1079 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
1080 *
1081 * Returns the sum of the rightmost extent rec logical offset and
1082 * cluster count.
1083 *
1084 * ocfs2_add_branch() uses this to determine what logical cluster
1085 * value should be populated into the leftmost new branch records.
1086 *
1087 * ocfs2_shift_tree_depth() uses this to determine the # clusters
1088 * value for the new topmost tree record.
1089 */
1090static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list *el)
1091{
1092 int i;
1093
1094 i = le16_to_cpu(el->l_next_free_rec) - 1;
1095
1096 return le32_to_cpu(el->l_recs[i].e_cpos) +
1097 ocfs2_rec_clusters(el, &el->l_recs[i]);
1098}
1099
1100/*
1101 * Change range of the branches in the right most path according to the leaf
1102 * extent block's rightmost record.
1103 */
1104static int ocfs2_adjust_rightmost_branch(handle_t *handle,
1105 struct ocfs2_extent_tree *et)
1106{
1107 int status;
1108 struct ocfs2_path *path = NULL;
1109 struct ocfs2_extent_list *el;
1110 struct ocfs2_extent_rec *rec;
1111
1112 path = ocfs2_new_path_from_et(et);
1113 if (!path) {
1114 status = -ENOMEM;
1115 return status;
1116 }
1117
1118 status = ocfs2_find_path(et->et_ci, path, UINT_MAX);
1119 if (status < 0) {
1120 mlog_errno(status);
1121 goto out;
1122 }
1123
1124 status = ocfs2_extend_trans(handle, path_num_items(path));
1125 if (status < 0) {
1126 mlog_errno(status);
1127 goto out;
1128 }
1129
1130 status = ocfs2_journal_access_path(et->et_ci, handle, path);
1131 if (status < 0) {
1132 mlog_errno(status);
1133 goto out;
1134 }
1135
1136 el = path_leaf_el(path);
1137 rec = &el->l_recs[le32_to_cpu(el->l_next_free_rec) - 1];
1138
1139 ocfs2_adjust_rightmost_records(handle, et, path, rec);
1140
1141out:
1142 ocfs2_free_path(path);
1143 return status;
1144}
1145
1146/*
1147 * Add an entire tree branch to our inode. eb_bh is the extent block
1148 * to start at, if we don't want to start the branch at the root
1149 * structure.
1150 *
1151 * last_eb_bh is required as we have to update it's next_leaf pointer
1152 * for the new last extent block.
1153 *
1154 * the new branch will be 'empty' in the sense that every block will
1155 * contain a single record with cluster count == 0.
1156 */
1157static int ocfs2_add_branch(handle_t *handle,
1158 struct ocfs2_extent_tree *et,
1159 struct buffer_head *eb_bh,
1160 struct buffer_head **last_eb_bh,
1161 struct ocfs2_alloc_context *meta_ac)
1162{
1163 int status, new_blocks, i;
1164 u64 next_blkno, new_last_eb_blk;
1165 struct buffer_head *bh;
1166 struct buffer_head **new_eb_bhs = NULL;
1167 struct ocfs2_extent_block *eb;
1168 struct ocfs2_extent_list *eb_el;
1169 struct ocfs2_extent_list *el;
1170 u32 new_cpos, root_end;
1171
1172 BUG_ON(!last_eb_bh || !*last_eb_bh);
1173
1174 if (eb_bh) {
1175 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1176 el = &eb->h_list;
1177 } else
1178 el = et->et_root_el;
1179
1180 /* we never add a branch to a leaf. */
1181 BUG_ON(!el->l_tree_depth);
1182
1183 new_blocks = le16_to_cpu(el->l_tree_depth);
1184
1185 eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
1186 new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
1187 root_end = ocfs2_sum_rightmost_rec(et->et_root_el);
1188
1189 /*
1190 * If there is a gap before the root end and the real end
1191 * of the righmost leaf block, we need to remove the gap
1192 * between new_cpos and root_end first so that the tree
1193 * is consistent after we add a new branch(it will start
1194 * from new_cpos).
1195 */
1196 if (root_end > new_cpos) {
1197 trace_ocfs2_adjust_rightmost_branch(
1198 (unsigned long long)
1199 ocfs2_metadata_cache_owner(et->et_ci),
1200 root_end, new_cpos);
1201
1202 status = ocfs2_adjust_rightmost_branch(handle, et);
1203 if (status) {
1204 mlog_errno(status);
1205 goto bail;
1206 }
1207 }
1208
1209 /* allocate the number of new eb blocks we need */
1210 new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
1211 GFP_KERNEL);
1212 if (!new_eb_bhs) {
1213 status = -ENOMEM;
1214 mlog_errno(status);
1215 goto bail;
1216 }
1217
1218 status = ocfs2_create_new_meta_bhs(handle, et, new_blocks,
1219 meta_ac, new_eb_bhs);
1220 if (status < 0) {
1221 mlog_errno(status);
1222 goto bail;
1223 }
1224
1225 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
1226 * linked with the rest of the tree.
1227 * conversly, new_eb_bhs[0] is the new bottommost leaf.
1228 *
1229 * when we leave the loop, new_last_eb_blk will point to the
1230 * newest leaf, and next_blkno will point to the topmost extent
1231 * block. */
1232 next_blkno = new_last_eb_blk = 0;
1233 for(i = 0; i < new_blocks; i++) {
1234 bh = new_eb_bhs[i];
1235 eb = (struct ocfs2_extent_block *) bh->b_data;
1236 /* ocfs2_create_new_meta_bhs() should create it right! */
1237 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1238 eb_el = &eb->h_list;
1239
1240 status = ocfs2_journal_access_eb(handle, et->et_ci, bh,
1241 OCFS2_JOURNAL_ACCESS_CREATE);
1242 if (status < 0) {
1243 mlog_errno(status);
1244 goto bail;
1245 }
1246
1247 eb->h_next_leaf_blk = 0;
1248 eb_el->l_tree_depth = cpu_to_le16(i);
1249 eb_el->l_next_free_rec = cpu_to_le16(1);
1250 /*
1251 * This actually counts as an empty extent as
1252 * c_clusters == 0
1253 */
1254 eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
1255 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
1256 /*
1257 * eb_el isn't always an interior node, but even leaf
1258 * nodes want a zero'd flags and reserved field so
1259 * this gets the whole 32 bits regardless of use.
1260 */
1261 eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
1262 if (!eb_el->l_tree_depth)
1263 new_last_eb_blk = le64_to_cpu(eb->h_blkno);
1264
1265 ocfs2_journal_dirty(handle, bh);
1266 next_blkno = le64_to_cpu(eb->h_blkno);
1267 }
1268
1269 /* This is a bit hairy. We want to update up to three blocks
1270 * here without leaving any of them in an inconsistent state
1271 * in case of error. We don't have to worry about
1272 * journal_dirty erroring as it won't unless we've aborted the
1273 * handle (in which case we would never be here) so reserving
1274 * the write with journal_access is all we need to do. */
1275 status = ocfs2_journal_access_eb(handle, et->et_ci, *last_eb_bh,
1276 OCFS2_JOURNAL_ACCESS_WRITE);
1277 if (status < 0) {
1278 mlog_errno(status);
1279 goto bail;
1280 }
1281 status = ocfs2_et_root_journal_access(handle, et,
1282 OCFS2_JOURNAL_ACCESS_WRITE);
1283 if (status < 0) {
1284 mlog_errno(status);
1285 goto bail;
1286 }
1287 if (eb_bh) {
1288 status = ocfs2_journal_access_eb(handle, et->et_ci, eb_bh,
1289 OCFS2_JOURNAL_ACCESS_WRITE);
1290 if (status < 0) {
1291 mlog_errno(status);
1292 goto bail;
1293 }
1294 }
1295
1296 /* Link the new branch into the rest of the tree (el will
1297 * either be on the root_bh, or the extent block passed in. */
1298 i = le16_to_cpu(el->l_next_free_rec);
1299 el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
1300 el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1301 el->l_recs[i].e_int_clusters = 0;
1302 le16_add_cpu(&el->l_next_free_rec, 1);
1303
1304 /* fe needs a new last extent block pointer, as does the
1305 * next_leaf on the previously last-extent-block. */
1306 ocfs2_et_set_last_eb_blk(et, new_last_eb_blk);
1307
1308 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
1309 eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);
1310
1311 ocfs2_journal_dirty(handle, *last_eb_bh);
1312 ocfs2_journal_dirty(handle, et->et_root_bh);
1313 if (eb_bh)
1314 ocfs2_journal_dirty(handle, eb_bh);
1315
1316 /*
1317 * Some callers want to track the rightmost leaf so pass it
1318 * back here.
1319 */
1320 brelse(*last_eb_bh);
1321 get_bh(new_eb_bhs[0]);
1322 *last_eb_bh = new_eb_bhs[0];
1323
1324 status = 0;
1325bail:
1326 if (new_eb_bhs) {
1327 for (i = 0; i < new_blocks; i++)
1328 brelse(new_eb_bhs[i]);
1329 kfree(new_eb_bhs);
1330 }
1331
1332 return status;
1333}
1334
1335/*
1336 * adds another level to the allocation tree.
1337 * returns back the new extent block so you can add a branch to it
1338 * after this call.
1339 */
1340static int ocfs2_shift_tree_depth(handle_t *handle,
1341 struct ocfs2_extent_tree *et,
1342 struct ocfs2_alloc_context *meta_ac,
1343 struct buffer_head **ret_new_eb_bh)
1344{
1345 int status, i;
1346 u32 new_clusters;
1347 struct buffer_head *new_eb_bh = NULL;
1348 struct ocfs2_extent_block *eb;
1349 struct ocfs2_extent_list *root_el;
1350 struct ocfs2_extent_list *eb_el;
1351
1352 status = ocfs2_create_new_meta_bhs(handle, et, 1, meta_ac,
1353 &new_eb_bh);
1354 if (status < 0) {
1355 mlog_errno(status);
1356 goto bail;
1357 }
1358
1359 eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
1360 /* ocfs2_create_new_meta_bhs() should create it right! */
1361 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1362
1363 eb_el = &eb->h_list;
1364 root_el = et->et_root_el;
1365
1366 status = ocfs2_journal_access_eb(handle, et->et_ci, new_eb_bh,
1367 OCFS2_JOURNAL_ACCESS_CREATE);
1368 if (status < 0) {
1369 mlog_errno(status);
1370 goto bail;
1371 }
1372
1373 /* copy the root extent list data into the new extent block */
1374 eb_el->l_tree_depth = root_el->l_tree_depth;
1375 eb_el->l_next_free_rec = root_el->l_next_free_rec;
1376 for (i = 0; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1377 eb_el->l_recs[i] = root_el->l_recs[i];
1378
1379 ocfs2_journal_dirty(handle, new_eb_bh);
1380
1381 status = ocfs2_et_root_journal_access(handle, et,
1382 OCFS2_JOURNAL_ACCESS_WRITE);
1383 if (status < 0) {
1384 mlog_errno(status);
1385 goto bail;
1386 }
1387
1388 new_clusters = ocfs2_sum_rightmost_rec(eb_el);
1389
1390 /* update root_bh now */
1391 le16_add_cpu(&root_el->l_tree_depth, 1);
1392 root_el->l_recs[0].e_cpos = 0;
1393 root_el->l_recs[0].e_blkno = eb->h_blkno;
1394 root_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
1395 for (i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1396 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
1397 root_el->l_next_free_rec = cpu_to_le16(1);
1398
1399 /* If this is our 1st tree depth shift, then last_eb_blk
1400 * becomes the allocated extent block */
1401 if (root_el->l_tree_depth == cpu_to_le16(1))
1402 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
1403
1404 ocfs2_journal_dirty(handle, et->et_root_bh);
1405
1406 *ret_new_eb_bh = new_eb_bh;
1407 new_eb_bh = NULL;
1408 status = 0;
1409bail:
1410 brelse(new_eb_bh);
1411
1412 return status;
1413}
1414
1415/*
1416 * Should only be called when there is no space left in any of the
1417 * leaf nodes. What we want to do is find the lowest tree depth
1418 * non-leaf extent block with room for new records. There are three
1419 * valid results of this search:
1420 *
1421 * 1) a lowest extent block is found, then we pass it back in
1422 * *lowest_eb_bh and return '0'
1423 *
1424 * 2) the search fails to find anything, but the root_el has room. We
1425 * pass NULL back in *lowest_eb_bh, but still return '0'
1426 *
1427 * 3) the search fails to find anything AND the root_el is full, in
1428 * which case we return > 0
1429 *
1430 * return status < 0 indicates an error.
1431 */
1432static int ocfs2_find_branch_target(struct ocfs2_extent_tree *et,
1433 struct buffer_head **target_bh)
1434{
1435 int status = 0, i;
1436 u64 blkno;
1437 struct ocfs2_extent_block *eb;
1438 struct ocfs2_extent_list *el;
1439 struct buffer_head *bh = NULL;
1440 struct buffer_head *lowest_bh = NULL;
1441
1442 *target_bh = NULL;
1443
1444 el = et->et_root_el;
1445
1446 while(le16_to_cpu(el->l_tree_depth) > 1) {
1447 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1448 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1449 "Owner %llu has empty "
1450 "extent list (next_free_rec == 0)",
1451 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
1452 status = -EIO;
1453 goto bail;
1454 }
1455 i = le16_to_cpu(el->l_next_free_rec) - 1;
1456 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1457 if (!blkno) {
1458 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1459 "Owner %llu has extent "
1460 "list where extent # %d has no physical "
1461 "block start",
1462 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), i);
1463 status = -EIO;
1464 goto bail;
1465 }
1466
1467 brelse(bh);
1468 bh = NULL;
1469
1470 status = ocfs2_read_extent_block(et->et_ci, blkno, &bh);
1471 if (status < 0) {
1472 mlog_errno(status);
1473 goto bail;
1474 }
1475
1476 eb = (struct ocfs2_extent_block *) bh->b_data;
1477 el = &eb->h_list;
1478
1479 if (le16_to_cpu(el->l_next_free_rec) <
1480 le16_to_cpu(el->l_count)) {
1481 brelse(lowest_bh);
1482 lowest_bh = bh;
1483 get_bh(lowest_bh);
1484 }
1485 }
1486
1487 /* If we didn't find one and the fe doesn't have any room,
1488 * then return '1' */
1489 el = et->et_root_el;
1490 if (!lowest_bh && (el->l_next_free_rec == el->l_count))
1491 status = 1;
1492
1493 *target_bh = lowest_bh;
1494bail:
1495 brelse(bh);
1496
1497 return status;
1498}
1499
1500/*
1501 * Grow a b-tree so that it has more records.
1502 *
1503 * We might shift the tree depth in which case existing paths should
1504 * be considered invalid.
1505 *
1506 * Tree depth after the grow is returned via *final_depth.
1507 *
1508 * *last_eb_bh will be updated by ocfs2_add_branch().
1509 */
1510static int ocfs2_grow_tree(handle_t *handle, struct ocfs2_extent_tree *et,
1511 int *final_depth, struct buffer_head **last_eb_bh,
1512 struct ocfs2_alloc_context *meta_ac)
1513{
1514 int ret, shift;
1515 struct ocfs2_extent_list *el = et->et_root_el;
1516 int depth = le16_to_cpu(el->l_tree_depth);
1517 struct buffer_head *bh = NULL;
1518
1519 BUG_ON(meta_ac == NULL);
1520
1521 shift = ocfs2_find_branch_target(et, &bh);
1522 if (shift < 0) {
1523 ret = shift;
1524 mlog_errno(ret);
1525 goto out;
1526 }
1527
1528 /* We traveled all the way to the bottom of the allocation tree
1529 * and didn't find room for any more extents - we need to add
1530 * another tree level */
1531 if (shift) {
1532 BUG_ON(bh);
1533 trace_ocfs2_grow_tree(
1534 (unsigned long long)
1535 ocfs2_metadata_cache_owner(et->et_ci),
1536 depth);
1537
1538 /* ocfs2_shift_tree_depth will return us a buffer with
1539 * the new extent block (so we can pass that to
1540 * ocfs2_add_branch). */
1541 ret = ocfs2_shift_tree_depth(handle, et, meta_ac, &bh);
1542 if (ret < 0) {
1543 mlog_errno(ret);
1544 goto out;
1545 }
1546 depth++;
1547 if (depth == 1) {
1548 /*
1549 * Special case: we have room now if we shifted from
1550 * tree_depth 0, so no more work needs to be done.
1551 *
1552 * We won't be calling add_branch, so pass
1553 * back *last_eb_bh as the new leaf. At depth
1554 * zero, it should always be null so there's
1555 * no reason to brelse.
1556 */
1557 BUG_ON(*last_eb_bh);
1558 get_bh(bh);
1559 *last_eb_bh = bh;
1560 goto out;
1561 }
1562 }
1563
1564 /* call ocfs2_add_branch to add the final part of the tree with
1565 * the new data. */
1566 ret = ocfs2_add_branch(handle, et, bh, last_eb_bh,
1567 meta_ac);
1568 if (ret < 0) {
1569 mlog_errno(ret);
1570 goto out;
1571 }
1572
1573out:
1574 if (final_depth)
1575 *final_depth = depth;
1576 brelse(bh);
1577 return ret;
1578}
1579
1580/*
1581 * This function will discard the rightmost extent record.
1582 */
1583static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
1584{
1585 int next_free = le16_to_cpu(el->l_next_free_rec);
1586 int count = le16_to_cpu(el->l_count);
1587 unsigned int num_bytes;
1588
1589 BUG_ON(!next_free);
1590 /* This will cause us to go off the end of our extent list. */
1591 BUG_ON(next_free >= count);
1592
1593 num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
1594
1595 memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
1596}
1597
1598static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
1599 struct ocfs2_extent_rec *insert_rec)
1600{
1601 int i, insert_index, next_free, has_empty, num_bytes;
1602 u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
1603 struct ocfs2_extent_rec *rec;
1604
1605 next_free = le16_to_cpu(el->l_next_free_rec);
1606 has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
1607
1608 BUG_ON(!next_free);
1609
1610 /* The tree code before us didn't allow enough room in the leaf. */
1611 BUG_ON(el->l_next_free_rec == el->l_count && !has_empty);
1612
1613 /*
1614 * The easiest way to approach this is to just remove the
1615 * empty extent and temporarily decrement next_free.
1616 */
1617 if (has_empty) {
1618 /*
1619 * If next_free was 1 (only an empty extent), this
1620 * loop won't execute, which is fine. We still want
1621 * the decrement above to happen.
1622 */
1623 for(i = 0; i < (next_free - 1); i++)
1624 el->l_recs[i] = el->l_recs[i+1];
1625
1626 next_free--;
1627 }
1628
1629 /*
1630 * Figure out what the new record index should be.
1631 */
1632 for(i = 0; i < next_free; i++) {
1633 rec = &el->l_recs[i];
1634
1635 if (insert_cpos < le32_to_cpu(rec->e_cpos))
1636 break;
1637 }
1638 insert_index = i;
1639
1640 trace_ocfs2_rotate_leaf(insert_cpos, insert_index,
1641 has_empty, next_free,
1642 le16_to_cpu(el->l_count));
1643
1644 BUG_ON(insert_index < 0);
1645 BUG_ON(insert_index >= le16_to_cpu(el->l_count));
1646 BUG_ON(insert_index > next_free);
1647
1648 /*
1649 * No need to memmove if we're just adding to the tail.
1650 */
1651 if (insert_index != next_free) {
1652 BUG_ON(next_free >= le16_to_cpu(el->l_count));
1653
1654 num_bytes = next_free - insert_index;
1655 num_bytes *= sizeof(struct ocfs2_extent_rec);
1656 memmove(&el->l_recs[insert_index + 1],
1657 &el->l_recs[insert_index],
1658 num_bytes);
1659 }
1660
1661 /*
1662 * Either we had an empty extent, and need to re-increment or
1663 * there was no empty extent on a non full rightmost leaf node,
1664 * in which case we still need to increment.
1665 */
1666 next_free++;
1667 el->l_next_free_rec = cpu_to_le16(next_free);
1668 /*
1669 * Make sure none of the math above just messed up our tree.
1670 */
1671 BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
1672
1673 el->l_recs[insert_index] = *insert_rec;
1674
1675}
1676
1677static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
1678{
1679 int size, num_recs = le16_to_cpu(el->l_next_free_rec);
1680
1681 BUG_ON(num_recs == 0);
1682
1683 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
1684 num_recs--;
1685 size = num_recs * sizeof(struct ocfs2_extent_rec);
1686 memmove(&el->l_recs[0], &el->l_recs[1], size);
1687 memset(&el->l_recs[num_recs], 0,
1688 sizeof(struct ocfs2_extent_rec));
1689 el->l_next_free_rec = cpu_to_le16(num_recs);
1690 }
1691}
1692
1693/*
1694 * Create an empty extent record .
1695 *
1696 * l_next_free_rec may be updated.
1697 *
1698 * If an empty extent already exists do nothing.
1699 */
1700static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
1701{
1702 int next_free = le16_to_cpu(el->l_next_free_rec);
1703
1704 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1705
1706 if (next_free == 0)
1707 goto set_and_inc;
1708
1709 if (ocfs2_is_empty_extent(&el->l_recs[0]))
1710 return;
1711
1712 mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
1713 "Asked to create an empty extent in a full list:\n"
1714 "count = %u, tree depth = %u",
1715 le16_to_cpu(el->l_count),
1716 le16_to_cpu(el->l_tree_depth));
1717
1718 ocfs2_shift_records_right(el);
1719
1720set_and_inc:
1721 le16_add_cpu(&el->l_next_free_rec, 1);
1722 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1723}
1724
1725/*
1726 * For a rotation which involves two leaf nodes, the "root node" is
1727 * the lowest level tree node which contains a path to both leafs. This
1728 * resulting set of information can be used to form a complete "subtree"
1729 *
1730 * This function is passed two full paths from the dinode down to a
1731 * pair of adjacent leaves. It's task is to figure out which path
1732 * index contains the subtree root - this can be the root index itself
1733 * in a worst-case rotation.
1734 *
1735 * The array index of the subtree root is passed back.
1736 */
1737int ocfs2_find_subtree_root(struct ocfs2_extent_tree *et,
1738 struct ocfs2_path *left,
1739 struct ocfs2_path *right)
1740{
1741 int i = 0;
1742
1743 /*
1744 * Check that the caller passed in two paths from the same tree.
1745 */
1746 BUG_ON(path_root_bh(left) != path_root_bh(right));
1747
1748 do {
1749 i++;
1750
1751 /*
1752 * The caller didn't pass two adjacent paths.
1753 */
1754 mlog_bug_on_msg(i > left->p_tree_depth,
1755 "Owner %llu, left depth %u, right depth %u\n"
1756 "left leaf blk %llu, right leaf blk %llu\n",
1757 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
1758 left->p_tree_depth, right->p_tree_depth,
1759 (unsigned long long)path_leaf_bh(left)->b_blocknr,
1760 (unsigned long long)path_leaf_bh(right)->b_blocknr);
1761 } while (left->p_node[i].bh->b_blocknr ==
1762 right->p_node[i].bh->b_blocknr);
1763
1764 return i - 1;
1765}
1766
1767typedef void (path_insert_t)(void *, struct buffer_head *);
1768
1769/*
1770 * Traverse a btree path in search of cpos, starting at root_el.
1771 *
1772 * This code can be called with a cpos larger than the tree, in which
1773 * case it will return the rightmost path.
1774 */
1775static int __ocfs2_find_path(struct ocfs2_caching_info *ci,
1776 struct ocfs2_extent_list *root_el, u32 cpos,
1777 path_insert_t *func, void *data)
1778{
1779 int i, ret = 0;
1780 u32 range;
1781 u64 blkno;
1782 struct buffer_head *bh = NULL;
1783 struct ocfs2_extent_block *eb;
1784 struct ocfs2_extent_list *el;
1785 struct ocfs2_extent_rec *rec;
1786
1787 el = root_el;
1788 while (el->l_tree_depth) {
1789 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1790 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1791 "Owner %llu has empty extent list at "
1792 "depth %u\n",
1793 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1794 le16_to_cpu(el->l_tree_depth));
1795 ret = -EROFS;
1796 goto out;
1797
1798 }
1799
1800 for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
1801 rec = &el->l_recs[i];
1802
1803 /*
1804 * In the case that cpos is off the allocation
1805 * tree, this should just wind up returning the
1806 * rightmost record.
1807 */
1808 range = le32_to_cpu(rec->e_cpos) +
1809 ocfs2_rec_clusters(el, rec);
1810 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1811 break;
1812 }
1813
1814 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1815 if (blkno == 0) {
1816 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1817 "Owner %llu has bad blkno in extent list "
1818 "at depth %u (index %d)\n",
1819 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1820 le16_to_cpu(el->l_tree_depth), i);
1821 ret = -EROFS;
1822 goto out;
1823 }
1824
1825 brelse(bh);
1826 bh = NULL;
1827 ret = ocfs2_read_extent_block(ci, blkno, &bh);
1828 if (ret) {
1829 mlog_errno(ret);
1830 goto out;
1831 }
1832
1833 eb = (struct ocfs2_extent_block *) bh->b_data;
1834 el = &eb->h_list;
1835
1836 if (le16_to_cpu(el->l_next_free_rec) >
1837 le16_to_cpu(el->l_count)) {
1838 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1839 "Owner %llu has bad count in extent list "
1840 "at block %llu (next free=%u, count=%u)\n",
1841 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1842 (unsigned long long)bh->b_blocknr,
1843 le16_to_cpu(el->l_next_free_rec),
1844 le16_to_cpu(el->l_count));
1845 ret = -EROFS;
1846 goto out;
1847 }
1848
1849 if (func)
1850 func(data, bh);
1851 }
1852
1853out:
1854 /*
1855 * Catch any trailing bh that the loop didn't handle.
1856 */
1857 brelse(bh);
1858
1859 return ret;
1860}
1861
1862/*
1863 * Given an initialized path (that is, it has a valid root extent
1864 * list), this function will traverse the btree in search of the path
1865 * which would contain cpos.
1866 *
1867 * The path traveled is recorded in the path structure.
1868 *
1869 * Note that this will not do any comparisons on leaf node extent
1870 * records, so it will work fine in the case that we just added a tree
1871 * branch.
1872 */
1873struct find_path_data {
1874 int index;
1875 struct ocfs2_path *path;
1876};
1877static void find_path_ins(void *data, struct buffer_head *bh)
1878{
1879 struct find_path_data *fp = data;
1880
1881 get_bh(bh);
1882 ocfs2_path_insert_eb(fp->path, fp->index, bh);
1883 fp->index++;
1884}
1885int ocfs2_find_path(struct ocfs2_caching_info *ci,
1886 struct ocfs2_path *path, u32 cpos)
1887{
1888 struct find_path_data data;
1889
1890 data.index = 1;
1891 data.path = path;
1892 return __ocfs2_find_path(ci, path_root_el(path), cpos,
1893 find_path_ins, &data);
1894}
1895
1896static void find_leaf_ins(void *data, struct buffer_head *bh)
1897{
1898 struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1899 struct ocfs2_extent_list *el = &eb->h_list;
1900 struct buffer_head **ret = data;
1901
1902 /* We want to retain only the leaf block. */
1903 if (le16_to_cpu(el->l_tree_depth) == 0) {
1904 get_bh(bh);
1905 *ret = bh;
1906 }
1907}
1908/*
1909 * Find the leaf block in the tree which would contain cpos. No
1910 * checking of the actual leaf is done.
1911 *
1912 * Some paths want to call this instead of allocating a path structure
1913 * and calling ocfs2_find_path().
1914 *
1915 * This function doesn't handle non btree extent lists.
1916 */
1917int ocfs2_find_leaf(struct ocfs2_caching_info *ci,
1918 struct ocfs2_extent_list *root_el, u32 cpos,
1919 struct buffer_head **leaf_bh)
1920{
1921 int ret;
1922 struct buffer_head *bh = NULL;
1923
1924 ret = __ocfs2_find_path(ci, root_el, cpos, find_leaf_ins, &bh);
1925 if (ret) {
1926 mlog_errno(ret);
1927 goto out;
1928 }
1929
1930 *leaf_bh = bh;
1931out:
1932 return ret;
1933}
1934
1935/*
1936 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1937 *
1938 * Basically, we've moved stuff around at the bottom of the tree and
1939 * we need to fix up the extent records above the changes to reflect
1940 * the new changes.
1941 *
1942 * left_rec: the record on the left.
1943 * left_child_el: is the child list pointed to by left_rec
1944 * right_rec: the record to the right of left_rec
1945 * right_child_el: is the child list pointed to by right_rec
1946 *
1947 * By definition, this only works on interior nodes.
1948 */
1949static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1950 struct ocfs2_extent_list *left_child_el,
1951 struct ocfs2_extent_rec *right_rec,
1952 struct ocfs2_extent_list *right_child_el)
1953{
1954 u32 left_clusters, right_end;
1955
1956 /*
1957 * Interior nodes never have holes. Their cpos is the cpos of
1958 * the leftmost record in their child list. Their cluster
1959 * count covers the full theoretical range of their child list
1960 * - the range between their cpos and the cpos of the record
1961 * immediately to their right.
1962 */
1963 left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1964 if (!ocfs2_rec_clusters(right_child_el, &right_child_el->l_recs[0])) {
1965 BUG_ON(right_child_el->l_tree_depth);
1966 BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
1967 left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
1968 }
1969 left_clusters -= le32_to_cpu(left_rec->e_cpos);
1970 left_rec->e_int_clusters = cpu_to_le32(left_clusters);
1971
1972 /*
1973 * Calculate the rightmost cluster count boundary before
1974 * moving cpos - we will need to adjust clusters after
1975 * updating e_cpos to keep the same highest cluster count.
1976 */
1977 right_end = le32_to_cpu(right_rec->e_cpos);
1978 right_end += le32_to_cpu(right_rec->e_int_clusters);
1979
1980 right_rec->e_cpos = left_rec->e_cpos;
1981 le32_add_cpu(&right_rec->e_cpos, left_clusters);
1982
1983 right_end -= le32_to_cpu(right_rec->e_cpos);
1984 right_rec->e_int_clusters = cpu_to_le32(right_end);
1985}
1986
1987/*
1988 * Adjust the adjacent root node records involved in a
1989 * rotation. left_el_blkno is passed in as a key so that we can easily
1990 * find it's index in the root list.
1991 */
1992static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
1993 struct ocfs2_extent_list *left_el,
1994 struct ocfs2_extent_list *right_el,
1995 u64 left_el_blkno)
1996{
1997 int i;
1998
1999 BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
2000 le16_to_cpu(left_el->l_tree_depth));
2001
2002 for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
2003 if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
2004 break;
2005 }
2006
2007 /*
2008 * The path walking code should have never returned a root and
2009 * two paths which are not adjacent.
2010 */
2011 BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
2012
2013 ocfs2_adjust_adjacent_records(&root_el->l_recs[i], left_el,
2014 &root_el->l_recs[i + 1], right_el);
2015}
2016
2017/*
2018 * We've changed a leaf block (in right_path) and need to reflect that
2019 * change back up the subtree.
2020 *
2021 * This happens in multiple places:
2022 * - When we've moved an extent record from the left path leaf to the right
2023 * path leaf to make room for an empty extent in the left path leaf.
2024 * - When our insert into the right path leaf is at the leftmost edge
2025 * and requires an update of the path immediately to it's left. This
2026 * can occur at the end of some types of rotation and appending inserts.
2027 * - When we've adjusted the last extent record in the left path leaf and the
2028 * 1st extent record in the right path leaf during cross extent block merge.
2029 */
2030static void ocfs2_complete_edge_insert(handle_t *handle,
2031 struct ocfs2_path *left_path,
2032 struct ocfs2_path *right_path,
2033 int subtree_index)
2034{
2035 int i, idx;
2036 struct ocfs2_extent_list *el, *left_el, *right_el;
2037 struct ocfs2_extent_rec *left_rec, *right_rec;
2038 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2039
2040 /*
2041 * Update the counts and position values within all the
2042 * interior nodes to reflect the leaf rotation we just did.
2043 *
2044 * The root node is handled below the loop.
2045 *
2046 * We begin the loop with right_el and left_el pointing to the
2047 * leaf lists and work our way up.
2048 *
2049 * NOTE: within this loop, left_el and right_el always refer
2050 * to the *child* lists.
2051 */
2052 left_el = path_leaf_el(left_path);
2053 right_el = path_leaf_el(right_path);
2054 for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
2055 trace_ocfs2_complete_edge_insert(i);
2056
2057 /*
2058 * One nice property of knowing that all of these
2059 * nodes are below the root is that we only deal with
2060 * the leftmost right node record and the rightmost
2061 * left node record.
2062 */
2063 el = left_path->p_node[i].el;
2064 idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
2065 left_rec = &el->l_recs[idx];
2066
2067 el = right_path->p_node[i].el;
2068 right_rec = &el->l_recs[0];
2069
2070 ocfs2_adjust_adjacent_records(left_rec, left_el, right_rec,
2071 right_el);
2072
2073 ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
2074 ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
2075
2076 /*
2077 * Setup our list pointers now so that the current
2078 * parents become children in the next iteration.
2079 */
2080 left_el = left_path->p_node[i].el;
2081 right_el = right_path->p_node[i].el;
2082 }
2083
2084 /*
2085 * At the root node, adjust the two adjacent records which
2086 * begin our path to the leaves.
2087 */
2088
2089 el = left_path->p_node[subtree_index].el;
2090 left_el = left_path->p_node[subtree_index + 1].el;
2091 right_el = right_path->p_node[subtree_index + 1].el;
2092
2093 ocfs2_adjust_root_records(el, left_el, right_el,
2094 left_path->p_node[subtree_index + 1].bh->b_blocknr);
2095
2096 root_bh = left_path->p_node[subtree_index].bh;
2097
2098 ocfs2_journal_dirty(handle, root_bh);
2099}
2100
2101static int ocfs2_rotate_subtree_right(handle_t *handle,
2102 struct ocfs2_extent_tree *et,
2103 struct ocfs2_path *left_path,
2104 struct ocfs2_path *right_path,
2105 int subtree_index)
2106{
2107 int ret, i;
2108 struct buffer_head *right_leaf_bh;
2109 struct buffer_head *left_leaf_bh = NULL;
2110 struct buffer_head *root_bh;
2111 struct ocfs2_extent_list *right_el, *left_el;
2112 struct ocfs2_extent_rec move_rec;
2113
2114 left_leaf_bh = path_leaf_bh(left_path);
2115 left_el = path_leaf_el(left_path);
2116
2117 if (left_el->l_next_free_rec != left_el->l_count) {
2118 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
2119 "Inode %llu has non-full interior leaf node %llu"
2120 "(next free = %u)",
2121 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2122 (unsigned long long)left_leaf_bh->b_blocknr,
2123 le16_to_cpu(left_el->l_next_free_rec));
2124 return -EROFS;
2125 }
2126
2127 /*
2128 * This extent block may already have an empty record, so we
2129 * return early if so.
2130 */
2131 if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
2132 return 0;
2133
2134 root_bh = left_path->p_node[subtree_index].bh;
2135 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2136
2137 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2138 subtree_index);
2139 if (ret) {
2140 mlog_errno(ret);
2141 goto out;
2142 }
2143
2144 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2145 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2146 right_path, i);
2147 if (ret) {
2148 mlog_errno(ret);
2149 goto out;
2150 }
2151
2152 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2153 left_path, i);
2154 if (ret) {
2155 mlog_errno(ret);
2156 goto out;
2157 }
2158 }
2159
2160 right_leaf_bh = path_leaf_bh(right_path);
2161 right_el = path_leaf_el(right_path);
2162
2163 /* This is a code error, not a disk corruption. */
2164 mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
2165 "because rightmost leaf block %llu is empty\n",
2166 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2167 (unsigned long long)right_leaf_bh->b_blocknr);
2168
2169 ocfs2_create_empty_extent(right_el);
2170
2171 ocfs2_journal_dirty(handle, right_leaf_bh);
2172
2173 /* Do the copy now. */
2174 i = le16_to_cpu(left_el->l_next_free_rec) - 1;
2175 move_rec = left_el->l_recs[i];
2176 right_el->l_recs[0] = move_rec;
2177
2178 /*
2179 * Clear out the record we just copied and shift everything
2180 * over, leaving an empty extent in the left leaf.
2181 *
2182 * We temporarily subtract from next_free_rec so that the
2183 * shift will lose the tail record (which is now defunct).
2184 */
2185 le16_add_cpu(&left_el->l_next_free_rec, -1);
2186 ocfs2_shift_records_right(left_el);
2187 memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2188 le16_add_cpu(&left_el->l_next_free_rec, 1);
2189
2190 ocfs2_journal_dirty(handle, left_leaf_bh);
2191
2192 ocfs2_complete_edge_insert(handle, left_path, right_path,
2193 subtree_index);
2194
2195out:
2196 return ret;
2197}
2198
2199/*
2200 * Given a full path, determine what cpos value would return us a path
2201 * containing the leaf immediately to the left of the current one.
2202 *
2203 * Will return zero if the path passed in is already the leftmost path.
2204 */
2205int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
2206 struct ocfs2_path *path, u32 *cpos)
2207{
2208 int i, j, ret = 0;
2209 u64 blkno;
2210 struct ocfs2_extent_list *el;
2211
2212 BUG_ON(path->p_tree_depth == 0);
2213
2214 *cpos = 0;
2215
2216 blkno = path_leaf_bh(path)->b_blocknr;
2217
2218 /* Start at the tree node just above the leaf and work our way up. */
2219 i = path->p_tree_depth - 1;
2220 while (i >= 0) {
2221 el = path->p_node[i].el;
2222
2223 /*
2224 * Find the extent record just before the one in our
2225 * path.
2226 */
2227 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2228 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2229 if (j == 0) {
2230 if (i == 0) {
2231 /*
2232 * We've determined that the
2233 * path specified is already
2234 * the leftmost one - return a
2235 * cpos of zero.
2236 */
2237 goto out;
2238 }
2239 /*
2240 * The leftmost record points to our
2241 * leaf - we need to travel up the
2242 * tree one level.
2243 */
2244 goto next_node;
2245 }
2246
2247 *cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
2248 *cpos = *cpos + ocfs2_rec_clusters(el,
2249 &el->l_recs[j - 1]);
2250 *cpos = *cpos - 1;
2251 goto out;
2252 }
2253 }
2254
2255 /*
2256 * If we got here, we never found a valid node where
2257 * the tree indicated one should be.
2258 */
2259 ocfs2_error(sb,
2260 "Invalid extent tree at extent block %llu\n",
2261 (unsigned long long)blkno);
2262 ret = -EROFS;
2263 goto out;
2264
2265next_node:
2266 blkno = path->p_node[i].bh->b_blocknr;
2267 i--;
2268 }
2269
2270out:
2271 return ret;
2272}
2273
2274/*
2275 * Extend the transaction by enough credits to complete the rotation,
2276 * and still leave at least the original number of credits allocated
2277 * to this transaction.
2278 */
2279static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
2280 int op_credits,
2281 struct ocfs2_path *path)
2282{
2283 int ret = 0;
2284 int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;
2285
2286 if (handle->h_buffer_credits < credits)
2287 ret = ocfs2_extend_trans(handle,
2288 credits - handle->h_buffer_credits);
2289
2290 return ret;
2291}
2292
2293/*
2294 * Trap the case where we're inserting into the theoretical range past
2295 * the _actual_ left leaf range. Otherwise, we'll rotate a record
2296 * whose cpos is less than ours into the right leaf.
2297 *
2298 * It's only necessary to look at the rightmost record of the left
2299 * leaf because the logic that calls us should ensure that the
2300 * theoretical ranges in the path components above the leaves are
2301 * correct.
2302 */
2303static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
2304 u32 insert_cpos)
2305{
2306 struct ocfs2_extent_list *left_el;
2307 struct ocfs2_extent_rec *rec;
2308 int next_free;
2309
2310 left_el = path_leaf_el(left_path);
2311 next_free = le16_to_cpu(left_el->l_next_free_rec);
2312 rec = &left_el->l_recs[next_free - 1];
2313
2314 if (insert_cpos > le32_to_cpu(rec->e_cpos))
2315 return 1;
2316 return 0;
2317}
2318
2319static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
2320{
2321 int next_free = le16_to_cpu(el->l_next_free_rec);
2322 unsigned int range;
2323 struct ocfs2_extent_rec *rec;
2324
2325 if (next_free == 0)
2326 return 0;
2327
2328 rec = &el->l_recs[0];
2329 if (ocfs2_is_empty_extent(rec)) {
2330 /* Empty list. */
2331 if (next_free == 1)
2332 return 0;
2333 rec = &el->l_recs[1];
2334 }
2335
2336 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2337 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
2338 return 1;
2339 return 0;
2340}
2341
2342/*
2343 * Rotate all the records in a btree right one record, starting at insert_cpos.
2344 *
2345 * The path to the rightmost leaf should be passed in.
2346 *
2347 * The array is assumed to be large enough to hold an entire path (tree depth).
2348 *
2349 * Upon successful return from this function:
2350 *
2351 * - The 'right_path' array will contain a path to the leaf block
2352 * whose range contains e_cpos.
2353 * - That leaf block will have a single empty extent in list index 0.
2354 * - In the case that the rotation requires a post-insert update,
2355 * *ret_left_path will contain a valid path which can be passed to
2356 * ocfs2_insert_path().
2357 */
2358static int ocfs2_rotate_tree_right(handle_t *handle,
2359 struct ocfs2_extent_tree *et,
2360 enum ocfs2_split_type split,
2361 u32 insert_cpos,
2362 struct ocfs2_path *right_path,
2363 struct ocfs2_path **ret_left_path)
2364{
2365 int ret, start, orig_credits = handle->h_buffer_credits;
2366 u32 cpos;
2367 struct ocfs2_path *left_path = NULL;
2368 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2369
2370 *ret_left_path = NULL;
2371
2372 left_path = ocfs2_new_path_from_path(right_path);
2373 if (!left_path) {
2374 ret = -ENOMEM;
2375 mlog_errno(ret);
2376 goto out;
2377 }
2378
2379 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2380 if (ret) {
2381 mlog_errno(ret);
2382 goto out;
2383 }
2384
2385 trace_ocfs2_rotate_tree_right(
2386 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2387 insert_cpos, cpos);
2388
2389 /*
2390 * What we want to do here is:
2391 *
2392 * 1) Start with the rightmost path.
2393 *
2394 * 2) Determine a path to the leaf block directly to the left
2395 * of that leaf.
2396 *
2397 * 3) Determine the 'subtree root' - the lowest level tree node
2398 * which contains a path to both leaves.
2399 *
2400 * 4) Rotate the subtree.
2401 *
2402 * 5) Find the next subtree by considering the left path to be
2403 * the new right path.
2404 *
2405 * The check at the top of this while loop also accepts
2406 * insert_cpos == cpos because cpos is only a _theoretical_
2407 * value to get us the left path - insert_cpos might very well
2408 * be filling that hole.
2409 *
2410 * Stop at a cpos of '0' because we either started at the
2411 * leftmost branch (i.e., a tree with one branch and a
2412 * rotation inside of it), or we've gone as far as we can in
2413 * rotating subtrees.
2414 */
2415 while (cpos && insert_cpos <= cpos) {
2416 trace_ocfs2_rotate_tree_right(
2417 (unsigned long long)
2418 ocfs2_metadata_cache_owner(et->et_ci),
2419 insert_cpos, cpos);
2420
2421 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
2422 if (ret) {
2423 mlog_errno(ret);
2424 goto out;
2425 }
2426
2427 mlog_bug_on_msg(path_leaf_bh(left_path) ==
2428 path_leaf_bh(right_path),
2429 "Owner %llu: error during insert of %u "
2430 "(left path cpos %u) results in two identical "
2431 "paths ending at %llu\n",
2432 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2433 insert_cpos, cpos,
2434 (unsigned long long)
2435 path_leaf_bh(left_path)->b_blocknr);
2436
2437 if (split == SPLIT_NONE &&
2438 ocfs2_rotate_requires_path_adjustment(left_path,
2439 insert_cpos)) {
2440
2441 /*
2442 * We've rotated the tree as much as we
2443 * should. The rest is up to
2444 * ocfs2_insert_path() to complete, after the
2445 * record insertion. We indicate this
2446 * situation by returning the left path.
2447 *
2448 * The reason we don't adjust the records here
2449 * before the record insert is that an error
2450 * later might break the rule where a parent
2451 * record e_cpos will reflect the actual
2452 * e_cpos of the 1st nonempty record of the
2453 * child list.
2454 */
2455 *ret_left_path = left_path;
2456 goto out_ret_path;
2457 }
2458
2459 start = ocfs2_find_subtree_root(et, left_path, right_path);
2460
2461 trace_ocfs2_rotate_subtree(start,
2462 (unsigned long long)
2463 right_path->p_node[start].bh->b_blocknr,
2464 right_path->p_tree_depth);
2465
2466 ret = ocfs2_extend_rotate_transaction(handle, start,
2467 orig_credits, right_path);
2468 if (ret) {
2469 mlog_errno(ret);
2470 goto out;
2471 }
2472
2473 ret = ocfs2_rotate_subtree_right(handle, et, left_path,
2474 right_path, start);
2475 if (ret) {
2476 mlog_errno(ret);
2477 goto out;
2478 }
2479
2480 if (split != SPLIT_NONE &&
2481 ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
2482 insert_cpos)) {
2483 /*
2484 * A rotate moves the rightmost left leaf
2485 * record over to the leftmost right leaf
2486 * slot. If we're doing an extent split
2487 * instead of a real insert, then we have to
2488 * check that the extent to be split wasn't
2489 * just moved over. If it was, then we can
2490 * exit here, passing left_path back -
2491 * ocfs2_split_extent() is smart enough to
2492 * search both leaves.
2493 */
2494 *ret_left_path = left_path;
2495 goto out_ret_path;
2496 }
2497
2498 /*
2499 * There is no need to re-read the next right path
2500 * as we know that it'll be our current left
2501 * path. Optimize by copying values instead.
2502 */
2503 ocfs2_mv_path(right_path, left_path);
2504
2505 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2506 if (ret) {
2507 mlog_errno(ret);
2508 goto out;
2509 }
2510 }
2511
2512out:
2513 ocfs2_free_path(left_path);
2514
2515out_ret_path:
2516 return ret;
2517}
2518
2519static int ocfs2_update_edge_lengths(handle_t *handle,
2520 struct ocfs2_extent_tree *et,
2521 int subtree_index, struct ocfs2_path *path)
2522{
2523 int i, idx, ret;
2524 struct ocfs2_extent_rec *rec;
2525 struct ocfs2_extent_list *el;
2526 struct ocfs2_extent_block *eb;
2527 u32 range;
2528
2529 /*
2530 * In normal tree rotation process, we will never touch the
2531 * tree branch above subtree_index and ocfs2_extend_rotate_transaction
2532 * doesn't reserve the credits for them either.
2533 *
2534 * But we do have a special case here which will update the rightmost
2535 * records for all the bh in the path.
2536 * So we have to allocate extra credits and access them.
2537 */
2538 ret = ocfs2_extend_trans(handle, subtree_index);
2539 if (ret) {
2540 mlog_errno(ret);
2541 goto out;
2542 }
2543
2544 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
2545 if (ret) {
2546 mlog_errno(ret);
2547 goto out;
2548 }
2549
2550 /* Path should always be rightmost. */
2551 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2552 BUG_ON(eb->h_next_leaf_blk != 0ULL);
2553
2554 el = &eb->h_list;
2555 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
2556 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2557 rec = &el->l_recs[idx];
2558 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2559
2560 for (i = 0; i < path->p_tree_depth; i++) {
2561 el = path->p_node[i].el;
2562 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2563 rec = &el->l_recs[idx];
2564
2565 rec->e_int_clusters = cpu_to_le32(range);
2566 le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));
2567
2568 ocfs2_journal_dirty(handle, path->p_node[i].bh);
2569 }
2570out:
2571 return ret;
2572}
2573
2574static void ocfs2_unlink_path(handle_t *handle,
2575 struct ocfs2_extent_tree *et,
2576 struct ocfs2_cached_dealloc_ctxt *dealloc,
2577 struct ocfs2_path *path, int unlink_start)
2578{
2579 int ret, i;
2580 struct ocfs2_extent_block *eb;
2581 struct ocfs2_extent_list *el;
2582 struct buffer_head *bh;
2583
2584 for(i = unlink_start; i < path_num_items(path); i++) {
2585 bh = path->p_node[i].bh;
2586
2587 eb = (struct ocfs2_extent_block *)bh->b_data;
2588 /*
2589 * Not all nodes might have had their final count
2590 * decremented by the caller - handle this here.
2591 */
2592 el = &eb->h_list;
2593 if (le16_to_cpu(el->l_next_free_rec) > 1) {
2594 mlog(ML_ERROR,
2595 "Inode %llu, attempted to remove extent block "
2596 "%llu with %u records\n",
2597 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2598 (unsigned long long)le64_to_cpu(eb->h_blkno),
2599 le16_to_cpu(el->l_next_free_rec));
2600
2601 ocfs2_journal_dirty(handle, bh);
2602 ocfs2_remove_from_cache(et->et_ci, bh);
2603 continue;
2604 }
2605
2606 el->l_next_free_rec = 0;
2607 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2608
2609 ocfs2_journal_dirty(handle, bh);
2610
2611 ret = ocfs2_cache_extent_block_free(dealloc, eb);
2612 if (ret)
2613 mlog_errno(ret);
2614
2615 ocfs2_remove_from_cache(et->et_ci, bh);
2616 }
2617}
2618
2619static void ocfs2_unlink_subtree(handle_t *handle,
2620 struct ocfs2_extent_tree *et,
2621 struct ocfs2_path *left_path,
2622 struct ocfs2_path *right_path,
2623 int subtree_index,
2624 struct ocfs2_cached_dealloc_ctxt *dealloc)
2625{
2626 int i;
2627 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2628 struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2629 struct ocfs2_extent_list *el;
2630 struct ocfs2_extent_block *eb;
2631
2632 el = path_leaf_el(left_path);
2633
2634 eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2635
2636 for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2637 if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2638 break;
2639
2640 BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2641
2642 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2643 le16_add_cpu(&root_el->l_next_free_rec, -1);
2644
2645 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2646 eb->h_next_leaf_blk = 0;
2647
2648 ocfs2_journal_dirty(handle, root_bh);
2649 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2650
2651 ocfs2_unlink_path(handle, et, dealloc, right_path,
2652 subtree_index + 1);
2653}
2654
2655static int ocfs2_rotate_subtree_left(handle_t *handle,
2656 struct ocfs2_extent_tree *et,
2657 struct ocfs2_path *left_path,
2658 struct ocfs2_path *right_path,
2659 int subtree_index,
2660 struct ocfs2_cached_dealloc_ctxt *dealloc,
2661 int *deleted)
2662{
2663 int ret, i, del_right_subtree = 0, right_has_empty = 0;
2664 struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path);
2665 struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2666 struct ocfs2_extent_block *eb;
2667
2668 *deleted = 0;
2669
2670 right_leaf_el = path_leaf_el(right_path);
2671 left_leaf_el = path_leaf_el(left_path);
2672 root_bh = left_path->p_node[subtree_index].bh;
2673 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2674
2675 if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2676 return 0;
2677
2678 eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2679 if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2680 /*
2681 * It's legal for us to proceed if the right leaf is
2682 * the rightmost one and it has an empty extent. There
2683 * are two cases to handle - whether the leaf will be
2684 * empty after removal or not. If the leaf isn't empty
2685 * then just remove the empty extent up front. The
2686 * next block will handle empty leaves by flagging
2687 * them for unlink.
2688 *
2689 * Non rightmost leaves will throw -EAGAIN and the
2690 * caller can manually move the subtree and retry.
2691 */
2692
2693 if (eb->h_next_leaf_blk != 0ULL)
2694 return -EAGAIN;
2695
2696 if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2697 ret = ocfs2_journal_access_eb(handle, et->et_ci,
2698 path_leaf_bh(right_path),
2699 OCFS2_JOURNAL_ACCESS_WRITE);
2700 if (ret) {
2701 mlog_errno(ret);
2702 goto out;
2703 }
2704
2705 ocfs2_remove_empty_extent(right_leaf_el);
2706 } else
2707 right_has_empty = 1;
2708 }
2709
2710 if (eb->h_next_leaf_blk == 0ULL &&
2711 le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2712 /*
2713 * We have to update i_last_eb_blk during the meta
2714 * data delete.
2715 */
2716 ret = ocfs2_et_root_journal_access(handle, et,
2717 OCFS2_JOURNAL_ACCESS_WRITE);
2718 if (ret) {
2719 mlog_errno(ret);
2720 goto out;
2721 }
2722
2723 del_right_subtree = 1;
2724 }
2725
2726 /*
2727 * Getting here with an empty extent in the right path implies
2728 * that it's the rightmost path and will be deleted.
2729 */
2730 BUG_ON(right_has_empty && !del_right_subtree);
2731
2732 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2733 subtree_index);
2734 if (ret) {
2735 mlog_errno(ret);
2736 goto out;
2737 }
2738
2739 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2740 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2741 right_path, i);
2742 if (ret) {
2743 mlog_errno(ret);
2744 goto out;
2745 }
2746
2747 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2748 left_path, i);
2749 if (ret) {
2750 mlog_errno(ret);
2751 goto out;
2752 }
2753 }
2754
2755 if (!right_has_empty) {
2756 /*
2757 * Only do this if we're moving a real
2758 * record. Otherwise, the action is delayed until
2759 * after removal of the right path in which case we
2760 * can do a simple shift to remove the empty extent.
2761 */
2762 ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2763 memset(&right_leaf_el->l_recs[0], 0,
2764 sizeof(struct ocfs2_extent_rec));
2765 }
2766 if (eb->h_next_leaf_blk == 0ULL) {
2767 /*
2768 * Move recs over to get rid of empty extent, decrease
2769 * next_free. This is allowed to remove the last
2770 * extent in our leaf (setting l_next_free_rec to
2771 * zero) - the delete code below won't care.
2772 */
2773 ocfs2_remove_empty_extent(right_leaf_el);
2774 }
2775
2776 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2777 ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2778
2779 if (del_right_subtree) {
2780 ocfs2_unlink_subtree(handle, et, left_path, right_path,
2781 subtree_index, dealloc);
2782 ret = ocfs2_update_edge_lengths(handle, et, subtree_index,
2783 left_path);
2784 if (ret) {
2785 mlog_errno(ret);
2786 goto out;
2787 }
2788
2789 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2790 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2791
2792 /*
2793 * Removal of the extent in the left leaf was skipped
2794 * above so we could delete the right path
2795 * 1st.
2796 */
2797 if (right_has_empty)
2798 ocfs2_remove_empty_extent(left_leaf_el);
2799
2800 ocfs2_journal_dirty(handle, et_root_bh);
2801
2802 *deleted = 1;
2803 } else
2804 ocfs2_complete_edge_insert(handle, left_path, right_path,
2805 subtree_index);
2806
2807out:
2808 return ret;
2809}
2810
2811/*
2812 * Given a full path, determine what cpos value would return us a path
2813 * containing the leaf immediately to the right of the current one.
2814 *
2815 * Will return zero if the path passed in is already the rightmost path.
2816 *
2817 * This looks similar, but is subtly different to
2818 * ocfs2_find_cpos_for_left_leaf().
2819 */
2820int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
2821 struct ocfs2_path *path, u32 *cpos)
2822{
2823 int i, j, ret = 0;
2824 u64 blkno;
2825 struct ocfs2_extent_list *el;
2826
2827 *cpos = 0;
2828
2829 if (path->p_tree_depth == 0)
2830 return 0;
2831
2832 blkno = path_leaf_bh(path)->b_blocknr;
2833
2834 /* Start at the tree node just above the leaf and work our way up. */
2835 i = path->p_tree_depth - 1;
2836 while (i >= 0) {
2837 int next_free;
2838
2839 el = path->p_node[i].el;
2840
2841 /*
2842 * Find the extent record just after the one in our
2843 * path.
2844 */
2845 next_free = le16_to_cpu(el->l_next_free_rec);
2846 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2847 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2848 if (j == (next_free - 1)) {
2849 if (i == 0) {
2850 /*
2851 * We've determined that the
2852 * path specified is already
2853 * the rightmost one - return a
2854 * cpos of zero.
2855 */
2856 goto out;
2857 }
2858 /*
2859 * The rightmost record points to our
2860 * leaf - we need to travel up the
2861 * tree one level.
2862 */
2863 goto next_node;
2864 }
2865
2866 *cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
2867 goto out;
2868 }
2869 }
2870
2871 /*
2872 * If we got here, we never found a valid node where
2873 * the tree indicated one should be.
2874 */
2875 ocfs2_error(sb,
2876 "Invalid extent tree at extent block %llu\n",
2877 (unsigned long long)blkno);
2878 ret = -EROFS;
2879 goto out;
2880
2881next_node:
2882 blkno = path->p_node[i].bh->b_blocknr;
2883 i--;
2884 }
2885
2886out:
2887 return ret;
2888}
2889
2890static int ocfs2_rotate_rightmost_leaf_left(handle_t *handle,
2891 struct ocfs2_extent_tree *et,
2892 struct ocfs2_path *path)
2893{
2894 int ret;
2895 struct buffer_head *bh = path_leaf_bh(path);
2896 struct ocfs2_extent_list *el = path_leaf_el(path);
2897
2898 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2899 return 0;
2900
2901 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
2902 path_num_items(path) - 1);
2903 if (ret) {
2904 mlog_errno(ret);
2905 goto out;
2906 }
2907
2908 ocfs2_remove_empty_extent(el);
2909 ocfs2_journal_dirty(handle, bh);
2910
2911out:
2912 return ret;
2913}
2914
2915static int __ocfs2_rotate_tree_left(handle_t *handle,
2916 struct ocfs2_extent_tree *et,
2917 int orig_credits,
2918 struct ocfs2_path *path,
2919 struct ocfs2_cached_dealloc_ctxt *dealloc,
2920 struct ocfs2_path **empty_extent_path)
2921{
2922 int ret, subtree_root, deleted;
2923 u32 right_cpos;
2924 struct ocfs2_path *left_path = NULL;
2925 struct ocfs2_path *right_path = NULL;
2926 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2927
2928 BUG_ON(!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])));
2929
2930 *empty_extent_path = NULL;
2931
2932 ret = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
2933 if (ret) {
2934 mlog_errno(ret);
2935 goto out;
2936 }
2937
2938 left_path = ocfs2_new_path_from_path(path);
2939 if (!left_path) {
2940 ret = -ENOMEM;
2941 mlog_errno(ret);
2942 goto out;
2943 }
2944
2945 ocfs2_cp_path(left_path, path);
2946
2947 right_path = ocfs2_new_path_from_path(path);
2948 if (!right_path) {
2949 ret = -ENOMEM;
2950 mlog_errno(ret);
2951 goto out;
2952 }
2953
2954 while (right_cpos) {
2955 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
2956 if (ret) {
2957 mlog_errno(ret);
2958 goto out;
2959 }
2960
2961 subtree_root = ocfs2_find_subtree_root(et, left_path,
2962 right_path);
2963
2964 trace_ocfs2_rotate_subtree(subtree_root,
2965 (unsigned long long)
2966 right_path->p_node[subtree_root].bh->b_blocknr,
2967 right_path->p_tree_depth);
2968
2969 ret = ocfs2_extend_rotate_transaction(handle, subtree_root,
2970 orig_credits, left_path);
2971 if (ret) {
2972 mlog_errno(ret);
2973 goto out;
2974 }
2975
2976 /*
2977 * Caller might still want to make changes to the
2978 * tree root, so re-add it to the journal here.
2979 */
2980 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2981 left_path, 0);
2982 if (ret) {
2983 mlog_errno(ret);
2984 goto out;
2985 }
2986
2987 ret = ocfs2_rotate_subtree_left(handle, et, left_path,
2988 right_path, subtree_root,
2989 dealloc, &deleted);
2990 if (ret == -EAGAIN) {
2991 /*
2992 * The rotation has to temporarily stop due to
2993 * the right subtree having an empty
2994 * extent. Pass it back to the caller for a
2995 * fixup.
2996 */
2997 *empty_extent_path = right_path;
2998 right_path = NULL;
2999 goto out;
3000 }
3001 if (ret) {
3002 mlog_errno(ret);
3003 goto out;
3004 }
3005
3006 /*
3007 * The subtree rotate might have removed records on
3008 * the rightmost edge. If so, then rotation is
3009 * complete.
3010 */
3011 if (deleted)
3012 break;
3013
3014 ocfs2_mv_path(left_path, right_path);
3015
3016 ret = ocfs2_find_cpos_for_right_leaf(sb, left_path,
3017 &right_cpos);
3018 if (ret) {
3019 mlog_errno(ret);
3020 goto out;
3021 }
3022 }
3023
3024out:
3025 ocfs2_free_path(right_path);
3026 ocfs2_free_path(left_path);
3027
3028 return ret;
3029}
3030
3031static int ocfs2_remove_rightmost_path(handle_t *handle,
3032 struct ocfs2_extent_tree *et,
3033 struct ocfs2_path *path,
3034 struct ocfs2_cached_dealloc_ctxt *dealloc)
3035{
3036 int ret, subtree_index;
3037 u32 cpos;
3038 struct ocfs2_path *left_path = NULL;
3039 struct ocfs2_extent_block *eb;
3040 struct ocfs2_extent_list *el;
3041
3042
3043 ret = ocfs2_et_sanity_check(et);
3044 if (ret)
3045 goto out;
3046 /*
3047 * There's two ways we handle this depending on
3048 * whether path is the only existing one.
3049 */
3050 ret = ocfs2_extend_rotate_transaction(handle, 0,
3051 handle->h_buffer_credits,
3052 path);
3053 if (ret) {
3054 mlog_errno(ret);
3055 goto out;
3056 }
3057
3058 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
3059 if (ret) {
3060 mlog_errno(ret);
3061 goto out;
3062 }
3063
3064 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3065 path, &cpos);
3066 if (ret) {
3067 mlog_errno(ret);
3068 goto out;
3069 }
3070
3071 if (cpos) {
3072 /*
3073 * We have a path to the left of this one - it needs
3074 * an update too.
3075 */
3076 left_path = ocfs2_new_path_from_path(path);
3077 if (!left_path) {
3078 ret = -ENOMEM;
3079 mlog_errno(ret);
3080 goto out;
3081 }
3082
3083 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
3084 if (ret) {
3085 mlog_errno(ret);
3086 goto out;
3087 }
3088
3089 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
3090 if (ret) {
3091 mlog_errno(ret);
3092 goto out;
3093 }
3094
3095 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
3096
3097 ocfs2_unlink_subtree(handle, et, left_path, path,
3098 subtree_index, dealloc);
3099 ret = ocfs2_update_edge_lengths(handle, et, subtree_index,
3100 left_path);
3101 if (ret) {
3102 mlog_errno(ret);
3103 goto out;
3104 }
3105
3106 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
3107 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
3108 } else {
3109 /*
3110 * 'path' is also the leftmost path which
3111 * means it must be the only one. This gets
3112 * handled differently because we want to
3113 * revert the root back to having extents
3114 * in-line.
3115 */
3116 ocfs2_unlink_path(handle, et, dealloc, path, 1);
3117
3118 el = et->et_root_el;
3119 el->l_tree_depth = 0;
3120 el->l_next_free_rec = 0;
3121 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3122
3123 ocfs2_et_set_last_eb_blk(et, 0);
3124 }
3125
3126 ocfs2_journal_dirty(handle, path_root_bh(path));
3127
3128out:
3129 ocfs2_free_path(left_path);
3130 return ret;
3131}
3132
3133/*
3134 * Left rotation of btree records.
3135 *
3136 * In many ways, this is (unsurprisingly) the opposite of right
3137 * rotation. We start at some non-rightmost path containing an empty
3138 * extent in the leaf block. The code works its way to the rightmost
3139 * path by rotating records to the left in every subtree.
3140 *
3141 * This is used by any code which reduces the number of extent records
3142 * in a leaf. After removal, an empty record should be placed in the
3143 * leftmost list position.
3144 *
3145 * This won't handle a length update of the rightmost path records if
3146 * the rightmost tree leaf record is removed so the caller is
3147 * responsible for detecting and correcting that.
3148 */
3149static int ocfs2_rotate_tree_left(handle_t *handle,
3150 struct ocfs2_extent_tree *et,
3151 struct ocfs2_path *path,
3152 struct ocfs2_cached_dealloc_ctxt *dealloc)
3153{
3154 int ret, orig_credits = handle->h_buffer_credits;
3155 struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
3156 struct ocfs2_extent_block *eb;
3157 struct ocfs2_extent_list *el;
3158
3159 el = path_leaf_el(path);
3160 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
3161 return 0;
3162
3163 if (path->p_tree_depth == 0) {
3164rightmost_no_delete:
3165 /*
3166 * Inline extents. This is trivially handled, so do
3167 * it up front.
3168 */
3169 ret = ocfs2_rotate_rightmost_leaf_left(handle, et, path);
3170 if (ret)
3171 mlog_errno(ret);
3172 goto out;
3173 }
3174
3175 /*
3176 * Handle rightmost branch now. There's several cases:
3177 * 1) simple rotation leaving records in there. That's trivial.
3178 * 2) rotation requiring a branch delete - there's no more
3179 * records left. Two cases of this:
3180 * a) There are branches to the left.
3181 * b) This is also the leftmost (the only) branch.
3182 *
3183 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
3184 * 2a) we need the left branch so that we can update it with the unlink
3185 * 2b) we need to bring the root back to inline extents.
3186 */
3187
3188 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
3189 el = &eb->h_list;
3190 if (eb->h_next_leaf_blk == 0) {
3191 /*
3192 * This gets a bit tricky if we're going to delete the
3193 * rightmost path. Get the other cases out of the way
3194 * 1st.
3195 */
3196 if (le16_to_cpu(el->l_next_free_rec) > 1)
3197 goto rightmost_no_delete;
3198
3199 if (le16_to_cpu(el->l_next_free_rec) == 0) {
3200 ret = -EIO;
3201 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3202 "Owner %llu has empty extent block at %llu",
3203 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
3204 (unsigned long long)le64_to_cpu(eb->h_blkno));
3205 goto out;
3206 }
3207
3208 /*
3209 * XXX: The caller can not trust "path" any more after
3210 * this as it will have been deleted. What do we do?
3211 *
3212 * In theory the rotate-for-merge code will never get
3213 * here because it'll always ask for a rotate in a
3214 * nonempty list.
3215 */
3216
3217 ret = ocfs2_remove_rightmost_path(handle, et, path,
3218 dealloc);
3219 if (ret)
3220 mlog_errno(ret);
3221 goto out;
3222 }
3223
3224 /*
3225 * Now we can loop, remembering the path we get from -EAGAIN
3226 * and restarting from there.
3227 */
3228try_rotate:
3229 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits, path,
3230 dealloc, &restart_path);
3231 if (ret && ret != -EAGAIN) {
3232 mlog_errno(ret);
3233 goto out;
3234 }
3235
3236 while (ret == -EAGAIN) {
3237 tmp_path = restart_path;
3238 restart_path = NULL;
3239
3240 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits,
3241 tmp_path, dealloc,
3242 &restart_path);
3243 if (ret && ret != -EAGAIN) {
3244 mlog_errno(ret);
3245 goto out;
3246 }
3247
3248 ocfs2_free_path(tmp_path);
3249 tmp_path = NULL;
3250
3251 if (ret == 0)
3252 goto try_rotate;
3253 }
3254
3255out:
3256 ocfs2_free_path(tmp_path);
3257 ocfs2_free_path(restart_path);
3258 return ret;
3259}
3260
3261static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
3262 int index)
3263{
3264 struct ocfs2_extent_rec *rec = &el->l_recs[index];
3265 unsigned int size;
3266
3267 if (rec->e_leaf_clusters == 0) {
3268 /*
3269 * We consumed all of the merged-from record. An empty
3270 * extent cannot exist anywhere but the 1st array
3271 * position, so move things over if the merged-from
3272 * record doesn't occupy that position.
3273 *
3274 * This creates a new empty extent so the caller
3275 * should be smart enough to have removed any existing
3276 * ones.
3277 */
3278 if (index > 0) {
3279 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3280 size = index * sizeof(struct ocfs2_extent_rec);
3281 memmove(&el->l_recs[1], &el->l_recs[0], size);
3282 }
3283
3284 /*
3285 * Always memset - the caller doesn't check whether it
3286 * created an empty extent, so there could be junk in
3287 * the other fields.
3288 */
3289 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3290 }
3291}
3292
3293static int ocfs2_get_right_path(struct ocfs2_extent_tree *et,
3294 struct ocfs2_path *left_path,
3295 struct ocfs2_path **ret_right_path)
3296{
3297 int ret;
3298 u32 right_cpos;
3299 struct ocfs2_path *right_path = NULL;
3300 struct ocfs2_extent_list *left_el;
3301
3302 *ret_right_path = NULL;
3303
3304 /* This function shouldn't be called for non-trees. */
3305 BUG_ON(left_path->p_tree_depth == 0);
3306
3307 left_el = path_leaf_el(left_path);
3308 BUG_ON(left_el->l_next_free_rec != left_el->l_count);
3309
3310 ret = ocfs2_find_cpos_for_right_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3311 left_path, &right_cpos);
3312 if (ret) {
3313 mlog_errno(ret);
3314 goto out;
3315 }
3316
3317 /* This function shouldn't be called for the rightmost leaf. */
3318 BUG_ON(right_cpos == 0);
3319
3320 right_path = ocfs2_new_path_from_path(left_path);
3321 if (!right_path) {
3322 ret = -ENOMEM;
3323 mlog_errno(ret);
3324 goto out;
3325 }
3326
3327 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
3328 if (ret) {
3329 mlog_errno(ret);
3330 goto out;
3331 }
3332
3333 *ret_right_path = right_path;
3334out:
3335 if (ret)
3336 ocfs2_free_path(right_path);
3337 return ret;
3338}
3339
3340/*
3341 * Remove split_rec clusters from the record at index and merge them
3342 * onto the beginning of the record "next" to it.
3343 * For index < l_count - 1, the next means the extent rec at index + 1.
3344 * For index == l_count - 1, the "next" means the 1st extent rec of the
3345 * next extent block.
3346 */
3347static int ocfs2_merge_rec_right(struct ocfs2_path *left_path,
3348 handle_t *handle,
3349 struct ocfs2_extent_tree *et,
3350 struct ocfs2_extent_rec *split_rec,
3351 int index)
3352{
3353 int ret, next_free, i;
3354 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3355 struct ocfs2_extent_rec *left_rec;
3356 struct ocfs2_extent_rec *right_rec;
3357 struct ocfs2_extent_list *right_el;
3358 struct ocfs2_path *right_path = NULL;
3359 int subtree_index = 0;
3360 struct ocfs2_extent_list *el = path_leaf_el(left_path);
3361 struct buffer_head *bh = path_leaf_bh(left_path);
3362 struct buffer_head *root_bh = NULL;
3363
3364 BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
3365 left_rec = &el->l_recs[index];
3366
3367 if (index == le16_to_cpu(el->l_next_free_rec) - 1 &&
3368 le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count)) {
3369 /* we meet with a cross extent block merge. */
3370 ret = ocfs2_get_right_path(et, left_path, &right_path);
3371 if (ret) {
3372 mlog_errno(ret);
3373 goto out;
3374 }
3375
3376 right_el = path_leaf_el(right_path);
3377 next_free = le16_to_cpu(right_el->l_next_free_rec);
3378 BUG_ON(next_free <= 0);
3379 right_rec = &right_el->l_recs[0];
3380 if (ocfs2_is_empty_extent(right_rec)) {
3381 BUG_ON(next_free <= 1);
3382 right_rec = &right_el->l_recs[1];
3383 }
3384
3385 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3386 le16_to_cpu(left_rec->e_leaf_clusters) !=
3387 le32_to_cpu(right_rec->e_cpos));
3388
3389 subtree_index = ocfs2_find_subtree_root(et, left_path,
3390 right_path);
3391
3392 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3393 handle->h_buffer_credits,
3394 right_path);
3395 if (ret) {
3396 mlog_errno(ret);
3397 goto out;
3398 }
3399
3400 root_bh = left_path->p_node[subtree_index].bh;
3401 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3402
3403 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3404 subtree_index);
3405 if (ret) {
3406 mlog_errno(ret);
3407 goto out;
3408 }
3409
3410 for (i = subtree_index + 1;
3411 i < path_num_items(right_path); i++) {
3412 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3413 right_path, i);
3414 if (ret) {
3415 mlog_errno(ret);
3416 goto out;
3417 }
3418
3419 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3420 left_path, i);
3421 if (ret) {
3422 mlog_errno(ret);
3423 goto out;
3424 }
3425 }
3426
3427 } else {
3428 BUG_ON(index == le16_to_cpu(el->l_next_free_rec) - 1);
3429 right_rec = &el->l_recs[index + 1];
3430 }
3431
3432 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, left_path,
3433 path_num_items(left_path) - 1);
3434 if (ret) {
3435 mlog_errno(ret);
3436 goto out;
3437 }
3438
3439 le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
3440
3441 le32_add_cpu(&right_rec->e_cpos, -split_clusters);
3442 le64_add_cpu(&right_rec->e_blkno,
3443 -ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3444 split_clusters));
3445 le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
3446
3447 ocfs2_cleanup_merge(el, index);
3448
3449 ocfs2_journal_dirty(handle, bh);
3450 if (right_path) {
3451 ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
3452 ocfs2_complete_edge_insert(handle, left_path, right_path,
3453 subtree_index);
3454 }
3455out:
3456 if (right_path)
3457 ocfs2_free_path(right_path);
3458 return ret;
3459}
3460
3461static int ocfs2_get_left_path(struct ocfs2_extent_tree *et,
3462 struct ocfs2_path *right_path,
3463 struct ocfs2_path **ret_left_path)
3464{
3465 int ret;
3466 u32 left_cpos;
3467 struct ocfs2_path *left_path = NULL;
3468
3469 *ret_left_path = NULL;
3470
3471 /* This function shouldn't be called for non-trees. */
3472 BUG_ON(right_path->p_tree_depth == 0);
3473
3474 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3475 right_path, &left_cpos);
3476 if (ret) {
3477 mlog_errno(ret);
3478 goto out;
3479 }
3480
3481 /* This function shouldn't be called for the leftmost leaf. */
3482 BUG_ON(left_cpos == 0);
3483
3484 left_path = ocfs2_new_path_from_path(right_path);
3485 if (!left_path) {
3486 ret = -ENOMEM;
3487 mlog_errno(ret);
3488 goto out;
3489 }
3490
3491 ret = ocfs2_find_path(et->et_ci, left_path, left_cpos);
3492 if (ret) {
3493 mlog_errno(ret);
3494 goto out;
3495 }
3496
3497 *ret_left_path = left_path;
3498out:
3499 if (ret)
3500 ocfs2_free_path(left_path);
3501 return ret;
3502}
3503
3504/*
3505 * Remove split_rec clusters from the record at index and merge them
3506 * onto the tail of the record "before" it.
3507 * For index > 0, the "before" means the extent rec at index - 1.
3508 *
3509 * For index == 0, the "before" means the last record of the previous
3510 * extent block. And there is also a situation that we may need to
3511 * remove the rightmost leaf extent block in the right_path and change
3512 * the right path to indicate the new rightmost path.
3513 */
3514static int ocfs2_merge_rec_left(struct ocfs2_path *right_path,
3515 handle_t *handle,
3516 struct ocfs2_extent_tree *et,
3517 struct ocfs2_extent_rec *split_rec,
3518 struct ocfs2_cached_dealloc_ctxt *dealloc,
3519 int index)
3520{
3521 int ret, i, subtree_index = 0, has_empty_extent = 0;
3522 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3523 struct ocfs2_extent_rec *left_rec;
3524 struct ocfs2_extent_rec *right_rec;
3525 struct ocfs2_extent_list *el = path_leaf_el(right_path);
3526 struct buffer_head *bh = path_leaf_bh(right_path);
3527 struct buffer_head *root_bh = NULL;
3528 struct ocfs2_path *left_path = NULL;
3529 struct ocfs2_extent_list *left_el;
3530
3531 BUG_ON(index < 0);
3532
3533 right_rec = &el->l_recs[index];
3534 if (index == 0) {
3535 /* we meet with a cross extent block merge. */
3536 ret = ocfs2_get_left_path(et, right_path, &left_path);
3537 if (ret) {
3538 mlog_errno(ret);
3539 goto out;
3540 }
3541
3542 left_el = path_leaf_el(left_path);
3543 BUG_ON(le16_to_cpu(left_el->l_next_free_rec) !=
3544 le16_to_cpu(left_el->l_count));
3545
3546 left_rec = &left_el->l_recs[
3547 le16_to_cpu(left_el->l_next_free_rec) - 1];
3548 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3549 le16_to_cpu(left_rec->e_leaf_clusters) !=
3550 le32_to_cpu(split_rec->e_cpos));
3551
3552 subtree_index = ocfs2_find_subtree_root(et, left_path,
3553 right_path);
3554
3555 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3556 handle->h_buffer_credits,
3557 left_path);
3558 if (ret) {
3559 mlog_errno(ret);
3560 goto out;
3561 }
3562
3563 root_bh = left_path->p_node[subtree_index].bh;
3564 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3565
3566 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3567 subtree_index);
3568 if (ret) {
3569 mlog_errno(ret);
3570 goto out;
3571 }
3572
3573 for (i = subtree_index + 1;
3574 i < path_num_items(right_path); i++) {
3575 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3576 right_path, i);
3577 if (ret) {
3578 mlog_errno(ret);
3579 goto out;
3580 }
3581
3582 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3583 left_path, i);
3584 if (ret) {
3585 mlog_errno(ret);
3586 goto out;
3587 }
3588 }
3589 } else {
3590 left_rec = &el->l_recs[index - 1];
3591 if (ocfs2_is_empty_extent(&el->l_recs[0]))
3592 has_empty_extent = 1;
3593 }
3594
3595 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3596 path_num_items(right_path) - 1);
3597 if (ret) {
3598 mlog_errno(ret);
3599 goto out;
3600 }
3601
3602 if (has_empty_extent && index == 1) {
3603 /*
3604 * The easy case - we can just plop the record right in.
3605 */
3606 *left_rec = *split_rec;
3607
3608 has_empty_extent = 0;
3609 } else
3610 le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
3611
3612 le32_add_cpu(&right_rec->e_cpos, split_clusters);
3613 le64_add_cpu(&right_rec->e_blkno,
3614 ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3615 split_clusters));
3616 le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
3617
3618 ocfs2_cleanup_merge(el, index);
3619
3620 ocfs2_journal_dirty(handle, bh);
3621 if (left_path) {
3622 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
3623
3624 /*
3625 * In the situation that the right_rec is empty and the extent
3626 * block is empty also, ocfs2_complete_edge_insert can't handle
3627 * it and we need to delete the right extent block.
3628 */
3629 if (le16_to_cpu(right_rec->e_leaf_clusters) == 0 &&
3630 le16_to_cpu(el->l_next_free_rec) == 1) {
3631
3632 ret = ocfs2_remove_rightmost_path(handle, et,
3633 right_path,
3634 dealloc);
3635 if (ret) {
3636 mlog_errno(ret);
3637 goto out;
3638 }
3639
3640 /* Now the rightmost extent block has been deleted.
3641 * So we use the new rightmost path.
3642 */
3643 ocfs2_mv_path(right_path, left_path);
3644 left_path = NULL;
3645 } else
3646 ocfs2_complete_edge_insert(handle, left_path,
3647 right_path, subtree_index);
3648 }
3649out:
3650 if (left_path)
3651 ocfs2_free_path(left_path);
3652 return ret;
3653}
3654
3655static int ocfs2_try_to_merge_extent(handle_t *handle,
3656 struct ocfs2_extent_tree *et,
3657 struct ocfs2_path *path,
3658 int split_index,
3659 struct ocfs2_extent_rec *split_rec,
3660 struct ocfs2_cached_dealloc_ctxt *dealloc,
3661 struct ocfs2_merge_ctxt *ctxt)
3662{
3663 int ret = 0;
3664 struct ocfs2_extent_list *el = path_leaf_el(path);
3665 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3666
3667 BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
3668
3669 if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) {
3670 /*
3671 * The merge code will need to create an empty
3672 * extent to take the place of the newly
3673 * emptied slot. Remove any pre-existing empty
3674 * extents - having more than one in a leaf is
3675 * illegal.
3676 */
3677 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3678 if (ret) {
3679 mlog_errno(ret);
3680 goto out;
3681 }
3682 split_index--;
3683 rec = &el->l_recs[split_index];
3684 }
3685
3686 if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
3687 /*
3688 * Left-right contig implies this.
3689 */
3690 BUG_ON(!ctxt->c_split_covers_rec);
3691
3692 /*
3693 * Since the leftright insert always covers the entire
3694 * extent, this call will delete the insert record
3695 * entirely, resulting in an empty extent record added to
3696 * the extent block.
3697 *
3698 * Since the adding of an empty extent shifts
3699 * everything back to the right, there's no need to
3700 * update split_index here.
3701 *
3702 * When the split_index is zero, we need to merge it to the
3703 * prevoius extent block. It is more efficient and easier
3704 * if we do merge_right first and merge_left later.
3705 */
3706 ret = ocfs2_merge_rec_right(path, handle, et, split_rec,
3707 split_index);
3708 if (ret) {
3709 mlog_errno(ret);
3710 goto out;
3711 }
3712
3713 /*
3714 * We can only get this from logic error above.
3715 */
3716 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
3717
3718 /* The merge left us with an empty extent, remove it. */
3719 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3720 if (ret) {
3721 mlog_errno(ret);
3722 goto out;
3723 }
3724
3725 rec = &el->l_recs[split_index];
3726
3727 /*
3728 * Note that we don't pass split_rec here on purpose -
3729 * we've merged it into the rec already.
3730 */
3731 ret = ocfs2_merge_rec_left(path, handle, et, rec,
3732 dealloc, split_index);
3733
3734 if (ret) {
3735 mlog_errno(ret);
3736 goto out;
3737 }
3738
3739 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3740 /*
3741 * Error from this last rotate is not critical, so
3742 * print but don't bubble it up.
3743 */
3744 if (ret)
3745 mlog_errno(ret);
3746 ret = 0;
3747 } else {
3748 /*
3749 * Merge a record to the left or right.
3750 *
3751 * 'contig_type' is relative to the existing record,
3752 * so for example, if we're "right contig", it's to
3753 * the record on the left (hence the left merge).
3754 */
3755 if (ctxt->c_contig_type == CONTIG_RIGHT) {
3756 ret = ocfs2_merge_rec_left(path, handle, et,
3757 split_rec, dealloc,
3758 split_index);
3759 if (ret) {
3760 mlog_errno(ret);
3761 goto out;
3762 }
3763 } else {
3764 ret = ocfs2_merge_rec_right(path, handle,
3765 et, split_rec,
3766 split_index);
3767 if (ret) {
3768 mlog_errno(ret);
3769 goto out;
3770 }
3771 }
3772
3773 if (ctxt->c_split_covers_rec) {
3774 /*
3775 * The merge may have left an empty extent in
3776 * our leaf. Try to rotate it away.
3777 */
3778 ret = ocfs2_rotate_tree_left(handle, et, path,
3779 dealloc);
3780 if (ret)
3781 mlog_errno(ret);
3782 ret = 0;
3783 }
3784 }
3785
3786out:
3787 return ret;
3788}
3789
3790static void ocfs2_subtract_from_rec(struct super_block *sb,
3791 enum ocfs2_split_type split,
3792 struct ocfs2_extent_rec *rec,
3793 struct ocfs2_extent_rec *split_rec)
3794{
3795 u64 len_blocks;
3796
3797 len_blocks = ocfs2_clusters_to_blocks(sb,
3798 le16_to_cpu(split_rec->e_leaf_clusters));
3799
3800 if (split == SPLIT_LEFT) {
3801 /*
3802 * Region is on the left edge of the existing
3803 * record.
3804 */
3805 le32_add_cpu(&rec->e_cpos,
3806 le16_to_cpu(split_rec->e_leaf_clusters));
3807 le64_add_cpu(&rec->e_blkno, len_blocks);
3808 le16_add_cpu(&rec->e_leaf_clusters,
3809 -le16_to_cpu(split_rec->e_leaf_clusters));
3810 } else {
3811 /*
3812 * Region is on the right edge of the existing
3813 * record.
3814 */
3815 le16_add_cpu(&rec->e_leaf_clusters,
3816 -le16_to_cpu(split_rec->e_leaf_clusters));
3817 }
3818}
3819
3820/*
3821 * Do the final bits of extent record insertion at the target leaf
3822 * list. If this leaf is part of an allocation tree, it is assumed
3823 * that the tree above has been prepared.
3824 */
3825static void ocfs2_insert_at_leaf(struct ocfs2_extent_tree *et,
3826 struct ocfs2_extent_rec *insert_rec,
3827 struct ocfs2_extent_list *el,
3828 struct ocfs2_insert_type *insert)
3829{
3830 int i = insert->ins_contig_index;
3831 unsigned int range;
3832 struct ocfs2_extent_rec *rec;
3833
3834 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3835
3836 if (insert->ins_split != SPLIT_NONE) {
3837 i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
3838 BUG_ON(i == -1);
3839 rec = &el->l_recs[i];
3840 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
3841 insert->ins_split, rec,
3842 insert_rec);
3843 goto rotate;
3844 }
3845
3846 /*
3847 * Contiguous insert - either left or right.
3848 */
3849 if (insert->ins_contig != CONTIG_NONE) {
3850 rec = &el->l_recs[i];
3851 if (insert->ins_contig == CONTIG_LEFT) {
3852 rec->e_blkno = insert_rec->e_blkno;
3853 rec->e_cpos = insert_rec->e_cpos;
3854 }
3855 le16_add_cpu(&rec->e_leaf_clusters,
3856 le16_to_cpu(insert_rec->e_leaf_clusters));
3857 return;
3858 }
3859
3860 /*
3861 * Handle insert into an empty leaf.
3862 */
3863 if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3864 ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3865 ocfs2_is_empty_extent(&el->l_recs[0]))) {
3866 el->l_recs[0] = *insert_rec;
3867 el->l_next_free_rec = cpu_to_le16(1);
3868 return;
3869 }
3870
3871 /*
3872 * Appending insert.
3873 */
3874 if (insert->ins_appending == APPEND_TAIL) {
3875 i = le16_to_cpu(el->l_next_free_rec) - 1;
3876 rec = &el->l_recs[i];
3877 range = le32_to_cpu(rec->e_cpos)
3878 + le16_to_cpu(rec->e_leaf_clusters);
3879 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3880
3881 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3882 le16_to_cpu(el->l_count),
3883 "owner %llu, depth %u, count %u, next free %u, "
3884 "rec.cpos %u, rec.clusters %u, "
3885 "insert.cpos %u, insert.clusters %u\n",
3886 ocfs2_metadata_cache_owner(et->et_ci),
3887 le16_to_cpu(el->l_tree_depth),
3888 le16_to_cpu(el->l_count),
3889 le16_to_cpu(el->l_next_free_rec),
3890 le32_to_cpu(el->l_recs[i].e_cpos),
3891 le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3892 le32_to_cpu(insert_rec->e_cpos),
3893 le16_to_cpu(insert_rec->e_leaf_clusters));
3894 i++;
3895 el->l_recs[i] = *insert_rec;
3896 le16_add_cpu(&el->l_next_free_rec, 1);
3897 return;
3898 }
3899
3900rotate:
3901 /*
3902 * Ok, we have to rotate.
3903 *
3904 * At this point, it is safe to assume that inserting into an
3905 * empty leaf and appending to a leaf have both been handled
3906 * above.
3907 *
3908 * This leaf needs to have space, either by the empty 1st
3909 * extent record, or by virtue of an l_next_rec < l_count.
3910 */
3911 ocfs2_rotate_leaf(el, insert_rec);
3912}
3913
3914static void ocfs2_adjust_rightmost_records(handle_t *handle,
3915 struct ocfs2_extent_tree *et,
3916 struct ocfs2_path *path,
3917 struct ocfs2_extent_rec *insert_rec)
3918{
3919 int ret, i, next_free;
3920 struct buffer_head *bh;
3921 struct ocfs2_extent_list *el;
3922 struct ocfs2_extent_rec *rec;
3923
3924 /*
3925 * Update everything except the leaf block.
3926 */
3927 for (i = 0; i < path->p_tree_depth; i++) {
3928 bh = path->p_node[i].bh;
3929 el = path->p_node[i].el;
3930
3931 next_free = le16_to_cpu(el->l_next_free_rec);
3932 if (next_free == 0) {
3933 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3934 "Owner %llu has a bad extent list",
3935 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
3936 ret = -EIO;
3937 return;
3938 }
3939
3940 rec = &el->l_recs[next_free - 1];
3941
3942 rec->e_int_clusters = insert_rec->e_cpos;
3943 le32_add_cpu(&rec->e_int_clusters,
3944 le16_to_cpu(insert_rec->e_leaf_clusters));
3945 le32_add_cpu(&rec->e_int_clusters,
3946 -le32_to_cpu(rec->e_cpos));
3947
3948 ocfs2_journal_dirty(handle, bh);
3949 }
3950}
3951
3952static int ocfs2_append_rec_to_path(handle_t *handle,
3953 struct ocfs2_extent_tree *et,
3954 struct ocfs2_extent_rec *insert_rec,
3955 struct ocfs2_path *right_path,
3956 struct ocfs2_path **ret_left_path)
3957{
3958 int ret, next_free;
3959 struct ocfs2_extent_list *el;
3960 struct ocfs2_path *left_path = NULL;
3961
3962 *ret_left_path = NULL;
3963
3964 /*
3965 * This shouldn't happen for non-trees. The extent rec cluster
3966 * count manipulation below only works for interior nodes.
3967 */
3968 BUG_ON(right_path->p_tree_depth == 0);
3969
3970 /*
3971 * If our appending insert is at the leftmost edge of a leaf,
3972 * then we might need to update the rightmost records of the
3973 * neighboring path.
3974 */
3975 el = path_leaf_el(right_path);
3976 next_free = le16_to_cpu(el->l_next_free_rec);
3977 if (next_free == 0 ||
3978 (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
3979 u32 left_cpos;
3980
3981 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3982 right_path, &left_cpos);
3983 if (ret) {
3984 mlog_errno(ret);
3985 goto out;
3986 }
3987
3988 trace_ocfs2_append_rec_to_path(
3989 (unsigned long long)
3990 ocfs2_metadata_cache_owner(et->et_ci),
3991 le32_to_cpu(insert_rec->e_cpos),
3992 left_cpos);
3993
3994 /*
3995 * No need to worry if the append is already in the
3996 * leftmost leaf.
3997 */
3998 if (left_cpos) {
3999 left_path = ocfs2_new_path_from_path(right_path);
4000 if (!left_path) {
4001 ret = -ENOMEM;
4002 mlog_errno(ret);
4003 goto out;
4004 }
4005
4006 ret = ocfs2_find_path(et->et_ci, left_path,
4007 left_cpos);
4008 if (ret) {
4009 mlog_errno(ret);
4010 goto out;
4011 }
4012
4013 /*
4014 * ocfs2_insert_path() will pass the left_path to the
4015 * journal for us.
4016 */
4017 }
4018 }
4019
4020 ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4021 if (ret) {
4022 mlog_errno(ret);
4023 goto out;
4024 }
4025
4026 ocfs2_adjust_rightmost_records(handle, et, right_path, insert_rec);
4027
4028 *ret_left_path = left_path;
4029 ret = 0;
4030out:
4031 if (ret != 0)
4032 ocfs2_free_path(left_path);
4033
4034 return ret;
4035}
4036
4037static void ocfs2_split_record(struct ocfs2_extent_tree *et,
4038 struct ocfs2_path *left_path,
4039 struct ocfs2_path *right_path,
4040 struct ocfs2_extent_rec *split_rec,
4041 enum ocfs2_split_type split)
4042{
4043 int index;
4044 u32 cpos = le32_to_cpu(split_rec->e_cpos);
4045 struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
4046 struct ocfs2_extent_rec *rec, *tmprec;
4047
4048 right_el = path_leaf_el(right_path);
4049 if (left_path)
4050 left_el = path_leaf_el(left_path);
4051
4052 el = right_el;
4053 insert_el = right_el;
4054 index = ocfs2_search_extent_list(el, cpos);
4055 if (index != -1) {
4056 if (index == 0 && left_path) {
4057 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
4058
4059 /*
4060 * This typically means that the record
4061 * started in the left path but moved to the
4062 * right as a result of rotation. We either
4063 * move the existing record to the left, or we
4064 * do the later insert there.
4065 *
4066 * In this case, the left path should always
4067 * exist as the rotate code will have passed
4068 * it back for a post-insert update.
4069 */
4070
4071 if (split == SPLIT_LEFT) {
4072 /*
4073 * It's a left split. Since we know
4074 * that the rotate code gave us an
4075 * empty extent in the left path, we
4076 * can just do the insert there.
4077 */
4078 insert_el = left_el;
4079 } else {
4080 /*
4081 * Right split - we have to move the
4082 * existing record over to the left
4083 * leaf. The insert will be into the
4084 * newly created empty extent in the
4085 * right leaf.
4086 */
4087 tmprec = &right_el->l_recs[index];
4088 ocfs2_rotate_leaf(left_el, tmprec);
4089 el = left_el;
4090
4091 memset(tmprec, 0, sizeof(*tmprec));
4092 index = ocfs2_search_extent_list(left_el, cpos);
4093 BUG_ON(index == -1);
4094 }
4095 }
4096 } else {
4097 BUG_ON(!left_path);
4098 BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
4099 /*
4100 * Left path is easy - we can just allow the insert to
4101 * happen.
4102 */
4103 el = left_el;
4104 insert_el = left_el;
4105 index = ocfs2_search_extent_list(el, cpos);
4106 BUG_ON(index == -1);
4107 }
4108
4109 rec = &el->l_recs[index];
4110 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4111 split, rec, split_rec);
4112 ocfs2_rotate_leaf(insert_el, split_rec);
4113}
4114
4115/*
4116 * This function only does inserts on an allocation b-tree. For tree
4117 * depth = 0, ocfs2_insert_at_leaf() is called directly.
4118 *
4119 * right_path is the path we want to do the actual insert
4120 * in. left_path should only be passed in if we need to update that
4121 * portion of the tree after an edge insert.
4122 */
4123static int ocfs2_insert_path(handle_t *handle,
4124 struct ocfs2_extent_tree *et,
4125 struct ocfs2_path *left_path,
4126 struct ocfs2_path *right_path,
4127 struct ocfs2_extent_rec *insert_rec,
4128 struct ocfs2_insert_type *insert)
4129{
4130 int ret, subtree_index;
4131 struct buffer_head *leaf_bh = path_leaf_bh(right_path);
4132
4133 if (left_path) {
4134 /*
4135 * There's a chance that left_path got passed back to
4136 * us without being accounted for in the
4137 * journal. Extend our transaction here to be sure we
4138 * can change those blocks.
4139 */
4140 ret = ocfs2_extend_trans(handle, left_path->p_tree_depth);
4141 if (ret < 0) {
4142 mlog_errno(ret);
4143 goto out;
4144 }
4145
4146 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
4147 if (ret < 0) {
4148 mlog_errno(ret);
4149 goto out;
4150 }
4151 }
4152
4153 /*
4154 * Pass both paths to the journal. The majority of inserts
4155 * will be touching all components anyway.
4156 */
4157 ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4158 if (ret < 0) {
4159 mlog_errno(ret);
4160 goto out;
4161 }
4162
4163 if (insert->ins_split != SPLIT_NONE) {
4164 /*
4165 * We could call ocfs2_insert_at_leaf() for some types
4166 * of splits, but it's easier to just let one separate
4167 * function sort it all out.
4168 */
4169 ocfs2_split_record(et, left_path, right_path,
4170 insert_rec, insert->ins_split);
4171
4172 /*
4173 * Split might have modified either leaf and we don't
4174 * have a guarantee that the later edge insert will
4175 * dirty this for us.
4176 */
4177 if (left_path)
4178 ocfs2_journal_dirty(handle,
4179 path_leaf_bh(left_path));
4180 } else
4181 ocfs2_insert_at_leaf(et, insert_rec, path_leaf_el(right_path),
4182 insert);
4183
4184 ocfs2_journal_dirty(handle, leaf_bh);
4185
4186 if (left_path) {
4187 /*
4188 * The rotate code has indicated that we need to fix
4189 * up portions of the tree after the insert.
4190 *
4191 * XXX: Should we extend the transaction here?
4192 */
4193 subtree_index = ocfs2_find_subtree_root(et, left_path,
4194 right_path);
4195 ocfs2_complete_edge_insert(handle, left_path, right_path,
4196 subtree_index);
4197 }
4198
4199 ret = 0;
4200out:
4201 return ret;
4202}
4203
4204static int ocfs2_do_insert_extent(handle_t *handle,
4205 struct ocfs2_extent_tree *et,
4206 struct ocfs2_extent_rec *insert_rec,
4207 struct ocfs2_insert_type *type)
4208{
4209 int ret, rotate = 0;
4210 u32 cpos;
4211 struct ocfs2_path *right_path = NULL;
4212 struct ocfs2_path *left_path = NULL;
4213 struct ocfs2_extent_list *el;
4214
4215 el = et->et_root_el;
4216
4217 ret = ocfs2_et_root_journal_access(handle, et,
4218 OCFS2_JOURNAL_ACCESS_WRITE);
4219 if (ret) {
4220 mlog_errno(ret);
4221 goto out;
4222 }
4223
4224 if (le16_to_cpu(el->l_tree_depth) == 0) {
4225 ocfs2_insert_at_leaf(et, insert_rec, el, type);
4226 goto out_update_clusters;
4227 }
4228
4229 right_path = ocfs2_new_path_from_et(et);
4230 if (!right_path) {
4231 ret = -ENOMEM;
4232 mlog_errno(ret);
4233 goto out;
4234 }
4235
4236 /*
4237 * Determine the path to start with. Rotations need the
4238 * rightmost path, everything else can go directly to the
4239 * target leaf.
4240 */
4241 cpos = le32_to_cpu(insert_rec->e_cpos);
4242 if (type->ins_appending == APPEND_NONE &&
4243 type->ins_contig == CONTIG_NONE) {
4244 rotate = 1;
4245 cpos = UINT_MAX;
4246 }
4247
4248 ret = ocfs2_find_path(et->et_ci, right_path, cpos);
4249 if (ret) {
4250 mlog_errno(ret);
4251 goto out;
4252 }
4253
4254 /*
4255 * Rotations and appends need special treatment - they modify
4256 * parts of the tree's above them.
4257 *
4258 * Both might pass back a path immediate to the left of the
4259 * one being inserted to. This will be cause
4260 * ocfs2_insert_path() to modify the rightmost records of
4261 * left_path to account for an edge insert.
4262 *
4263 * XXX: When modifying this code, keep in mind that an insert
4264 * can wind up skipping both of these two special cases...
4265 */
4266 if (rotate) {
4267 ret = ocfs2_rotate_tree_right(handle, et, type->ins_split,
4268 le32_to_cpu(insert_rec->e_cpos),
4269 right_path, &left_path);
4270 if (ret) {
4271 mlog_errno(ret);
4272 goto out;
4273 }
4274
4275 /*
4276 * ocfs2_rotate_tree_right() might have extended the
4277 * transaction without re-journaling our tree root.
4278 */
4279 ret = ocfs2_et_root_journal_access(handle, et,
4280 OCFS2_JOURNAL_ACCESS_WRITE);
4281 if (ret) {
4282 mlog_errno(ret);
4283 goto out;
4284 }
4285 } else if (type->ins_appending == APPEND_TAIL
4286 && type->ins_contig != CONTIG_LEFT) {
4287 ret = ocfs2_append_rec_to_path(handle, et, insert_rec,
4288 right_path, &left_path);
4289 if (ret) {
4290 mlog_errno(ret);
4291 goto out;
4292 }
4293 }
4294
4295 ret = ocfs2_insert_path(handle, et, left_path, right_path,
4296 insert_rec, type);
4297 if (ret) {
4298 mlog_errno(ret);
4299 goto out;
4300 }
4301
4302out_update_clusters:
4303 if (type->ins_split == SPLIT_NONE)
4304 ocfs2_et_update_clusters(et,
4305 le16_to_cpu(insert_rec->e_leaf_clusters));
4306
4307 ocfs2_journal_dirty(handle, et->et_root_bh);
4308
4309out:
4310 ocfs2_free_path(left_path);
4311 ocfs2_free_path(right_path);
4312
4313 return ret;
4314}
4315
4316static enum ocfs2_contig_type
4317ocfs2_figure_merge_contig_type(struct ocfs2_extent_tree *et,
4318 struct ocfs2_path *path,
4319 struct ocfs2_extent_list *el, int index,
4320 struct ocfs2_extent_rec *split_rec)
4321{
4322 int status;
4323 enum ocfs2_contig_type ret = CONTIG_NONE;
4324 u32 left_cpos, right_cpos;
4325 struct ocfs2_extent_rec *rec = NULL;
4326 struct ocfs2_extent_list *new_el;
4327 struct ocfs2_path *left_path = NULL, *right_path = NULL;
4328 struct buffer_head *bh;
4329 struct ocfs2_extent_block *eb;
4330 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
4331
4332 if (index > 0) {
4333 rec = &el->l_recs[index - 1];
4334 } else if (path->p_tree_depth > 0) {
4335 status = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
4336 if (status)
4337 goto out;
4338
4339 if (left_cpos != 0) {
4340 left_path = ocfs2_new_path_from_path(path);
4341 if (!left_path)
4342 goto out;
4343
4344 status = ocfs2_find_path(et->et_ci, left_path,
4345 left_cpos);
4346 if (status)
4347 goto out;
4348
4349 new_el = path_leaf_el(left_path);
4350
4351 if (le16_to_cpu(new_el->l_next_free_rec) !=
4352 le16_to_cpu(new_el->l_count)) {
4353 bh = path_leaf_bh(left_path);
4354 eb = (struct ocfs2_extent_block *)bh->b_data;
4355 ocfs2_error(sb,
4356 "Extent block #%llu has an "
4357 "invalid l_next_free_rec of "
4358 "%d. It should have "
4359 "matched the l_count of %d",
4360 (unsigned long long)le64_to_cpu(eb->h_blkno),
4361 le16_to_cpu(new_el->l_next_free_rec),
4362 le16_to_cpu(new_el->l_count));
4363 status = -EINVAL;
4364 goto out;
4365 }
4366 rec = &new_el->l_recs[
4367 le16_to_cpu(new_el->l_next_free_rec) - 1];
4368 }
4369 }
4370
4371 /*
4372 * We're careful to check for an empty extent record here -
4373 * the merge code will know what to do if it sees one.
4374 */
4375 if (rec) {
4376 if (index == 1 && ocfs2_is_empty_extent(rec)) {
4377 if (split_rec->e_cpos == el->l_recs[index].e_cpos)
4378 ret = CONTIG_RIGHT;
4379 } else {
4380 ret = ocfs2_et_extent_contig(et, rec, split_rec);
4381 }
4382 }
4383
4384 rec = NULL;
4385 if (index < (le16_to_cpu(el->l_next_free_rec) - 1))
4386 rec = &el->l_recs[index + 1];
4387 else if (le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count) &&
4388 path->p_tree_depth > 0) {
4389 status = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
4390 if (status)
4391 goto out;
4392
4393 if (right_cpos == 0)
4394 goto out;
4395
4396 right_path = ocfs2_new_path_from_path(path);
4397 if (!right_path)
4398 goto out;
4399
4400 status = ocfs2_find_path(et->et_ci, right_path, right_cpos);
4401 if (status)
4402 goto out;
4403
4404 new_el = path_leaf_el(right_path);
4405 rec = &new_el->l_recs[0];
4406 if (ocfs2_is_empty_extent(rec)) {
4407 if (le16_to_cpu(new_el->l_next_free_rec) <= 1) {
4408 bh = path_leaf_bh(right_path);
4409 eb = (struct ocfs2_extent_block *)bh->b_data;
4410 ocfs2_error(sb,
4411 "Extent block #%llu has an "
4412 "invalid l_next_free_rec of %d",
4413 (unsigned long long)le64_to_cpu(eb->h_blkno),
4414 le16_to_cpu(new_el->l_next_free_rec));
4415 status = -EINVAL;
4416 goto out;
4417 }
4418 rec = &new_el->l_recs[1];
4419 }
4420 }
4421
4422 if (rec) {
4423 enum ocfs2_contig_type contig_type;
4424
4425 contig_type = ocfs2_et_extent_contig(et, rec, split_rec);
4426
4427 if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
4428 ret = CONTIG_LEFTRIGHT;
4429 else if (ret == CONTIG_NONE)
4430 ret = contig_type;
4431 }
4432
4433out:
4434 if (left_path)
4435 ocfs2_free_path(left_path);
4436 if (right_path)
4437 ocfs2_free_path(right_path);
4438
4439 return ret;
4440}
4441
4442static void ocfs2_figure_contig_type(struct ocfs2_extent_tree *et,
4443 struct ocfs2_insert_type *insert,
4444 struct ocfs2_extent_list *el,
4445 struct ocfs2_extent_rec *insert_rec)
4446{
4447 int i;
4448 enum ocfs2_contig_type contig_type = CONTIG_NONE;
4449
4450 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4451
4452 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
4453 contig_type = ocfs2_et_extent_contig(et, &el->l_recs[i],
4454 insert_rec);
4455 if (contig_type != CONTIG_NONE) {
4456 insert->ins_contig_index = i;
4457 break;
4458 }
4459 }
4460 insert->ins_contig = contig_type;
4461
4462 if (insert->ins_contig != CONTIG_NONE) {
4463 struct ocfs2_extent_rec *rec =
4464 &el->l_recs[insert->ins_contig_index];
4465 unsigned int len = le16_to_cpu(rec->e_leaf_clusters) +
4466 le16_to_cpu(insert_rec->e_leaf_clusters);
4467
4468 /*
4469 * Caller might want us to limit the size of extents, don't
4470 * calculate contiguousness if we might exceed that limit.
4471 */
4472 if (et->et_max_leaf_clusters &&
4473 (len > et->et_max_leaf_clusters))
4474 insert->ins_contig = CONTIG_NONE;
4475 }
4476}
4477
4478/*
4479 * This should only be called against the righmost leaf extent list.
4480 *
4481 * ocfs2_figure_appending_type() will figure out whether we'll have to
4482 * insert at the tail of the rightmost leaf.
4483 *
4484 * This should also work against the root extent list for tree's with 0
4485 * depth. If we consider the root extent list to be the rightmost leaf node
4486 * then the logic here makes sense.
4487 */
4488static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
4489 struct ocfs2_extent_list *el,
4490 struct ocfs2_extent_rec *insert_rec)
4491{
4492 int i;
4493 u32 cpos = le32_to_cpu(insert_rec->e_cpos);
4494 struct ocfs2_extent_rec *rec;
4495
4496 insert->ins_appending = APPEND_NONE;
4497
4498 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4499
4500 if (!el->l_next_free_rec)
4501 goto set_tail_append;
4502
4503 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
4504 /* Were all records empty? */
4505 if (le16_to_cpu(el->l_next_free_rec) == 1)
4506 goto set_tail_append;
4507 }
4508
4509 i = le16_to_cpu(el->l_next_free_rec) - 1;
4510 rec = &el->l_recs[i];
4511
4512 if (cpos >=
4513 (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
4514 goto set_tail_append;
4515
4516 return;
4517
4518set_tail_append:
4519 insert->ins_appending = APPEND_TAIL;
4520}
4521
4522/*
4523 * Helper function called at the beginning of an insert.
4524 *
4525 * This computes a few things that are commonly used in the process of
4526 * inserting into the btree:
4527 * - Whether the new extent is contiguous with an existing one.
4528 * - The current tree depth.
4529 * - Whether the insert is an appending one.
4530 * - The total # of free records in the tree.
4531 *
4532 * All of the information is stored on the ocfs2_insert_type
4533 * structure.
4534 */
4535static int ocfs2_figure_insert_type(struct ocfs2_extent_tree *et,
4536 struct buffer_head **last_eb_bh,
4537 struct ocfs2_extent_rec *insert_rec,
4538 int *free_records,
4539 struct ocfs2_insert_type *insert)
4540{
4541 int ret;
4542 struct ocfs2_extent_block *eb;
4543 struct ocfs2_extent_list *el;
4544 struct ocfs2_path *path = NULL;
4545 struct buffer_head *bh = NULL;
4546
4547 insert->ins_split = SPLIT_NONE;
4548
4549 el = et->et_root_el;
4550 insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
4551
4552 if (el->l_tree_depth) {
4553 /*
4554 * If we have tree depth, we read in the
4555 * rightmost extent block ahead of time as
4556 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4557 * may want it later.
4558 */
4559 ret = ocfs2_read_extent_block(et->et_ci,
4560 ocfs2_et_get_last_eb_blk(et),
4561 &bh);
4562 if (ret) {
4563 mlog_errno(ret);
4564 goto out;
4565 }
4566 eb = (struct ocfs2_extent_block *) bh->b_data;
4567 el = &eb->h_list;
4568 }
4569
4570 /*
4571 * Unless we have a contiguous insert, we'll need to know if
4572 * there is room left in our allocation tree for another
4573 * extent record.
4574 *
4575 * XXX: This test is simplistic, we can search for empty
4576 * extent records too.
4577 */
4578 *free_records = le16_to_cpu(el->l_count) -
4579 le16_to_cpu(el->l_next_free_rec);
4580
4581 if (!insert->ins_tree_depth) {
4582 ocfs2_figure_contig_type(et, insert, el, insert_rec);
4583 ocfs2_figure_appending_type(insert, el, insert_rec);
4584 return 0;
4585 }
4586
4587 path = ocfs2_new_path_from_et(et);
4588 if (!path) {
4589 ret = -ENOMEM;
4590 mlog_errno(ret);
4591 goto out;
4592 }
4593
4594 /*
4595 * In the case that we're inserting past what the tree
4596 * currently accounts for, ocfs2_find_path() will return for
4597 * us the rightmost tree path. This is accounted for below in
4598 * the appending code.
4599 */
4600 ret = ocfs2_find_path(et->et_ci, path, le32_to_cpu(insert_rec->e_cpos));
4601 if (ret) {
4602 mlog_errno(ret);
4603 goto out;
4604 }
4605
4606 el = path_leaf_el(path);
4607
4608 /*
4609 * Now that we have the path, there's two things we want to determine:
4610 * 1) Contiguousness (also set contig_index if this is so)
4611 *
4612 * 2) Are we doing an append? We can trivially break this up
4613 * into two types of appends: simple record append, or a
4614 * rotate inside the tail leaf.
4615 */
4616 ocfs2_figure_contig_type(et, insert, el, insert_rec);
4617
4618 /*
4619 * The insert code isn't quite ready to deal with all cases of
4620 * left contiguousness. Specifically, if it's an insert into
4621 * the 1st record in a leaf, it will require the adjustment of
4622 * cluster count on the last record of the path directly to it's
4623 * left. For now, just catch that case and fool the layers
4624 * above us. This works just fine for tree_depth == 0, which
4625 * is why we allow that above.
4626 */
4627 if (insert->ins_contig == CONTIG_LEFT &&
4628 insert->ins_contig_index == 0)
4629 insert->ins_contig = CONTIG_NONE;
4630
4631 /*
4632 * Ok, so we can simply compare against last_eb to figure out
4633 * whether the path doesn't exist. This will only happen in
4634 * the case that we're doing a tail append, so maybe we can
4635 * take advantage of that information somehow.
4636 */
4637 if (ocfs2_et_get_last_eb_blk(et) ==
4638 path_leaf_bh(path)->b_blocknr) {
4639 /*
4640 * Ok, ocfs2_find_path() returned us the rightmost
4641 * tree path. This might be an appending insert. There are
4642 * two cases:
4643 * 1) We're doing a true append at the tail:
4644 * -This might even be off the end of the leaf
4645 * 2) We're "appending" by rotating in the tail
4646 */
4647 ocfs2_figure_appending_type(insert, el, insert_rec);
4648 }
4649
4650out:
4651 ocfs2_free_path(path);
4652
4653 if (ret == 0)
4654 *last_eb_bh = bh;
4655 else
4656 brelse(bh);
4657 return ret;
4658}
4659
4660/*
4661 * Insert an extent into a btree.
4662 *
4663 * The caller needs to update the owning btree's cluster count.
4664 */
4665int ocfs2_insert_extent(handle_t *handle,
4666 struct ocfs2_extent_tree *et,
4667 u32 cpos,
4668 u64 start_blk,
4669 u32 new_clusters,
4670 u8 flags,
4671 struct ocfs2_alloc_context *meta_ac)
4672{
4673 int status;
4674 int uninitialized_var(free_records);
4675 struct buffer_head *last_eb_bh = NULL;
4676 struct ocfs2_insert_type insert = {0, };
4677 struct ocfs2_extent_rec rec;
4678
4679 trace_ocfs2_insert_extent_start(
4680 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4681 cpos, new_clusters);
4682
4683 memset(&rec, 0, sizeof(rec));
4684 rec.e_cpos = cpu_to_le32(cpos);
4685 rec.e_blkno = cpu_to_le64(start_blk);
4686 rec.e_leaf_clusters = cpu_to_le16(new_clusters);
4687 rec.e_flags = flags;
4688 status = ocfs2_et_insert_check(et, &rec);
4689 if (status) {
4690 mlog_errno(status);
4691 goto bail;
4692 }
4693
4694 status = ocfs2_figure_insert_type(et, &last_eb_bh, &rec,
4695 &free_records, &insert);
4696 if (status < 0) {
4697 mlog_errno(status);
4698 goto bail;
4699 }
4700
4701 trace_ocfs2_insert_extent(insert.ins_appending, insert.ins_contig,
4702 insert.ins_contig_index, free_records,
4703 insert.ins_tree_depth);
4704
4705 if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
4706 status = ocfs2_grow_tree(handle, et,
4707 &insert.ins_tree_depth, &last_eb_bh,
4708 meta_ac);
4709 if (status) {
4710 mlog_errno(status);
4711 goto bail;
4712 }
4713 }
4714
4715 /* Finally, we can add clusters. This might rotate the tree for us. */
4716 status = ocfs2_do_insert_extent(handle, et, &rec, &insert);
4717 if (status < 0)
4718 mlog_errno(status);
4719 else
4720 ocfs2_et_extent_map_insert(et, &rec);
4721
4722bail:
4723 brelse(last_eb_bh);
4724
4725 return status;
4726}
4727
4728/*
4729 * Allcate and add clusters into the extent b-tree.
4730 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4731 * The extent b-tree's root is specified by et, and
4732 * it is not limited to the file storage. Any extent tree can use this
4733 * function if it implements the proper ocfs2_extent_tree.
4734 */
4735int ocfs2_add_clusters_in_btree(handle_t *handle,
4736 struct ocfs2_extent_tree *et,
4737 u32 *logical_offset,
4738 u32 clusters_to_add,
4739 int mark_unwritten,
4740 struct ocfs2_alloc_context *data_ac,
4741 struct ocfs2_alloc_context *meta_ac,
4742 enum ocfs2_alloc_restarted *reason_ret)
4743{
4744 int status = 0, err = 0;
4745 int free_extents;
4746 enum ocfs2_alloc_restarted reason = RESTART_NONE;
4747 u32 bit_off, num_bits;
4748 u64 block;
4749 u8 flags = 0;
4750 struct ocfs2_super *osb =
4751 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
4752
4753 BUG_ON(!clusters_to_add);
4754
4755 if (mark_unwritten)
4756 flags = OCFS2_EXT_UNWRITTEN;
4757
4758 free_extents = ocfs2_num_free_extents(osb, et);
4759 if (free_extents < 0) {
4760 status = free_extents;
4761 mlog_errno(status);
4762 goto leave;
4763 }
4764
4765 /* there are two cases which could cause us to EAGAIN in the
4766 * we-need-more-metadata case:
4767 * 1) we haven't reserved *any*
4768 * 2) we are so fragmented, we've needed to add metadata too
4769 * many times. */
4770 if (!free_extents && !meta_ac) {
4771 err = -1;
4772 status = -EAGAIN;
4773 reason = RESTART_META;
4774 goto leave;
4775 } else if ((!free_extents)
4776 && (ocfs2_alloc_context_bits_left(meta_ac)
4777 < ocfs2_extend_meta_needed(et->et_root_el))) {
4778 err = -2;
4779 status = -EAGAIN;
4780 reason = RESTART_META;
4781 goto leave;
4782 }
4783
4784 status = __ocfs2_claim_clusters(handle, data_ac, 1,
4785 clusters_to_add, &bit_off, &num_bits);
4786 if (status < 0) {
4787 if (status != -ENOSPC)
4788 mlog_errno(status);
4789 goto leave;
4790 }
4791
4792 BUG_ON(num_bits > clusters_to_add);
4793
4794 /* reserve our write early -- insert_extent may update the tree root */
4795 status = ocfs2_et_root_journal_access(handle, et,
4796 OCFS2_JOURNAL_ACCESS_WRITE);
4797 if (status < 0) {
4798 mlog_errno(status);
4799 goto leave;
4800 }
4801
4802 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
4803 trace_ocfs2_add_clusters_in_btree(
4804 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4805 bit_off, num_bits);
4806 status = ocfs2_insert_extent(handle, et, *logical_offset, block,
4807 num_bits, flags, meta_ac);
4808 if (status < 0) {
4809 mlog_errno(status);
4810 goto leave;
4811 }
4812
4813 ocfs2_journal_dirty(handle, et->et_root_bh);
4814
4815 clusters_to_add -= num_bits;
4816 *logical_offset += num_bits;
4817
4818 if (clusters_to_add) {
4819 err = clusters_to_add;
4820 status = -EAGAIN;
4821 reason = RESTART_TRANS;
4822 }
4823
4824leave:
4825 if (reason_ret)
4826 *reason_ret = reason;
4827 trace_ocfs2_add_clusters_in_btree_ret(status, reason, err);
4828 return status;
4829}
4830
4831static void ocfs2_make_right_split_rec(struct super_block *sb,
4832 struct ocfs2_extent_rec *split_rec,
4833 u32 cpos,
4834 struct ocfs2_extent_rec *rec)
4835{
4836 u32 rec_cpos = le32_to_cpu(rec->e_cpos);
4837 u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
4838
4839 memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
4840
4841 split_rec->e_cpos = cpu_to_le32(cpos);
4842 split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
4843
4844 split_rec->e_blkno = rec->e_blkno;
4845 le64_add_cpu(&split_rec->e_blkno,
4846 ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
4847
4848 split_rec->e_flags = rec->e_flags;
4849}
4850
4851static int ocfs2_split_and_insert(handle_t *handle,
4852 struct ocfs2_extent_tree *et,
4853 struct ocfs2_path *path,
4854 struct buffer_head **last_eb_bh,
4855 int split_index,
4856 struct ocfs2_extent_rec *orig_split_rec,
4857 struct ocfs2_alloc_context *meta_ac)
4858{
4859 int ret = 0, depth;
4860 unsigned int insert_range, rec_range, do_leftright = 0;
4861 struct ocfs2_extent_rec tmprec;
4862 struct ocfs2_extent_list *rightmost_el;
4863 struct ocfs2_extent_rec rec;
4864 struct ocfs2_extent_rec split_rec = *orig_split_rec;
4865 struct ocfs2_insert_type insert;
4866 struct ocfs2_extent_block *eb;
4867
4868leftright:
4869 /*
4870 * Store a copy of the record on the stack - it might move
4871 * around as the tree is manipulated below.
4872 */
4873 rec = path_leaf_el(path)->l_recs[split_index];
4874
4875 rightmost_el = et->et_root_el;
4876
4877 depth = le16_to_cpu(rightmost_el->l_tree_depth);
4878 if (depth) {
4879 BUG_ON(!(*last_eb_bh));
4880 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
4881 rightmost_el = &eb->h_list;
4882 }
4883
4884 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4885 le16_to_cpu(rightmost_el->l_count)) {
4886 ret = ocfs2_grow_tree(handle, et,
4887 &depth, last_eb_bh, meta_ac);
4888 if (ret) {
4889 mlog_errno(ret);
4890 goto out;
4891 }
4892 }
4893
4894 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4895 insert.ins_appending = APPEND_NONE;
4896 insert.ins_contig = CONTIG_NONE;
4897 insert.ins_tree_depth = depth;
4898
4899 insert_range = le32_to_cpu(split_rec.e_cpos) +
4900 le16_to_cpu(split_rec.e_leaf_clusters);
4901 rec_range = le32_to_cpu(rec.e_cpos) +
4902 le16_to_cpu(rec.e_leaf_clusters);
4903
4904 if (split_rec.e_cpos == rec.e_cpos) {
4905 insert.ins_split = SPLIT_LEFT;
4906 } else if (insert_range == rec_range) {
4907 insert.ins_split = SPLIT_RIGHT;
4908 } else {
4909 /*
4910 * Left/right split. We fake this as a right split
4911 * first and then make a second pass as a left split.
4912 */
4913 insert.ins_split = SPLIT_RIGHT;
4914
4915 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4916 &tmprec, insert_range, &rec);
4917
4918 split_rec = tmprec;
4919
4920 BUG_ON(do_leftright);
4921 do_leftright = 1;
4922 }
4923
4924 ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
4925 if (ret) {
4926 mlog_errno(ret);
4927 goto out;
4928 }
4929
4930 if (do_leftright == 1) {
4931 u32 cpos;
4932 struct ocfs2_extent_list *el;
4933
4934 do_leftright++;
4935 split_rec = *orig_split_rec;
4936
4937 ocfs2_reinit_path(path, 1);
4938
4939 cpos = le32_to_cpu(split_rec.e_cpos);
4940 ret = ocfs2_find_path(et->et_ci, path, cpos);
4941 if (ret) {
4942 mlog_errno(ret);
4943 goto out;
4944 }
4945
4946 el = path_leaf_el(path);
4947 split_index = ocfs2_search_extent_list(el, cpos);
4948 goto leftright;
4949 }
4950out:
4951
4952 return ret;
4953}
4954
4955static int ocfs2_replace_extent_rec(handle_t *handle,
4956 struct ocfs2_extent_tree *et,
4957 struct ocfs2_path *path,
4958 struct ocfs2_extent_list *el,
4959 int split_index,
4960 struct ocfs2_extent_rec *split_rec)
4961{
4962 int ret;
4963
4964 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
4965 path_num_items(path) - 1);
4966 if (ret) {
4967 mlog_errno(ret);
4968 goto out;
4969 }
4970
4971 el->l_recs[split_index] = *split_rec;
4972
4973 ocfs2_journal_dirty(handle, path_leaf_bh(path));
4974out:
4975 return ret;
4976}
4977
4978/*
4979 * Split part or all of the extent record at split_index in the leaf
4980 * pointed to by path. Merge with the contiguous extent record if needed.
4981 *
4982 * Care is taken to handle contiguousness so as to not grow the tree.
4983 *
4984 * meta_ac is not strictly necessary - we only truly need it if growth
4985 * of the tree is required. All other cases will degrade into a less
4986 * optimal tree layout.
4987 *
4988 * last_eb_bh should be the rightmost leaf block for any extent
4989 * btree. Since a split may grow the tree or a merge might shrink it,
4990 * the caller cannot trust the contents of that buffer after this call.
4991 *
4992 * This code is optimized for readability - several passes might be
4993 * made over certain portions of the tree. All of those blocks will
4994 * have been brought into cache (and pinned via the journal), so the
4995 * extra overhead is not expressed in terms of disk reads.
4996 */
4997int ocfs2_split_extent(handle_t *handle,
4998 struct ocfs2_extent_tree *et,
4999 struct ocfs2_path *path,
5000 int split_index,
5001 struct ocfs2_extent_rec *split_rec,
5002 struct ocfs2_alloc_context *meta_ac,
5003 struct ocfs2_cached_dealloc_ctxt *dealloc)
5004{
5005 int ret = 0;
5006 struct ocfs2_extent_list *el = path_leaf_el(path);
5007 struct buffer_head *last_eb_bh = NULL;
5008 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
5009 struct ocfs2_merge_ctxt ctxt;
5010 struct ocfs2_extent_list *rightmost_el;
5011
5012 if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) ||
5013 ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) <
5014 (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) {
5015 ret = -EIO;
5016 mlog_errno(ret);
5017 goto out;
5018 }
5019
5020 ctxt.c_contig_type = ocfs2_figure_merge_contig_type(et, path, el,
5021 split_index,
5022 split_rec);
5023
5024 /*
5025 * The core merge / split code wants to know how much room is
5026 * left in this allocation tree, so we pass the
5027 * rightmost extent list.
5028 */
5029 if (path->p_tree_depth) {
5030 struct ocfs2_extent_block *eb;
5031
5032 ret = ocfs2_read_extent_block(et->et_ci,
5033 ocfs2_et_get_last_eb_blk(et),
5034 &last_eb_bh);
5035 if (ret) {
5036 mlog_errno(ret);
5037 goto out;
5038 }
5039
5040 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5041 rightmost_el = &eb->h_list;
5042 } else
5043 rightmost_el = path_root_el(path);
5044
5045 if (rec->e_cpos == split_rec->e_cpos &&
5046 rec->e_leaf_clusters == split_rec->e_leaf_clusters)
5047 ctxt.c_split_covers_rec = 1;
5048 else
5049 ctxt.c_split_covers_rec = 0;
5050
5051 ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
5052
5053 trace_ocfs2_split_extent(split_index, ctxt.c_contig_type,
5054 ctxt.c_has_empty_extent,
5055 ctxt.c_split_covers_rec);
5056
5057 if (ctxt.c_contig_type == CONTIG_NONE) {
5058 if (ctxt.c_split_covers_rec)
5059 ret = ocfs2_replace_extent_rec(handle, et, path, el,
5060 split_index, split_rec);
5061 else
5062 ret = ocfs2_split_and_insert(handle, et, path,
5063 &last_eb_bh, split_index,
5064 split_rec, meta_ac);
5065 if (ret)
5066 mlog_errno(ret);
5067 } else {
5068 ret = ocfs2_try_to_merge_extent(handle, et, path,
5069 split_index, split_rec,
5070 dealloc, &ctxt);
5071 if (ret)
5072 mlog_errno(ret);
5073 }
5074
5075out:
5076 brelse(last_eb_bh);
5077 return ret;
5078}
5079
5080/*
5081 * Change the flags of the already-existing extent at cpos for len clusters.
5082 *
5083 * new_flags: the flags we want to set.
5084 * clear_flags: the flags we want to clear.
5085 * phys: the new physical offset we want this new extent starts from.
5086 *
5087 * If the existing extent is larger than the request, initiate a
5088 * split. An attempt will be made at merging with adjacent extents.
5089 *
5090 * The caller is responsible for passing down meta_ac if we'll need it.
5091 */
5092int ocfs2_change_extent_flag(handle_t *handle,
5093 struct ocfs2_extent_tree *et,
5094 u32 cpos, u32 len, u32 phys,
5095 struct ocfs2_alloc_context *meta_ac,
5096 struct ocfs2_cached_dealloc_ctxt *dealloc,
5097 int new_flags, int clear_flags)
5098{
5099 int ret, index;
5100 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5101 u64 start_blkno = ocfs2_clusters_to_blocks(sb, phys);
5102 struct ocfs2_extent_rec split_rec;
5103 struct ocfs2_path *left_path = NULL;
5104 struct ocfs2_extent_list *el;
5105 struct ocfs2_extent_rec *rec;
5106
5107 left_path = ocfs2_new_path_from_et(et);
5108 if (!left_path) {
5109 ret = -ENOMEM;
5110 mlog_errno(ret);
5111 goto out;
5112 }
5113
5114 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
5115 if (ret) {
5116 mlog_errno(ret);
5117 goto out;
5118 }
5119 el = path_leaf_el(left_path);
5120
5121 index = ocfs2_search_extent_list(el, cpos);
5122 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
5123 ocfs2_error(sb,
5124 "Owner %llu has an extent at cpos %u which can no "
5125 "longer be found.\n",
5126 (unsigned long long)
5127 ocfs2_metadata_cache_owner(et->et_ci), cpos);
5128 ret = -EROFS;
5129 goto out;
5130 }
5131
5132 ret = -EIO;
5133 rec = &el->l_recs[index];
5134 if (new_flags && (rec->e_flags & new_flags)) {
5135 mlog(ML_ERROR, "Owner %llu tried to set %d flags on an "
5136 "extent that already had them",
5137 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5138 new_flags);
5139 goto out;
5140 }
5141
5142 if (clear_flags && !(rec->e_flags & clear_flags)) {
5143 mlog(ML_ERROR, "Owner %llu tried to clear %d flags on an "
5144 "extent that didn't have them",
5145 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5146 clear_flags);
5147 goto out;
5148 }
5149
5150 memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
5151 split_rec.e_cpos = cpu_to_le32(cpos);
5152 split_rec.e_leaf_clusters = cpu_to_le16(len);
5153 split_rec.e_blkno = cpu_to_le64(start_blkno);
5154 split_rec.e_flags = rec->e_flags;
5155 if (new_flags)
5156 split_rec.e_flags |= new_flags;
5157 if (clear_flags)
5158 split_rec.e_flags &= ~clear_flags;
5159
5160 ret = ocfs2_split_extent(handle, et, left_path,
5161 index, &split_rec, meta_ac,
5162 dealloc);
5163 if (ret)
5164 mlog_errno(ret);
5165
5166out:
5167 ocfs2_free_path(left_path);
5168 return ret;
5169
5170}
5171
5172/*
5173 * Mark the already-existing extent at cpos as written for len clusters.
5174 * This removes the unwritten extent flag.
5175 *
5176 * If the existing extent is larger than the request, initiate a
5177 * split. An attempt will be made at merging with adjacent extents.
5178 *
5179 * The caller is responsible for passing down meta_ac if we'll need it.
5180 */
5181int ocfs2_mark_extent_written(struct inode *inode,
5182 struct ocfs2_extent_tree *et,
5183 handle_t *handle, u32 cpos, u32 len, u32 phys,
5184 struct ocfs2_alloc_context *meta_ac,
5185 struct ocfs2_cached_dealloc_ctxt *dealloc)
5186{
5187 int ret;
5188
5189 trace_ocfs2_mark_extent_written(
5190 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5191 cpos, len, phys);
5192
5193 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
5194 ocfs2_error(inode->i_sb, "Inode %llu has unwritten extents "
5195 "that are being written to, but the feature bit "
5196 "is not set in the super block.",
5197 (unsigned long long)OCFS2_I(inode)->ip_blkno);
5198 ret = -EROFS;
5199 goto out;
5200 }
5201
5202 /*
5203 * XXX: This should be fixed up so that we just re-insert the
5204 * next extent records.
5205 */
5206 ocfs2_et_extent_map_truncate(et, 0);
5207
5208 ret = ocfs2_change_extent_flag(handle, et, cpos,
5209 len, phys, meta_ac, dealloc,
5210 0, OCFS2_EXT_UNWRITTEN);
5211 if (ret)
5212 mlog_errno(ret);
5213
5214out:
5215 return ret;
5216}
5217
5218static int ocfs2_split_tree(handle_t *handle, struct ocfs2_extent_tree *et,
5219 struct ocfs2_path *path,
5220 int index, u32 new_range,
5221 struct ocfs2_alloc_context *meta_ac)
5222{
5223 int ret, depth, credits;
5224 struct buffer_head *last_eb_bh = NULL;
5225 struct ocfs2_extent_block *eb;
5226 struct ocfs2_extent_list *rightmost_el, *el;
5227 struct ocfs2_extent_rec split_rec;
5228 struct ocfs2_extent_rec *rec;
5229 struct ocfs2_insert_type insert;
5230
5231 /*
5232 * Setup the record to split before we grow the tree.
5233 */
5234 el = path_leaf_el(path);
5235 rec = &el->l_recs[index];
5236 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
5237 &split_rec, new_range, rec);
5238
5239 depth = path->p_tree_depth;
5240 if (depth > 0) {
5241 ret = ocfs2_read_extent_block(et->et_ci,
5242 ocfs2_et_get_last_eb_blk(et),
5243 &last_eb_bh);
5244 if (ret < 0) {
5245 mlog_errno(ret);
5246 goto out;
5247 }
5248
5249 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5250 rightmost_el = &eb->h_list;
5251 } else
5252 rightmost_el = path_leaf_el(path);
5253
5254 credits = path->p_tree_depth +
5255 ocfs2_extend_meta_needed(et->et_root_el);
5256 ret = ocfs2_extend_trans(handle, credits);
5257 if (ret) {
5258 mlog_errno(ret);
5259 goto out;
5260 }
5261
5262 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
5263 le16_to_cpu(rightmost_el->l_count)) {
5264 ret = ocfs2_grow_tree(handle, et, &depth, &last_eb_bh,
5265 meta_ac);
5266 if (ret) {
5267 mlog_errno(ret);
5268 goto out;
5269 }
5270 }
5271
5272 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
5273 insert.ins_appending = APPEND_NONE;
5274 insert.ins_contig = CONTIG_NONE;
5275 insert.ins_split = SPLIT_RIGHT;
5276 insert.ins_tree_depth = depth;
5277
5278 ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
5279 if (ret)
5280 mlog_errno(ret);
5281
5282out:
5283 brelse(last_eb_bh);
5284 return ret;
5285}
5286
5287static int ocfs2_truncate_rec(handle_t *handle,
5288 struct ocfs2_extent_tree *et,
5289 struct ocfs2_path *path, int index,
5290 struct ocfs2_cached_dealloc_ctxt *dealloc,
5291 u32 cpos, u32 len)
5292{
5293 int ret;
5294 u32 left_cpos, rec_range, trunc_range;
5295 int wants_rotate = 0, is_rightmost_tree_rec = 0;
5296 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5297 struct ocfs2_path *left_path = NULL;
5298 struct ocfs2_extent_list *el = path_leaf_el(path);
5299 struct ocfs2_extent_rec *rec;
5300 struct ocfs2_extent_block *eb;
5301
5302 if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
5303 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5304 if (ret) {
5305 mlog_errno(ret);
5306 goto out;
5307 }
5308
5309 index--;
5310 }
5311
5312 if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
5313 path->p_tree_depth) {
5314 /*
5315 * Check whether this is the rightmost tree record. If
5316 * we remove all of this record or part of its right
5317 * edge then an update of the record lengths above it
5318 * will be required.
5319 */
5320 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
5321 if (eb->h_next_leaf_blk == 0)
5322 is_rightmost_tree_rec = 1;
5323 }
5324
5325 rec = &el->l_recs[index];
5326 if (index == 0 && path->p_tree_depth &&
5327 le32_to_cpu(rec->e_cpos) == cpos) {
5328 /*
5329 * Changing the leftmost offset (via partial or whole
5330 * record truncate) of an interior (or rightmost) path
5331 * means we have to update the subtree that is formed
5332 * by this leaf and the one to it's left.
5333 *
5334 * There are two cases we can skip:
5335 * 1) Path is the leftmost one in our btree.
5336 * 2) The leaf is rightmost and will be empty after
5337 * we remove the extent record - the rotate code
5338 * knows how to update the newly formed edge.
5339 */
5340
5341 ret = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
5342 if (ret) {
5343 mlog_errno(ret);
5344 goto out;
5345 }
5346
5347 if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
5348 left_path = ocfs2_new_path_from_path(path);
5349 if (!left_path) {
5350 ret = -ENOMEM;
5351 mlog_errno(ret);
5352 goto out;
5353 }
5354
5355 ret = ocfs2_find_path(et->et_ci, left_path,
5356 left_cpos);
5357 if (ret) {
5358 mlog_errno(ret);
5359 goto out;
5360 }
5361 }
5362 }
5363
5364 ret = ocfs2_extend_rotate_transaction(handle, 0,
5365 handle->h_buffer_credits,
5366 path);
5367 if (ret) {
5368 mlog_errno(ret);
5369 goto out;
5370 }
5371
5372 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
5373 if (ret) {
5374 mlog_errno(ret);
5375 goto out;
5376 }
5377
5378 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
5379 if (ret) {
5380 mlog_errno(ret);
5381 goto out;
5382 }
5383
5384 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5385 trunc_range = cpos + len;
5386
5387 if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
5388 int next_free;
5389
5390 memset(rec, 0, sizeof(*rec));
5391 ocfs2_cleanup_merge(el, index);
5392 wants_rotate = 1;
5393
5394 next_free = le16_to_cpu(el->l_next_free_rec);
5395 if (is_rightmost_tree_rec && next_free > 1) {
5396 /*
5397 * We skip the edge update if this path will
5398 * be deleted by the rotate code.
5399 */
5400 rec = &el->l_recs[next_free - 1];
5401 ocfs2_adjust_rightmost_records(handle, et, path,
5402 rec);
5403 }
5404 } else if (le32_to_cpu(rec->e_cpos) == cpos) {
5405 /* Remove leftmost portion of the record. */
5406 le32_add_cpu(&rec->e_cpos, len);
5407 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
5408 le16_add_cpu(&rec->e_leaf_clusters, -len);
5409 } else if (rec_range == trunc_range) {
5410 /* Remove rightmost portion of the record */
5411 le16_add_cpu(&rec->e_leaf_clusters, -len);
5412 if (is_rightmost_tree_rec)
5413 ocfs2_adjust_rightmost_records(handle, et, path, rec);
5414 } else {
5415 /* Caller should have trapped this. */
5416 mlog(ML_ERROR, "Owner %llu: Invalid record truncate: (%u, %u) "
5417 "(%u, %u)\n",
5418 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5419 le32_to_cpu(rec->e_cpos),
5420 le16_to_cpu(rec->e_leaf_clusters), cpos, len);
5421 BUG();
5422 }
5423
5424 if (left_path) {
5425 int subtree_index;
5426
5427 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
5428 ocfs2_complete_edge_insert(handle, left_path, path,
5429 subtree_index);
5430 }
5431
5432 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5433
5434 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5435 if (ret) {
5436 mlog_errno(ret);
5437 goto out;
5438 }
5439
5440out:
5441 ocfs2_free_path(left_path);
5442 return ret;
5443}
5444
5445int ocfs2_remove_extent(handle_t *handle,
5446 struct ocfs2_extent_tree *et,
5447 u32 cpos, u32 len,
5448 struct ocfs2_alloc_context *meta_ac,
5449 struct ocfs2_cached_dealloc_ctxt *dealloc)
5450{
5451 int ret, index;
5452 u32 rec_range, trunc_range;
5453 struct ocfs2_extent_rec *rec;
5454 struct ocfs2_extent_list *el;
5455 struct ocfs2_path *path = NULL;
5456
5457 /*
5458 * XXX: Why are we truncating to 0 instead of wherever this
5459 * affects us?
5460 */
5461 ocfs2_et_extent_map_truncate(et, 0);
5462
5463 path = ocfs2_new_path_from_et(et);
5464 if (!path) {
5465 ret = -ENOMEM;
5466 mlog_errno(ret);
5467 goto out;
5468 }
5469
5470 ret = ocfs2_find_path(et->et_ci, path, cpos);
5471 if (ret) {
5472 mlog_errno(ret);
5473 goto out;
5474 }
5475
5476 el = path_leaf_el(path);
5477 index = ocfs2_search_extent_list(el, cpos);
5478 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
5479 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5480 "Owner %llu has an extent at cpos %u which can no "
5481 "longer be found.\n",
5482 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5483 cpos);
5484 ret = -EROFS;
5485 goto out;
5486 }
5487
5488 /*
5489 * We have 3 cases of extent removal:
5490 * 1) Range covers the entire extent rec
5491 * 2) Range begins or ends on one edge of the extent rec
5492 * 3) Range is in the middle of the extent rec (no shared edges)
5493 *
5494 * For case 1 we remove the extent rec and left rotate to
5495 * fill the hole.
5496 *
5497 * For case 2 we just shrink the existing extent rec, with a
5498 * tree update if the shrinking edge is also the edge of an
5499 * extent block.
5500 *
5501 * For case 3 we do a right split to turn the extent rec into
5502 * something case 2 can handle.
5503 */
5504 rec = &el->l_recs[index];
5505 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5506 trunc_range = cpos + len;
5507
5508 BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
5509
5510 trace_ocfs2_remove_extent(
5511 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5512 cpos, len, index, le32_to_cpu(rec->e_cpos),
5513 ocfs2_rec_clusters(el, rec));
5514
5515 if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
5516 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5517 cpos, len);
5518 if (ret) {
5519 mlog_errno(ret);
5520 goto out;
5521 }
5522 } else {
5523 ret = ocfs2_split_tree(handle, et, path, index,
5524 trunc_range, meta_ac);
5525 if (ret) {
5526 mlog_errno(ret);
5527 goto out;
5528 }
5529
5530 /*
5531 * The split could have manipulated the tree enough to
5532 * move the record location, so we have to look for it again.
5533 */
5534 ocfs2_reinit_path(path, 1);
5535
5536 ret = ocfs2_find_path(et->et_ci, path, cpos);
5537 if (ret) {
5538 mlog_errno(ret);
5539 goto out;
5540 }
5541
5542 el = path_leaf_el(path);
5543 index = ocfs2_search_extent_list(el, cpos);
5544 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
5545 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5546 "Owner %llu: split at cpos %u lost record.",
5547 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5548 cpos);
5549 ret = -EROFS;
5550 goto out;
5551 }
5552
5553 /*
5554 * Double check our values here. If anything is fishy,
5555 * it's easier to catch it at the top level.
5556 */
5557 rec = &el->l_recs[index];
5558 rec_range = le32_to_cpu(rec->e_cpos) +
5559 ocfs2_rec_clusters(el, rec);
5560 if (rec_range != trunc_range) {
5561 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5562 "Owner %llu: error after split at cpos %u"
5563 "trunc len %u, existing record is (%u,%u)",
5564 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5565 cpos, len, le32_to_cpu(rec->e_cpos),
5566 ocfs2_rec_clusters(el, rec));
5567 ret = -EROFS;
5568 goto out;
5569 }
5570
5571 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5572 cpos, len);
5573 if (ret) {
5574 mlog_errno(ret);
5575 goto out;
5576 }
5577 }
5578
5579out:
5580 ocfs2_free_path(path);
5581 return ret;
5582}
5583
5584/*
5585 * ocfs2_reserve_blocks_for_rec_trunc() would look basically the
5586 * same as ocfs2_lock_alloctors(), except for it accepts a blocks
5587 * number to reserve some extra blocks, and it only handles meta
5588 * data allocations.
5589 *
5590 * Currently, only ocfs2_remove_btree_range() uses it for truncating
5591 * and punching holes.
5592 */
5593static int ocfs2_reserve_blocks_for_rec_trunc(struct inode *inode,
5594 struct ocfs2_extent_tree *et,
5595 u32 extents_to_split,
5596 struct ocfs2_alloc_context **ac,
5597 int extra_blocks)
5598{
5599 int ret = 0, num_free_extents;
5600 unsigned int max_recs_needed = 2 * extents_to_split;
5601 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5602
5603 *ac = NULL;
5604
5605 num_free_extents = ocfs2_num_free_extents(osb, et);
5606 if (num_free_extents < 0) {
5607 ret = num_free_extents;
5608 mlog_errno(ret);
5609 goto out;
5610 }
5611
5612 if (!num_free_extents ||
5613 (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
5614 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
5615
5616 if (extra_blocks) {
5617 ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, ac);
5618 if (ret < 0) {
5619 if (ret != -ENOSPC)
5620 mlog_errno(ret);
5621 goto out;
5622 }
5623 }
5624
5625out:
5626 if (ret) {
5627 if (*ac) {
5628 ocfs2_free_alloc_context(*ac);
5629 *ac = NULL;
5630 }
5631 }
5632
5633 return ret;
5634}
5635
5636int ocfs2_remove_btree_range(struct inode *inode,
5637 struct ocfs2_extent_tree *et,
5638 u32 cpos, u32 phys_cpos, u32 len, int flags,
5639 struct ocfs2_cached_dealloc_ctxt *dealloc,
5640 u64 refcount_loc)
5641{
5642 int ret, credits = 0, extra_blocks = 0;
5643 u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
5644 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5645 struct inode *tl_inode = osb->osb_tl_inode;
5646 handle_t *handle;
5647 struct ocfs2_alloc_context *meta_ac = NULL;
5648 struct ocfs2_refcount_tree *ref_tree = NULL;
5649
5650 if ((flags & OCFS2_EXT_REFCOUNTED) && len) {
5651 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
5652 OCFS2_HAS_REFCOUNT_FL));
5653
5654 ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
5655 &ref_tree, NULL);
5656 if (ret) {
5657 mlog_errno(ret);
5658 goto out;
5659 }
5660
5661 ret = ocfs2_prepare_refcount_change_for_del(inode,
5662 refcount_loc,
5663 phys_blkno,
5664 len,
5665 &credits,
5666 &extra_blocks);
5667 if (ret < 0) {
5668 mlog_errno(ret);
5669 goto out;
5670 }
5671 }
5672
5673 ret = ocfs2_reserve_blocks_for_rec_trunc(inode, et, 1, &meta_ac,
5674 extra_blocks);
5675 if (ret) {
5676 mlog_errno(ret);
5677 return ret;
5678 }
5679
5680 mutex_lock(&tl_inode->i_mutex);
5681
5682 if (ocfs2_truncate_log_needs_flush(osb)) {
5683 ret = __ocfs2_flush_truncate_log(osb);
5684 if (ret < 0) {
5685 mlog_errno(ret);
5686 goto out;
5687 }
5688 }
5689
5690 handle = ocfs2_start_trans(osb,
5691 ocfs2_remove_extent_credits(osb->sb) + credits);
5692 if (IS_ERR(handle)) {
5693 ret = PTR_ERR(handle);
5694 mlog_errno(ret);
5695 goto out;
5696 }
5697
5698 ret = ocfs2_et_root_journal_access(handle, et,
5699 OCFS2_JOURNAL_ACCESS_WRITE);
5700 if (ret) {
5701 mlog_errno(ret);
5702 goto out;
5703 }
5704
5705 dquot_free_space_nodirty(inode,
5706 ocfs2_clusters_to_bytes(inode->i_sb, len));
5707
5708 ret = ocfs2_remove_extent(handle, et, cpos, len, meta_ac, dealloc);
5709 if (ret) {
5710 mlog_errno(ret);
5711 goto out_commit;
5712 }
5713
5714 ocfs2_et_update_clusters(et, -len);
5715
5716 ocfs2_journal_dirty(handle, et->et_root_bh);
5717
5718 if (phys_blkno) {
5719 if (flags & OCFS2_EXT_REFCOUNTED)
5720 ret = ocfs2_decrease_refcount(inode, handle,
5721 ocfs2_blocks_to_clusters(osb->sb,
5722 phys_blkno),
5723 len, meta_ac,
5724 dealloc, 1);
5725 else
5726 ret = ocfs2_truncate_log_append(osb, handle,
5727 phys_blkno, len);
5728 if (ret)
5729 mlog_errno(ret);
5730
5731 }
5732
5733out_commit:
5734 ocfs2_commit_trans(osb, handle);
5735out:
5736 mutex_unlock(&tl_inode->i_mutex);
5737
5738 if (meta_ac)
5739 ocfs2_free_alloc_context(meta_ac);
5740
5741 if (ref_tree)
5742 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
5743
5744 return ret;
5745}
5746
5747int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
5748{
5749 struct buffer_head *tl_bh = osb->osb_tl_bh;
5750 struct ocfs2_dinode *di;
5751 struct ocfs2_truncate_log *tl;
5752
5753 di = (struct ocfs2_dinode *) tl_bh->b_data;
5754 tl = &di->id2.i_dealloc;
5755
5756 mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
5757 "slot %d, invalid truncate log parameters: used = "
5758 "%u, count = %u\n", osb->slot_num,
5759 le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
5760 return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
5761}
5762
5763static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
5764 unsigned int new_start)
5765{
5766 unsigned int tail_index;
5767 unsigned int current_tail;
5768
5769 /* No records, nothing to coalesce */
5770 if (!le16_to_cpu(tl->tl_used))
5771 return 0;
5772
5773 tail_index = le16_to_cpu(tl->tl_used) - 1;
5774 current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
5775 current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
5776
5777 return current_tail == new_start;
5778}
5779
5780int ocfs2_truncate_log_append(struct ocfs2_super *osb,
5781 handle_t *handle,
5782 u64 start_blk,
5783 unsigned int num_clusters)
5784{
5785 int status, index;
5786 unsigned int start_cluster, tl_count;
5787 struct inode *tl_inode = osb->osb_tl_inode;
5788 struct buffer_head *tl_bh = osb->osb_tl_bh;
5789 struct ocfs2_dinode *di;
5790 struct ocfs2_truncate_log *tl;
5791
5792 BUG_ON(mutex_trylock(&tl_inode->i_mutex));
5793
5794 start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
5795
5796 di = (struct ocfs2_dinode *) tl_bh->b_data;
5797
5798 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5799 * by the underlying call to ocfs2_read_inode_block(), so any
5800 * corruption is a code bug */
5801 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5802
5803 tl = &di->id2.i_dealloc;
5804 tl_count = le16_to_cpu(tl->tl_count);
5805 mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
5806 tl_count == 0,
5807 "Truncate record count on #%llu invalid "
5808 "wanted %u, actual %u\n",
5809 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5810 ocfs2_truncate_recs_per_inode(osb->sb),
5811 le16_to_cpu(tl->tl_count));
5812
5813 /* Caller should have known to flush before calling us. */
5814 index = le16_to_cpu(tl->tl_used);
5815 if (index >= tl_count) {
5816 status = -ENOSPC;
5817 mlog_errno(status);
5818 goto bail;
5819 }
5820
5821 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5822 OCFS2_JOURNAL_ACCESS_WRITE);
5823 if (status < 0) {
5824 mlog_errno(status);
5825 goto bail;
5826 }
5827
5828 trace_ocfs2_truncate_log_append(
5829 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index,
5830 start_cluster, num_clusters);
5831 if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
5832 /*
5833 * Move index back to the record we are coalescing with.
5834 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5835 */
5836 index--;
5837
5838 num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
5839 trace_ocfs2_truncate_log_append(
5840 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5841 index, le32_to_cpu(tl->tl_recs[index].t_start),
5842 num_clusters);
5843 } else {
5844 tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
5845 tl->tl_used = cpu_to_le16(index + 1);
5846 }
5847 tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
5848
5849 ocfs2_journal_dirty(handle, tl_bh);
5850
5851 osb->truncated_clusters += num_clusters;
5852bail:
5853 return status;
5854}
5855
5856static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
5857 handle_t *handle,
5858 struct inode *data_alloc_inode,
5859 struct buffer_head *data_alloc_bh)
5860{
5861 int status = 0;
5862 int i;
5863 unsigned int num_clusters;
5864 u64 start_blk;
5865 struct ocfs2_truncate_rec rec;
5866 struct ocfs2_dinode *di;
5867 struct ocfs2_truncate_log *tl;
5868 struct inode *tl_inode = osb->osb_tl_inode;
5869 struct buffer_head *tl_bh = osb->osb_tl_bh;
5870
5871 di = (struct ocfs2_dinode *) tl_bh->b_data;
5872 tl = &di->id2.i_dealloc;
5873 i = le16_to_cpu(tl->tl_used) - 1;
5874 while (i >= 0) {
5875 /* Caller has given us at least enough credits to
5876 * update the truncate log dinode */
5877 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5878 OCFS2_JOURNAL_ACCESS_WRITE);
5879 if (status < 0) {
5880 mlog_errno(status);
5881 goto bail;
5882 }
5883
5884 tl->tl_used = cpu_to_le16(i);
5885
5886 ocfs2_journal_dirty(handle, tl_bh);
5887
5888 /* TODO: Perhaps we can calculate the bulk of the
5889 * credits up front rather than extending like
5890 * this. */
5891 status = ocfs2_extend_trans(handle,
5892 OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
5893 if (status < 0) {
5894 mlog_errno(status);
5895 goto bail;
5896 }
5897
5898 rec = tl->tl_recs[i];
5899 start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
5900 le32_to_cpu(rec.t_start));
5901 num_clusters = le32_to_cpu(rec.t_clusters);
5902
5903 /* if start_blk is not set, we ignore the record as
5904 * invalid. */
5905 if (start_blk) {
5906 trace_ocfs2_replay_truncate_records(
5907 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5908 i, le32_to_cpu(rec.t_start), num_clusters);
5909
5910 status = ocfs2_free_clusters(handle, data_alloc_inode,
5911 data_alloc_bh, start_blk,
5912 num_clusters);
5913 if (status < 0) {
5914 mlog_errno(status);
5915 goto bail;
5916 }
5917 }
5918 i--;
5919 }
5920
5921 osb->truncated_clusters = 0;
5922
5923bail:
5924 return status;
5925}
5926
5927/* Expects you to already be holding tl_inode->i_mutex */
5928int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5929{
5930 int status;
5931 unsigned int num_to_flush;
5932 handle_t *handle;
5933 struct inode *tl_inode = osb->osb_tl_inode;
5934 struct inode *data_alloc_inode = NULL;
5935 struct buffer_head *tl_bh = osb->osb_tl_bh;
5936 struct buffer_head *data_alloc_bh = NULL;
5937 struct ocfs2_dinode *di;
5938 struct ocfs2_truncate_log *tl;
5939
5940 BUG_ON(mutex_trylock(&tl_inode->i_mutex));
5941
5942 di = (struct ocfs2_dinode *) tl_bh->b_data;
5943
5944 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5945 * by the underlying call to ocfs2_read_inode_block(), so any
5946 * corruption is a code bug */
5947 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5948
5949 tl = &di->id2.i_dealloc;
5950 num_to_flush = le16_to_cpu(tl->tl_used);
5951 trace_ocfs2_flush_truncate_log(
5952 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5953 num_to_flush);
5954 if (!num_to_flush) {
5955 status = 0;
5956 goto out;
5957 }
5958
5959 data_alloc_inode = ocfs2_get_system_file_inode(osb,
5960 GLOBAL_BITMAP_SYSTEM_INODE,
5961 OCFS2_INVALID_SLOT);
5962 if (!data_alloc_inode) {
5963 status = -EINVAL;
5964 mlog(ML_ERROR, "Could not get bitmap inode!\n");
5965 goto out;
5966 }
5967
5968 mutex_lock(&data_alloc_inode->i_mutex);
5969
5970 status = ocfs2_inode_lock(data_alloc_inode, &data_alloc_bh, 1);
5971 if (status < 0) {
5972 mlog_errno(status);
5973 goto out_mutex;
5974 }
5975
5976 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
5977 if (IS_ERR(handle)) {
5978 status = PTR_ERR(handle);
5979 mlog_errno(status);
5980 goto out_unlock;
5981 }
5982
5983 status = ocfs2_replay_truncate_records(osb, handle, data_alloc_inode,
5984 data_alloc_bh);
5985 if (status < 0)
5986 mlog_errno(status);
5987
5988 ocfs2_commit_trans(osb, handle);
5989
5990out_unlock:
5991 brelse(data_alloc_bh);
5992 ocfs2_inode_unlock(data_alloc_inode, 1);
5993
5994out_mutex:
5995 mutex_unlock(&data_alloc_inode->i_mutex);
5996 iput(data_alloc_inode);
5997
5998out:
5999 return status;
6000}
6001
6002int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
6003{
6004 int status;
6005 struct inode *tl_inode = osb->osb_tl_inode;
6006
6007 mutex_lock(&tl_inode->i_mutex);
6008 status = __ocfs2_flush_truncate_log(osb);
6009 mutex_unlock(&tl_inode->i_mutex);
6010
6011 return status;
6012}
6013
6014static void ocfs2_truncate_log_worker(struct work_struct *work)
6015{
6016 int status;
6017 struct ocfs2_super *osb =
6018 container_of(work, struct ocfs2_super,
6019 osb_truncate_log_wq.work);
6020
6021 status = ocfs2_flush_truncate_log(osb);
6022 if (status < 0)
6023 mlog_errno(status);
6024 else
6025 ocfs2_init_steal_slots(osb);
6026}
6027
6028#define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
6029void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
6030 int cancel)
6031{
6032 if (osb->osb_tl_inode) {
6033 /* We want to push off log flushes while truncates are
6034 * still running. */
6035 if (cancel)
6036 cancel_delayed_work(&osb->osb_truncate_log_wq);
6037
6038 queue_delayed_work(ocfs2_wq, &osb->osb_truncate_log_wq,
6039 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
6040 }
6041}
6042
6043static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
6044 int slot_num,
6045 struct inode **tl_inode,
6046 struct buffer_head **tl_bh)
6047{
6048 int status;
6049 struct inode *inode = NULL;
6050 struct buffer_head *bh = NULL;
6051
6052 inode = ocfs2_get_system_file_inode(osb,
6053 TRUNCATE_LOG_SYSTEM_INODE,
6054 slot_num);
6055 if (!inode) {
6056 status = -EINVAL;
6057 mlog(ML_ERROR, "Could not get load truncate log inode!\n");
6058 goto bail;
6059 }
6060
6061 status = ocfs2_read_inode_block(inode, &bh);
6062 if (status < 0) {
6063 iput(inode);
6064 mlog_errno(status);
6065 goto bail;
6066 }
6067
6068 *tl_inode = inode;
6069 *tl_bh = bh;
6070bail:
6071 return status;
6072}
6073
6074/* called during the 1st stage of node recovery. we stamp a clean
6075 * truncate log and pass back a copy for processing later. if the
6076 * truncate log does not require processing, a *tl_copy is set to
6077 * NULL. */
6078int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
6079 int slot_num,
6080 struct ocfs2_dinode **tl_copy)
6081{
6082 int status;
6083 struct inode *tl_inode = NULL;
6084 struct buffer_head *tl_bh = NULL;
6085 struct ocfs2_dinode *di;
6086 struct ocfs2_truncate_log *tl;
6087
6088 *tl_copy = NULL;
6089
6090 trace_ocfs2_begin_truncate_log_recovery(slot_num);
6091
6092 status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
6093 if (status < 0) {
6094 mlog_errno(status);
6095 goto bail;
6096 }
6097
6098 di = (struct ocfs2_dinode *) tl_bh->b_data;
6099
6100 /* tl_bh is loaded from ocfs2_get_truncate_log_info(). It's
6101 * validated by the underlying call to ocfs2_read_inode_block(),
6102 * so any corruption is a code bug */
6103 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6104
6105 tl = &di->id2.i_dealloc;
6106 if (le16_to_cpu(tl->tl_used)) {
6107 trace_ocfs2_truncate_log_recovery_num(le16_to_cpu(tl->tl_used));
6108
6109 *tl_copy = kmalloc(tl_bh->b_size, GFP_KERNEL);
6110 if (!(*tl_copy)) {
6111 status = -ENOMEM;
6112 mlog_errno(status);
6113 goto bail;
6114 }
6115
6116 /* Assuming the write-out below goes well, this copy
6117 * will be passed back to recovery for processing. */
6118 memcpy(*tl_copy, tl_bh->b_data, tl_bh->b_size);
6119
6120 /* All we need to do to clear the truncate log is set
6121 * tl_used. */
6122 tl->tl_used = 0;
6123
6124 ocfs2_compute_meta_ecc(osb->sb, tl_bh->b_data, &di->i_check);
6125 status = ocfs2_write_block(osb, tl_bh, INODE_CACHE(tl_inode));
6126 if (status < 0) {
6127 mlog_errno(status);
6128 goto bail;
6129 }
6130 }
6131
6132bail:
6133 if (tl_inode)
6134 iput(tl_inode);
6135 brelse(tl_bh);
6136
6137 if (status < 0 && (*tl_copy)) {
6138 kfree(*tl_copy);
6139 *tl_copy = NULL;
6140 mlog_errno(status);
6141 }
6142
6143 return status;
6144}
6145
6146int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
6147 struct ocfs2_dinode *tl_copy)
6148{
6149 int status = 0;
6150 int i;
6151 unsigned int clusters, num_recs, start_cluster;
6152 u64 start_blk;
6153 handle_t *handle;
6154 struct inode *tl_inode = osb->osb_tl_inode;
6155 struct ocfs2_truncate_log *tl;
6156
6157 if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
6158 mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
6159 return -EINVAL;
6160 }
6161
6162 tl = &tl_copy->id2.i_dealloc;
6163 num_recs = le16_to_cpu(tl->tl_used);
6164 trace_ocfs2_complete_truncate_log_recovery(
6165 (unsigned long long)le64_to_cpu(tl_copy->i_blkno),
6166 num_recs);
6167
6168 mutex_lock(&tl_inode->i_mutex);
6169 for(i = 0; i < num_recs; i++) {
6170 if (ocfs2_truncate_log_needs_flush(osb)) {
6171 status = __ocfs2_flush_truncate_log(osb);
6172 if (status < 0) {
6173 mlog_errno(status);
6174 goto bail_up;
6175 }
6176 }
6177
6178 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6179 if (IS_ERR(handle)) {
6180 status = PTR_ERR(handle);
6181 mlog_errno(status);
6182 goto bail_up;
6183 }
6184
6185 clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
6186 start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
6187 start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
6188
6189 status = ocfs2_truncate_log_append(osb, handle,
6190 start_blk, clusters);
6191 ocfs2_commit_trans(osb, handle);
6192 if (status < 0) {
6193 mlog_errno(status);
6194 goto bail_up;
6195 }
6196 }
6197
6198bail_up:
6199 mutex_unlock(&tl_inode->i_mutex);
6200
6201 return status;
6202}
6203
6204void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
6205{
6206 int status;
6207 struct inode *tl_inode = osb->osb_tl_inode;
6208
6209 if (tl_inode) {
6210 cancel_delayed_work(&osb->osb_truncate_log_wq);
6211 flush_workqueue(ocfs2_wq);
6212
6213 status = ocfs2_flush_truncate_log(osb);
6214 if (status < 0)
6215 mlog_errno(status);
6216
6217 brelse(osb->osb_tl_bh);
6218 iput(osb->osb_tl_inode);
6219 }
6220}
6221
6222int ocfs2_truncate_log_init(struct ocfs2_super *osb)
6223{
6224 int status;
6225 struct inode *tl_inode = NULL;
6226 struct buffer_head *tl_bh = NULL;
6227
6228 status = ocfs2_get_truncate_log_info(osb,
6229 osb->slot_num,
6230 &tl_inode,
6231 &tl_bh);
6232 if (status < 0)
6233 mlog_errno(status);
6234
6235 /* ocfs2_truncate_log_shutdown keys on the existence of
6236 * osb->osb_tl_inode so we don't set any of the osb variables
6237 * until we're sure all is well. */
6238 INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
6239 ocfs2_truncate_log_worker);
6240 osb->osb_tl_bh = tl_bh;
6241 osb->osb_tl_inode = tl_inode;
6242
6243 return status;
6244}
6245
6246/*
6247 * Delayed de-allocation of suballocator blocks.
6248 *
6249 * Some sets of block de-allocations might involve multiple suballocator inodes.
6250 *
6251 * The locking for this can get extremely complicated, especially when
6252 * the suballocator inodes to delete from aren't known until deep
6253 * within an unrelated codepath.
6254 *
6255 * ocfs2_extent_block structures are a good example of this - an inode
6256 * btree could have been grown by any number of nodes each allocating
6257 * out of their own suballoc inode.
6258 *
6259 * These structures allow the delay of block de-allocation until a
6260 * later time, when locking of multiple cluster inodes won't cause
6261 * deadlock.
6262 */
6263
6264/*
6265 * Describe a single bit freed from a suballocator. For the block
6266 * suballocators, it represents one block. For the global cluster
6267 * allocator, it represents some clusters and free_bit indicates
6268 * clusters number.
6269 */
6270struct ocfs2_cached_block_free {
6271 struct ocfs2_cached_block_free *free_next;
6272 u64 free_bg;
6273 u64 free_blk;
6274 unsigned int free_bit;
6275};
6276
6277struct ocfs2_per_slot_free_list {
6278 struct ocfs2_per_slot_free_list *f_next_suballocator;
6279 int f_inode_type;
6280 int f_slot;
6281 struct ocfs2_cached_block_free *f_first;
6282};
6283
6284static int ocfs2_free_cached_blocks(struct ocfs2_super *osb,
6285 int sysfile_type,
6286 int slot,
6287 struct ocfs2_cached_block_free *head)
6288{
6289 int ret;
6290 u64 bg_blkno;
6291 handle_t *handle;
6292 struct inode *inode;
6293 struct buffer_head *di_bh = NULL;
6294 struct ocfs2_cached_block_free *tmp;
6295
6296 inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
6297 if (!inode) {
6298 ret = -EINVAL;
6299 mlog_errno(ret);
6300 goto out;
6301 }
6302
6303 mutex_lock(&inode->i_mutex);
6304
6305 ret = ocfs2_inode_lock(inode, &di_bh, 1);
6306 if (ret) {
6307 mlog_errno(ret);
6308 goto out_mutex;
6309 }
6310
6311 handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
6312 if (IS_ERR(handle)) {
6313 ret = PTR_ERR(handle);
6314 mlog_errno(ret);
6315 goto out_unlock;
6316 }
6317
6318 while (head) {
6319 if (head->free_bg)
6320 bg_blkno = head->free_bg;
6321 else
6322 bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
6323 head->free_bit);
6324 trace_ocfs2_free_cached_blocks(
6325 (unsigned long long)head->free_blk, head->free_bit);
6326
6327 ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
6328 head->free_bit, bg_blkno, 1);
6329 if (ret) {
6330 mlog_errno(ret);
6331 goto out_journal;
6332 }
6333
6334 ret = ocfs2_extend_trans(handle, OCFS2_SUBALLOC_FREE);
6335 if (ret) {
6336 mlog_errno(ret);
6337 goto out_journal;
6338 }
6339
6340 tmp = head;
6341 head = head->free_next;
6342 kfree(tmp);
6343 }
6344
6345out_journal:
6346 ocfs2_commit_trans(osb, handle);
6347
6348out_unlock:
6349 ocfs2_inode_unlock(inode, 1);
6350 brelse(di_bh);
6351out_mutex:
6352 mutex_unlock(&inode->i_mutex);
6353 iput(inode);
6354out:
6355 while(head) {
6356 /* Premature exit may have left some dangling items. */
6357 tmp = head;
6358 head = head->free_next;
6359 kfree(tmp);
6360 }
6361
6362 return ret;
6363}
6364
6365int ocfs2_cache_cluster_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6366 u64 blkno, unsigned int bit)
6367{
6368 int ret = 0;
6369 struct ocfs2_cached_block_free *item;
6370
6371 item = kzalloc(sizeof(*item), GFP_NOFS);
6372 if (item == NULL) {
6373 ret = -ENOMEM;
6374 mlog_errno(ret);
6375 return ret;
6376 }
6377
6378 trace_ocfs2_cache_cluster_dealloc((unsigned long long)blkno, bit);
6379
6380 item->free_blk = blkno;
6381 item->free_bit = bit;
6382 item->free_next = ctxt->c_global_allocator;
6383
6384 ctxt->c_global_allocator = item;
6385 return ret;
6386}
6387
6388static int ocfs2_free_cached_clusters(struct ocfs2_super *osb,
6389 struct ocfs2_cached_block_free *head)
6390{
6391 struct ocfs2_cached_block_free *tmp;
6392 struct inode *tl_inode = osb->osb_tl_inode;
6393 handle_t *handle;
6394 int ret = 0;
6395
6396 mutex_lock(&tl_inode->i_mutex);
6397
6398 while (head) {
6399 if (ocfs2_truncate_log_needs_flush(osb)) {
6400 ret = __ocfs2_flush_truncate_log(osb);
6401 if (ret < 0) {
6402 mlog_errno(ret);
6403 break;
6404 }
6405 }
6406
6407 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6408 if (IS_ERR(handle)) {
6409 ret = PTR_ERR(handle);
6410 mlog_errno(ret);
6411 break;
6412 }
6413
6414 ret = ocfs2_truncate_log_append(osb, handle, head->free_blk,
6415 head->free_bit);
6416
6417 ocfs2_commit_trans(osb, handle);
6418 tmp = head;
6419 head = head->free_next;
6420 kfree(tmp);
6421
6422 if (ret < 0) {
6423 mlog_errno(ret);
6424 break;
6425 }
6426 }
6427
6428 mutex_unlock(&tl_inode->i_mutex);
6429
6430 while (head) {
6431 /* Premature exit may have left some dangling items. */
6432 tmp = head;
6433 head = head->free_next;
6434 kfree(tmp);
6435 }
6436
6437 return ret;
6438}
6439
6440int ocfs2_run_deallocs(struct ocfs2_super *osb,
6441 struct ocfs2_cached_dealloc_ctxt *ctxt)
6442{
6443 int ret = 0, ret2;
6444 struct ocfs2_per_slot_free_list *fl;
6445
6446 if (!ctxt)
6447 return 0;
6448
6449 while (ctxt->c_first_suballocator) {
6450 fl = ctxt->c_first_suballocator;
6451
6452 if (fl->f_first) {
6453 trace_ocfs2_run_deallocs(fl->f_inode_type,
6454 fl->f_slot);
6455 ret2 = ocfs2_free_cached_blocks(osb,
6456 fl->f_inode_type,
6457 fl->f_slot,
6458 fl->f_first);
6459 if (ret2)
6460 mlog_errno(ret2);
6461 if (!ret)
6462 ret = ret2;
6463 }
6464
6465 ctxt->c_first_suballocator = fl->f_next_suballocator;
6466 kfree(fl);
6467 }
6468
6469 if (ctxt->c_global_allocator) {
6470 ret2 = ocfs2_free_cached_clusters(osb,
6471 ctxt->c_global_allocator);
6472 if (ret2)
6473 mlog_errno(ret2);
6474 if (!ret)
6475 ret = ret2;
6476
6477 ctxt->c_global_allocator = NULL;
6478 }
6479
6480 return ret;
6481}
6482
6483static struct ocfs2_per_slot_free_list *
6484ocfs2_find_per_slot_free_list(int type,
6485 int slot,
6486 struct ocfs2_cached_dealloc_ctxt *ctxt)
6487{
6488 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6489
6490 while (fl) {
6491 if (fl->f_inode_type == type && fl->f_slot == slot)
6492 return fl;
6493
6494 fl = fl->f_next_suballocator;
6495 }
6496
6497 fl = kmalloc(sizeof(*fl), GFP_NOFS);
6498 if (fl) {
6499 fl->f_inode_type = type;
6500 fl->f_slot = slot;
6501 fl->f_first = NULL;
6502 fl->f_next_suballocator = ctxt->c_first_suballocator;
6503
6504 ctxt->c_first_suballocator = fl;
6505 }
6506 return fl;
6507}
6508
6509int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6510 int type, int slot, u64 suballoc,
6511 u64 blkno, unsigned int bit)
6512{
6513 int ret;
6514 struct ocfs2_per_slot_free_list *fl;
6515 struct ocfs2_cached_block_free *item;
6516
6517 fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
6518 if (fl == NULL) {
6519 ret = -ENOMEM;
6520 mlog_errno(ret);
6521 goto out;
6522 }
6523
6524 item = kzalloc(sizeof(*item), GFP_NOFS);
6525 if (item == NULL) {
6526 ret = -ENOMEM;
6527 mlog_errno(ret);
6528 goto out;
6529 }
6530
6531 trace_ocfs2_cache_block_dealloc(type, slot,
6532 (unsigned long long)suballoc,
6533 (unsigned long long)blkno, bit);
6534
6535 item->free_bg = suballoc;
6536 item->free_blk = blkno;
6537 item->free_bit = bit;
6538 item->free_next = fl->f_first;
6539
6540 fl->f_first = item;
6541
6542 ret = 0;
6543out:
6544 return ret;
6545}
6546
6547static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
6548 struct ocfs2_extent_block *eb)
6549{
6550 return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
6551 le16_to_cpu(eb->h_suballoc_slot),
6552 le64_to_cpu(eb->h_suballoc_loc),
6553 le64_to_cpu(eb->h_blkno),
6554 le16_to_cpu(eb->h_suballoc_bit));
6555}
6556
6557static int ocfs2_zero_func(handle_t *handle, struct buffer_head *bh)
6558{
6559 set_buffer_uptodate(bh);
6560 mark_buffer_dirty(bh);
6561 return 0;
6562}
6563
6564void ocfs2_map_and_dirty_page(struct inode *inode, handle_t *handle,
6565 unsigned int from, unsigned int to,
6566 struct page *page, int zero, u64 *phys)
6567{
6568 int ret, partial = 0;
6569
6570 ret = ocfs2_map_page_blocks(page, phys, inode, from, to, 0);
6571 if (ret)
6572 mlog_errno(ret);
6573
6574 if (zero)
6575 zero_user_segment(page, from, to);
6576
6577 /*
6578 * Need to set the buffers we zero'd into uptodate
6579 * here if they aren't - ocfs2_map_page_blocks()
6580 * might've skipped some
6581 */
6582 ret = walk_page_buffers(handle, page_buffers(page),
6583 from, to, &partial,
6584 ocfs2_zero_func);
6585 if (ret < 0)
6586 mlog_errno(ret);
6587 else if (ocfs2_should_order_data(inode)) {
6588 ret = ocfs2_jbd2_file_inode(handle, inode);
6589 if (ret < 0)
6590 mlog_errno(ret);
6591 }
6592
6593 if (!partial)
6594 SetPageUptodate(page);
6595
6596 flush_dcache_page(page);
6597}
6598
6599static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start,
6600 loff_t end, struct page **pages,
6601 int numpages, u64 phys, handle_t *handle)
6602{
6603 int i;
6604 struct page *page;
6605 unsigned int from, to = PAGE_CACHE_SIZE;
6606 struct super_block *sb = inode->i_sb;
6607
6608 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
6609
6610 if (numpages == 0)
6611 goto out;
6612
6613 to = PAGE_CACHE_SIZE;
6614 for(i = 0; i < numpages; i++) {
6615 page = pages[i];
6616
6617 from = start & (PAGE_CACHE_SIZE - 1);
6618 if ((end >> PAGE_CACHE_SHIFT) == page->index)
6619 to = end & (PAGE_CACHE_SIZE - 1);
6620
6621 BUG_ON(from > PAGE_CACHE_SIZE);
6622 BUG_ON(to > PAGE_CACHE_SIZE);
6623
6624 ocfs2_map_and_dirty_page(inode, handle, from, to, page, 1,
6625 &phys);
6626
6627 start = (page->index + 1) << PAGE_CACHE_SHIFT;
6628 }
6629out:
6630 if (pages)
6631 ocfs2_unlock_and_free_pages(pages, numpages);
6632}
6633
6634int ocfs2_grab_pages(struct inode *inode, loff_t start, loff_t end,
6635 struct page **pages, int *num)
6636{
6637 int numpages, ret = 0;
6638 struct address_space *mapping = inode->i_mapping;
6639 unsigned long index;
6640 loff_t last_page_bytes;
6641
6642 BUG_ON(start > end);
6643
6644 numpages = 0;
6645 last_page_bytes = PAGE_ALIGN(end);
6646 index = start >> PAGE_CACHE_SHIFT;
6647 do {
6648 pages[numpages] = find_or_create_page(mapping, index, GFP_NOFS);
6649 if (!pages[numpages]) {
6650 ret = -ENOMEM;
6651 mlog_errno(ret);
6652 goto out;
6653 }
6654
6655 numpages++;
6656 index++;
6657 } while (index < (last_page_bytes >> PAGE_CACHE_SHIFT));
6658
6659out:
6660 if (ret != 0) {
6661 if (pages)
6662 ocfs2_unlock_and_free_pages(pages, numpages);
6663 numpages = 0;
6664 }
6665
6666 *num = numpages;
6667
6668 return ret;
6669}
6670
6671static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end,
6672 struct page **pages, int *num)
6673{
6674 struct super_block *sb = inode->i_sb;
6675
6676 BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
6677 (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
6678
6679 return ocfs2_grab_pages(inode, start, end, pages, num);
6680}
6681
6682/*
6683 * Zero the area past i_size but still within an allocated
6684 * cluster. This avoids exposing nonzero data on subsequent file
6685 * extends.
6686 *
6687 * We need to call this before i_size is updated on the inode because
6688 * otherwise block_write_full_page() will skip writeout of pages past
6689 * i_size. The new_i_size parameter is passed for this reason.
6690 */
6691int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
6692 u64 range_start, u64 range_end)
6693{
6694 int ret = 0, numpages;
6695 struct page **pages = NULL;
6696 u64 phys;
6697 unsigned int ext_flags;
6698 struct super_block *sb = inode->i_sb;
6699
6700 /*
6701 * File systems which don't support sparse files zero on every
6702 * extend.
6703 */
6704 if (!ocfs2_sparse_alloc(OCFS2_SB(sb)))
6705 return 0;
6706
6707 pages = kcalloc(ocfs2_pages_per_cluster(sb),
6708 sizeof(struct page *), GFP_NOFS);
6709 if (pages == NULL) {
6710 ret = -ENOMEM;
6711 mlog_errno(ret);
6712 goto out;
6713 }
6714
6715 if (range_start == range_end)
6716 goto out;
6717
6718 ret = ocfs2_extent_map_get_blocks(inode,
6719 range_start >> sb->s_blocksize_bits,
6720 &phys, NULL, &ext_flags);
6721 if (ret) {
6722 mlog_errno(ret);
6723 goto out;
6724 }
6725
6726 /*
6727 * Tail is a hole, or is marked unwritten. In either case, we
6728 * can count on read and write to return/push zero's.
6729 */
6730 if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN)
6731 goto out;
6732
6733 ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages,
6734 &numpages);
6735 if (ret) {
6736 mlog_errno(ret);
6737 goto out;
6738 }
6739
6740 ocfs2_zero_cluster_pages(inode, range_start, range_end, pages,
6741 numpages, phys, handle);
6742
6743 /*
6744 * Initiate writeout of the pages we zero'd here. We don't
6745 * wait on them - the truncate_inode_pages() call later will
6746 * do that for us.
6747 */
6748 ret = filemap_fdatawrite_range(inode->i_mapping, range_start,
6749 range_end - 1);
6750 if (ret)
6751 mlog_errno(ret);
6752
6753out:
6754 if (pages)
6755 kfree(pages);
6756
6757 return ret;
6758}
6759
6760static void ocfs2_zero_dinode_id2_with_xattr(struct inode *inode,
6761 struct ocfs2_dinode *di)
6762{
6763 unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits;
6764 unsigned int xattrsize = le16_to_cpu(di->i_xattr_inline_size);
6765
6766 if (le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_XATTR_FL)
6767 memset(&di->id2, 0, blocksize -
6768 offsetof(struct ocfs2_dinode, id2) -
6769 xattrsize);
6770 else
6771 memset(&di->id2, 0, blocksize -
6772 offsetof(struct ocfs2_dinode, id2));
6773}
6774
6775void ocfs2_dinode_new_extent_list(struct inode *inode,
6776 struct ocfs2_dinode *di)
6777{
6778 ocfs2_zero_dinode_id2_with_xattr(inode, di);
6779 di->id2.i_list.l_tree_depth = 0;
6780 di->id2.i_list.l_next_free_rec = 0;
6781 di->id2.i_list.l_count = cpu_to_le16(
6782 ocfs2_extent_recs_per_inode_with_xattr(inode->i_sb, di));
6783}
6784
6785void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di)
6786{
6787 struct ocfs2_inode_info *oi = OCFS2_I(inode);
6788 struct ocfs2_inline_data *idata = &di->id2.i_data;
6789
6790 spin_lock(&oi->ip_lock);
6791 oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
6792 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
6793 spin_unlock(&oi->ip_lock);
6794
6795 /*
6796 * We clear the entire i_data structure here so that all
6797 * fields can be properly initialized.
6798 */
6799 ocfs2_zero_dinode_id2_with_xattr(inode, di);
6800
6801 idata->id_count = cpu_to_le16(
6802 ocfs2_max_inline_data_with_xattr(inode->i_sb, di));
6803}
6804
6805int ocfs2_convert_inline_data_to_extents(struct inode *inode,
6806 struct buffer_head *di_bh)
6807{
6808 int ret, i, has_data, num_pages = 0;
6809 handle_t *handle;
6810 u64 uninitialized_var(block);
6811 struct ocfs2_inode_info *oi = OCFS2_I(inode);
6812 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
6813 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
6814 struct ocfs2_alloc_context *data_ac = NULL;
6815 struct page **pages = NULL;
6816 loff_t end = osb->s_clustersize;
6817 struct ocfs2_extent_tree et;
6818 int did_quota = 0;
6819
6820 has_data = i_size_read(inode) ? 1 : 0;
6821
6822 if (has_data) {
6823 pages = kcalloc(ocfs2_pages_per_cluster(osb->sb),
6824 sizeof(struct page *), GFP_NOFS);
6825 if (pages == NULL) {
6826 ret = -ENOMEM;
6827 mlog_errno(ret);
6828 goto out;
6829 }
6830
6831 ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
6832 if (ret) {
6833 mlog_errno(ret);
6834 goto out;
6835 }
6836 }
6837
6838 handle = ocfs2_start_trans(osb,
6839 ocfs2_inline_to_extents_credits(osb->sb));
6840 if (IS_ERR(handle)) {
6841 ret = PTR_ERR(handle);
6842 mlog_errno(ret);
6843 goto out_unlock;
6844 }
6845
6846 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
6847 OCFS2_JOURNAL_ACCESS_WRITE);
6848 if (ret) {
6849 mlog_errno(ret);
6850 goto out_commit;
6851 }
6852
6853 if (has_data) {
6854 u32 bit_off, num;
6855 unsigned int page_end;
6856 u64 phys;
6857
6858 ret = dquot_alloc_space_nodirty(inode,
6859 ocfs2_clusters_to_bytes(osb->sb, 1));
6860 if (ret)
6861 goto out_commit;
6862 did_quota = 1;
6863
6864 data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
6865
6866 ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off,
6867 &num);
6868 if (ret) {
6869 mlog_errno(ret);
6870 goto out_commit;
6871 }
6872
6873 /*
6874 * Save two copies, one for insert, and one that can
6875 * be changed by ocfs2_map_and_dirty_page() below.
6876 */
6877 block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off);
6878
6879 /*
6880 * Non sparse file systems zero on extend, so no need
6881 * to do that now.
6882 */
6883 if (!ocfs2_sparse_alloc(osb) &&
6884 PAGE_CACHE_SIZE < osb->s_clustersize)
6885 end = PAGE_CACHE_SIZE;
6886
6887 ret = ocfs2_grab_eof_pages(inode, 0, end, pages, &num_pages);
6888 if (ret) {
6889 mlog_errno(ret);
6890 goto out_commit;
6891 }
6892
6893 /*
6894 * This should populate the 1st page for us and mark
6895 * it up to date.
6896 */
6897 ret = ocfs2_read_inline_data(inode, pages[0], di_bh);
6898 if (ret) {
6899 mlog_errno(ret);
6900 goto out_commit;
6901 }
6902
6903 page_end = PAGE_CACHE_SIZE;
6904 if (PAGE_CACHE_SIZE > osb->s_clustersize)
6905 page_end = osb->s_clustersize;
6906
6907 for (i = 0; i < num_pages; i++)
6908 ocfs2_map_and_dirty_page(inode, handle, 0, page_end,
6909 pages[i], i > 0, &phys);
6910 }
6911
6912 spin_lock(&oi->ip_lock);
6913 oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
6914 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
6915 spin_unlock(&oi->ip_lock);
6916
6917 ocfs2_dinode_new_extent_list(inode, di);
6918
6919 ocfs2_journal_dirty(handle, di_bh);
6920
6921 if (has_data) {
6922 /*
6923 * An error at this point should be extremely rare. If
6924 * this proves to be false, we could always re-build
6925 * the in-inode data from our pages.
6926 */
6927 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
6928 ret = ocfs2_insert_extent(handle, &et, 0, block, 1, 0, NULL);
6929 if (ret) {
6930 mlog_errno(ret);
6931 goto out_commit;
6932 }
6933
6934 inode->i_blocks = ocfs2_inode_sector_count(inode);
6935 }
6936
6937out_commit:
6938 if (ret < 0 && did_quota)
6939 dquot_free_space_nodirty(inode,
6940 ocfs2_clusters_to_bytes(osb->sb, 1));
6941
6942 ocfs2_commit_trans(osb, handle);
6943
6944out_unlock:
6945 if (data_ac)
6946 ocfs2_free_alloc_context(data_ac);
6947
6948out:
6949 if (pages) {
6950 ocfs2_unlock_and_free_pages(pages, num_pages);
6951 kfree(pages);
6952 }
6953
6954 return ret;
6955}
6956
6957/*
6958 * It is expected, that by the time you call this function,
6959 * inode->i_size and fe->i_size have been adjusted.
6960 *
6961 * WARNING: This will kfree the truncate context
6962 */
6963int ocfs2_commit_truncate(struct ocfs2_super *osb,
6964 struct inode *inode,
6965 struct buffer_head *di_bh)
6966{
6967 int status = 0, i, flags = 0;
6968 u32 new_highest_cpos, range, trunc_cpos, trunc_len, phys_cpos, coff;
6969 u64 blkno = 0;
6970 struct ocfs2_extent_list *el;
6971 struct ocfs2_extent_rec *rec;
6972 struct ocfs2_path *path = NULL;
6973 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
6974 struct ocfs2_extent_list *root_el = &(di->id2.i_list);
6975 u64 refcount_loc = le64_to_cpu(di->i_refcount_loc);
6976 struct ocfs2_extent_tree et;
6977 struct ocfs2_cached_dealloc_ctxt dealloc;
6978
6979 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
6980 ocfs2_init_dealloc_ctxt(&dealloc);
6981
6982 new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
6983 i_size_read(inode));
6984
6985 path = ocfs2_new_path(di_bh, &di->id2.i_list,
6986 ocfs2_journal_access_di);
6987 if (!path) {
6988 status = -ENOMEM;
6989 mlog_errno(status);
6990 goto bail;
6991 }
6992
6993 ocfs2_extent_map_trunc(inode, new_highest_cpos);
6994
6995start:
6996 /*
6997 * Check that we still have allocation to delete.
6998 */
6999 if (OCFS2_I(inode)->ip_clusters == 0) {
7000 status = 0;
7001 goto bail;
7002 }
7003
7004 /*
7005 * Truncate always works against the rightmost tree branch.
7006 */
7007 status = ocfs2_find_path(INODE_CACHE(inode), path, UINT_MAX);
7008 if (status) {
7009 mlog_errno(status);
7010 goto bail;
7011 }
7012
7013 trace_ocfs2_commit_truncate(
7014 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7015 new_highest_cpos,
7016 OCFS2_I(inode)->ip_clusters,
7017 path->p_tree_depth);
7018
7019 /*
7020 * By now, el will point to the extent list on the bottom most
7021 * portion of this tree. Only the tail record is considered in
7022 * each pass.
7023 *
7024 * We handle the following cases, in order:
7025 * - empty extent: delete the remaining branch
7026 * - remove the entire record
7027 * - remove a partial record
7028 * - no record needs to be removed (truncate has completed)
7029 */
7030 el = path_leaf_el(path);
7031 if (le16_to_cpu(el->l_next_free_rec) == 0) {
7032 ocfs2_error(inode->i_sb,
7033 "Inode %llu has empty extent block at %llu\n",
7034 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7035 (unsigned long long)path_leaf_bh(path)->b_blocknr);
7036 status = -EROFS;
7037 goto bail;
7038 }
7039
7040 i = le16_to_cpu(el->l_next_free_rec) - 1;
7041 rec = &el->l_recs[i];
7042 flags = rec->e_flags;
7043 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
7044
7045 if (i == 0 && ocfs2_is_empty_extent(rec)) {
7046 /*
7047 * Lower levels depend on this never happening, but it's best
7048 * to check it up here before changing the tree.
7049 */
7050 if (root_el->l_tree_depth && rec->e_int_clusters == 0) {
7051 ocfs2_error(inode->i_sb, "Inode %lu has an empty "
7052 "extent record, depth %u\n", inode->i_ino,
7053 le16_to_cpu(root_el->l_tree_depth));
7054 status = -EROFS;
7055 goto bail;
7056 }
7057 trunc_cpos = le32_to_cpu(rec->e_cpos);
7058 trunc_len = 0;
7059 blkno = 0;
7060 } else if (le32_to_cpu(rec->e_cpos) >= new_highest_cpos) {
7061 /*
7062 * Truncate entire record.
7063 */
7064 trunc_cpos = le32_to_cpu(rec->e_cpos);
7065 trunc_len = ocfs2_rec_clusters(el, rec);
7066 blkno = le64_to_cpu(rec->e_blkno);
7067 } else if (range > new_highest_cpos) {
7068 /*
7069 * Partial truncate. it also should be
7070 * the last truncate we're doing.
7071 */
7072 trunc_cpos = new_highest_cpos;
7073 trunc_len = range - new_highest_cpos;
7074 coff = new_highest_cpos - le32_to_cpu(rec->e_cpos);
7075 blkno = le64_to_cpu(rec->e_blkno) +
7076 ocfs2_clusters_to_blocks(inode->i_sb, coff);
7077 } else {
7078 /*
7079 * Truncate completed, leave happily.
7080 */
7081 status = 0;
7082 goto bail;
7083 }
7084
7085 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
7086
7087 status = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
7088 phys_cpos, trunc_len, flags, &dealloc,
7089 refcount_loc);
7090 if (status < 0) {
7091 mlog_errno(status);
7092 goto bail;
7093 }
7094
7095 ocfs2_reinit_path(path, 1);
7096
7097 /*
7098 * The check above will catch the case where we've truncated
7099 * away all allocation.
7100 */
7101 goto start;
7102
7103bail:
7104
7105 ocfs2_schedule_truncate_log_flush(osb, 1);
7106
7107 ocfs2_run_deallocs(osb, &dealloc);
7108
7109 ocfs2_free_path(path);
7110
7111 return status;
7112}
7113
7114/*
7115 * 'start' is inclusive, 'end' is not.
7116 */
7117int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh,
7118 unsigned int start, unsigned int end, int trunc)
7119{
7120 int ret;
7121 unsigned int numbytes;
7122 handle_t *handle;
7123 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7124 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7125 struct ocfs2_inline_data *idata = &di->id2.i_data;
7126
7127 if (end > i_size_read(inode))
7128 end = i_size_read(inode);
7129
7130 BUG_ON(start >= end);
7131
7132 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
7133 !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) ||
7134 !ocfs2_supports_inline_data(osb)) {
7135 ocfs2_error(inode->i_sb,
7136 "Inline data flags for inode %llu don't agree! "
7137 "Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7138 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7139 le16_to_cpu(di->i_dyn_features),
7140 OCFS2_I(inode)->ip_dyn_features,
7141 osb->s_feature_incompat);
7142 ret = -EROFS;
7143 goto out;
7144 }
7145
7146 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
7147 if (IS_ERR(handle)) {
7148 ret = PTR_ERR(handle);
7149 mlog_errno(ret);
7150 goto out;
7151 }
7152
7153 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7154 OCFS2_JOURNAL_ACCESS_WRITE);
7155 if (ret) {
7156 mlog_errno(ret);
7157 goto out_commit;
7158 }
7159
7160 numbytes = end - start;
7161 memset(idata->id_data + start, 0, numbytes);
7162
7163 /*
7164 * No need to worry about the data page here - it's been
7165 * truncated already and inline data doesn't need it for
7166 * pushing zero's to disk, so we'll let readpage pick it up
7167 * later.
7168 */
7169 if (trunc) {
7170 i_size_write(inode, start);
7171 di->i_size = cpu_to_le64(start);
7172 }
7173
7174 inode->i_blocks = ocfs2_inode_sector_count(inode);
7175 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
7176
7177 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
7178 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
7179
7180 ocfs2_journal_dirty(handle, di_bh);
7181
7182out_commit:
7183 ocfs2_commit_trans(osb, handle);
7184
7185out:
7186 return ret;
7187}
7188
7189static int ocfs2_trim_extent(struct super_block *sb,
7190 struct ocfs2_group_desc *gd,
7191 u32 start, u32 count)
7192{
7193 u64 discard, bcount;
7194
7195 bcount = ocfs2_clusters_to_blocks(sb, count);
7196 discard = le64_to_cpu(gd->bg_blkno) +
7197 ocfs2_clusters_to_blocks(sb, start);
7198
7199 trace_ocfs2_trim_extent(sb, (unsigned long long)discard, bcount);
7200
7201 return sb_issue_discard(sb, discard, bcount, GFP_NOFS, 0);
7202}
7203
7204static int ocfs2_trim_group(struct super_block *sb,
7205 struct ocfs2_group_desc *gd,
7206 u32 start, u32 max, u32 minbits)
7207{
7208 int ret = 0, count = 0, next;
7209 void *bitmap = gd->bg_bitmap;
7210
7211 if (le16_to_cpu(gd->bg_free_bits_count) < minbits)
7212 return 0;
7213
7214 trace_ocfs2_trim_group((unsigned long long)le64_to_cpu(gd->bg_blkno),
7215 start, max, minbits);
7216
7217 while (start < max) {
7218 start = ocfs2_find_next_zero_bit(bitmap, max, start);
7219 if (start >= max)
7220 break;
7221 next = ocfs2_find_next_bit(bitmap, max, start);
7222
7223 if ((next - start) >= minbits) {
7224 ret = ocfs2_trim_extent(sb, gd,
7225 start, next - start);
7226 if (ret < 0) {
7227 mlog_errno(ret);
7228 break;
7229 }
7230 count += next - start;
7231 }
7232 start = next + 1;
7233
7234 if (fatal_signal_pending(current)) {
7235 count = -ERESTARTSYS;
7236 break;
7237 }
7238
7239 if ((le16_to_cpu(gd->bg_free_bits_count) - count) < minbits)
7240 break;
7241 }
7242
7243 if (ret < 0)
7244 count = ret;
7245
7246 return count;
7247}
7248
7249int ocfs2_trim_fs(struct super_block *sb, struct fstrim_range *range)
7250{
7251 struct ocfs2_super *osb = OCFS2_SB(sb);
7252 u64 start, len, trimmed, first_group, last_group, group;
7253 int ret, cnt;
7254 u32 first_bit, last_bit, minlen;
7255 struct buffer_head *main_bm_bh = NULL;
7256 struct inode *main_bm_inode = NULL;
7257 struct buffer_head *gd_bh = NULL;
7258 struct ocfs2_dinode *main_bm;
7259 struct ocfs2_group_desc *gd = NULL;
7260
7261 start = range->start >> osb->s_clustersize_bits;
7262 len = range->len >> osb->s_clustersize_bits;
7263 minlen = range->minlen >> osb->s_clustersize_bits;
7264 trimmed = 0;
7265
7266 if (!len) {
7267 range->len = 0;
7268 return 0;
7269 }
7270
7271 if (minlen >= osb->bitmap_cpg)
7272 return -EINVAL;
7273
7274 main_bm_inode = ocfs2_get_system_file_inode(osb,
7275 GLOBAL_BITMAP_SYSTEM_INODE,
7276 OCFS2_INVALID_SLOT);
7277 if (!main_bm_inode) {
7278 ret = -EIO;
7279 mlog_errno(ret);
7280 goto out;
7281 }
7282
7283 mutex_lock(&main_bm_inode->i_mutex);
7284
7285 ret = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 0);
7286 if (ret < 0) {
7287 mlog_errno(ret);
7288 goto out_mutex;
7289 }
7290 main_bm = (struct ocfs2_dinode *)main_bm_bh->b_data;
7291
7292 if (start >= le32_to_cpu(main_bm->i_clusters)) {
7293 ret = -EINVAL;
7294 goto out_unlock;
7295 }
7296
7297 if (start + len > le32_to_cpu(main_bm->i_clusters))
7298 len = le32_to_cpu(main_bm->i_clusters) - start;
7299
7300 trace_ocfs2_trim_fs(start, len, minlen);
7301
7302 /* Determine first and last group to examine based on start and len */
7303 first_group = ocfs2_which_cluster_group(main_bm_inode, start);
7304 if (first_group == osb->first_cluster_group_blkno)
7305 first_bit = start;
7306 else
7307 first_bit = start - ocfs2_blocks_to_clusters(sb, first_group);
7308 last_group = ocfs2_which_cluster_group(main_bm_inode, start + len - 1);
7309 last_bit = osb->bitmap_cpg;
7310
7311 for (group = first_group; group <= last_group;) {
7312 if (first_bit + len >= osb->bitmap_cpg)
7313 last_bit = osb->bitmap_cpg;
7314 else
7315 last_bit = first_bit + len;
7316
7317 ret = ocfs2_read_group_descriptor(main_bm_inode,
7318 main_bm, group,
7319 &gd_bh);
7320 if (ret < 0) {
7321 mlog_errno(ret);
7322 break;
7323 }
7324
7325 gd = (struct ocfs2_group_desc *)gd_bh->b_data;
7326 cnt = ocfs2_trim_group(sb, gd, first_bit, last_bit, minlen);
7327 brelse(gd_bh);
7328 gd_bh = NULL;
7329 if (cnt < 0) {
7330 ret = cnt;
7331 mlog_errno(ret);
7332 break;
7333 }
7334
7335 trimmed += cnt;
7336 len -= osb->bitmap_cpg - first_bit;
7337 first_bit = 0;
7338 if (group == osb->first_cluster_group_blkno)
7339 group = ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7340 else
7341 group += ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7342 }
7343 range->len = trimmed * sb->s_blocksize;
7344out_unlock:
7345 ocfs2_inode_unlock(main_bm_inode, 0);
7346 brelse(main_bm_bh);
7347out_mutex:
7348 mutex_unlock(&main_bm_inode->i_mutex);
7349 iput(main_bm_inode);
7350out:
7351 return ret;
7352}