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",
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",
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 wants_rotate = 0, 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 wants_rotate = 1;
5461
5462 next_free = le16_to_cpu(el->l_next_free_rec);
5463 if (is_rightmost_tree_rec && next_free > 1) {
5464 /*
5465 * We skip the edge update if this path will
5466 * be deleted by the rotate code.
5467 */
5468 rec = &el->l_recs[next_free - 1];
5469 ocfs2_adjust_rightmost_records(handle, et, path,
5470 rec);
5471 }
5472 } else if (le32_to_cpu(rec->e_cpos) == cpos) {
5473 /* Remove leftmost portion of the record. */
5474 le32_add_cpu(&rec->e_cpos, len);
5475 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
5476 le16_add_cpu(&rec->e_leaf_clusters, -len);
5477 } else if (rec_range == trunc_range) {
5478 /* Remove rightmost portion of the record */
5479 le16_add_cpu(&rec->e_leaf_clusters, -len);
5480 if (is_rightmost_tree_rec)
5481 ocfs2_adjust_rightmost_records(handle, et, path, rec);
5482 } else {
5483 /* Caller should have trapped this. */
5484 mlog(ML_ERROR, "Owner %llu: Invalid record truncate: (%u, %u) "
5485 "(%u, %u)\n",
5486 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5487 le32_to_cpu(rec->e_cpos),
5488 le16_to_cpu(rec->e_leaf_clusters), cpos, len);
5489 BUG();
5490 }
5491
5492 if (left_path) {
5493 int subtree_index;
5494
5495 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
5496 ocfs2_complete_edge_insert(handle, left_path, path,
5497 subtree_index);
5498 }
5499
5500 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5501
5502 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5503 if (ret) {
5504 mlog_errno(ret);
5505 goto out;
5506 }
5507
5508out:
5509 ocfs2_free_path(left_path);
5510 return ret;
5511}
5512
5513int ocfs2_remove_extent(handle_t *handle,
5514 struct ocfs2_extent_tree *et,
5515 u32 cpos, u32 len,
5516 struct ocfs2_alloc_context *meta_ac,
5517 struct ocfs2_cached_dealloc_ctxt *dealloc)
5518{
5519 int ret, index;
5520 u32 rec_range, trunc_range;
5521 struct ocfs2_extent_rec *rec;
5522 struct ocfs2_extent_list *el;
5523 struct ocfs2_path *path = NULL;
5524
5525 /*
5526 * XXX: Why are we truncating to 0 instead of wherever this
5527 * affects us?
5528 */
5529 ocfs2_et_extent_map_truncate(et, 0);
5530
5531 path = ocfs2_new_path_from_et(et);
5532 if (!path) {
5533 ret = -ENOMEM;
5534 mlog_errno(ret);
5535 goto out;
5536 }
5537
5538 ret = ocfs2_find_path(et->et_ci, path, cpos);
5539 if (ret) {
5540 mlog_errno(ret);
5541 goto out;
5542 }
5543
5544 el = path_leaf_el(path);
5545 index = ocfs2_search_extent_list(el, cpos);
5546 if (index == -1) {
5547 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5548 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5549 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5550 cpos);
5551 ret = -EROFS;
5552 goto out;
5553 }
5554
5555 /*
5556 * We have 3 cases of extent removal:
5557 * 1) Range covers the entire extent rec
5558 * 2) Range begins or ends on one edge of the extent rec
5559 * 3) Range is in the middle of the extent rec (no shared edges)
5560 *
5561 * For case 1 we remove the extent rec and left rotate to
5562 * fill the hole.
5563 *
5564 * For case 2 we just shrink the existing extent rec, with a
5565 * tree update if the shrinking edge is also the edge of an
5566 * extent block.
5567 *
5568 * For case 3 we do a right split to turn the extent rec into
5569 * something case 2 can handle.
5570 */
5571 rec = &el->l_recs[index];
5572 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5573 trunc_range = cpos + len;
5574
5575 BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
5576
5577 trace_ocfs2_remove_extent(
5578 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5579 cpos, len, index, le32_to_cpu(rec->e_cpos),
5580 ocfs2_rec_clusters(el, rec));
5581
5582 if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
5583 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5584 cpos, len);
5585 if (ret) {
5586 mlog_errno(ret);
5587 goto out;
5588 }
5589 } else {
5590 ret = ocfs2_split_tree(handle, et, path, index,
5591 trunc_range, meta_ac);
5592 if (ret) {
5593 mlog_errno(ret);
5594 goto out;
5595 }
5596
5597 /*
5598 * The split could have manipulated the tree enough to
5599 * move the record location, so we have to look for it again.
5600 */
5601 ocfs2_reinit_path(path, 1);
5602
5603 ret = ocfs2_find_path(et->et_ci, path, cpos);
5604 if (ret) {
5605 mlog_errno(ret);
5606 goto out;
5607 }
5608
5609 el = path_leaf_el(path);
5610 index = ocfs2_search_extent_list(el, cpos);
5611 if (index == -1) {
5612 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5613 "Owner %llu: split at cpos %u lost record\n",
5614 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5615 cpos);
5616 ret = -EROFS;
5617 goto out;
5618 }
5619
5620 /*
5621 * Double check our values here. If anything is fishy,
5622 * it's easier to catch it at the top level.
5623 */
5624 rec = &el->l_recs[index];
5625 rec_range = le32_to_cpu(rec->e_cpos) +
5626 ocfs2_rec_clusters(el, rec);
5627 if (rec_range != trunc_range) {
5628 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5629 "Owner %llu: error after split at cpos %u trunc len %u, existing record is (%u,%u)\n",
5630 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5631 cpos, len, le32_to_cpu(rec->e_cpos),
5632 ocfs2_rec_clusters(el, rec));
5633 ret = -EROFS;
5634 goto out;
5635 }
5636
5637 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5638 cpos, len);
5639 if (ret) {
5640 mlog_errno(ret);
5641 goto out;
5642 }
5643 }
5644
5645out:
5646 ocfs2_free_path(path);
5647 return ret;
5648}
5649
5650/*
5651 * ocfs2_reserve_blocks_for_rec_trunc() would look basically the
5652 * same as ocfs2_lock_alloctors(), except for it accepts a blocks
5653 * number to reserve some extra blocks, and it only handles meta
5654 * data allocations.
5655 *
5656 * Currently, only ocfs2_remove_btree_range() uses it for truncating
5657 * and punching holes.
5658 */
5659static int ocfs2_reserve_blocks_for_rec_trunc(struct inode *inode,
5660 struct ocfs2_extent_tree *et,
5661 u32 extents_to_split,
5662 struct ocfs2_alloc_context **ac,
5663 int extra_blocks)
5664{
5665 int ret = 0, num_free_extents;
5666 unsigned int max_recs_needed = 2 * extents_to_split;
5667 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5668
5669 *ac = NULL;
5670
5671 num_free_extents = ocfs2_num_free_extents(osb, et);
5672 if (num_free_extents < 0) {
5673 ret = num_free_extents;
5674 mlog_errno(ret);
5675 goto out;
5676 }
5677
5678 if (!num_free_extents ||
5679 (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
5680 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
5681
5682 if (extra_blocks) {
5683 ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, ac);
5684 if (ret < 0) {
5685 if (ret != -ENOSPC)
5686 mlog_errno(ret);
5687 goto out;
5688 }
5689 }
5690
5691out:
5692 if (ret) {
5693 if (*ac) {
5694 ocfs2_free_alloc_context(*ac);
5695 *ac = NULL;
5696 }
5697 }
5698
5699 return ret;
5700}
5701
5702int ocfs2_remove_btree_range(struct inode *inode,
5703 struct ocfs2_extent_tree *et,
5704 u32 cpos, u32 phys_cpos, u32 len, int flags,
5705 struct ocfs2_cached_dealloc_ctxt *dealloc,
5706 u64 refcount_loc, bool refcount_tree_locked)
5707{
5708 int ret, credits = 0, extra_blocks = 0;
5709 u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
5710 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5711 struct inode *tl_inode = osb->osb_tl_inode;
5712 handle_t *handle;
5713 struct ocfs2_alloc_context *meta_ac = NULL;
5714 struct ocfs2_refcount_tree *ref_tree = NULL;
5715
5716 if ((flags & OCFS2_EXT_REFCOUNTED) && len) {
5717 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
5718 OCFS2_HAS_REFCOUNT_FL));
5719
5720 if (!refcount_tree_locked) {
5721 ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
5722 &ref_tree, NULL);
5723 if (ret) {
5724 mlog_errno(ret);
5725 goto bail;
5726 }
5727 }
5728
5729 ret = ocfs2_prepare_refcount_change_for_del(inode,
5730 refcount_loc,
5731 phys_blkno,
5732 len,
5733 &credits,
5734 &extra_blocks);
5735 if (ret < 0) {
5736 mlog_errno(ret);
5737 goto bail;
5738 }
5739 }
5740
5741 ret = ocfs2_reserve_blocks_for_rec_trunc(inode, et, 1, &meta_ac,
5742 extra_blocks);
5743 if (ret) {
5744 mlog_errno(ret);
5745 goto bail;
5746 }
5747
5748 inode_lock(tl_inode);
5749
5750 if (ocfs2_truncate_log_needs_flush(osb)) {
5751 ret = __ocfs2_flush_truncate_log(osb);
5752 if (ret < 0) {
5753 mlog_errno(ret);
5754 goto out;
5755 }
5756 }
5757
5758 handle = ocfs2_start_trans(osb,
5759 ocfs2_remove_extent_credits(osb->sb) + credits);
5760 if (IS_ERR(handle)) {
5761 ret = PTR_ERR(handle);
5762 mlog_errno(ret);
5763 goto out;
5764 }
5765
5766 ret = ocfs2_et_root_journal_access(handle, et,
5767 OCFS2_JOURNAL_ACCESS_WRITE);
5768 if (ret) {
5769 mlog_errno(ret);
5770 goto out_commit;
5771 }
5772
5773 dquot_free_space_nodirty(inode,
5774 ocfs2_clusters_to_bytes(inode->i_sb, len));
5775
5776 ret = ocfs2_remove_extent(handle, et, cpos, len, meta_ac, dealloc);
5777 if (ret) {
5778 mlog_errno(ret);
5779 goto out_commit;
5780 }
5781
5782 ocfs2_et_update_clusters(et, -len);
5783 ocfs2_update_inode_fsync_trans(handle, inode, 1);
5784
5785 ocfs2_journal_dirty(handle, et->et_root_bh);
5786
5787 if (phys_blkno) {
5788 if (flags & OCFS2_EXT_REFCOUNTED)
5789 ret = ocfs2_decrease_refcount(inode, handle,
5790 ocfs2_blocks_to_clusters(osb->sb,
5791 phys_blkno),
5792 len, meta_ac,
5793 dealloc, 1);
5794 else
5795 ret = ocfs2_truncate_log_append(osb, handle,
5796 phys_blkno, len);
5797 if (ret)
5798 mlog_errno(ret);
5799
5800 }
5801
5802out_commit:
5803 ocfs2_commit_trans(osb, handle);
5804out:
5805 inode_unlock(tl_inode);
5806bail:
5807 if (meta_ac)
5808 ocfs2_free_alloc_context(meta_ac);
5809
5810 if (ref_tree)
5811 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
5812
5813 return ret;
5814}
5815
5816int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
5817{
5818 struct buffer_head *tl_bh = osb->osb_tl_bh;
5819 struct ocfs2_dinode *di;
5820 struct ocfs2_truncate_log *tl;
5821
5822 di = (struct ocfs2_dinode *) tl_bh->b_data;
5823 tl = &di->id2.i_dealloc;
5824
5825 mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
5826 "slot %d, invalid truncate log parameters: used = "
5827 "%u, count = %u\n", osb->slot_num,
5828 le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
5829 return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
5830}
5831
5832static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
5833 unsigned int new_start)
5834{
5835 unsigned int tail_index;
5836 unsigned int current_tail;
5837
5838 /* No records, nothing to coalesce */
5839 if (!le16_to_cpu(tl->tl_used))
5840 return 0;
5841
5842 tail_index = le16_to_cpu(tl->tl_used) - 1;
5843 current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
5844 current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
5845
5846 return current_tail == new_start;
5847}
5848
5849int ocfs2_truncate_log_append(struct ocfs2_super *osb,
5850 handle_t *handle,
5851 u64 start_blk,
5852 unsigned int num_clusters)
5853{
5854 int status, index;
5855 unsigned int start_cluster, tl_count;
5856 struct inode *tl_inode = osb->osb_tl_inode;
5857 struct buffer_head *tl_bh = osb->osb_tl_bh;
5858 struct ocfs2_dinode *di;
5859 struct ocfs2_truncate_log *tl;
5860
5861 BUG_ON(inode_trylock(tl_inode));
5862
5863 start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
5864
5865 di = (struct ocfs2_dinode *) tl_bh->b_data;
5866
5867 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5868 * by the underlying call to ocfs2_read_inode_block(), so any
5869 * corruption is a code bug */
5870 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5871
5872 tl = &di->id2.i_dealloc;
5873 tl_count = le16_to_cpu(tl->tl_count);
5874 mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
5875 tl_count == 0,
5876 "Truncate record count on #%llu invalid "
5877 "wanted %u, actual %u\n",
5878 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5879 ocfs2_truncate_recs_per_inode(osb->sb),
5880 le16_to_cpu(tl->tl_count));
5881
5882 /* Caller should have known to flush before calling us. */
5883 index = le16_to_cpu(tl->tl_used);
5884 if (index >= tl_count) {
5885 status = -ENOSPC;
5886 mlog_errno(status);
5887 goto bail;
5888 }
5889
5890 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5891 OCFS2_JOURNAL_ACCESS_WRITE);
5892 if (status < 0) {
5893 mlog_errno(status);
5894 goto bail;
5895 }
5896
5897 trace_ocfs2_truncate_log_append(
5898 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index,
5899 start_cluster, num_clusters);
5900 if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
5901 /*
5902 * Move index back to the record we are coalescing with.
5903 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5904 */
5905 index--;
5906
5907 num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
5908 trace_ocfs2_truncate_log_append(
5909 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5910 index, le32_to_cpu(tl->tl_recs[index].t_start),
5911 num_clusters);
5912 } else {
5913 tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
5914 tl->tl_used = cpu_to_le16(index + 1);
5915 }
5916 tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
5917
5918 ocfs2_journal_dirty(handle, tl_bh);
5919
5920 osb->truncated_clusters += num_clusters;
5921bail:
5922 return status;
5923}
5924
5925static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
5926 handle_t *handle,
5927 struct inode *data_alloc_inode,
5928 struct buffer_head *data_alloc_bh)
5929{
5930 int status = 0;
5931 int i;
5932 unsigned int num_clusters;
5933 u64 start_blk;
5934 struct ocfs2_truncate_rec rec;
5935 struct ocfs2_dinode *di;
5936 struct ocfs2_truncate_log *tl;
5937 struct inode *tl_inode = osb->osb_tl_inode;
5938 struct buffer_head *tl_bh = osb->osb_tl_bh;
5939
5940 di = (struct ocfs2_dinode *) tl_bh->b_data;
5941 tl = &di->id2.i_dealloc;
5942 i = le16_to_cpu(tl->tl_used) - 1;
5943 while (i >= 0) {
5944 /* Caller has given us at least enough credits to
5945 * update the truncate log dinode */
5946 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5947 OCFS2_JOURNAL_ACCESS_WRITE);
5948 if (status < 0) {
5949 mlog_errno(status);
5950 goto bail;
5951 }
5952
5953 tl->tl_used = cpu_to_le16(i);
5954
5955 ocfs2_journal_dirty(handle, tl_bh);
5956
5957 rec = tl->tl_recs[i];
5958 start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
5959 le32_to_cpu(rec.t_start));
5960 num_clusters = le32_to_cpu(rec.t_clusters);
5961
5962 /* if start_blk is not set, we ignore the record as
5963 * invalid. */
5964 if (start_blk) {
5965 trace_ocfs2_replay_truncate_records(
5966 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5967 i, le32_to_cpu(rec.t_start), num_clusters);
5968
5969 status = ocfs2_free_clusters(handle, data_alloc_inode,
5970 data_alloc_bh, start_blk,
5971 num_clusters);
5972 if (status < 0) {
5973 mlog_errno(status);
5974 goto bail;
5975 }
5976 }
5977
5978 status = ocfs2_extend_trans(handle,
5979 OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
5980 if (status < 0) {
5981 mlog_errno(status);
5982 goto bail;
5983 }
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 handle_t *handle;
5999 struct inode *tl_inode = osb->osb_tl_inode;
6000 struct inode *data_alloc_inode = NULL;
6001 struct buffer_head *tl_bh = osb->osb_tl_bh;
6002 struct buffer_head *data_alloc_bh = NULL;
6003 struct ocfs2_dinode *di;
6004 struct ocfs2_truncate_log *tl;
6005
6006 BUG_ON(inode_trylock(tl_inode));
6007
6008 di = (struct ocfs2_dinode *) tl_bh->b_data;
6009
6010 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
6011 * by the underlying call to ocfs2_read_inode_block(), so any
6012 * corruption is a code bug */
6013 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6014
6015 tl = &di->id2.i_dealloc;
6016 num_to_flush = le16_to_cpu(tl->tl_used);
6017 trace_ocfs2_flush_truncate_log(
6018 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
6019 num_to_flush);
6020 if (!num_to_flush) {
6021 status = 0;
6022 goto out;
6023 }
6024
6025 data_alloc_inode = ocfs2_get_system_file_inode(osb,
6026 GLOBAL_BITMAP_SYSTEM_INODE,
6027 OCFS2_INVALID_SLOT);
6028 if (!data_alloc_inode) {
6029 status = -EINVAL;
6030 mlog(ML_ERROR, "Could not get bitmap inode!\n");
6031 goto out;
6032 }
6033
6034 inode_lock(data_alloc_inode);
6035
6036 status = ocfs2_inode_lock(data_alloc_inode, &data_alloc_bh, 1);
6037 if (status < 0) {
6038 mlog_errno(status);
6039 goto out_mutex;
6040 }
6041
6042 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
6043 if (IS_ERR(handle)) {
6044 status = PTR_ERR(handle);
6045 mlog_errno(status);
6046 goto out_unlock;
6047 }
6048
6049 status = ocfs2_replay_truncate_records(osb, handle, data_alloc_inode,
6050 data_alloc_bh);
6051 if (status < 0)
6052 mlog_errno(status);
6053
6054 ocfs2_commit_trans(osb, handle);
6055
6056out_unlock:
6057 brelse(data_alloc_bh);
6058 ocfs2_inode_unlock(data_alloc_inode, 1);
6059
6060out_mutex:
6061 inode_unlock(data_alloc_inode);
6062 iput(data_alloc_inode);
6063
6064out:
6065 return status;
6066}
6067
6068int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
6069{
6070 int status;
6071 struct inode *tl_inode = osb->osb_tl_inode;
6072
6073 inode_lock(tl_inode);
6074 status = __ocfs2_flush_truncate_log(osb);
6075 inode_unlock(tl_inode);
6076
6077 return status;
6078}
6079
6080static void ocfs2_truncate_log_worker(struct work_struct *work)
6081{
6082 int status;
6083 struct ocfs2_super *osb =
6084 container_of(work, struct ocfs2_super,
6085 osb_truncate_log_wq.work);
6086
6087 status = ocfs2_flush_truncate_log(osb);
6088 if (status < 0)
6089 mlog_errno(status);
6090 else
6091 ocfs2_init_steal_slots(osb);
6092}
6093
6094#define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
6095void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
6096 int cancel)
6097{
6098 if (osb->osb_tl_inode &&
6099 atomic_read(&osb->osb_tl_disable) == 0) {
6100 /* We want to push off log flushes while truncates are
6101 * still running. */
6102 if (cancel)
6103 cancel_delayed_work(&osb->osb_truncate_log_wq);
6104
6105 queue_delayed_work(osb->ocfs2_wq, &osb->osb_truncate_log_wq,
6106 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
6107 }
6108}
6109
6110static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
6111 int slot_num,
6112 struct inode **tl_inode,
6113 struct buffer_head **tl_bh)
6114{
6115 int status;
6116 struct inode *inode = NULL;
6117 struct buffer_head *bh = NULL;
6118
6119 inode = ocfs2_get_system_file_inode(osb,
6120 TRUNCATE_LOG_SYSTEM_INODE,
6121 slot_num);
6122 if (!inode) {
6123 status = -EINVAL;
6124 mlog(ML_ERROR, "Could not get load truncate log inode!\n");
6125 goto bail;
6126 }
6127
6128 status = ocfs2_read_inode_block(inode, &bh);
6129 if (status < 0) {
6130 iput(inode);
6131 mlog_errno(status);
6132 goto bail;
6133 }
6134
6135 *tl_inode = inode;
6136 *tl_bh = bh;
6137bail:
6138 return status;
6139}
6140
6141/* called during the 1st stage of node recovery. we stamp a clean
6142 * truncate log and pass back a copy for processing later. if the
6143 * truncate log does not require processing, a *tl_copy is set to
6144 * NULL. */
6145int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
6146 int slot_num,
6147 struct ocfs2_dinode **tl_copy)
6148{
6149 int status;
6150 struct inode *tl_inode = NULL;
6151 struct buffer_head *tl_bh = NULL;
6152 struct ocfs2_dinode *di;
6153 struct ocfs2_truncate_log *tl;
6154
6155 *tl_copy = NULL;
6156
6157 trace_ocfs2_begin_truncate_log_recovery(slot_num);
6158
6159 status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
6160 if (status < 0) {
6161 mlog_errno(status);
6162 goto bail;
6163 }
6164
6165 di = (struct ocfs2_dinode *) tl_bh->b_data;
6166
6167 /* tl_bh is loaded from ocfs2_get_truncate_log_info(). It's
6168 * validated by the underlying call to ocfs2_read_inode_block(),
6169 * so any corruption is a code bug */
6170 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6171
6172 tl = &di->id2.i_dealloc;
6173 if (le16_to_cpu(tl->tl_used)) {
6174 trace_ocfs2_truncate_log_recovery_num(le16_to_cpu(tl->tl_used));
6175
6176 *tl_copy = kmalloc(tl_bh->b_size, GFP_KERNEL);
6177 if (!(*tl_copy)) {
6178 status = -ENOMEM;
6179 mlog_errno(status);
6180 goto bail;
6181 }
6182
6183 /* Assuming the write-out below goes well, this copy
6184 * will be passed back to recovery for processing. */
6185 memcpy(*tl_copy, tl_bh->b_data, tl_bh->b_size);
6186
6187 /* All we need to do to clear the truncate log is set
6188 * tl_used. */
6189 tl->tl_used = 0;
6190
6191 ocfs2_compute_meta_ecc(osb->sb, tl_bh->b_data, &di->i_check);
6192 status = ocfs2_write_block(osb, tl_bh, INODE_CACHE(tl_inode));
6193 if (status < 0) {
6194 mlog_errno(status);
6195 goto bail;
6196 }
6197 }
6198
6199bail:
6200 iput(tl_inode);
6201 brelse(tl_bh);
6202
6203 if (status < 0) {
6204 kfree(*tl_copy);
6205 *tl_copy = NULL;
6206 mlog_errno(status);
6207 }
6208
6209 return status;
6210}
6211
6212int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
6213 struct ocfs2_dinode *tl_copy)
6214{
6215 int status = 0;
6216 int i;
6217 unsigned int clusters, num_recs, start_cluster;
6218 u64 start_blk;
6219 handle_t *handle;
6220 struct inode *tl_inode = osb->osb_tl_inode;
6221 struct ocfs2_truncate_log *tl;
6222
6223 if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
6224 mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
6225 return -EINVAL;
6226 }
6227
6228 tl = &tl_copy->id2.i_dealloc;
6229 num_recs = le16_to_cpu(tl->tl_used);
6230 trace_ocfs2_complete_truncate_log_recovery(
6231 (unsigned long long)le64_to_cpu(tl_copy->i_blkno),
6232 num_recs);
6233
6234 inode_lock(tl_inode);
6235 for(i = 0; i < num_recs; i++) {
6236 if (ocfs2_truncate_log_needs_flush(osb)) {
6237 status = __ocfs2_flush_truncate_log(osb);
6238 if (status < 0) {
6239 mlog_errno(status);
6240 goto bail_up;
6241 }
6242 }
6243
6244 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6245 if (IS_ERR(handle)) {
6246 status = PTR_ERR(handle);
6247 mlog_errno(status);
6248 goto bail_up;
6249 }
6250
6251 clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
6252 start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
6253 start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
6254
6255 status = ocfs2_truncate_log_append(osb, handle,
6256 start_blk, clusters);
6257 ocfs2_commit_trans(osb, handle);
6258 if (status < 0) {
6259 mlog_errno(status);
6260 goto bail_up;
6261 }
6262 }
6263
6264bail_up:
6265 inode_unlock(tl_inode);
6266
6267 return status;
6268}
6269
6270void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
6271{
6272 int status;
6273 struct inode *tl_inode = osb->osb_tl_inode;
6274
6275 atomic_set(&osb->osb_tl_disable, 1);
6276
6277 if (tl_inode) {
6278 cancel_delayed_work(&osb->osb_truncate_log_wq);
6279 flush_workqueue(osb->ocfs2_wq);
6280
6281 status = ocfs2_flush_truncate_log(osb);
6282 if (status < 0)
6283 mlog_errno(status);
6284
6285 brelse(osb->osb_tl_bh);
6286 iput(osb->osb_tl_inode);
6287 }
6288}
6289
6290int ocfs2_truncate_log_init(struct ocfs2_super *osb)
6291{
6292 int status;
6293 struct inode *tl_inode = NULL;
6294 struct buffer_head *tl_bh = NULL;
6295
6296 status = ocfs2_get_truncate_log_info(osb,
6297 osb->slot_num,
6298 &tl_inode,
6299 &tl_bh);
6300 if (status < 0)
6301 mlog_errno(status);
6302
6303 /* ocfs2_truncate_log_shutdown keys on the existence of
6304 * osb->osb_tl_inode so we don't set any of the osb variables
6305 * until we're sure all is well. */
6306 INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
6307 ocfs2_truncate_log_worker);
6308 atomic_set(&osb->osb_tl_disable, 0);
6309 osb->osb_tl_bh = tl_bh;
6310 osb->osb_tl_inode = tl_inode;
6311
6312 return status;
6313}
6314
6315/*
6316 * Delayed de-allocation of suballocator blocks.
6317 *
6318 * Some sets of block de-allocations might involve multiple suballocator inodes.
6319 *
6320 * The locking for this can get extremely complicated, especially when
6321 * the suballocator inodes to delete from aren't known until deep
6322 * within an unrelated codepath.
6323 *
6324 * ocfs2_extent_block structures are a good example of this - an inode
6325 * btree could have been grown by any number of nodes each allocating
6326 * out of their own suballoc inode.
6327 *
6328 * These structures allow the delay of block de-allocation until a
6329 * later time, when locking of multiple cluster inodes won't cause
6330 * deadlock.
6331 */
6332
6333/*
6334 * Describe a single bit freed from a suballocator. For the block
6335 * suballocators, it represents one block. For the global cluster
6336 * allocator, it represents some clusters and free_bit indicates
6337 * clusters number.
6338 */
6339struct ocfs2_cached_block_free {
6340 struct ocfs2_cached_block_free *free_next;
6341 u64 free_bg;
6342 u64 free_blk;
6343 unsigned int free_bit;
6344};
6345
6346struct ocfs2_per_slot_free_list {
6347 struct ocfs2_per_slot_free_list *f_next_suballocator;
6348 int f_inode_type;
6349 int f_slot;
6350 struct ocfs2_cached_block_free *f_first;
6351};
6352
6353static int ocfs2_free_cached_blocks(struct ocfs2_super *osb,
6354 int sysfile_type,
6355 int slot,
6356 struct ocfs2_cached_block_free *head)
6357{
6358 int ret;
6359 u64 bg_blkno;
6360 handle_t *handle;
6361 struct inode *inode;
6362 struct buffer_head *di_bh = NULL;
6363 struct ocfs2_cached_block_free *tmp;
6364
6365 inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
6366 if (!inode) {
6367 ret = -EINVAL;
6368 mlog_errno(ret);
6369 goto out;
6370 }
6371
6372 inode_lock(inode);
6373
6374 ret = ocfs2_inode_lock(inode, &di_bh, 1);
6375 if (ret) {
6376 mlog_errno(ret);
6377 goto out_mutex;
6378 }
6379
6380 handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
6381 if (IS_ERR(handle)) {
6382 ret = PTR_ERR(handle);
6383 mlog_errno(ret);
6384 goto out_unlock;
6385 }
6386
6387 while (head) {
6388 if (head->free_bg)
6389 bg_blkno = head->free_bg;
6390 else
6391 bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
6392 head->free_bit);
6393 trace_ocfs2_free_cached_blocks(
6394 (unsigned long long)head->free_blk, head->free_bit);
6395
6396 ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
6397 head->free_bit, bg_blkno, 1);
6398 if (ret) {
6399 mlog_errno(ret);
6400 goto out_journal;
6401 }
6402
6403 ret = ocfs2_extend_trans(handle, OCFS2_SUBALLOC_FREE);
6404 if (ret) {
6405 mlog_errno(ret);
6406 goto out_journal;
6407 }
6408
6409 tmp = head;
6410 head = head->free_next;
6411 kfree(tmp);
6412 }
6413
6414out_journal:
6415 ocfs2_commit_trans(osb, handle);
6416
6417out_unlock:
6418 ocfs2_inode_unlock(inode, 1);
6419 brelse(di_bh);
6420out_mutex:
6421 inode_unlock(inode);
6422 iput(inode);
6423out:
6424 while(head) {
6425 /* Premature exit may have left some dangling items. */
6426 tmp = head;
6427 head = head->free_next;
6428 kfree(tmp);
6429 }
6430
6431 return ret;
6432}
6433
6434int ocfs2_cache_cluster_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6435 u64 blkno, unsigned int bit)
6436{
6437 int ret = 0;
6438 struct ocfs2_cached_block_free *item;
6439
6440 item = kzalloc(sizeof(*item), GFP_NOFS);
6441 if (item == NULL) {
6442 ret = -ENOMEM;
6443 mlog_errno(ret);
6444 return ret;
6445 }
6446
6447 trace_ocfs2_cache_cluster_dealloc((unsigned long long)blkno, bit);
6448
6449 item->free_blk = blkno;
6450 item->free_bit = bit;
6451 item->free_next = ctxt->c_global_allocator;
6452
6453 ctxt->c_global_allocator = item;
6454 return ret;
6455}
6456
6457static int ocfs2_free_cached_clusters(struct ocfs2_super *osb,
6458 struct ocfs2_cached_block_free *head)
6459{
6460 struct ocfs2_cached_block_free *tmp;
6461 struct inode *tl_inode = osb->osb_tl_inode;
6462 handle_t *handle;
6463 int ret = 0;
6464
6465 inode_lock(tl_inode);
6466
6467 while (head) {
6468 if (ocfs2_truncate_log_needs_flush(osb)) {
6469 ret = __ocfs2_flush_truncate_log(osb);
6470 if (ret < 0) {
6471 mlog_errno(ret);
6472 break;
6473 }
6474 }
6475
6476 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6477 if (IS_ERR(handle)) {
6478 ret = PTR_ERR(handle);
6479 mlog_errno(ret);
6480 break;
6481 }
6482
6483 ret = ocfs2_truncate_log_append(osb, handle, head->free_blk,
6484 head->free_bit);
6485
6486 ocfs2_commit_trans(osb, handle);
6487 tmp = head;
6488 head = head->free_next;
6489 kfree(tmp);
6490
6491 if (ret < 0) {
6492 mlog_errno(ret);
6493 break;
6494 }
6495 }
6496
6497 inode_unlock(tl_inode);
6498
6499 while (head) {
6500 /* Premature exit may have left some dangling items. */
6501 tmp = head;
6502 head = head->free_next;
6503 kfree(tmp);
6504 }
6505
6506 return ret;
6507}
6508
6509int ocfs2_run_deallocs(struct ocfs2_super *osb,
6510 struct ocfs2_cached_dealloc_ctxt *ctxt)
6511{
6512 int ret = 0, ret2;
6513 struct ocfs2_per_slot_free_list *fl;
6514
6515 if (!ctxt)
6516 return 0;
6517
6518 while (ctxt->c_first_suballocator) {
6519 fl = ctxt->c_first_suballocator;
6520
6521 if (fl->f_first) {
6522 trace_ocfs2_run_deallocs(fl->f_inode_type,
6523 fl->f_slot);
6524 ret2 = ocfs2_free_cached_blocks(osb,
6525 fl->f_inode_type,
6526 fl->f_slot,
6527 fl->f_first);
6528 if (ret2)
6529 mlog_errno(ret2);
6530 if (!ret)
6531 ret = ret2;
6532 }
6533
6534 ctxt->c_first_suballocator = fl->f_next_suballocator;
6535 kfree(fl);
6536 }
6537
6538 if (ctxt->c_global_allocator) {
6539 ret2 = ocfs2_free_cached_clusters(osb,
6540 ctxt->c_global_allocator);
6541 if (ret2)
6542 mlog_errno(ret2);
6543 if (!ret)
6544 ret = ret2;
6545
6546 ctxt->c_global_allocator = NULL;
6547 }
6548
6549 return ret;
6550}
6551
6552static struct ocfs2_per_slot_free_list *
6553ocfs2_find_per_slot_free_list(int type,
6554 int slot,
6555 struct ocfs2_cached_dealloc_ctxt *ctxt)
6556{
6557 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6558
6559 while (fl) {
6560 if (fl->f_inode_type == type && fl->f_slot == slot)
6561 return fl;
6562
6563 fl = fl->f_next_suballocator;
6564 }
6565
6566 fl = kmalloc(sizeof(*fl), GFP_NOFS);
6567 if (fl) {
6568 fl->f_inode_type = type;
6569 fl->f_slot = slot;
6570 fl->f_first = NULL;
6571 fl->f_next_suballocator = ctxt->c_first_suballocator;
6572
6573 ctxt->c_first_suballocator = fl;
6574 }
6575 return fl;
6576}
6577
6578int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6579 int type, int slot, u64 suballoc,
6580 u64 blkno, unsigned int bit)
6581{
6582 int ret;
6583 struct ocfs2_per_slot_free_list *fl;
6584 struct ocfs2_cached_block_free *item;
6585
6586 fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
6587 if (fl == NULL) {
6588 ret = -ENOMEM;
6589 mlog_errno(ret);
6590 goto out;
6591 }
6592
6593 item = kzalloc(sizeof(*item), GFP_NOFS);
6594 if (item == NULL) {
6595 ret = -ENOMEM;
6596 mlog_errno(ret);
6597 goto out;
6598 }
6599
6600 trace_ocfs2_cache_block_dealloc(type, slot,
6601 (unsigned long long)suballoc,
6602 (unsigned long long)blkno, bit);
6603
6604 item->free_bg = suballoc;
6605 item->free_blk = blkno;
6606 item->free_bit = bit;
6607 item->free_next = fl->f_first;
6608
6609 fl->f_first = item;
6610
6611 ret = 0;
6612out:
6613 return ret;
6614}
6615
6616static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
6617 struct ocfs2_extent_block *eb)
6618{
6619 return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
6620 le16_to_cpu(eb->h_suballoc_slot),
6621 le64_to_cpu(eb->h_suballoc_loc),
6622 le64_to_cpu(eb->h_blkno),
6623 le16_to_cpu(eb->h_suballoc_bit));
6624}
6625
6626static int ocfs2_zero_func(handle_t *handle, struct buffer_head *bh)
6627{
6628 set_buffer_uptodate(bh);
6629 mark_buffer_dirty(bh);
6630 return 0;
6631}
6632
6633void ocfs2_map_and_dirty_page(struct inode *inode, handle_t *handle,
6634 unsigned int from, unsigned int to,
6635 struct page *page, int zero, u64 *phys)
6636{
6637 int ret, partial = 0;
6638
6639 ret = ocfs2_map_page_blocks(page, phys, inode, from, to, 0);
6640 if (ret)
6641 mlog_errno(ret);
6642
6643 if (zero)
6644 zero_user_segment(page, from, to);
6645
6646 /*
6647 * Need to set the buffers we zero'd into uptodate
6648 * here if they aren't - ocfs2_map_page_blocks()
6649 * might've skipped some
6650 */
6651 ret = walk_page_buffers(handle, page_buffers(page),
6652 from, to, &partial,
6653 ocfs2_zero_func);
6654 if (ret < 0)
6655 mlog_errno(ret);
6656 else if (ocfs2_should_order_data(inode)) {
6657 ret = ocfs2_jbd2_file_inode(handle, inode);
6658 if (ret < 0)
6659 mlog_errno(ret);
6660 }
6661
6662 if (!partial)
6663 SetPageUptodate(page);
6664
6665 flush_dcache_page(page);
6666}
6667
6668static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start,
6669 loff_t end, struct page **pages,
6670 int numpages, u64 phys, handle_t *handle)
6671{
6672 int i;
6673 struct page *page;
6674 unsigned int from, to = PAGE_SIZE;
6675 struct super_block *sb = inode->i_sb;
6676
6677 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
6678
6679 if (numpages == 0)
6680 goto out;
6681
6682 to = PAGE_SIZE;
6683 for(i = 0; i < numpages; i++) {
6684 page = pages[i];
6685
6686 from = start & (PAGE_SIZE - 1);
6687 if ((end >> PAGE_SHIFT) == page->index)
6688 to = end & (PAGE_SIZE - 1);
6689
6690 BUG_ON(from > PAGE_SIZE);
6691 BUG_ON(to > PAGE_SIZE);
6692
6693 ocfs2_map_and_dirty_page(inode, handle, from, to, page, 1,
6694 &phys);
6695
6696 start = (page->index + 1) << PAGE_SHIFT;
6697 }
6698out:
6699 if (pages)
6700 ocfs2_unlock_and_free_pages(pages, numpages);
6701}
6702
6703int ocfs2_grab_pages(struct inode *inode, loff_t start, loff_t end,
6704 struct page **pages, int *num)
6705{
6706 int numpages, ret = 0;
6707 struct address_space *mapping = inode->i_mapping;
6708 unsigned long index;
6709 loff_t last_page_bytes;
6710
6711 BUG_ON(start > end);
6712
6713 numpages = 0;
6714 last_page_bytes = PAGE_ALIGN(end);
6715 index = start >> PAGE_SHIFT;
6716 do {
6717 pages[numpages] = find_or_create_page(mapping, index, GFP_NOFS);
6718 if (!pages[numpages]) {
6719 ret = -ENOMEM;
6720 mlog_errno(ret);
6721 goto out;
6722 }
6723
6724 numpages++;
6725 index++;
6726 } while (index < (last_page_bytes >> PAGE_SHIFT));
6727
6728out:
6729 if (ret != 0) {
6730 if (pages)
6731 ocfs2_unlock_and_free_pages(pages, numpages);
6732 numpages = 0;
6733 }
6734
6735 *num = numpages;
6736
6737 return ret;
6738}
6739
6740static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end,
6741 struct page **pages, int *num)
6742{
6743 struct super_block *sb = inode->i_sb;
6744
6745 BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
6746 (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
6747
6748 return ocfs2_grab_pages(inode, start, end, pages, num);
6749}
6750
6751/*
6752 * Zero the area past i_size but still within an allocated
6753 * cluster. This avoids exposing nonzero data on subsequent file
6754 * extends.
6755 *
6756 * We need to call this before i_size is updated on the inode because
6757 * otherwise block_write_full_page() will skip writeout of pages past
6758 * i_size. The new_i_size parameter is passed for this reason.
6759 */
6760int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
6761 u64 range_start, u64 range_end)
6762{
6763 int ret = 0, numpages;
6764 struct page **pages = NULL;
6765 u64 phys;
6766 unsigned int ext_flags;
6767 struct super_block *sb = inode->i_sb;
6768
6769 /*
6770 * File systems which don't support sparse files zero on every
6771 * extend.
6772 */
6773 if (!ocfs2_sparse_alloc(OCFS2_SB(sb)))
6774 return 0;
6775
6776 pages = kcalloc(ocfs2_pages_per_cluster(sb),
6777 sizeof(struct page *), GFP_NOFS);
6778 if (pages == NULL) {
6779 ret = -ENOMEM;
6780 mlog_errno(ret);
6781 goto out;
6782 }
6783
6784 if (range_start == range_end)
6785 goto out;
6786
6787 ret = ocfs2_extent_map_get_blocks(inode,
6788 range_start >> sb->s_blocksize_bits,
6789 &phys, NULL, &ext_flags);
6790 if (ret) {
6791 mlog_errno(ret);
6792 goto out;
6793 }
6794
6795 /*
6796 * Tail is a hole, or is marked unwritten. In either case, we
6797 * can count on read and write to return/push zero's.
6798 */
6799 if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN)
6800 goto out;
6801
6802 ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages,
6803 &numpages);
6804 if (ret) {
6805 mlog_errno(ret);
6806 goto out;
6807 }
6808
6809 ocfs2_zero_cluster_pages(inode, range_start, range_end, pages,
6810 numpages, phys, handle);
6811
6812 /*
6813 * Initiate writeout of the pages we zero'd here. We don't
6814 * wait on them - the truncate_inode_pages() call later will
6815 * do that for us.
6816 */
6817 ret = filemap_fdatawrite_range(inode->i_mapping, range_start,
6818 range_end - 1);
6819 if (ret)
6820 mlog_errno(ret);
6821
6822out:
6823 kfree(pages);
6824
6825 return ret;
6826}
6827
6828static void ocfs2_zero_dinode_id2_with_xattr(struct inode *inode,
6829 struct ocfs2_dinode *di)
6830{
6831 unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits;
6832 unsigned int xattrsize = le16_to_cpu(di->i_xattr_inline_size);
6833
6834 if (le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_XATTR_FL)
6835 memset(&di->id2, 0, blocksize -
6836 offsetof(struct ocfs2_dinode, id2) -
6837 xattrsize);
6838 else
6839 memset(&di->id2, 0, blocksize -
6840 offsetof(struct ocfs2_dinode, id2));
6841}
6842
6843void ocfs2_dinode_new_extent_list(struct inode *inode,
6844 struct ocfs2_dinode *di)
6845{
6846 ocfs2_zero_dinode_id2_with_xattr(inode, di);
6847 di->id2.i_list.l_tree_depth = 0;
6848 di->id2.i_list.l_next_free_rec = 0;
6849 di->id2.i_list.l_count = cpu_to_le16(
6850 ocfs2_extent_recs_per_inode_with_xattr(inode->i_sb, di));
6851}
6852
6853void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di)
6854{
6855 struct ocfs2_inode_info *oi = OCFS2_I(inode);
6856 struct ocfs2_inline_data *idata = &di->id2.i_data;
6857
6858 spin_lock(&oi->ip_lock);
6859 oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
6860 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
6861 spin_unlock(&oi->ip_lock);
6862
6863 /*
6864 * We clear the entire i_data structure here so that all
6865 * fields can be properly initialized.
6866 */
6867 ocfs2_zero_dinode_id2_with_xattr(inode, di);
6868
6869 idata->id_count = cpu_to_le16(
6870 ocfs2_max_inline_data_with_xattr(inode->i_sb, di));
6871}
6872
6873int ocfs2_convert_inline_data_to_extents(struct inode *inode,
6874 struct buffer_head *di_bh)
6875{
6876 int ret, i, has_data, num_pages = 0;
6877 int need_free = 0;
6878 u32 bit_off, num;
6879 handle_t *handle;
6880 u64 uninitialized_var(block);
6881 struct ocfs2_inode_info *oi = OCFS2_I(inode);
6882 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
6883 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
6884 struct ocfs2_alloc_context *data_ac = NULL;
6885 struct page **pages = NULL;
6886 loff_t end = osb->s_clustersize;
6887 struct ocfs2_extent_tree et;
6888 int did_quota = 0;
6889
6890 has_data = i_size_read(inode) ? 1 : 0;
6891
6892 if (has_data) {
6893 pages = kcalloc(ocfs2_pages_per_cluster(osb->sb),
6894 sizeof(struct page *), GFP_NOFS);
6895 if (pages == NULL) {
6896 ret = -ENOMEM;
6897 mlog_errno(ret);
6898 return ret;
6899 }
6900
6901 ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
6902 if (ret) {
6903 mlog_errno(ret);
6904 goto free_pages;
6905 }
6906 }
6907
6908 handle = ocfs2_start_trans(osb,
6909 ocfs2_inline_to_extents_credits(osb->sb));
6910 if (IS_ERR(handle)) {
6911 ret = PTR_ERR(handle);
6912 mlog_errno(ret);
6913 goto out;
6914 }
6915
6916 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
6917 OCFS2_JOURNAL_ACCESS_WRITE);
6918 if (ret) {
6919 mlog_errno(ret);
6920 goto out_commit;
6921 }
6922
6923 if (has_data) {
6924 unsigned int page_end;
6925 u64 phys;
6926
6927 ret = dquot_alloc_space_nodirty(inode,
6928 ocfs2_clusters_to_bytes(osb->sb, 1));
6929 if (ret)
6930 goto out_commit;
6931 did_quota = 1;
6932
6933 data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
6934
6935 ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off,
6936 &num);
6937 if (ret) {
6938 mlog_errno(ret);
6939 goto out_commit;
6940 }
6941
6942 /*
6943 * Save two copies, one for insert, and one that can
6944 * be changed by ocfs2_map_and_dirty_page() below.
6945 */
6946 block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off);
6947
6948 /*
6949 * Non sparse file systems zero on extend, so no need
6950 * to do that now.
6951 */
6952 if (!ocfs2_sparse_alloc(osb) &&
6953 PAGE_SIZE < osb->s_clustersize)
6954 end = PAGE_SIZE;
6955
6956 ret = ocfs2_grab_eof_pages(inode, 0, end, pages, &num_pages);
6957 if (ret) {
6958 mlog_errno(ret);
6959 need_free = 1;
6960 goto out_commit;
6961 }
6962
6963 /*
6964 * This should populate the 1st page for us and mark
6965 * it up to date.
6966 */
6967 ret = ocfs2_read_inline_data(inode, pages[0], di_bh);
6968 if (ret) {
6969 mlog_errno(ret);
6970 need_free = 1;
6971 goto out_unlock;
6972 }
6973
6974 page_end = PAGE_SIZE;
6975 if (PAGE_SIZE > osb->s_clustersize)
6976 page_end = osb->s_clustersize;
6977
6978 for (i = 0; i < num_pages; i++)
6979 ocfs2_map_and_dirty_page(inode, handle, 0, page_end,
6980 pages[i], i > 0, &phys);
6981 }
6982
6983 spin_lock(&oi->ip_lock);
6984 oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
6985 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
6986 spin_unlock(&oi->ip_lock);
6987
6988 ocfs2_update_inode_fsync_trans(handle, inode, 1);
6989 ocfs2_dinode_new_extent_list(inode, di);
6990
6991 ocfs2_journal_dirty(handle, di_bh);
6992
6993 if (has_data) {
6994 /*
6995 * An error at this point should be extremely rare. If
6996 * this proves to be false, we could always re-build
6997 * the in-inode data from our pages.
6998 */
6999 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7000 ret = ocfs2_insert_extent(handle, &et, 0, block, 1, 0, NULL);
7001 if (ret) {
7002 mlog_errno(ret);
7003 need_free = 1;
7004 goto out_unlock;
7005 }
7006
7007 inode->i_blocks = ocfs2_inode_sector_count(inode);
7008 }
7009
7010out_unlock:
7011 if (pages)
7012 ocfs2_unlock_and_free_pages(pages, num_pages);
7013
7014out_commit:
7015 if (ret < 0 && did_quota)
7016 dquot_free_space_nodirty(inode,
7017 ocfs2_clusters_to_bytes(osb->sb, 1));
7018
7019 if (need_free) {
7020 if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
7021 ocfs2_free_local_alloc_bits(osb, handle, data_ac,
7022 bit_off, num);
7023 else
7024 ocfs2_free_clusters(handle,
7025 data_ac->ac_inode,
7026 data_ac->ac_bh,
7027 ocfs2_clusters_to_blocks(osb->sb, bit_off),
7028 num);
7029 }
7030
7031 ocfs2_commit_trans(osb, handle);
7032
7033out:
7034 if (data_ac)
7035 ocfs2_free_alloc_context(data_ac);
7036free_pages:
7037 kfree(pages);
7038 return ret;
7039}
7040
7041/*
7042 * It is expected, that by the time you call this function,
7043 * inode->i_size and fe->i_size have been adjusted.
7044 *
7045 * WARNING: This will kfree the truncate context
7046 */
7047int ocfs2_commit_truncate(struct ocfs2_super *osb,
7048 struct inode *inode,
7049 struct buffer_head *di_bh)
7050{
7051 int status = 0, i, flags = 0;
7052 u32 new_highest_cpos, range, trunc_cpos, trunc_len, phys_cpos, coff;
7053 u64 blkno = 0;
7054 struct ocfs2_extent_list *el;
7055 struct ocfs2_extent_rec *rec;
7056 struct ocfs2_path *path = NULL;
7057 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7058 struct ocfs2_extent_list *root_el = &(di->id2.i_list);
7059 u64 refcount_loc = le64_to_cpu(di->i_refcount_loc);
7060 struct ocfs2_extent_tree et;
7061 struct ocfs2_cached_dealloc_ctxt dealloc;
7062 struct ocfs2_refcount_tree *ref_tree = NULL;
7063
7064 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7065 ocfs2_init_dealloc_ctxt(&dealloc);
7066
7067 new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
7068 i_size_read(inode));
7069
7070 path = ocfs2_new_path(di_bh, &di->id2.i_list,
7071 ocfs2_journal_access_di);
7072 if (!path) {
7073 status = -ENOMEM;
7074 mlog_errno(status);
7075 goto bail;
7076 }
7077
7078 ocfs2_extent_map_trunc(inode, new_highest_cpos);
7079
7080start:
7081 /*
7082 * Check that we still have allocation to delete.
7083 */
7084 if (OCFS2_I(inode)->ip_clusters == 0) {
7085 status = 0;
7086 goto bail;
7087 }
7088
7089 /*
7090 * Truncate always works against the rightmost tree branch.
7091 */
7092 status = ocfs2_find_path(INODE_CACHE(inode), path, UINT_MAX);
7093 if (status) {
7094 mlog_errno(status);
7095 goto bail;
7096 }
7097
7098 trace_ocfs2_commit_truncate(
7099 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7100 new_highest_cpos,
7101 OCFS2_I(inode)->ip_clusters,
7102 path->p_tree_depth);
7103
7104 /*
7105 * By now, el will point to the extent list on the bottom most
7106 * portion of this tree. Only the tail record is considered in
7107 * each pass.
7108 *
7109 * We handle the following cases, in order:
7110 * - empty extent: delete the remaining branch
7111 * - remove the entire record
7112 * - remove a partial record
7113 * - no record needs to be removed (truncate has completed)
7114 */
7115 el = path_leaf_el(path);
7116 if (le16_to_cpu(el->l_next_free_rec) == 0) {
7117 ocfs2_error(inode->i_sb,
7118 "Inode %llu has empty extent block at %llu\n",
7119 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7120 (unsigned long long)path_leaf_bh(path)->b_blocknr);
7121 status = -EROFS;
7122 goto bail;
7123 }
7124
7125 i = le16_to_cpu(el->l_next_free_rec) - 1;
7126 rec = &el->l_recs[i];
7127 flags = rec->e_flags;
7128 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
7129
7130 if (i == 0 && ocfs2_is_empty_extent(rec)) {
7131 /*
7132 * Lower levels depend on this never happening, but it's best
7133 * to check it up here before changing the tree.
7134 */
7135 if (root_el->l_tree_depth && rec->e_int_clusters == 0) {
7136 mlog(ML_ERROR, "Inode %lu has an empty "
7137 "extent record, depth %u\n", inode->i_ino,
7138 le16_to_cpu(root_el->l_tree_depth));
7139 status = ocfs2_remove_rightmost_empty_extent(osb,
7140 &et, path, &dealloc);
7141 if (status) {
7142 mlog_errno(status);
7143 goto bail;
7144 }
7145
7146 ocfs2_reinit_path(path, 1);
7147 goto start;
7148 } else {
7149 trunc_cpos = le32_to_cpu(rec->e_cpos);
7150 trunc_len = 0;
7151 blkno = 0;
7152 }
7153 } else if (le32_to_cpu(rec->e_cpos) >= new_highest_cpos) {
7154 /*
7155 * Truncate entire record.
7156 */
7157 trunc_cpos = le32_to_cpu(rec->e_cpos);
7158 trunc_len = ocfs2_rec_clusters(el, rec);
7159 blkno = le64_to_cpu(rec->e_blkno);
7160 } else if (range > new_highest_cpos) {
7161 /*
7162 * Partial truncate. it also should be
7163 * the last truncate we're doing.
7164 */
7165 trunc_cpos = new_highest_cpos;
7166 trunc_len = range - new_highest_cpos;
7167 coff = new_highest_cpos - le32_to_cpu(rec->e_cpos);
7168 blkno = le64_to_cpu(rec->e_blkno) +
7169 ocfs2_clusters_to_blocks(inode->i_sb, coff);
7170 } else {
7171 /*
7172 * Truncate completed, leave happily.
7173 */
7174 status = 0;
7175 goto bail;
7176 }
7177
7178 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
7179
7180 if ((flags & OCFS2_EXT_REFCOUNTED) && trunc_len && !ref_tree) {
7181 status = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
7182 &ref_tree, NULL);
7183 if (status) {
7184 mlog_errno(status);
7185 goto bail;
7186 }
7187 }
7188
7189 status = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
7190 phys_cpos, trunc_len, flags, &dealloc,
7191 refcount_loc, true);
7192 if (status < 0) {
7193 mlog_errno(status);
7194 goto bail;
7195 }
7196
7197 ocfs2_reinit_path(path, 1);
7198
7199 /*
7200 * The check above will catch the case where we've truncated
7201 * away all allocation.
7202 */
7203 goto start;
7204
7205bail:
7206 if (ref_tree)
7207 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
7208
7209 ocfs2_schedule_truncate_log_flush(osb, 1);
7210
7211 ocfs2_run_deallocs(osb, &dealloc);
7212
7213 ocfs2_free_path(path);
7214
7215 return status;
7216}
7217
7218/*
7219 * 'start' is inclusive, 'end' is not.
7220 */
7221int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh,
7222 unsigned int start, unsigned int end, int trunc)
7223{
7224 int ret;
7225 unsigned int numbytes;
7226 handle_t *handle;
7227 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7228 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7229 struct ocfs2_inline_data *idata = &di->id2.i_data;
7230
7231 if (end > i_size_read(inode))
7232 end = i_size_read(inode);
7233
7234 BUG_ON(start > end);
7235
7236 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
7237 !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) ||
7238 !ocfs2_supports_inline_data(osb)) {
7239 ocfs2_error(inode->i_sb,
7240 "Inline data flags for inode %llu don't agree! Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7241 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7242 le16_to_cpu(di->i_dyn_features),
7243 OCFS2_I(inode)->ip_dyn_features,
7244 osb->s_feature_incompat);
7245 ret = -EROFS;
7246 goto out;
7247 }
7248
7249 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
7250 if (IS_ERR(handle)) {
7251 ret = PTR_ERR(handle);
7252 mlog_errno(ret);
7253 goto out;
7254 }
7255
7256 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7257 OCFS2_JOURNAL_ACCESS_WRITE);
7258 if (ret) {
7259 mlog_errno(ret);
7260 goto out_commit;
7261 }
7262
7263 numbytes = end - start;
7264 memset(idata->id_data + start, 0, numbytes);
7265
7266 /*
7267 * No need to worry about the data page here - it's been
7268 * truncated already and inline data doesn't need it for
7269 * pushing zero's to disk, so we'll let readpage pick it up
7270 * later.
7271 */
7272 if (trunc) {
7273 i_size_write(inode, start);
7274 di->i_size = cpu_to_le64(start);
7275 }
7276
7277 inode->i_blocks = ocfs2_inode_sector_count(inode);
7278 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
7279
7280 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
7281 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
7282
7283 ocfs2_update_inode_fsync_trans(handle, inode, 1);
7284 ocfs2_journal_dirty(handle, di_bh);
7285
7286out_commit:
7287 ocfs2_commit_trans(osb, handle);
7288
7289out:
7290 return ret;
7291}
7292
7293static int ocfs2_trim_extent(struct super_block *sb,
7294 struct ocfs2_group_desc *gd,
7295 u32 start, u32 count)
7296{
7297 u64 discard, bcount;
7298
7299 bcount = ocfs2_clusters_to_blocks(sb, count);
7300 discard = le64_to_cpu(gd->bg_blkno) +
7301 ocfs2_clusters_to_blocks(sb, start);
7302
7303 trace_ocfs2_trim_extent(sb, (unsigned long long)discard, bcount);
7304
7305 return sb_issue_discard(sb, discard, bcount, GFP_NOFS, 0);
7306}
7307
7308static int ocfs2_trim_group(struct super_block *sb,
7309 struct ocfs2_group_desc *gd,
7310 u32 start, u32 max, u32 minbits)
7311{
7312 int ret = 0, count = 0, next;
7313 void *bitmap = gd->bg_bitmap;
7314
7315 if (le16_to_cpu(gd->bg_free_bits_count) < minbits)
7316 return 0;
7317
7318 trace_ocfs2_trim_group((unsigned long long)le64_to_cpu(gd->bg_blkno),
7319 start, max, minbits);
7320
7321 while (start < max) {
7322 start = ocfs2_find_next_zero_bit(bitmap, max, start);
7323 if (start >= max)
7324 break;
7325 next = ocfs2_find_next_bit(bitmap, max, start);
7326
7327 if ((next - start) >= minbits) {
7328 ret = ocfs2_trim_extent(sb, gd,
7329 start, next - start);
7330 if (ret < 0) {
7331 mlog_errno(ret);
7332 break;
7333 }
7334 count += next - start;
7335 }
7336 start = next + 1;
7337
7338 if (fatal_signal_pending(current)) {
7339 count = -ERESTARTSYS;
7340 break;
7341 }
7342
7343 if ((le16_to_cpu(gd->bg_free_bits_count) - count) < minbits)
7344 break;
7345 }
7346
7347 if (ret < 0)
7348 count = ret;
7349
7350 return count;
7351}
7352
7353int ocfs2_trim_fs(struct super_block *sb, struct fstrim_range *range)
7354{
7355 struct ocfs2_super *osb = OCFS2_SB(sb);
7356 u64 start, len, trimmed, first_group, last_group, group;
7357 int ret, cnt;
7358 u32 first_bit, last_bit, minlen;
7359 struct buffer_head *main_bm_bh = NULL;
7360 struct inode *main_bm_inode = NULL;
7361 struct buffer_head *gd_bh = NULL;
7362 struct ocfs2_dinode *main_bm;
7363 struct ocfs2_group_desc *gd = NULL;
7364
7365 start = range->start >> osb->s_clustersize_bits;
7366 len = range->len >> osb->s_clustersize_bits;
7367 minlen = range->minlen >> osb->s_clustersize_bits;
7368
7369 if (minlen >= osb->bitmap_cpg || range->len < sb->s_blocksize)
7370 return -EINVAL;
7371
7372 main_bm_inode = ocfs2_get_system_file_inode(osb,
7373 GLOBAL_BITMAP_SYSTEM_INODE,
7374 OCFS2_INVALID_SLOT);
7375 if (!main_bm_inode) {
7376 ret = -EIO;
7377 mlog_errno(ret);
7378 goto out;
7379 }
7380
7381 inode_lock(main_bm_inode);
7382
7383 ret = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 0);
7384 if (ret < 0) {
7385 mlog_errno(ret);
7386 goto out_mutex;
7387 }
7388 main_bm = (struct ocfs2_dinode *)main_bm_bh->b_data;
7389
7390 if (start >= le32_to_cpu(main_bm->i_clusters)) {
7391 ret = -EINVAL;
7392 goto out_unlock;
7393 }
7394
7395 len = range->len >> osb->s_clustersize_bits;
7396 if (start + len > le32_to_cpu(main_bm->i_clusters))
7397 len = le32_to_cpu(main_bm->i_clusters) - start;
7398
7399 trace_ocfs2_trim_fs(start, len, minlen);
7400
7401 /* Determine first and last group to examine based on start and len */
7402 first_group = ocfs2_which_cluster_group(main_bm_inode, start);
7403 if (first_group == osb->first_cluster_group_blkno)
7404 first_bit = start;
7405 else
7406 first_bit = start - ocfs2_blocks_to_clusters(sb, first_group);
7407 last_group = ocfs2_which_cluster_group(main_bm_inode, start + len - 1);
7408 last_bit = osb->bitmap_cpg;
7409
7410 trimmed = 0;
7411 for (group = first_group; group <= last_group;) {
7412 if (first_bit + len >= osb->bitmap_cpg)
7413 last_bit = osb->bitmap_cpg;
7414 else
7415 last_bit = first_bit + len;
7416
7417 ret = ocfs2_read_group_descriptor(main_bm_inode,
7418 main_bm, group,
7419 &gd_bh);
7420 if (ret < 0) {
7421 mlog_errno(ret);
7422 break;
7423 }
7424
7425 gd = (struct ocfs2_group_desc *)gd_bh->b_data;
7426 cnt = ocfs2_trim_group(sb, gd, first_bit, last_bit, minlen);
7427 brelse(gd_bh);
7428 gd_bh = NULL;
7429 if (cnt < 0) {
7430 ret = cnt;
7431 mlog_errno(ret);
7432 break;
7433 }
7434
7435 trimmed += cnt;
7436 len -= osb->bitmap_cpg - first_bit;
7437 first_bit = 0;
7438 if (group == osb->first_cluster_group_blkno)
7439 group = ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7440 else
7441 group += ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7442 }
7443 range->len = trimmed * sb->s_blocksize;
7444out_unlock:
7445 ocfs2_inode_unlock(main_bm_inode, 0);
7446 brelse(main_bm_bh);
7447out_mutex:
7448 inode_unlock(main_bm_inode);
7449 iput(main_bm_inode);
7450out:
7451 return ret;
7452}
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * alloc.c
4 *
5 * Extent allocs and frees
6 *
7 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
8 */
9
10#include <linux/fs.h>
11#include <linux/types.h>
12#include <linux/slab.h>
13#include <linux/highmem.h>
14#include <linux/swap.h>
15#include <linux/quotaops.h>
16#include <linux/blkdev.h>
17#include <linux/sched/signal.h>
18
19#include <cluster/masklog.h>
20
21#include "ocfs2.h"
22
23#include "alloc.h"
24#include "aops.h"
25#include "blockcheck.h"
26#include "dlmglue.h"
27#include "extent_map.h"
28#include "inode.h"
29#include "journal.h"
30#include "localalloc.h"
31#include "suballoc.h"
32#include "sysfile.h"
33#include "file.h"
34#include "super.h"
35#include "uptodate.h"
36#include "xattr.h"
37#include "refcounttree.h"
38#include "ocfs2_trace.h"
39
40#include "buffer_head_io.h"
41
42enum ocfs2_contig_type {
43 CONTIG_NONE = 0,
44 CONTIG_LEFT,
45 CONTIG_RIGHT,
46 CONTIG_LEFTRIGHT,
47};
48
49static enum ocfs2_contig_type
50 ocfs2_extent_rec_contig(struct super_block *sb,
51 struct ocfs2_extent_rec *ext,
52 struct ocfs2_extent_rec *insert_rec);
53/*
54 * Operations for a specific extent tree type.
55 *
56 * To implement an on-disk btree (extent tree) type in ocfs2, add
57 * an ocfs2_extent_tree_operations structure and the matching
58 * ocfs2_init_<thingy>_extent_tree() function. That's pretty much it
59 * for the allocation portion of the extent tree.
60 */
61struct ocfs2_extent_tree_operations {
62 /*
63 * last_eb_blk is the block number of the right most leaf extent
64 * block. Most on-disk structures containing an extent tree store
65 * this value for fast access. The ->eo_set_last_eb_blk() and
66 * ->eo_get_last_eb_blk() operations access this value. They are
67 * both required.
68 */
69 void (*eo_set_last_eb_blk)(struct ocfs2_extent_tree *et,
70 u64 blkno);
71 u64 (*eo_get_last_eb_blk)(struct ocfs2_extent_tree *et);
72
73 /*
74 * The on-disk structure usually keeps track of how many total
75 * clusters are stored in this extent tree. This function updates
76 * that value. new_clusters is the delta, and must be
77 * added to the total. Required.
78 */
79 void (*eo_update_clusters)(struct ocfs2_extent_tree *et,
80 u32 new_clusters);
81
82 /*
83 * If this extent tree is supported by an extent map, insert
84 * a record into the map.
85 */
86 void (*eo_extent_map_insert)(struct ocfs2_extent_tree *et,
87 struct ocfs2_extent_rec *rec);
88
89 /*
90 * If this extent tree is supported by an extent map, truncate the
91 * map to clusters,
92 */
93 void (*eo_extent_map_truncate)(struct ocfs2_extent_tree *et,
94 u32 clusters);
95
96 /*
97 * If ->eo_insert_check() exists, it is called before rec is
98 * inserted into the extent tree. It is optional.
99 */
100 int (*eo_insert_check)(struct ocfs2_extent_tree *et,
101 struct ocfs2_extent_rec *rec);
102 int (*eo_sanity_check)(struct ocfs2_extent_tree *et);
103
104 /*
105 * --------------------------------------------------------------
106 * The remaining are internal to ocfs2_extent_tree and don't have
107 * accessor functions
108 */
109
110 /*
111 * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el.
112 * It is required.
113 */
114 void (*eo_fill_root_el)(struct ocfs2_extent_tree *et);
115
116 /*
117 * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if
118 * it exists. If it does not, et->et_max_leaf_clusters is set
119 * to 0 (unlimited). Optional.
120 */
121 void (*eo_fill_max_leaf_clusters)(struct ocfs2_extent_tree *et);
122
123 /*
124 * ->eo_extent_contig test whether the 2 ocfs2_extent_rec
125 * are contiguous or not. Optional. Don't need to set it if use
126 * ocfs2_extent_rec as the tree leaf.
127 */
128 enum ocfs2_contig_type
129 (*eo_extent_contig)(struct ocfs2_extent_tree *et,
130 struct ocfs2_extent_rec *ext,
131 struct ocfs2_extent_rec *insert_rec);
132};
133
134
135/*
136 * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check
137 * in the methods.
138 */
139static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et);
140static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
141 u64 blkno);
142static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
143 u32 clusters);
144static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
145 struct ocfs2_extent_rec *rec);
146static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
147 u32 clusters);
148static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
149 struct ocfs2_extent_rec *rec);
150static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et);
151static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et);
152
153static int ocfs2_reuse_blk_from_dealloc(handle_t *handle,
154 struct ocfs2_extent_tree *et,
155 struct buffer_head **new_eb_bh,
156 int blk_wanted, int *blk_given);
157static int ocfs2_is_dealloc_empty(struct ocfs2_extent_tree *et);
158
159static const struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops = {
160 .eo_set_last_eb_blk = ocfs2_dinode_set_last_eb_blk,
161 .eo_get_last_eb_blk = ocfs2_dinode_get_last_eb_blk,
162 .eo_update_clusters = ocfs2_dinode_update_clusters,
163 .eo_extent_map_insert = ocfs2_dinode_extent_map_insert,
164 .eo_extent_map_truncate = ocfs2_dinode_extent_map_truncate,
165 .eo_insert_check = ocfs2_dinode_insert_check,
166 .eo_sanity_check = ocfs2_dinode_sanity_check,
167 .eo_fill_root_el = ocfs2_dinode_fill_root_el,
168};
169
170static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
171 u64 blkno)
172{
173 struct ocfs2_dinode *di = et->et_object;
174
175 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
176 di->i_last_eb_blk = cpu_to_le64(blkno);
177}
178
179static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et)
180{
181 struct ocfs2_dinode *di = et->et_object;
182
183 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
184 return le64_to_cpu(di->i_last_eb_blk);
185}
186
187static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
188 u32 clusters)
189{
190 struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
191 struct ocfs2_dinode *di = et->et_object;
192
193 le32_add_cpu(&di->i_clusters, clusters);
194 spin_lock(&oi->ip_lock);
195 oi->ip_clusters = le32_to_cpu(di->i_clusters);
196 spin_unlock(&oi->ip_lock);
197}
198
199static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
200 struct ocfs2_extent_rec *rec)
201{
202 struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
203
204 ocfs2_extent_map_insert_rec(inode, rec);
205}
206
207static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
208 u32 clusters)
209{
210 struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
211
212 ocfs2_extent_map_trunc(inode, clusters);
213}
214
215static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
216 struct ocfs2_extent_rec *rec)
217{
218 struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
219 struct ocfs2_super *osb = OCFS2_SB(oi->vfs_inode.i_sb);
220
221 BUG_ON(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL);
222 mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) &&
223 (oi->ip_clusters != le32_to_cpu(rec->e_cpos)),
224 "Device %s, asking for sparse allocation: inode %llu, "
225 "cpos %u, clusters %u\n",
226 osb->dev_str,
227 (unsigned long long)oi->ip_blkno,
228 rec->e_cpos, oi->ip_clusters);
229
230 return 0;
231}
232
233static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et)
234{
235 struct ocfs2_dinode *di = et->et_object;
236
237 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
238 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
239
240 return 0;
241}
242
243static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et)
244{
245 struct ocfs2_dinode *di = et->et_object;
246
247 et->et_root_el = &di->id2.i_list;
248}
249
250
251static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree *et)
252{
253 struct ocfs2_xattr_value_buf *vb = et->et_object;
254
255 et->et_root_el = &vb->vb_xv->xr_list;
256}
257
258static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree *et,
259 u64 blkno)
260{
261 struct ocfs2_xattr_value_buf *vb = et->et_object;
262
263 vb->vb_xv->xr_last_eb_blk = cpu_to_le64(blkno);
264}
265
266static u64 ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree *et)
267{
268 struct ocfs2_xattr_value_buf *vb = et->et_object;
269
270 return le64_to_cpu(vb->vb_xv->xr_last_eb_blk);
271}
272
273static void ocfs2_xattr_value_update_clusters(struct ocfs2_extent_tree *et,
274 u32 clusters)
275{
276 struct ocfs2_xattr_value_buf *vb = et->et_object;
277
278 le32_add_cpu(&vb->vb_xv->xr_clusters, clusters);
279}
280
281static const struct ocfs2_extent_tree_operations ocfs2_xattr_value_et_ops = {
282 .eo_set_last_eb_blk = ocfs2_xattr_value_set_last_eb_blk,
283 .eo_get_last_eb_blk = ocfs2_xattr_value_get_last_eb_blk,
284 .eo_update_clusters = ocfs2_xattr_value_update_clusters,
285 .eo_fill_root_el = ocfs2_xattr_value_fill_root_el,
286};
287
288static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree *et)
289{
290 struct ocfs2_xattr_block *xb = et->et_object;
291
292 et->et_root_el = &xb->xb_attrs.xb_root.xt_list;
293}
294
295static void ocfs2_xattr_tree_fill_max_leaf_clusters(struct ocfs2_extent_tree *et)
296{
297 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
298 et->et_max_leaf_clusters =
299 ocfs2_clusters_for_bytes(sb, OCFS2_MAX_XATTR_TREE_LEAF_SIZE);
300}
301
302static void ocfs2_xattr_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
303 u64 blkno)
304{
305 struct ocfs2_xattr_block *xb = et->et_object;
306 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
307
308 xt->xt_last_eb_blk = cpu_to_le64(blkno);
309}
310
311static u64 ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
312{
313 struct ocfs2_xattr_block *xb = et->et_object;
314 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
315
316 return le64_to_cpu(xt->xt_last_eb_blk);
317}
318
319static void ocfs2_xattr_tree_update_clusters(struct ocfs2_extent_tree *et,
320 u32 clusters)
321{
322 struct ocfs2_xattr_block *xb = et->et_object;
323
324 le32_add_cpu(&xb->xb_attrs.xb_root.xt_clusters, clusters);
325}
326
327static const struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops = {
328 .eo_set_last_eb_blk = ocfs2_xattr_tree_set_last_eb_blk,
329 .eo_get_last_eb_blk = ocfs2_xattr_tree_get_last_eb_blk,
330 .eo_update_clusters = ocfs2_xattr_tree_update_clusters,
331 .eo_fill_root_el = ocfs2_xattr_tree_fill_root_el,
332 .eo_fill_max_leaf_clusters = ocfs2_xattr_tree_fill_max_leaf_clusters,
333};
334
335static void ocfs2_dx_root_set_last_eb_blk(struct ocfs2_extent_tree *et,
336 u64 blkno)
337{
338 struct ocfs2_dx_root_block *dx_root = et->et_object;
339
340 dx_root->dr_last_eb_blk = cpu_to_le64(blkno);
341}
342
343static u64 ocfs2_dx_root_get_last_eb_blk(struct ocfs2_extent_tree *et)
344{
345 struct ocfs2_dx_root_block *dx_root = et->et_object;
346
347 return le64_to_cpu(dx_root->dr_last_eb_blk);
348}
349
350static void ocfs2_dx_root_update_clusters(struct ocfs2_extent_tree *et,
351 u32 clusters)
352{
353 struct ocfs2_dx_root_block *dx_root = et->et_object;
354
355 le32_add_cpu(&dx_root->dr_clusters, clusters);
356}
357
358static int ocfs2_dx_root_sanity_check(struct ocfs2_extent_tree *et)
359{
360 struct ocfs2_dx_root_block *dx_root = et->et_object;
361
362 BUG_ON(!OCFS2_IS_VALID_DX_ROOT(dx_root));
363
364 return 0;
365}
366
367static void ocfs2_dx_root_fill_root_el(struct ocfs2_extent_tree *et)
368{
369 struct ocfs2_dx_root_block *dx_root = et->et_object;
370
371 et->et_root_el = &dx_root->dr_list;
372}
373
374static const struct ocfs2_extent_tree_operations ocfs2_dx_root_et_ops = {
375 .eo_set_last_eb_blk = ocfs2_dx_root_set_last_eb_blk,
376 .eo_get_last_eb_blk = ocfs2_dx_root_get_last_eb_blk,
377 .eo_update_clusters = ocfs2_dx_root_update_clusters,
378 .eo_sanity_check = ocfs2_dx_root_sanity_check,
379 .eo_fill_root_el = ocfs2_dx_root_fill_root_el,
380};
381
382static void ocfs2_refcount_tree_fill_root_el(struct ocfs2_extent_tree *et)
383{
384 struct ocfs2_refcount_block *rb = et->et_object;
385
386 et->et_root_el = &rb->rf_list;
387}
388
389static void ocfs2_refcount_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
390 u64 blkno)
391{
392 struct ocfs2_refcount_block *rb = et->et_object;
393
394 rb->rf_last_eb_blk = cpu_to_le64(blkno);
395}
396
397static u64 ocfs2_refcount_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
398{
399 struct ocfs2_refcount_block *rb = et->et_object;
400
401 return le64_to_cpu(rb->rf_last_eb_blk);
402}
403
404static void ocfs2_refcount_tree_update_clusters(struct ocfs2_extent_tree *et,
405 u32 clusters)
406{
407 struct ocfs2_refcount_block *rb = et->et_object;
408
409 le32_add_cpu(&rb->rf_clusters, clusters);
410}
411
412static enum ocfs2_contig_type
413ocfs2_refcount_tree_extent_contig(struct ocfs2_extent_tree *et,
414 struct ocfs2_extent_rec *ext,
415 struct ocfs2_extent_rec *insert_rec)
416{
417 return CONTIG_NONE;
418}
419
420static const struct ocfs2_extent_tree_operations ocfs2_refcount_tree_et_ops = {
421 .eo_set_last_eb_blk = ocfs2_refcount_tree_set_last_eb_blk,
422 .eo_get_last_eb_blk = ocfs2_refcount_tree_get_last_eb_blk,
423 .eo_update_clusters = ocfs2_refcount_tree_update_clusters,
424 .eo_fill_root_el = ocfs2_refcount_tree_fill_root_el,
425 .eo_extent_contig = ocfs2_refcount_tree_extent_contig,
426};
427
428static void __ocfs2_init_extent_tree(struct ocfs2_extent_tree *et,
429 struct ocfs2_caching_info *ci,
430 struct buffer_head *bh,
431 ocfs2_journal_access_func access,
432 void *obj,
433 const struct ocfs2_extent_tree_operations *ops)
434{
435 et->et_ops = ops;
436 et->et_root_bh = bh;
437 et->et_ci = ci;
438 et->et_root_journal_access = access;
439 if (!obj)
440 obj = (void *)bh->b_data;
441 et->et_object = obj;
442 et->et_dealloc = NULL;
443
444 et->et_ops->eo_fill_root_el(et);
445 if (!et->et_ops->eo_fill_max_leaf_clusters)
446 et->et_max_leaf_clusters = 0;
447 else
448 et->et_ops->eo_fill_max_leaf_clusters(et);
449}
450
451void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree *et,
452 struct ocfs2_caching_info *ci,
453 struct buffer_head *bh)
454{
455 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_di,
456 NULL, &ocfs2_dinode_et_ops);
457}
458
459void ocfs2_init_xattr_tree_extent_tree(struct ocfs2_extent_tree *et,
460 struct ocfs2_caching_info *ci,
461 struct buffer_head *bh)
462{
463 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_xb,
464 NULL, &ocfs2_xattr_tree_et_ops);
465}
466
467void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree *et,
468 struct ocfs2_caching_info *ci,
469 struct ocfs2_xattr_value_buf *vb)
470{
471 __ocfs2_init_extent_tree(et, ci, vb->vb_bh, vb->vb_access, vb,
472 &ocfs2_xattr_value_et_ops);
473}
474
475void ocfs2_init_dx_root_extent_tree(struct ocfs2_extent_tree *et,
476 struct ocfs2_caching_info *ci,
477 struct buffer_head *bh)
478{
479 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_dr,
480 NULL, &ocfs2_dx_root_et_ops);
481}
482
483void ocfs2_init_refcount_extent_tree(struct ocfs2_extent_tree *et,
484 struct ocfs2_caching_info *ci,
485 struct buffer_head *bh)
486{
487 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_rb,
488 NULL, &ocfs2_refcount_tree_et_ops);
489}
490
491static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree *et,
492 u64 new_last_eb_blk)
493{
494 et->et_ops->eo_set_last_eb_blk(et, new_last_eb_blk);
495}
496
497static inline u64 ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree *et)
498{
499 return et->et_ops->eo_get_last_eb_blk(et);
500}
501
502static inline void ocfs2_et_update_clusters(struct ocfs2_extent_tree *et,
503 u32 clusters)
504{
505 et->et_ops->eo_update_clusters(et, clusters);
506}
507
508static inline void ocfs2_et_extent_map_insert(struct ocfs2_extent_tree *et,
509 struct ocfs2_extent_rec *rec)
510{
511 if (et->et_ops->eo_extent_map_insert)
512 et->et_ops->eo_extent_map_insert(et, rec);
513}
514
515static inline void ocfs2_et_extent_map_truncate(struct ocfs2_extent_tree *et,
516 u32 clusters)
517{
518 if (et->et_ops->eo_extent_map_truncate)
519 et->et_ops->eo_extent_map_truncate(et, clusters);
520}
521
522static inline int ocfs2_et_root_journal_access(handle_t *handle,
523 struct ocfs2_extent_tree *et,
524 int type)
525{
526 return et->et_root_journal_access(handle, et->et_ci, et->et_root_bh,
527 type);
528}
529
530static inline enum ocfs2_contig_type
531 ocfs2_et_extent_contig(struct ocfs2_extent_tree *et,
532 struct ocfs2_extent_rec *rec,
533 struct ocfs2_extent_rec *insert_rec)
534{
535 if (et->et_ops->eo_extent_contig)
536 return et->et_ops->eo_extent_contig(et, rec, insert_rec);
537
538 return ocfs2_extent_rec_contig(
539 ocfs2_metadata_cache_get_super(et->et_ci),
540 rec, insert_rec);
541}
542
543static inline int ocfs2_et_insert_check(struct ocfs2_extent_tree *et,
544 struct ocfs2_extent_rec *rec)
545{
546 int ret = 0;
547
548 if (et->et_ops->eo_insert_check)
549 ret = et->et_ops->eo_insert_check(et, rec);
550 return ret;
551}
552
553static inline int ocfs2_et_sanity_check(struct ocfs2_extent_tree *et)
554{
555 int ret = 0;
556
557 if (et->et_ops->eo_sanity_check)
558 ret = et->et_ops->eo_sanity_check(et);
559 return ret;
560}
561
562static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
563 struct ocfs2_extent_block *eb);
564static void ocfs2_adjust_rightmost_records(handle_t *handle,
565 struct ocfs2_extent_tree *et,
566 struct ocfs2_path *path,
567 struct ocfs2_extent_rec *insert_rec);
568/*
569 * Reset the actual path elements so that we can re-use the structure
570 * to build another path. Generally, this involves freeing the buffer
571 * heads.
572 */
573void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root)
574{
575 int i, start = 0, depth = 0;
576 struct ocfs2_path_item *node;
577
578 if (keep_root)
579 start = 1;
580
581 for(i = start; i < path_num_items(path); i++) {
582 node = &path->p_node[i];
583
584 brelse(node->bh);
585 node->bh = NULL;
586 node->el = NULL;
587 }
588
589 /*
590 * Tree depth may change during truncate, or insert. If we're
591 * keeping the root extent list, then make sure that our path
592 * structure reflects the proper depth.
593 */
594 if (keep_root)
595 depth = le16_to_cpu(path_root_el(path)->l_tree_depth);
596 else
597 path_root_access(path) = NULL;
598
599 path->p_tree_depth = depth;
600}
601
602void ocfs2_free_path(struct ocfs2_path *path)
603{
604 if (path) {
605 ocfs2_reinit_path(path, 0);
606 kfree(path);
607 }
608}
609
610/*
611 * All the elements of src into dest. After this call, src could be freed
612 * without affecting dest.
613 *
614 * Both paths should have the same root. Any non-root elements of dest
615 * will be freed.
616 */
617static void ocfs2_cp_path(struct ocfs2_path *dest, struct ocfs2_path *src)
618{
619 int i;
620
621 BUG_ON(path_root_bh(dest) != path_root_bh(src));
622 BUG_ON(path_root_el(dest) != path_root_el(src));
623 BUG_ON(path_root_access(dest) != path_root_access(src));
624
625 ocfs2_reinit_path(dest, 1);
626
627 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
628 dest->p_node[i].bh = src->p_node[i].bh;
629 dest->p_node[i].el = src->p_node[i].el;
630
631 if (dest->p_node[i].bh)
632 get_bh(dest->p_node[i].bh);
633 }
634}
635
636/*
637 * Make the *dest path the same as src and re-initialize src path to
638 * have a root only.
639 */
640static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src)
641{
642 int i;
643
644 BUG_ON(path_root_bh(dest) != path_root_bh(src));
645 BUG_ON(path_root_access(dest) != path_root_access(src));
646
647 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
648 brelse(dest->p_node[i].bh);
649
650 dest->p_node[i].bh = src->p_node[i].bh;
651 dest->p_node[i].el = src->p_node[i].el;
652
653 src->p_node[i].bh = NULL;
654 src->p_node[i].el = NULL;
655 }
656}
657
658/*
659 * Insert an extent block at given index.
660 *
661 * This will not take an additional reference on eb_bh.
662 */
663static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index,
664 struct buffer_head *eb_bh)
665{
666 struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data;
667
668 /*
669 * Right now, no root bh is an extent block, so this helps
670 * catch code errors with dinode trees. The assertion can be
671 * safely removed if we ever need to insert extent block
672 * structures at the root.
673 */
674 BUG_ON(index == 0);
675
676 path->p_node[index].bh = eb_bh;
677 path->p_node[index].el = &eb->h_list;
678}
679
680static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh,
681 struct ocfs2_extent_list *root_el,
682 ocfs2_journal_access_func access)
683{
684 struct ocfs2_path *path;
685
686 BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH);
687
688 path = kzalloc(sizeof(*path), GFP_NOFS);
689 if (path) {
690 path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth);
691 get_bh(root_bh);
692 path_root_bh(path) = root_bh;
693 path_root_el(path) = root_el;
694 path_root_access(path) = access;
695 }
696
697 return path;
698}
699
700struct ocfs2_path *ocfs2_new_path_from_path(struct ocfs2_path *path)
701{
702 return ocfs2_new_path(path_root_bh(path), path_root_el(path),
703 path_root_access(path));
704}
705
706struct ocfs2_path *ocfs2_new_path_from_et(struct ocfs2_extent_tree *et)
707{
708 return ocfs2_new_path(et->et_root_bh, et->et_root_el,
709 et->et_root_journal_access);
710}
711
712/*
713 * Journal the buffer at depth idx. All idx>0 are extent_blocks,
714 * otherwise it's the root_access function.
715 *
716 * I don't like the way this function's name looks next to
717 * ocfs2_journal_access_path(), but I don't have a better one.
718 */
719int ocfs2_path_bh_journal_access(handle_t *handle,
720 struct ocfs2_caching_info *ci,
721 struct ocfs2_path *path,
722 int idx)
723{
724 ocfs2_journal_access_func access = path_root_access(path);
725
726 if (!access)
727 access = ocfs2_journal_access;
728
729 if (idx)
730 access = ocfs2_journal_access_eb;
731
732 return access(handle, ci, path->p_node[idx].bh,
733 OCFS2_JOURNAL_ACCESS_WRITE);
734}
735
736/*
737 * Convenience function to journal all components in a path.
738 */
739int ocfs2_journal_access_path(struct ocfs2_caching_info *ci,
740 handle_t *handle,
741 struct ocfs2_path *path)
742{
743 int i, ret = 0;
744
745 if (!path)
746 goto out;
747
748 for(i = 0; i < path_num_items(path); i++) {
749 ret = ocfs2_path_bh_journal_access(handle, ci, path, i);
750 if (ret < 0) {
751 mlog_errno(ret);
752 goto out;
753 }
754 }
755
756out:
757 return ret;
758}
759
760/*
761 * Return the index of the extent record which contains cluster #v_cluster.
762 * -1 is returned if it was not found.
763 *
764 * Should work fine on interior and exterior nodes.
765 */
766int ocfs2_search_extent_list(struct ocfs2_extent_list *el, u32 v_cluster)
767{
768 int ret = -1;
769 int i;
770 struct ocfs2_extent_rec *rec;
771 u32 rec_end, rec_start, clusters;
772
773 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
774 rec = &el->l_recs[i];
775
776 rec_start = le32_to_cpu(rec->e_cpos);
777 clusters = ocfs2_rec_clusters(el, rec);
778
779 rec_end = rec_start + clusters;
780
781 if (v_cluster >= rec_start && v_cluster < rec_end) {
782 ret = i;
783 break;
784 }
785 }
786
787 return ret;
788}
789
790/*
791 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
792 * ocfs2_extent_rec_contig only work properly against leaf nodes!
793 */
794static int ocfs2_block_extent_contig(struct super_block *sb,
795 struct ocfs2_extent_rec *ext,
796 u64 blkno)
797{
798 u64 blk_end = le64_to_cpu(ext->e_blkno);
799
800 blk_end += ocfs2_clusters_to_blocks(sb,
801 le16_to_cpu(ext->e_leaf_clusters));
802
803 return blkno == blk_end;
804}
805
806static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
807 struct ocfs2_extent_rec *right)
808{
809 u32 left_range;
810
811 left_range = le32_to_cpu(left->e_cpos) +
812 le16_to_cpu(left->e_leaf_clusters);
813
814 return (left_range == le32_to_cpu(right->e_cpos));
815}
816
817static enum ocfs2_contig_type
818 ocfs2_extent_rec_contig(struct super_block *sb,
819 struct ocfs2_extent_rec *ext,
820 struct ocfs2_extent_rec *insert_rec)
821{
822 u64 blkno = le64_to_cpu(insert_rec->e_blkno);
823
824 /*
825 * Refuse to coalesce extent records with different flag
826 * fields - we don't want to mix unwritten extents with user
827 * data.
828 */
829 if (ext->e_flags != insert_rec->e_flags)
830 return CONTIG_NONE;
831
832 if (ocfs2_extents_adjacent(ext, insert_rec) &&
833 ocfs2_block_extent_contig(sb, ext, blkno))
834 return CONTIG_RIGHT;
835
836 blkno = le64_to_cpu(ext->e_blkno);
837 if (ocfs2_extents_adjacent(insert_rec, ext) &&
838 ocfs2_block_extent_contig(sb, insert_rec, blkno))
839 return CONTIG_LEFT;
840
841 return CONTIG_NONE;
842}
843
844/*
845 * NOTE: We can have pretty much any combination of contiguousness and
846 * appending.
847 *
848 * The usefulness of APPEND_TAIL is more in that it lets us know that
849 * we'll have to update the path to that leaf.
850 */
851enum ocfs2_append_type {
852 APPEND_NONE = 0,
853 APPEND_TAIL,
854};
855
856enum ocfs2_split_type {
857 SPLIT_NONE = 0,
858 SPLIT_LEFT,
859 SPLIT_RIGHT,
860};
861
862struct ocfs2_insert_type {
863 enum ocfs2_split_type ins_split;
864 enum ocfs2_append_type ins_appending;
865 enum ocfs2_contig_type ins_contig;
866 int ins_contig_index;
867 int ins_tree_depth;
868};
869
870struct ocfs2_merge_ctxt {
871 enum ocfs2_contig_type c_contig_type;
872 int c_has_empty_extent;
873 int c_split_covers_rec;
874};
875
876static int ocfs2_validate_extent_block(struct super_block *sb,
877 struct buffer_head *bh)
878{
879 int rc;
880 struct ocfs2_extent_block *eb =
881 (struct ocfs2_extent_block *)bh->b_data;
882
883 trace_ocfs2_validate_extent_block((unsigned long long)bh->b_blocknr);
884
885 BUG_ON(!buffer_uptodate(bh));
886
887 /*
888 * If the ecc fails, we return the error but otherwise
889 * leave the filesystem running. We know any error is
890 * local to this block.
891 */
892 rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &eb->h_check);
893 if (rc) {
894 mlog(ML_ERROR, "Checksum failed for extent block %llu\n",
895 (unsigned long long)bh->b_blocknr);
896 return rc;
897 }
898
899 /*
900 * Errors after here are fatal.
901 */
902
903 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
904 rc = ocfs2_error(sb,
905 "Extent block #%llu has bad signature %.*s\n",
906 (unsigned long long)bh->b_blocknr, 7,
907 eb->h_signature);
908 goto bail;
909 }
910
911 if (le64_to_cpu(eb->h_blkno) != bh->b_blocknr) {
912 rc = ocfs2_error(sb,
913 "Extent block #%llu has an invalid h_blkno of %llu\n",
914 (unsigned long long)bh->b_blocknr,
915 (unsigned long long)le64_to_cpu(eb->h_blkno));
916 goto bail;
917 }
918
919 if (le32_to_cpu(eb->h_fs_generation) != OCFS2_SB(sb)->fs_generation)
920 rc = ocfs2_error(sb,
921 "Extent block #%llu has an invalid h_fs_generation of #%u\n",
922 (unsigned long long)bh->b_blocknr,
923 le32_to_cpu(eb->h_fs_generation));
924bail:
925 return rc;
926}
927
928int ocfs2_read_extent_block(struct ocfs2_caching_info *ci, u64 eb_blkno,
929 struct buffer_head **bh)
930{
931 int rc;
932 struct buffer_head *tmp = *bh;
933
934 rc = ocfs2_read_block(ci, eb_blkno, &tmp,
935 ocfs2_validate_extent_block);
936
937 /* If ocfs2_read_block() got us a new bh, pass it up. */
938 if (!rc && !*bh)
939 *bh = tmp;
940
941 return rc;
942}
943
944
945/*
946 * How many free extents have we got before we need more meta data?
947 */
948int ocfs2_num_free_extents(struct ocfs2_extent_tree *et)
949{
950 int retval;
951 struct ocfs2_extent_list *el = NULL;
952 struct ocfs2_extent_block *eb;
953 struct buffer_head *eb_bh = NULL;
954 u64 last_eb_blk = 0;
955
956 el = et->et_root_el;
957 last_eb_blk = ocfs2_et_get_last_eb_blk(et);
958
959 if (last_eb_blk) {
960 retval = ocfs2_read_extent_block(et->et_ci, last_eb_blk,
961 &eb_bh);
962 if (retval < 0) {
963 mlog_errno(retval);
964 goto bail;
965 }
966 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
967 el = &eb->h_list;
968 }
969
970 BUG_ON(el->l_tree_depth != 0);
971
972 retval = le16_to_cpu(el->l_count) - le16_to_cpu(el->l_next_free_rec);
973bail:
974 brelse(eb_bh);
975
976 trace_ocfs2_num_free_extents(retval);
977 return retval;
978}
979
980/* expects array to already be allocated
981 *
982 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
983 * l_count for you
984 */
985static int ocfs2_create_new_meta_bhs(handle_t *handle,
986 struct ocfs2_extent_tree *et,
987 int wanted,
988 struct ocfs2_alloc_context *meta_ac,
989 struct buffer_head *bhs[])
990{
991 int count, status, i;
992 u16 suballoc_bit_start;
993 u32 num_got;
994 u64 suballoc_loc, first_blkno;
995 struct ocfs2_super *osb =
996 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
997 struct ocfs2_extent_block *eb;
998
999 count = 0;
1000 while (count < wanted) {
1001 status = ocfs2_claim_metadata(handle,
1002 meta_ac,
1003 wanted - count,
1004 &suballoc_loc,
1005 &suballoc_bit_start,
1006 &num_got,
1007 &first_blkno);
1008 if (status < 0) {
1009 mlog_errno(status);
1010 goto bail;
1011 }
1012
1013 for(i = count; i < (num_got + count); i++) {
1014 bhs[i] = sb_getblk(osb->sb, first_blkno);
1015 if (bhs[i] == NULL) {
1016 status = -ENOMEM;
1017 mlog_errno(status);
1018 goto bail;
1019 }
1020 ocfs2_set_new_buffer_uptodate(et->et_ci, bhs[i]);
1021
1022 status = ocfs2_journal_access_eb(handle, et->et_ci,
1023 bhs[i],
1024 OCFS2_JOURNAL_ACCESS_CREATE);
1025 if (status < 0) {
1026 mlog_errno(status);
1027 goto bail;
1028 }
1029
1030 memset(bhs[i]->b_data, 0, osb->sb->s_blocksize);
1031 eb = (struct ocfs2_extent_block *) bhs[i]->b_data;
1032 /* Ok, setup the minimal stuff here. */
1033 strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
1034 eb->h_blkno = cpu_to_le64(first_blkno);
1035 eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
1036 eb->h_suballoc_slot =
1037 cpu_to_le16(meta_ac->ac_alloc_slot);
1038 eb->h_suballoc_loc = cpu_to_le64(suballoc_loc);
1039 eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1040 eb->h_list.l_count =
1041 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
1042
1043 suballoc_bit_start++;
1044 first_blkno++;
1045
1046 /* We'll also be dirtied by the caller, so
1047 * this isn't absolutely necessary. */
1048 ocfs2_journal_dirty(handle, bhs[i]);
1049 }
1050
1051 count += num_got;
1052 }
1053
1054 status = 0;
1055bail:
1056 if (status < 0) {
1057 for(i = 0; i < wanted; i++) {
1058 brelse(bhs[i]);
1059 bhs[i] = NULL;
1060 }
1061 }
1062 return status;
1063}
1064
1065/*
1066 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
1067 *
1068 * Returns the sum of the rightmost extent rec logical offset and
1069 * cluster count.
1070 *
1071 * ocfs2_add_branch() uses this to determine what logical cluster
1072 * value should be populated into the leftmost new branch records.
1073 *
1074 * ocfs2_shift_tree_depth() uses this to determine the # clusters
1075 * value for the new topmost tree record.
1076 */
1077static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list *el)
1078{
1079 int i;
1080
1081 i = le16_to_cpu(el->l_next_free_rec) - 1;
1082
1083 return le32_to_cpu(el->l_recs[i].e_cpos) +
1084 ocfs2_rec_clusters(el, &el->l_recs[i]);
1085}
1086
1087/*
1088 * Change range of the branches in the right most path according to the leaf
1089 * extent block's rightmost record.
1090 */
1091static int ocfs2_adjust_rightmost_branch(handle_t *handle,
1092 struct ocfs2_extent_tree *et)
1093{
1094 int status;
1095 struct ocfs2_path *path = NULL;
1096 struct ocfs2_extent_list *el;
1097 struct ocfs2_extent_rec *rec;
1098
1099 path = ocfs2_new_path_from_et(et);
1100 if (!path) {
1101 status = -ENOMEM;
1102 return status;
1103 }
1104
1105 status = ocfs2_find_path(et->et_ci, path, UINT_MAX);
1106 if (status < 0) {
1107 mlog_errno(status);
1108 goto out;
1109 }
1110
1111 status = ocfs2_extend_trans(handle, path_num_items(path));
1112 if (status < 0) {
1113 mlog_errno(status);
1114 goto out;
1115 }
1116
1117 status = ocfs2_journal_access_path(et->et_ci, handle, path);
1118 if (status < 0) {
1119 mlog_errno(status);
1120 goto out;
1121 }
1122
1123 el = path_leaf_el(path);
1124 rec = &el->l_recs[le16_to_cpu(el->l_next_free_rec) - 1];
1125
1126 ocfs2_adjust_rightmost_records(handle, et, path, rec);
1127
1128out:
1129 ocfs2_free_path(path);
1130 return status;
1131}
1132
1133/*
1134 * Add an entire tree branch to our inode. eb_bh is the extent block
1135 * to start at, if we don't want to start the branch at the root
1136 * structure.
1137 *
1138 * last_eb_bh is required as we have to update it's next_leaf pointer
1139 * for the new last extent block.
1140 *
1141 * the new branch will be 'empty' in the sense that every block will
1142 * contain a single record with cluster count == 0.
1143 */
1144static int ocfs2_add_branch(handle_t *handle,
1145 struct ocfs2_extent_tree *et,
1146 struct buffer_head *eb_bh,
1147 struct buffer_head **last_eb_bh,
1148 struct ocfs2_alloc_context *meta_ac)
1149{
1150 int status, new_blocks, i, block_given = 0;
1151 u64 next_blkno, new_last_eb_blk;
1152 struct buffer_head *bh;
1153 struct buffer_head **new_eb_bhs = NULL;
1154 struct ocfs2_extent_block *eb;
1155 struct ocfs2_extent_list *eb_el;
1156 struct ocfs2_extent_list *el;
1157 u32 new_cpos, root_end;
1158
1159 BUG_ON(!last_eb_bh || !*last_eb_bh);
1160
1161 if (eb_bh) {
1162 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1163 el = &eb->h_list;
1164 } else
1165 el = et->et_root_el;
1166
1167 /* we never add a branch to a leaf. */
1168 BUG_ON(!el->l_tree_depth);
1169
1170 new_blocks = le16_to_cpu(el->l_tree_depth);
1171
1172 eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
1173 new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
1174 root_end = ocfs2_sum_rightmost_rec(et->et_root_el);
1175
1176 /*
1177 * If there is a gap before the root end and the real end
1178 * of the righmost leaf block, we need to remove the gap
1179 * between new_cpos and root_end first so that the tree
1180 * is consistent after we add a new branch(it will start
1181 * from new_cpos).
1182 */
1183 if (root_end > new_cpos) {
1184 trace_ocfs2_adjust_rightmost_branch(
1185 (unsigned long long)
1186 ocfs2_metadata_cache_owner(et->et_ci),
1187 root_end, new_cpos);
1188
1189 status = ocfs2_adjust_rightmost_branch(handle, et);
1190 if (status) {
1191 mlog_errno(status);
1192 goto bail;
1193 }
1194 }
1195
1196 /* allocate the number of new eb blocks we need */
1197 new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
1198 GFP_KERNEL);
1199 if (!new_eb_bhs) {
1200 status = -ENOMEM;
1201 mlog_errno(status);
1202 goto bail;
1203 }
1204
1205 /* Firstyly, try to reuse dealloc since we have already estimated how
1206 * many extent blocks we may use.
1207 */
1208 if (!ocfs2_is_dealloc_empty(et)) {
1209 status = ocfs2_reuse_blk_from_dealloc(handle, et,
1210 new_eb_bhs, new_blocks,
1211 &block_given);
1212 if (status < 0) {
1213 mlog_errno(status);
1214 goto bail;
1215 }
1216 }
1217
1218 BUG_ON(block_given > new_blocks);
1219
1220 if (block_given < new_blocks) {
1221 BUG_ON(!meta_ac);
1222 status = ocfs2_create_new_meta_bhs(handle, et,
1223 new_blocks - block_given,
1224 meta_ac,
1225 &new_eb_bhs[block_given]);
1226 if (status < 0) {
1227 mlog_errno(status);
1228 goto bail;
1229 }
1230 }
1231
1232 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
1233 * linked with the rest of the tree.
1234 * conversly, new_eb_bhs[0] is the new bottommost leaf.
1235 *
1236 * when we leave the loop, new_last_eb_blk will point to the
1237 * newest leaf, and next_blkno will point to the topmost extent
1238 * block. */
1239 next_blkno = new_last_eb_blk = 0;
1240 for(i = 0; i < new_blocks; i++) {
1241 bh = new_eb_bhs[i];
1242 eb = (struct ocfs2_extent_block *) bh->b_data;
1243 /* ocfs2_create_new_meta_bhs() should create it right! */
1244 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1245 eb_el = &eb->h_list;
1246
1247 status = ocfs2_journal_access_eb(handle, et->et_ci, bh,
1248 OCFS2_JOURNAL_ACCESS_CREATE);
1249 if (status < 0) {
1250 mlog_errno(status);
1251 goto bail;
1252 }
1253
1254 eb->h_next_leaf_blk = 0;
1255 eb_el->l_tree_depth = cpu_to_le16(i);
1256 eb_el->l_next_free_rec = cpu_to_le16(1);
1257 /*
1258 * This actually counts as an empty extent as
1259 * c_clusters == 0
1260 */
1261 eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
1262 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
1263 /*
1264 * eb_el isn't always an interior node, but even leaf
1265 * nodes want a zero'd flags and reserved field so
1266 * this gets the whole 32 bits regardless of use.
1267 */
1268 eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
1269 if (!eb_el->l_tree_depth)
1270 new_last_eb_blk = le64_to_cpu(eb->h_blkno);
1271
1272 ocfs2_journal_dirty(handle, bh);
1273 next_blkno = le64_to_cpu(eb->h_blkno);
1274 }
1275
1276 /* This is a bit hairy. We want to update up to three blocks
1277 * here without leaving any of them in an inconsistent state
1278 * in case of error. We don't have to worry about
1279 * journal_dirty erroring as it won't unless we've aborted the
1280 * handle (in which case we would never be here) so reserving
1281 * the write with journal_access is all we need to do. */
1282 status = ocfs2_journal_access_eb(handle, et->et_ci, *last_eb_bh,
1283 OCFS2_JOURNAL_ACCESS_WRITE);
1284 if (status < 0) {
1285 mlog_errno(status);
1286 goto bail;
1287 }
1288 status = ocfs2_et_root_journal_access(handle, et,
1289 OCFS2_JOURNAL_ACCESS_WRITE);
1290 if (status < 0) {
1291 mlog_errno(status);
1292 goto bail;
1293 }
1294 if (eb_bh) {
1295 status = ocfs2_journal_access_eb(handle, et->et_ci, eb_bh,
1296 OCFS2_JOURNAL_ACCESS_WRITE);
1297 if (status < 0) {
1298 mlog_errno(status);
1299 goto bail;
1300 }
1301 }
1302
1303 /* Link the new branch into the rest of the tree (el will
1304 * either be on the root_bh, or the extent block passed in. */
1305 i = le16_to_cpu(el->l_next_free_rec);
1306 el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
1307 el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1308 el->l_recs[i].e_int_clusters = 0;
1309 le16_add_cpu(&el->l_next_free_rec, 1);
1310
1311 /* fe needs a new last extent block pointer, as does the
1312 * next_leaf on the previously last-extent-block. */
1313 ocfs2_et_set_last_eb_blk(et, new_last_eb_blk);
1314
1315 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
1316 eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);
1317
1318 ocfs2_journal_dirty(handle, *last_eb_bh);
1319 ocfs2_journal_dirty(handle, et->et_root_bh);
1320 if (eb_bh)
1321 ocfs2_journal_dirty(handle, eb_bh);
1322
1323 /*
1324 * Some callers want to track the rightmost leaf so pass it
1325 * back here.
1326 */
1327 brelse(*last_eb_bh);
1328 get_bh(new_eb_bhs[0]);
1329 *last_eb_bh = new_eb_bhs[0];
1330
1331 status = 0;
1332bail:
1333 if (new_eb_bhs) {
1334 for (i = 0; i < new_blocks; i++)
1335 brelse(new_eb_bhs[i]);
1336 kfree(new_eb_bhs);
1337 }
1338
1339 return status;
1340}
1341
1342/*
1343 * adds another level to the allocation tree.
1344 * returns back the new extent block so you can add a branch to it
1345 * after this call.
1346 */
1347static int ocfs2_shift_tree_depth(handle_t *handle,
1348 struct ocfs2_extent_tree *et,
1349 struct ocfs2_alloc_context *meta_ac,
1350 struct buffer_head **ret_new_eb_bh)
1351{
1352 int status, i, block_given = 0;
1353 u32 new_clusters;
1354 struct buffer_head *new_eb_bh = NULL;
1355 struct ocfs2_extent_block *eb;
1356 struct ocfs2_extent_list *root_el;
1357 struct ocfs2_extent_list *eb_el;
1358
1359 if (!ocfs2_is_dealloc_empty(et)) {
1360 status = ocfs2_reuse_blk_from_dealloc(handle, et,
1361 &new_eb_bh, 1,
1362 &block_given);
1363 } else if (meta_ac) {
1364 status = ocfs2_create_new_meta_bhs(handle, et, 1, meta_ac,
1365 &new_eb_bh);
1366
1367 } else {
1368 BUG();
1369 }
1370
1371 if (status < 0) {
1372 mlog_errno(status);
1373 goto bail;
1374 }
1375
1376 eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
1377 /* ocfs2_create_new_meta_bhs() should create it right! */
1378 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1379
1380 eb_el = &eb->h_list;
1381 root_el = et->et_root_el;
1382
1383 status = ocfs2_journal_access_eb(handle, et->et_ci, new_eb_bh,
1384 OCFS2_JOURNAL_ACCESS_CREATE);
1385 if (status < 0) {
1386 mlog_errno(status);
1387 goto bail;
1388 }
1389
1390 /* copy the root extent list data into the new extent block */
1391 eb_el->l_tree_depth = root_el->l_tree_depth;
1392 eb_el->l_next_free_rec = root_el->l_next_free_rec;
1393 for (i = 0; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1394 eb_el->l_recs[i] = root_el->l_recs[i];
1395
1396 ocfs2_journal_dirty(handle, new_eb_bh);
1397
1398 status = ocfs2_et_root_journal_access(handle, et,
1399 OCFS2_JOURNAL_ACCESS_WRITE);
1400 if (status < 0) {
1401 mlog_errno(status);
1402 goto bail;
1403 }
1404
1405 new_clusters = ocfs2_sum_rightmost_rec(eb_el);
1406
1407 /* update root_bh now */
1408 le16_add_cpu(&root_el->l_tree_depth, 1);
1409 root_el->l_recs[0].e_cpos = 0;
1410 root_el->l_recs[0].e_blkno = eb->h_blkno;
1411 root_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
1412 for (i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1413 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
1414 root_el->l_next_free_rec = cpu_to_le16(1);
1415
1416 /* If this is our 1st tree depth shift, then last_eb_blk
1417 * becomes the allocated extent block */
1418 if (root_el->l_tree_depth == cpu_to_le16(1))
1419 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
1420
1421 ocfs2_journal_dirty(handle, et->et_root_bh);
1422
1423 *ret_new_eb_bh = new_eb_bh;
1424 new_eb_bh = NULL;
1425 status = 0;
1426bail:
1427 brelse(new_eb_bh);
1428
1429 return status;
1430}
1431
1432/*
1433 * Should only be called when there is no space left in any of the
1434 * leaf nodes. What we want to do is find the lowest tree depth
1435 * non-leaf extent block with room for new records. There are three
1436 * valid results of this search:
1437 *
1438 * 1) a lowest extent block is found, then we pass it back in
1439 * *lowest_eb_bh and return '0'
1440 *
1441 * 2) the search fails to find anything, but the root_el has room. We
1442 * pass NULL back in *lowest_eb_bh, but still return '0'
1443 *
1444 * 3) the search fails to find anything AND the root_el is full, in
1445 * which case we return > 0
1446 *
1447 * return status < 0 indicates an error.
1448 */
1449static int ocfs2_find_branch_target(struct ocfs2_extent_tree *et,
1450 struct buffer_head **target_bh)
1451{
1452 int status = 0, i;
1453 u64 blkno;
1454 struct ocfs2_extent_block *eb;
1455 struct ocfs2_extent_list *el;
1456 struct buffer_head *bh = NULL;
1457 struct buffer_head *lowest_bh = NULL;
1458
1459 *target_bh = NULL;
1460
1461 el = et->et_root_el;
1462
1463 while(le16_to_cpu(el->l_tree_depth) > 1) {
1464 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1465 status = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1466 "Owner %llu has empty extent list (next_free_rec == 0)\n",
1467 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
1468 goto bail;
1469 }
1470 i = le16_to_cpu(el->l_next_free_rec) - 1;
1471 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1472 if (!blkno) {
1473 status = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1474 "Owner %llu has extent list where extent # %d has no physical block start\n",
1475 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), i);
1476 goto bail;
1477 }
1478
1479 brelse(bh);
1480 bh = NULL;
1481
1482 status = ocfs2_read_extent_block(et->et_ci, blkno, &bh);
1483 if (status < 0) {
1484 mlog_errno(status);
1485 goto bail;
1486 }
1487
1488 eb = (struct ocfs2_extent_block *) bh->b_data;
1489 el = &eb->h_list;
1490
1491 if (le16_to_cpu(el->l_next_free_rec) <
1492 le16_to_cpu(el->l_count)) {
1493 brelse(lowest_bh);
1494 lowest_bh = bh;
1495 get_bh(lowest_bh);
1496 }
1497 }
1498
1499 /* If we didn't find one and the fe doesn't have any room,
1500 * then return '1' */
1501 el = et->et_root_el;
1502 if (!lowest_bh && (el->l_next_free_rec == el->l_count))
1503 status = 1;
1504
1505 *target_bh = lowest_bh;
1506bail:
1507 brelse(bh);
1508
1509 return status;
1510}
1511
1512/*
1513 * Grow a b-tree so that it has more records.
1514 *
1515 * We might shift the tree depth in which case existing paths should
1516 * be considered invalid.
1517 *
1518 * Tree depth after the grow is returned via *final_depth.
1519 *
1520 * *last_eb_bh will be updated by ocfs2_add_branch().
1521 */
1522static int ocfs2_grow_tree(handle_t *handle, struct ocfs2_extent_tree *et,
1523 int *final_depth, struct buffer_head **last_eb_bh,
1524 struct ocfs2_alloc_context *meta_ac)
1525{
1526 int ret, shift;
1527 struct ocfs2_extent_list *el = et->et_root_el;
1528 int depth = le16_to_cpu(el->l_tree_depth);
1529 struct buffer_head *bh = NULL;
1530
1531 BUG_ON(meta_ac == NULL && ocfs2_is_dealloc_empty(et));
1532
1533 shift = ocfs2_find_branch_target(et, &bh);
1534 if (shift < 0) {
1535 ret = shift;
1536 mlog_errno(ret);
1537 goto out;
1538 }
1539
1540 /* We traveled all the way to the bottom of the allocation tree
1541 * and didn't find room for any more extents - we need to add
1542 * another tree level */
1543 if (shift) {
1544 BUG_ON(bh);
1545 trace_ocfs2_grow_tree(
1546 (unsigned long long)
1547 ocfs2_metadata_cache_owner(et->et_ci),
1548 depth);
1549
1550 /* ocfs2_shift_tree_depth will return us a buffer with
1551 * the new extent block (so we can pass that to
1552 * ocfs2_add_branch). */
1553 ret = ocfs2_shift_tree_depth(handle, et, meta_ac, &bh);
1554 if (ret < 0) {
1555 mlog_errno(ret);
1556 goto out;
1557 }
1558 depth++;
1559 if (depth == 1) {
1560 /*
1561 * Special case: we have room now if we shifted from
1562 * tree_depth 0, so no more work needs to be done.
1563 *
1564 * We won't be calling add_branch, so pass
1565 * back *last_eb_bh as the new leaf. At depth
1566 * zero, it should always be null so there's
1567 * no reason to brelse.
1568 */
1569 BUG_ON(*last_eb_bh);
1570 get_bh(bh);
1571 *last_eb_bh = bh;
1572 goto out;
1573 }
1574 }
1575
1576 /* call ocfs2_add_branch to add the final part of the tree with
1577 * the new data. */
1578 ret = ocfs2_add_branch(handle, et, bh, last_eb_bh,
1579 meta_ac);
1580 if (ret < 0)
1581 mlog_errno(ret);
1582
1583out:
1584 if (final_depth)
1585 *final_depth = depth;
1586 brelse(bh);
1587 return ret;
1588}
1589
1590/*
1591 * This function will discard the rightmost extent record.
1592 */
1593static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
1594{
1595 int next_free = le16_to_cpu(el->l_next_free_rec);
1596 int count = le16_to_cpu(el->l_count);
1597 unsigned int num_bytes;
1598
1599 BUG_ON(!next_free);
1600 /* This will cause us to go off the end of our extent list. */
1601 BUG_ON(next_free >= count);
1602
1603 num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
1604
1605 memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
1606}
1607
1608static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
1609 struct ocfs2_extent_rec *insert_rec)
1610{
1611 int i, insert_index, next_free, has_empty, num_bytes;
1612 u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
1613 struct ocfs2_extent_rec *rec;
1614
1615 next_free = le16_to_cpu(el->l_next_free_rec);
1616 has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
1617
1618 BUG_ON(!next_free);
1619
1620 /* The tree code before us didn't allow enough room in the leaf. */
1621 BUG_ON(el->l_next_free_rec == el->l_count && !has_empty);
1622
1623 /*
1624 * The easiest way to approach this is to just remove the
1625 * empty extent and temporarily decrement next_free.
1626 */
1627 if (has_empty) {
1628 /*
1629 * If next_free was 1 (only an empty extent), this
1630 * loop won't execute, which is fine. We still want
1631 * the decrement above to happen.
1632 */
1633 for(i = 0; i < (next_free - 1); i++)
1634 el->l_recs[i] = el->l_recs[i+1];
1635
1636 next_free--;
1637 }
1638
1639 /*
1640 * Figure out what the new record index should be.
1641 */
1642 for(i = 0; i < next_free; i++) {
1643 rec = &el->l_recs[i];
1644
1645 if (insert_cpos < le32_to_cpu(rec->e_cpos))
1646 break;
1647 }
1648 insert_index = i;
1649
1650 trace_ocfs2_rotate_leaf(insert_cpos, insert_index,
1651 has_empty, next_free,
1652 le16_to_cpu(el->l_count));
1653
1654 BUG_ON(insert_index < 0);
1655 BUG_ON(insert_index >= le16_to_cpu(el->l_count));
1656 BUG_ON(insert_index > next_free);
1657
1658 /*
1659 * No need to memmove if we're just adding to the tail.
1660 */
1661 if (insert_index != next_free) {
1662 BUG_ON(next_free >= le16_to_cpu(el->l_count));
1663
1664 num_bytes = next_free - insert_index;
1665 num_bytes *= sizeof(struct ocfs2_extent_rec);
1666 memmove(&el->l_recs[insert_index + 1],
1667 &el->l_recs[insert_index],
1668 num_bytes);
1669 }
1670
1671 /*
1672 * Either we had an empty extent, and need to re-increment or
1673 * there was no empty extent on a non full rightmost leaf node,
1674 * in which case we still need to increment.
1675 */
1676 next_free++;
1677 el->l_next_free_rec = cpu_to_le16(next_free);
1678 /*
1679 * Make sure none of the math above just messed up our tree.
1680 */
1681 BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
1682
1683 el->l_recs[insert_index] = *insert_rec;
1684
1685}
1686
1687static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
1688{
1689 int size, num_recs = le16_to_cpu(el->l_next_free_rec);
1690
1691 BUG_ON(num_recs == 0);
1692
1693 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
1694 num_recs--;
1695 size = num_recs * sizeof(struct ocfs2_extent_rec);
1696 memmove(&el->l_recs[0], &el->l_recs[1], size);
1697 memset(&el->l_recs[num_recs], 0,
1698 sizeof(struct ocfs2_extent_rec));
1699 el->l_next_free_rec = cpu_to_le16(num_recs);
1700 }
1701}
1702
1703/*
1704 * Create an empty extent record .
1705 *
1706 * l_next_free_rec may be updated.
1707 *
1708 * If an empty extent already exists do nothing.
1709 */
1710static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
1711{
1712 int next_free = le16_to_cpu(el->l_next_free_rec);
1713
1714 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1715
1716 if (next_free == 0)
1717 goto set_and_inc;
1718
1719 if (ocfs2_is_empty_extent(&el->l_recs[0]))
1720 return;
1721
1722 mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
1723 "Asked to create an empty extent in a full list:\n"
1724 "count = %u, tree depth = %u",
1725 le16_to_cpu(el->l_count),
1726 le16_to_cpu(el->l_tree_depth));
1727
1728 ocfs2_shift_records_right(el);
1729
1730set_and_inc:
1731 le16_add_cpu(&el->l_next_free_rec, 1);
1732 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1733}
1734
1735/*
1736 * For a rotation which involves two leaf nodes, the "root node" is
1737 * the lowest level tree node which contains a path to both leafs. This
1738 * resulting set of information can be used to form a complete "subtree"
1739 *
1740 * This function is passed two full paths from the dinode down to a
1741 * pair of adjacent leaves. It's task is to figure out which path
1742 * index contains the subtree root - this can be the root index itself
1743 * in a worst-case rotation.
1744 *
1745 * The array index of the subtree root is passed back.
1746 */
1747int ocfs2_find_subtree_root(struct ocfs2_extent_tree *et,
1748 struct ocfs2_path *left,
1749 struct ocfs2_path *right)
1750{
1751 int i = 0;
1752
1753 /*
1754 * Check that the caller passed in two paths from the same tree.
1755 */
1756 BUG_ON(path_root_bh(left) != path_root_bh(right));
1757
1758 do {
1759 i++;
1760
1761 /*
1762 * The caller didn't pass two adjacent paths.
1763 */
1764 mlog_bug_on_msg(i > left->p_tree_depth,
1765 "Owner %llu, left depth %u, right depth %u\n"
1766 "left leaf blk %llu, right leaf blk %llu\n",
1767 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
1768 left->p_tree_depth, right->p_tree_depth,
1769 (unsigned long long)path_leaf_bh(left)->b_blocknr,
1770 (unsigned long long)path_leaf_bh(right)->b_blocknr);
1771 } while (left->p_node[i].bh->b_blocknr ==
1772 right->p_node[i].bh->b_blocknr);
1773
1774 return i - 1;
1775}
1776
1777typedef void (path_insert_t)(void *, struct buffer_head *);
1778
1779/*
1780 * Traverse a btree path in search of cpos, starting at root_el.
1781 *
1782 * This code can be called with a cpos larger than the tree, in which
1783 * case it will return the rightmost path.
1784 */
1785static int __ocfs2_find_path(struct ocfs2_caching_info *ci,
1786 struct ocfs2_extent_list *root_el, u32 cpos,
1787 path_insert_t *func, void *data)
1788{
1789 int i, ret = 0;
1790 u32 range;
1791 u64 blkno;
1792 struct buffer_head *bh = NULL;
1793 struct ocfs2_extent_block *eb;
1794 struct ocfs2_extent_list *el;
1795 struct ocfs2_extent_rec *rec;
1796
1797 el = root_el;
1798 while (el->l_tree_depth) {
1799 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1800 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1801 "Owner %llu has empty extent list at depth %u\n",
1802 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1803 le16_to_cpu(el->l_tree_depth));
1804 ret = -EROFS;
1805 goto out;
1806
1807 }
1808
1809 for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
1810 rec = &el->l_recs[i];
1811
1812 /*
1813 * In the case that cpos is off the allocation
1814 * tree, this should just wind up returning the
1815 * rightmost record.
1816 */
1817 range = le32_to_cpu(rec->e_cpos) +
1818 ocfs2_rec_clusters(el, rec);
1819 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1820 break;
1821 }
1822
1823 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1824 if (blkno == 0) {
1825 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1826 "Owner %llu has bad blkno in extent list at depth %u (index %d)\n",
1827 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1828 le16_to_cpu(el->l_tree_depth), i);
1829 ret = -EROFS;
1830 goto out;
1831 }
1832
1833 brelse(bh);
1834 bh = NULL;
1835 ret = ocfs2_read_extent_block(ci, blkno, &bh);
1836 if (ret) {
1837 mlog_errno(ret);
1838 goto out;
1839 }
1840
1841 eb = (struct ocfs2_extent_block *) bh->b_data;
1842 el = &eb->h_list;
1843
1844 if (le16_to_cpu(el->l_next_free_rec) >
1845 le16_to_cpu(el->l_count)) {
1846 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1847 "Owner %llu has bad count in extent list at block %llu (next free=%u, count=%u)\n",
1848 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1849 (unsigned long long)bh->b_blocknr,
1850 le16_to_cpu(el->l_next_free_rec),
1851 le16_to_cpu(el->l_count));
1852 ret = -EROFS;
1853 goto out;
1854 }
1855
1856 if (func)
1857 func(data, bh);
1858 }
1859
1860out:
1861 /*
1862 * Catch any trailing bh that the loop didn't handle.
1863 */
1864 brelse(bh);
1865
1866 return ret;
1867}
1868
1869/*
1870 * Given an initialized path (that is, it has a valid root extent
1871 * list), this function will traverse the btree in search of the path
1872 * which would contain cpos.
1873 *
1874 * The path traveled is recorded in the path structure.
1875 *
1876 * Note that this will not do any comparisons on leaf node extent
1877 * records, so it will work fine in the case that we just added a tree
1878 * branch.
1879 */
1880struct find_path_data {
1881 int index;
1882 struct ocfs2_path *path;
1883};
1884static void find_path_ins(void *data, struct buffer_head *bh)
1885{
1886 struct find_path_data *fp = data;
1887
1888 get_bh(bh);
1889 ocfs2_path_insert_eb(fp->path, fp->index, bh);
1890 fp->index++;
1891}
1892int ocfs2_find_path(struct ocfs2_caching_info *ci,
1893 struct ocfs2_path *path, u32 cpos)
1894{
1895 struct find_path_data data;
1896
1897 data.index = 1;
1898 data.path = path;
1899 return __ocfs2_find_path(ci, path_root_el(path), cpos,
1900 find_path_ins, &data);
1901}
1902
1903static void find_leaf_ins(void *data, struct buffer_head *bh)
1904{
1905 struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1906 struct ocfs2_extent_list *el = &eb->h_list;
1907 struct buffer_head **ret = data;
1908
1909 /* We want to retain only the leaf block. */
1910 if (le16_to_cpu(el->l_tree_depth) == 0) {
1911 get_bh(bh);
1912 *ret = bh;
1913 }
1914}
1915/*
1916 * Find the leaf block in the tree which would contain cpos. No
1917 * checking of the actual leaf is done.
1918 *
1919 * Some paths want to call this instead of allocating a path structure
1920 * and calling ocfs2_find_path().
1921 *
1922 * This function doesn't handle non btree extent lists.
1923 */
1924int ocfs2_find_leaf(struct ocfs2_caching_info *ci,
1925 struct ocfs2_extent_list *root_el, u32 cpos,
1926 struct buffer_head **leaf_bh)
1927{
1928 int ret;
1929 struct buffer_head *bh = NULL;
1930
1931 ret = __ocfs2_find_path(ci, root_el, cpos, find_leaf_ins, &bh);
1932 if (ret) {
1933 mlog_errno(ret);
1934 goto out;
1935 }
1936
1937 *leaf_bh = bh;
1938out:
1939 return ret;
1940}
1941
1942/*
1943 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1944 *
1945 * Basically, we've moved stuff around at the bottom of the tree and
1946 * we need to fix up the extent records above the changes to reflect
1947 * the new changes.
1948 *
1949 * left_rec: the record on the left.
1950 * right_rec: the record to the right of left_rec
1951 * right_child_el: is the child list pointed to by right_rec
1952 *
1953 * By definition, this only works on interior nodes.
1954 */
1955static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1956 struct ocfs2_extent_rec *right_rec,
1957 struct ocfs2_extent_list *right_child_el)
1958{
1959 u32 left_clusters, right_end;
1960
1961 /*
1962 * Interior nodes never have holes. Their cpos is the cpos of
1963 * the leftmost record in their child list. Their cluster
1964 * count covers the full theoretical range of their child list
1965 * - the range between their cpos and the cpos of the record
1966 * immediately to their right.
1967 */
1968 left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1969 if (!ocfs2_rec_clusters(right_child_el, &right_child_el->l_recs[0])) {
1970 BUG_ON(right_child_el->l_tree_depth);
1971 BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
1972 left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
1973 }
1974 left_clusters -= le32_to_cpu(left_rec->e_cpos);
1975 left_rec->e_int_clusters = cpu_to_le32(left_clusters);
1976
1977 /*
1978 * Calculate the rightmost cluster count boundary before
1979 * moving cpos - we will need to adjust clusters after
1980 * updating e_cpos to keep the same highest cluster count.
1981 */
1982 right_end = le32_to_cpu(right_rec->e_cpos);
1983 right_end += le32_to_cpu(right_rec->e_int_clusters);
1984
1985 right_rec->e_cpos = left_rec->e_cpos;
1986 le32_add_cpu(&right_rec->e_cpos, left_clusters);
1987
1988 right_end -= le32_to_cpu(right_rec->e_cpos);
1989 right_rec->e_int_clusters = cpu_to_le32(right_end);
1990}
1991
1992/*
1993 * Adjust the adjacent root node records involved in a
1994 * rotation. left_el_blkno is passed in as a key so that we can easily
1995 * find it's index in the root list.
1996 */
1997static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
1998 struct ocfs2_extent_list *left_el,
1999 struct ocfs2_extent_list *right_el,
2000 u64 left_el_blkno)
2001{
2002 int i;
2003
2004 BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
2005 le16_to_cpu(left_el->l_tree_depth));
2006
2007 for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
2008 if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
2009 break;
2010 }
2011
2012 /*
2013 * The path walking code should have never returned a root and
2014 * two paths which are not adjacent.
2015 */
2016 BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
2017
2018 ocfs2_adjust_adjacent_records(&root_el->l_recs[i],
2019 &root_el->l_recs[i + 1], right_el);
2020}
2021
2022/*
2023 * We've changed a leaf block (in right_path) and need to reflect that
2024 * change back up the subtree.
2025 *
2026 * This happens in multiple places:
2027 * - When we've moved an extent record from the left path leaf to the right
2028 * path leaf to make room for an empty extent in the left path leaf.
2029 * - When our insert into the right path leaf is at the leftmost edge
2030 * and requires an update of the path immediately to it's left. This
2031 * can occur at the end of some types of rotation and appending inserts.
2032 * - When we've adjusted the last extent record in the left path leaf and the
2033 * 1st extent record in the right path leaf during cross extent block merge.
2034 */
2035static void ocfs2_complete_edge_insert(handle_t *handle,
2036 struct ocfs2_path *left_path,
2037 struct ocfs2_path *right_path,
2038 int subtree_index)
2039{
2040 int i, idx;
2041 struct ocfs2_extent_list *el, *left_el, *right_el;
2042 struct ocfs2_extent_rec *left_rec, *right_rec;
2043 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2044
2045 /*
2046 * Update the counts and position values within all the
2047 * interior nodes to reflect the leaf rotation we just did.
2048 *
2049 * The root node is handled below the loop.
2050 *
2051 * We begin the loop with right_el and left_el pointing to the
2052 * leaf lists and work our way up.
2053 *
2054 * NOTE: within this loop, left_el and right_el always refer
2055 * to the *child* lists.
2056 */
2057 left_el = path_leaf_el(left_path);
2058 right_el = path_leaf_el(right_path);
2059 for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
2060 trace_ocfs2_complete_edge_insert(i);
2061
2062 /*
2063 * One nice property of knowing that all of these
2064 * nodes are below the root is that we only deal with
2065 * the leftmost right node record and the rightmost
2066 * left node record.
2067 */
2068 el = left_path->p_node[i].el;
2069 idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
2070 left_rec = &el->l_recs[idx];
2071
2072 el = right_path->p_node[i].el;
2073 right_rec = &el->l_recs[0];
2074
2075 ocfs2_adjust_adjacent_records(left_rec, right_rec, right_el);
2076
2077 ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
2078 ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
2079
2080 /*
2081 * Setup our list pointers now so that the current
2082 * parents become children in the next iteration.
2083 */
2084 left_el = left_path->p_node[i].el;
2085 right_el = right_path->p_node[i].el;
2086 }
2087
2088 /*
2089 * At the root node, adjust the two adjacent records which
2090 * begin our path to the leaves.
2091 */
2092
2093 el = left_path->p_node[subtree_index].el;
2094 left_el = left_path->p_node[subtree_index + 1].el;
2095 right_el = right_path->p_node[subtree_index + 1].el;
2096
2097 ocfs2_adjust_root_records(el, left_el, right_el,
2098 left_path->p_node[subtree_index + 1].bh->b_blocknr);
2099
2100 root_bh = left_path->p_node[subtree_index].bh;
2101
2102 ocfs2_journal_dirty(handle, root_bh);
2103}
2104
2105static int ocfs2_rotate_subtree_right(handle_t *handle,
2106 struct ocfs2_extent_tree *et,
2107 struct ocfs2_path *left_path,
2108 struct ocfs2_path *right_path,
2109 int subtree_index)
2110{
2111 int ret, i;
2112 struct buffer_head *right_leaf_bh;
2113 struct buffer_head *left_leaf_bh = NULL;
2114 struct buffer_head *root_bh;
2115 struct ocfs2_extent_list *right_el, *left_el;
2116 struct ocfs2_extent_rec move_rec;
2117
2118 left_leaf_bh = path_leaf_bh(left_path);
2119 left_el = path_leaf_el(left_path);
2120
2121 if (left_el->l_next_free_rec != left_el->l_count) {
2122 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
2123 "Inode %llu has non-full interior leaf node %llu (next free = %u)\n",
2124 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2125 (unsigned long long)left_leaf_bh->b_blocknr,
2126 le16_to_cpu(left_el->l_next_free_rec));
2127 return -EROFS;
2128 }
2129
2130 /*
2131 * This extent block may already have an empty record, so we
2132 * return early if so.
2133 */
2134 if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
2135 return 0;
2136
2137 root_bh = left_path->p_node[subtree_index].bh;
2138 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2139
2140 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2141 subtree_index);
2142 if (ret) {
2143 mlog_errno(ret);
2144 goto out;
2145 }
2146
2147 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2148 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2149 right_path, i);
2150 if (ret) {
2151 mlog_errno(ret);
2152 goto out;
2153 }
2154
2155 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2156 left_path, i);
2157 if (ret) {
2158 mlog_errno(ret);
2159 goto out;
2160 }
2161 }
2162
2163 right_leaf_bh = path_leaf_bh(right_path);
2164 right_el = path_leaf_el(right_path);
2165
2166 /* This is a code error, not a disk corruption. */
2167 mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
2168 "because rightmost leaf block %llu is empty\n",
2169 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2170 (unsigned long long)right_leaf_bh->b_blocknr);
2171
2172 ocfs2_create_empty_extent(right_el);
2173
2174 ocfs2_journal_dirty(handle, right_leaf_bh);
2175
2176 /* Do the copy now. */
2177 i = le16_to_cpu(left_el->l_next_free_rec) - 1;
2178 move_rec = left_el->l_recs[i];
2179 right_el->l_recs[0] = move_rec;
2180
2181 /*
2182 * Clear out the record we just copied and shift everything
2183 * over, leaving an empty extent in the left leaf.
2184 *
2185 * We temporarily subtract from next_free_rec so that the
2186 * shift will lose the tail record (which is now defunct).
2187 */
2188 le16_add_cpu(&left_el->l_next_free_rec, -1);
2189 ocfs2_shift_records_right(left_el);
2190 memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2191 le16_add_cpu(&left_el->l_next_free_rec, 1);
2192
2193 ocfs2_journal_dirty(handle, left_leaf_bh);
2194
2195 ocfs2_complete_edge_insert(handle, left_path, right_path,
2196 subtree_index);
2197
2198out:
2199 return ret;
2200}
2201
2202/*
2203 * Given a full path, determine what cpos value would return us a path
2204 * containing the leaf immediately to the left of the current one.
2205 *
2206 * Will return zero if the path passed in is already the leftmost path.
2207 */
2208int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
2209 struct ocfs2_path *path, u32 *cpos)
2210{
2211 int i, j, ret = 0;
2212 u64 blkno;
2213 struct ocfs2_extent_list *el;
2214
2215 BUG_ON(path->p_tree_depth == 0);
2216
2217 *cpos = 0;
2218
2219 blkno = path_leaf_bh(path)->b_blocknr;
2220
2221 /* Start at the tree node just above the leaf and work our way up. */
2222 i = path->p_tree_depth - 1;
2223 while (i >= 0) {
2224 el = path->p_node[i].el;
2225
2226 /*
2227 * Find the extent record just before the one in our
2228 * path.
2229 */
2230 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2231 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2232 if (j == 0) {
2233 if (i == 0) {
2234 /*
2235 * We've determined that the
2236 * path specified is already
2237 * the leftmost one - return a
2238 * cpos of zero.
2239 */
2240 goto out;
2241 }
2242 /*
2243 * The leftmost record points to our
2244 * leaf - we need to travel up the
2245 * tree one level.
2246 */
2247 goto next_node;
2248 }
2249
2250 *cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
2251 *cpos = *cpos + ocfs2_rec_clusters(el,
2252 &el->l_recs[j - 1]);
2253 *cpos = *cpos - 1;
2254 goto out;
2255 }
2256 }
2257
2258 /*
2259 * If we got here, we never found a valid node where
2260 * the tree indicated one should be.
2261 */
2262 ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
2263 (unsigned long long)blkno);
2264 ret = -EROFS;
2265 goto out;
2266
2267next_node:
2268 blkno = path->p_node[i].bh->b_blocknr;
2269 i--;
2270 }
2271
2272out:
2273 return ret;
2274}
2275
2276/*
2277 * Extend the transaction by enough credits to complete the rotation,
2278 * and still leave at least the original number of credits allocated
2279 * to this transaction.
2280 */
2281static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
2282 int op_credits,
2283 struct ocfs2_path *path)
2284{
2285 int ret = 0;
2286 int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;
2287
2288 if (jbd2_handle_buffer_credits(handle) < credits)
2289 ret = ocfs2_extend_trans(handle,
2290 credits - jbd2_handle_buffer_credits(handle));
2291
2292 return ret;
2293}
2294
2295/*
2296 * Trap the case where we're inserting into the theoretical range past
2297 * the _actual_ left leaf range. Otherwise, we'll rotate a record
2298 * whose cpos is less than ours into the right leaf.
2299 *
2300 * It's only necessary to look at the rightmost record of the left
2301 * leaf because the logic that calls us should ensure that the
2302 * theoretical ranges in the path components above the leaves are
2303 * correct.
2304 */
2305static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
2306 u32 insert_cpos)
2307{
2308 struct ocfs2_extent_list *left_el;
2309 struct ocfs2_extent_rec *rec;
2310 int next_free;
2311
2312 left_el = path_leaf_el(left_path);
2313 next_free = le16_to_cpu(left_el->l_next_free_rec);
2314 rec = &left_el->l_recs[next_free - 1];
2315
2316 if (insert_cpos > le32_to_cpu(rec->e_cpos))
2317 return 1;
2318 return 0;
2319}
2320
2321static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
2322{
2323 int next_free = le16_to_cpu(el->l_next_free_rec);
2324 unsigned int range;
2325 struct ocfs2_extent_rec *rec;
2326
2327 if (next_free == 0)
2328 return 0;
2329
2330 rec = &el->l_recs[0];
2331 if (ocfs2_is_empty_extent(rec)) {
2332 /* Empty list. */
2333 if (next_free == 1)
2334 return 0;
2335 rec = &el->l_recs[1];
2336 }
2337
2338 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2339 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
2340 return 1;
2341 return 0;
2342}
2343
2344/*
2345 * Rotate all the records in a btree right one record, starting at insert_cpos.
2346 *
2347 * The path to the rightmost leaf should be passed in.
2348 *
2349 * The array is assumed to be large enough to hold an entire path (tree depth).
2350 *
2351 * Upon successful return from this function:
2352 *
2353 * - The 'right_path' array will contain a path to the leaf block
2354 * whose range contains e_cpos.
2355 * - That leaf block will have a single empty extent in list index 0.
2356 * - In the case that the rotation requires a post-insert update,
2357 * *ret_left_path will contain a valid path which can be passed to
2358 * ocfs2_insert_path().
2359 */
2360static int ocfs2_rotate_tree_right(handle_t *handle,
2361 struct ocfs2_extent_tree *et,
2362 enum ocfs2_split_type split,
2363 u32 insert_cpos,
2364 struct ocfs2_path *right_path,
2365 struct ocfs2_path **ret_left_path)
2366{
2367 int ret, start, orig_credits = jbd2_handle_buffer_credits(handle);
2368 u32 cpos;
2369 struct ocfs2_path *left_path = NULL;
2370 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2371
2372 *ret_left_path = NULL;
2373
2374 left_path = ocfs2_new_path_from_path(right_path);
2375 if (!left_path) {
2376 ret = -ENOMEM;
2377 mlog_errno(ret);
2378 goto out;
2379 }
2380
2381 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2382 if (ret) {
2383 mlog_errno(ret);
2384 goto out;
2385 }
2386
2387 trace_ocfs2_rotate_tree_right(
2388 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2389 insert_cpos, cpos);
2390
2391 /*
2392 * What we want to do here is:
2393 *
2394 * 1) Start with the rightmost path.
2395 *
2396 * 2) Determine a path to the leaf block directly to the left
2397 * of that leaf.
2398 *
2399 * 3) Determine the 'subtree root' - the lowest level tree node
2400 * which contains a path to both leaves.
2401 *
2402 * 4) Rotate the subtree.
2403 *
2404 * 5) Find the next subtree by considering the left path to be
2405 * the new right path.
2406 *
2407 * The check at the top of this while loop also accepts
2408 * insert_cpos == cpos because cpos is only a _theoretical_
2409 * value to get us the left path - insert_cpos might very well
2410 * be filling that hole.
2411 *
2412 * Stop at a cpos of '0' because we either started at the
2413 * leftmost branch (i.e., a tree with one branch and a
2414 * rotation inside of it), or we've gone as far as we can in
2415 * rotating subtrees.
2416 */
2417 while (cpos && insert_cpos <= cpos) {
2418 trace_ocfs2_rotate_tree_right(
2419 (unsigned long long)
2420 ocfs2_metadata_cache_owner(et->et_ci),
2421 insert_cpos, cpos);
2422
2423 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
2424 if (ret) {
2425 mlog_errno(ret);
2426 goto out;
2427 }
2428
2429 mlog_bug_on_msg(path_leaf_bh(left_path) ==
2430 path_leaf_bh(right_path),
2431 "Owner %llu: error during insert of %u "
2432 "(left path cpos %u) results in two identical "
2433 "paths ending at %llu\n",
2434 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2435 insert_cpos, cpos,
2436 (unsigned long long)
2437 path_leaf_bh(left_path)->b_blocknr);
2438
2439 if (split == SPLIT_NONE &&
2440 ocfs2_rotate_requires_path_adjustment(left_path,
2441 insert_cpos)) {
2442
2443 /*
2444 * We've rotated the tree as much as we
2445 * should. The rest is up to
2446 * ocfs2_insert_path() to complete, after the
2447 * record insertion. We indicate this
2448 * situation by returning the left path.
2449 *
2450 * The reason we don't adjust the records here
2451 * before the record insert is that an error
2452 * later might break the rule where a parent
2453 * record e_cpos will reflect the actual
2454 * e_cpos of the 1st nonempty record of the
2455 * child list.
2456 */
2457 *ret_left_path = left_path;
2458 goto out_ret_path;
2459 }
2460
2461 start = ocfs2_find_subtree_root(et, left_path, right_path);
2462
2463 trace_ocfs2_rotate_subtree(start,
2464 (unsigned long long)
2465 right_path->p_node[start].bh->b_blocknr,
2466 right_path->p_tree_depth);
2467
2468 ret = ocfs2_extend_rotate_transaction(handle, start,
2469 orig_credits, right_path);
2470 if (ret) {
2471 mlog_errno(ret);
2472 goto out;
2473 }
2474
2475 ret = ocfs2_rotate_subtree_right(handle, et, left_path,
2476 right_path, start);
2477 if (ret) {
2478 mlog_errno(ret);
2479 goto out;
2480 }
2481
2482 if (split != SPLIT_NONE &&
2483 ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
2484 insert_cpos)) {
2485 /*
2486 * A rotate moves the rightmost left leaf
2487 * record over to the leftmost right leaf
2488 * slot. If we're doing an extent split
2489 * instead of a real insert, then we have to
2490 * check that the extent to be split wasn't
2491 * just moved over. If it was, then we can
2492 * exit here, passing left_path back -
2493 * ocfs2_split_extent() is smart enough to
2494 * search both leaves.
2495 */
2496 *ret_left_path = left_path;
2497 goto out_ret_path;
2498 }
2499
2500 /*
2501 * There is no need to re-read the next right path
2502 * as we know that it'll be our current left
2503 * path. Optimize by copying values instead.
2504 */
2505 ocfs2_mv_path(right_path, left_path);
2506
2507 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2508 if (ret) {
2509 mlog_errno(ret);
2510 goto out;
2511 }
2512 }
2513
2514out:
2515 ocfs2_free_path(left_path);
2516
2517out_ret_path:
2518 return ret;
2519}
2520
2521static int ocfs2_update_edge_lengths(handle_t *handle,
2522 struct ocfs2_extent_tree *et,
2523 struct ocfs2_path *path)
2524{
2525 int i, idx, ret;
2526 struct ocfs2_extent_rec *rec;
2527 struct ocfs2_extent_list *el;
2528 struct ocfs2_extent_block *eb;
2529 u32 range;
2530
2531 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
2532 if (ret) {
2533 mlog_errno(ret);
2534 goto out;
2535 }
2536
2537 /* Path should always be rightmost. */
2538 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2539 BUG_ON(eb->h_next_leaf_blk != 0ULL);
2540
2541 el = &eb->h_list;
2542 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
2543 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2544 rec = &el->l_recs[idx];
2545 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2546
2547 for (i = 0; i < path->p_tree_depth; i++) {
2548 el = path->p_node[i].el;
2549 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2550 rec = &el->l_recs[idx];
2551
2552 rec->e_int_clusters = cpu_to_le32(range);
2553 le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));
2554
2555 ocfs2_journal_dirty(handle, path->p_node[i].bh);
2556 }
2557out:
2558 return ret;
2559}
2560
2561static void ocfs2_unlink_path(handle_t *handle,
2562 struct ocfs2_extent_tree *et,
2563 struct ocfs2_cached_dealloc_ctxt *dealloc,
2564 struct ocfs2_path *path, int unlink_start)
2565{
2566 int ret, i;
2567 struct ocfs2_extent_block *eb;
2568 struct ocfs2_extent_list *el;
2569 struct buffer_head *bh;
2570
2571 for(i = unlink_start; i < path_num_items(path); i++) {
2572 bh = path->p_node[i].bh;
2573
2574 eb = (struct ocfs2_extent_block *)bh->b_data;
2575 /*
2576 * Not all nodes might have had their final count
2577 * decremented by the caller - handle this here.
2578 */
2579 el = &eb->h_list;
2580 if (le16_to_cpu(el->l_next_free_rec) > 1) {
2581 mlog(ML_ERROR,
2582 "Inode %llu, attempted to remove extent block "
2583 "%llu with %u records\n",
2584 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2585 (unsigned long long)le64_to_cpu(eb->h_blkno),
2586 le16_to_cpu(el->l_next_free_rec));
2587
2588 ocfs2_journal_dirty(handle, bh);
2589 ocfs2_remove_from_cache(et->et_ci, bh);
2590 continue;
2591 }
2592
2593 el->l_next_free_rec = 0;
2594 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2595
2596 ocfs2_journal_dirty(handle, bh);
2597
2598 ret = ocfs2_cache_extent_block_free(dealloc, eb);
2599 if (ret)
2600 mlog_errno(ret);
2601
2602 ocfs2_remove_from_cache(et->et_ci, bh);
2603 }
2604}
2605
2606static void ocfs2_unlink_subtree(handle_t *handle,
2607 struct ocfs2_extent_tree *et,
2608 struct ocfs2_path *left_path,
2609 struct ocfs2_path *right_path,
2610 int subtree_index,
2611 struct ocfs2_cached_dealloc_ctxt *dealloc)
2612{
2613 int i;
2614 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2615 struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2616 struct ocfs2_extent_block *eb;
2617
2618 eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2619
2620 for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2621 if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2622 break;
2623
2624 BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2625
2626 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2627 le16_add_cpu(&root_el->l_next_free_rec, -1);
2628
2629 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2630 eb->h_next_leaf_blk = 0;
2631
2632 ocfs2_journal_dirty(handle, root_bh);
2633 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2634
2635 ocfs2_unlink_path(handle, et, dealloc, right_path,
2636 subtree_index + 1);
2637}
2638
2639static int ocfs2_rotate_subtree_left(handle_t *handle,
2640 struct ocfs2_extent_tree *et,
2641 struct ocfs2_path *left_path,
2642 struct ocfs2_path *right_path,
2643 int subtree_index,
2644 struct ocfs2_cached_dealloc_ctxt *dealloc,
2645 int *deleted)
2646{
2647 int ret, i, del_right_subtree = 0, right_has_empty = 0;
2648 struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path);
2649 struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2650 struct ocfs2_extent_block *eb;
2651
2652 *deleted = 0;
2653
2654 right_leaf_el = path_leaf_el(right_path);
2655 left_leaf_el = path_leaf_el(left_path);
2656 root_bh = left_path->p_node[subtree_index].bh;
2657 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2658
2659 if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2660 return 0;
2661
2662 eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2663 if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2664 /*
2665 * It's legal for us to proceed if the right leaf is
2666 * the rightmost one and it has an empty extent. There
2667 * are two cases to handle - whether the leaf will be
2668 * empty after removal or not. If the leaf isn't empty
2669 * then just remove the empty extent up front. The
2670 * next block will handle empty leaves by flagging
2671 * them for unlink.
2672 *
2673 * Non rightmost leaves will throw -EAGAIN and the
2674 * caller can manually move the subtree and retry.
2675 */
2676
2677 if (eb->h_next_leaf_blk != 0ULL)
2678 return -EAGAIN;
2679
2680 if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2681 ret = ocfs2_journal_access_eb(handle, et->et_ci,
2682 path_leaf_bh(right_path),
2683 OCFS2_JOURNAL_ACCESS_WRITE);
2684 if (ret) {
2685 mlog_errno(ret);
2686 goto out;
2687 }
2688
2689 ocfs2_remove_empty_extent(right_leaf_el);
2690 } else
2691 right_has_empty = 1;
2692 }
2693
2694 if (eb->h_next_leaf_blk == 0ULL &&
2695 le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2696 /*
2697 * We have to update i_last_eb_blk during the meta
2698 * data delete.
2699 */
2700 ret = ocfs2_et_root_journal_access(handle, et,
2701 OCFS2_JOURNAL_ACCESS_WRITE);
2702 if (ret) {
2703 mlog_errno(ret);
2704 goto out;
2705 }
2706
2707 del_right_subtree = 1;
2708 }
2709
2710 /*
2711 * Getting here with an empty extent in the right path implies
2712 * that it's the rightmost path and will be deleted.
2713 */
2714 BUG_ON(right_has_empty && !del_right_subtree);
2715
2716 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2717 subtree_index);
2718 if (ret) {
2719 mlog_errno(ret);
2720 goto out;
2721 }
2722
2723 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2724 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2725 right_path, i);
2726 if (ret) {
2727 mlog_errno(ret);
2728 goto out;
2729 }
2730
2731 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2732 left_path, i);
2733 if (ret) {
2734 mlog_errno(ret);
2735 goto out;
2736 }
2737 }
2738
2739 if (!right_has_empty) {
2740 /*
2741 * Only do this if we're moving a real
2742 * record. Otherwise, the action is delayed until
2743 * after removal of the right path in which case we
2744 * can do a simple shift to remove the empty extent.
2745 */
2746 ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2747 memset(&right_leaf_el->l_recs[0], 0,
2748 sizeof(struct ocfs2_extent_rec));
2749 }
2750 if (eb->h_next_leaf_blk == 0ULL) {
2751 /*
2752 * Move recs over to get rid of empty extent, decrease
2753 * next_free. This is allowed to remove the last
2754 * extent in our leaf (setting l_next_free_rec to
2755 * zero) - the delete code below won't care.
2756 */
2757 ocfs2_remove_empty_extent(right_leaf_el);
2758 }
2759
2760 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2761 ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2762
2763 if (del_right_subtree) {
2764 ocfs2_unlink_subtree(handle, et, left_path, right_path,
2765 subtree_index, dealloc);
2766 ret = ocfs2_update_edge_lengths(handle, et, left_path);
2767 if (ret) {
2768 mlog_errno(ret);
2769 goto out;
2770 }
2771
2772 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2773 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2774
2775 /*
2776 * Removal of the extent in the left leaf was skipped
2777 * above so we could delete the right path
2778 * 1st.
2779 */
2780 if (right_has_empty)
2781 ocfs2_remove_empty_extent(left_leaf_el);
2782
2783 ocfs2_journal_dirty(handle, et_root_bh);
2784
2785 *deleted = 1;
2786 } else
2787 ocfs2_complete_edge_insert(handle, left_path, right_path,
2788 subtree_index);
2789
2790out:
2791 return ret;
2792}
2793
2794/*
2795 * Given a full path, determine what cpos value would return us a path
2796 * containing the leaf immediately to the right of the current one.
2797 *
2798 * Will return zero if the path passed in is already the rightmost path.
2799 *
2800 * This looks similar, but is subtly different to
2801 * ocfs2_find_cpos_for_left_leaf().
2802 */
2803int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
2804 struct ocfs2_path *path, u32 *cpos)
2805{
2806 int i, j, ret = 0;
2807 u64 blkno;
2808 struct ocfs2_extent_list *el;
2809
2810 *cpos = 0;
2811
2812 if (path->p_tree_depth == 0)
2813 return 0;
2814
2815 blkno = path_leaf_bh(path)->b_blocknr;
2816
2817 /* Start at the tree node just above the leaf and work our way up. */
2818 i = path->p_tree_depth - 1;
2819 while (i >= 0) {
2820 int next_free;
2821
2822 el = path->p_node[i].el;
2823
2824 /*
2825 * Find the extent record just after the one in our
2826 * path.
2827 */
2828 next_free = le16_to_cpu(el->l_next_free_rec);
2829 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2830 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2831 if (j == (next_free - 1)) {
2832 if (i == 0) {
2833 /*
2834 * We've determined that the
2835 * path specified is already
2836 * the rightmost one - return a
2837 * cpos of zero.
2838 */
2839 goto out;
2840 }
2841 /*
2842 * The rightmost record points to our
2843 * leaf - we need to travel up the
2844 * tree one level.
2845 */
2846 goto next_node;
2847 }
2848
2849 *cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
2850 goto out;
2851 }
2852 }
2853
2854 /*
2855 * If we got here, we never found a valid node where
2856 * the tree indicated one should be.
2857 */
2858 ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
2859 (unsigned long long)blkno);
2860 ret = -EROFS;
2861 goto out;
2862
2863next_node:
2864 blkno = path->p_node[i].bh->b_blocknr;
2865 i--;
2866 }
2867
2868out:
2869 return ret;
2870}
2871
2872static int ocfs2_rotate_rightmost_leaf_left(handle_t *handle,
2873 struct ocfs2_extent_tree *et,
2874 struct ocfs2_path *path)
2875{
2876 int ret;
2877 struct buffer_head *bh = path_leaf_bh(path);
2878 struct ocfs2_extent_list *el = path_leaf_el(path);
2879
2880 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2881 return 0;
2882
2883 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
2884 path_num_items(path) - 1);
2885 if (ret) {
2886 mlog_errno(ret);
2887 goto out;
2888 }
2889
2890 ocfs2_remove_empty_extent(el);
2891 ocfs2_journal_dirty(handle, bh);
2892
2893out:
2894 return ret;
2895}
2896
2897static int __ocfs2_rotate_tree_left(handle_t *handle,
2898 struct ocfs2_extent_tree *et,
2899 int orig_credits,
2900 struct ocfs2_path *path,
2901 struct ocfs2_cached_dealloc_ctxt *dealloc,
2902 struct ocfs2_path **empty_extent_path)
2903{
2904 int ret, subtree_root, deleted;
2905 u32 right_cpos;
2906 struct ocfs2_path *left_path = NULL;
2907 struct ocfs2_path *right_path = NULL;
2908 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2909
2910 if (!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])))
2911 return 0;
2912
2913 *empty_extent_path = NULL;
2914
2915 ret = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
2916 if (ret) {
2917 mlog_errno(ret);
2918 goto out;
2919 }
2920
2921 left_path = ocfs2_new_path_from_path(path);
2922 if (!left_path) {
2923 ret = -ENOMEM;
2924 mlog_errno(ret);
2925 goto out;
2926 }
2927
2928 ocfs2_cp_path(left_path, path);
2929
2930 right_path = ocfs2_new_path_from_path(path);
2931 if (!right_path) {
2932 ret = -ENOMEM;
2933 mlog_errno(ret);
2934 goto out;
2935 }
2936
2937 while (right_cpos) {
2938 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
2939 if (ret) {
2940 mlog_errno(ret);
2941 goto out;
2942 }
2943
2944 subtree_root = ocfs2_find_subtree_root(et, left_path,
2945 right_path);
2946
2947 trace_ocfs2_rotate_subtree(subtree_root,
2948 (unsigned long long)
2949 right_path->p_node[subtree_root].bh->b_blocknr,
2950 right_path->p_tree_depth);
2951
2952 ret = ocfs2_extend_rotate_transaction(handle, 0,
2953 orig_credits, left_path);
2954 if (ret) {
2955 mlog_errno(ret);
2956 goto out;
2957 }
2958
2959 /*
2960 * Caller might still want to make changes to the
2961 * tree root, so re-add it to the journal here.
2962 */
2963 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2964 left_path, 0);
2965 if (ret) {
2966 mlog_errno(ret);
2967 goto out;
2968 }
2969
2970 ret = ocfs2_rotate_subtree_left(handle, et, left_path,
2971 right_path, subtree_root,
2972 dealloc, &deleted);
2973 if (ret == -EAGAIN) {
2974 /*
2975 * The rotation has to temporarily stop due to
2976 * the right subtree having an empty
2977 * extent. Pass it back to the caller for a
2978 * fixup.
2979 */
2980 *empty_extent_path = right_path;
2981 right_path = NULL;
2982 goto out;
2983 }
2984 if (ret) {
2985 mlog_errno(ret);
2986 goto out;
2987 }
2988
2989 /*
2990 * The subtree rotate might have removed records on
2991 * the rightmost edge. If so, then rotation is
2992 * complete.
2993 */
2994 if (deleted)
2995 break;
2996
2997 ocfs2_mv_path(left_path, right_path);
2998
2999 ret = ocfs2_find_cpos_for_right_leaf(sb, left_path,
3000 &right_cpos);
3001 if (ret) {
3002 mlog_errno(ret);
3003 goto out;
3004 }
3005 }
3006
3007out:
3008 ocfs2_free_path(right_path);
3009 ocfs2_free_path(left_path);
3010
3011 return ret;
3012}
3013
3014static int ocfs2_remove_rightmost_path(handle_t *handle,
3015 struct ocfs2_extent_tree *et,
3016 struct ocfs2_path *path,
3017 struct ocfs2_cached_dealloc_ctxt *dealloc)
3018{
3019 int ret, subtree_index;
3020 u32 cpos;
3021 struct ocfs2_path *left_path = NULL;
3022 struct ocfs2_extent_block *eb;
3023 struct ocfs2_extent_list *el;
3024
3025 ret = ocfs2_et_sanity_check(et);
3026 if (ret)
3027 goto out;
3028
3029 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
3030 if (ret) {
3031 mlog_errno(ret);
3032 goto out;
3033 }
3034
3035 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3036 path, &cpos);
3037 if (ret) {
3038 mlog_errno(ret);
3039 goto out;
3040 }
3041
3042 if (cpos) {
3043 /*
3044 * We have a path to the left of this one - it needs
3045 * an update too.
3046 */
3047 left_path = ocfs2_new_path_from_path(path);
3048 if (!left_path) {
3049 ret = -ENOMEM;
3050 mlog_errno(ret);
3051 goto out;
3052 }
3053
3054 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
3055 if (ret) {
3056 mlog_errno(ret);
3057 goto out;
3058 }
3059
3060 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
3061 if (ret) {
3062 mlog_errno(ret);
3063 goto out;
3064 }
3065
3066 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
3067
3068 ocfs2_unlink_subtree(handle, et, left_path, path,
3069 subtree_index, dealloc);
3070 ret = ocfs2_update_edge_lengths(handle, et, left_path);
3071 if (ret) {
3072 mlog_errno(ret);
3073 goto out;
3074 }
3075
3076 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
3077 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
3078 } else {
3079 /*
3080 * 'path' is also the leftmost path which
3081 * means it must be the only one. This gets
3082 * handled differently because we want to
3083 * revert the root back to having extents
3084 * in-line.
3085 */
3086 ocfs2_unlink_path(handle, et, dealloc, path, 1);
3087
3088 el = et->et_root_el;
3089 el->l_tree_depth = 0;
3090 el->l_next_free_rec = 0;
3091 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3092
3093 ocfs2_et_set_last_eb_blk(et, 0);
3094 }
3095
3096 ocfs2_journal_dirty(handle, path_root_bh(path));
3097
3098out:
3099 ocfs2_free_path(left_path);
3100 return ret;
3101}
3102
3103static int ocfs2_remove_rightmost_empty_extent(struct ocfs2_super *osb,
3104 struct ocfs2_extent_tree *et,
3105 struct ocfs2_path *path,
3106 struct ocfs2_cached_dealloc_ctxt *dealloc)
3107{
3108 handle_t *handle;
3109 int ret;
3110 int credits = path->p_tree_depth * 2 + 1;
3111
3112 handle = ocfs2_start_trans(osb, credits);
3113 if (IS_ERR(handle)) {
3114 ret = PTR_ERR(handle);
3115 mlog_errno(ret);
3116 return ret;
3117 }
3118
3119 ret = ocfs2_remove_rightmost_path(handle, et, path, dealloc);
3120 if (ret)
3121 mlog_errno(ret);
3122
3123 ocfs2_commit_trans(osb, handle);
3124 return ret;
3125}
3126
3127/*
3128 * Left rotation of btree records.
3129 *
3130 * In many ways, this is (unsurprisingly) the opposite of right
3131 * rotation. We start at some non-rightmost path containing an empty
3132 * extent in the leaf block. The code works its way to the rightmost
3133 * path by rotating records to the left in every subtree.
3134 *
3135 * This is used by any code which reduces the number of extent records
3136 * in a leaf. After removal, an empty record should be placed in the
3137 * leftmost list position.
3138 *
3139 * This won't handle a length update of the rightmost path records if
3140 * the rightmost tree leaf record is removed so the caller is
3141 * responsible for detecting and correcting that.
3142 */
3143static int ocfs2_rotate_tree_left(handle_t *handle,
3144 struct ocfs2_extent_tree *et,
3145 struct ocfs2_path *path,
3146 struct ocfs2_cached_dealloc_ctxt *dealloc)
3147{
3148 int ret, orig_credits = jbd2_handle_buffer_credits(handle);
3149 struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
3150 struct ocfs2_extent_block *eb;
3151 struct ocfs2_extent_list *el;
3152
3153 el = path_leaf_el(path);
3154 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
3155 return 0;
3156
3157 if (path->p_tree_depth == 0) {
3158rightmost_no_delete:
3159 /*
3160 * Inline extents. This is trivially handled, so do
3161 * it up front.
3162 */
3163 ret = ocfs2_rotate_rightmost_leaf_left(handle, et, path);
3164 if (ret)
3165 mlog_errno(ret);
3166 goto out;
3167 }
3168
3169 /*
3170 * Handle rightmost branch now. There's several cases:
3171 * 1) simple rotation leaving records in there. That's trivial.
3172 * 2) rotation requiring a branch delete - there's no more
3173 * records left. Two cases of this:
3174 * a) There are branches to the left.
3175 * b) This is also the leftmost (the only) branch.
3176 *
3177 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
3178 * 2a) we need the left branch so that we can update it with the unlink
3179 * 2b) we need to bring the root back to inline extents.
3180 */
3181
3182 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
3183 el = &eb->h_list;
3184 if (eb->h_next_leaf_blk == 0) {
3185 /*
3186 * This gets a bit tricky if we're going to delete the
3187 * rightmost path. Get the other cases out of the way
3188 * 1st.
3189 */
3190 if (le16_to_cpu(el->l_next_free_rec) > 1)
3191 goto rightmost_no_delete;
3192
3193 if (le16_to_cpu(el->l_next_free_rec) == 0) {
3194 ret = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3195 "Owner %llu has empty extent block at %llu\n",
3196 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
3197 (unsigned long long)le64_to_cpu(eb->h_blkno));
3198 goto out;
3199 }
3200
3201 /*
3202 * XXX: The caller can not trust "path" any more after
3203 * this as it will have been deleted. What do we do?
3204 *
3205 * In theory the rotate-for-merge code will never get
3206 * here because it'll always ask for a rotate in a
3207 * nonempty list.
3208 */
3209
3210 ret = ocfs2_remove_rightmost_path(handle, et, path,
3211 dealloc);
3212 if (ret)
3213 mlog_errno(ret);
3214 goto out;
3215 }
3216
3217 /*
3218 * Now we can loop, remembering the path we get from -EAGAIN
3219 * and restarting from there.
3220 */
3221try_rotate:
3222 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits, path,
3223 dealloc, &restart_path);
3224 if (ret && ret != -EAGAIN) {
3225 mlog_errno(ret);
3226 goto out;
3227 }
3228
3229 while (ret == -EAGAIN) {
3230 tmp_path = restart_path;
3231 restart_path = NULL;
3232
3233 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits,
3234 tmp_path, dealloc,
3235 &restart_path);
3236 if (ret && ret != -EAGAIN) {
3237 mlog_errno(ret);
3238 goto out;
3239 }
3240
3241 ocfs2_free_path(tmp_path);
3242 tmp_path = NULL;
3243
3244 if (ret == 0)
3245 goto try_rotate;
3246 }
3247
3248out:
3249 ocfs2_free_path(tmp_path);
3250 ocfs2_free_path(restart_path);
3251 return ret;
3252}
3253
3254static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
3255 int index)
3256{
3257 struct ocfs2_extent_rec *rec = &el->l_recs[index];
3258 unsigned int size;
3259
3260 if (rec->e_leaf_clusters == 0) {
3261 /*
3262 * We consumed all of the merged-from record. An empty
3263 * extent cannot exist anywhere but the 1st array
3264 * position, so move things over if the merged-from
3265 * record doesn't occupy that position.
3266 *
3267 * This creates a new empty extent so the caller
3268 * should be smart enough to have removed any existing
3269 * ones.
3270 */
3271 if (index > 0) {
3272 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3273 size = index * sizeof(struct ocfs2_extent_rec);
3274 memmove(&el->l_recs[1], &el->l_recs[0], size);
3275 }
3276
3277 /*
3278 * Always memset - the caller doesn't check whether it
3279 * created an empty extent, so there could be junk in
3280 * the other fields.
3281 */
3282 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3283 }
3284}
3285
3286static int ocfs2_get_right_path(struct ocfs2_extent_tree *et,
3287 struct ocfs2_path *left_path,
3288 struct ocfs2_path **ret_right_path)
3289{
3290 int ret;
3291 u32 right_cpos;
3292 struct ocfs2_path *right_path = NULL;
3293 struct ocfs2_extent_list *left_el;
3294
3295 *ret_right_path = NULL;
3296
3297 /* This function shouldn't be called for non-trees. */
3298 BUG_ON(left_path->p_tree_depth == 0);
3299
3300 left_el = path_leaf_el(left_path);
3301 BUG_ON(left_el->l_next_free_rec != left_el->l_count);
3302
3303 ret = ocfs2_find_cpos_for_right_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3304 left_path, &right_cpos);
3305 if (ret) {
3306 mlog_errno(ret);
3307 goto out;
3308 }
3309
3310 /* This function shouldn't be called for the rightmost leaf. */
3311 BUG_ON(right_cpos == 0);
3312
3313 right_path = ocfs2_new_path_from_path(left_path);
3314 if (!right_path) {
3315 ret = -ENOMEM;
3316 mlog_errno(ret);
3317 goto out;
3318 }
3319
3320 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
3321 if (ret) {
3322 mlog_errno(ret);
3323 goto out;
3324 }
3325
3326 *ret_right_path = right_path;
3327out:
3328 if (ret)
3329 ocfs2_free_path(right_path);
3330 return ret;
3331}
3332
3333/*
3334 * Remove split_rec clusters from the record at index and merge them
3335 * onto the beginning of the record "next" to it.
3336 * For index < l_count - 1, the next means the extent rec at index + 1.
3337 * For index == l_count - 1, the "next" means the 1st extent rec of the
3338 * next extent block.
3339 */
3340static int ocfs2_merge_rec_right(struct ocfs2_path *left_path,
3341 handle_t *handle,
3342 struct ocfs2_extent_tree *et,
3343 struct ocfs2_extent_rec *split_rec,
3344 int index)
3345{
3346 int ret, next_free, i;
3347 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3348 struct ocfs2_extent_rec *left_rec;
3349 struct ocfs2_extent_rec *right_rec;
3350 struct ocfs2_extent_list *right_el;
3351 struct ocfs2_path *right_path = NULL;
3352 int subtree_index = 0;
3353 struct ocfs2_extent_list *el = path_leaf_el(left_path);
3354 struct buffer_head *bh = path_leaf_bh(left_path);
3355 struct buffer_head *root_bh = NULL;
3356
3357 BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
3358 left_rec = &el->l_recs[index];
3359
3360 if (index == le16_to_cpu(el->l_next_free_rec) - 1 &&
3361 le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count)) {
3362 /* we meet with a cross extent block merge. */
3363 ret = ocfs2_get_right_path(et, left_path, &right_path);
3364 if (ret) {
3365 mlog_errno(ret);
3366 return ret;
3367 }
3368
3369 right_el = path_leaf_el(right_path);
3370 next_free = le16_to_cpu(right_el->l_next_free_rec);
3371 BUG_ON(next_free <= 0);
3372 right_rec = &right_el->l_recs[0];
3373 if (ocfs2_is_empty_extent(right_rec)) {
3374 BUG_ON(next_free <= 1);
3375 right_rec = &right_el->l_recs[1];
3376 }
3377
3378 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3379 le16_to_cpu(left_rec->e_leaf_clusters) !=
3380 le32_to_cpu(right_rec->e_cpos));
3381
3382 subtree_index = ocfs2_find_subtree_root(et, left_path,
3383 right_path);
3384
3385 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3386 jbd2_handle_buffer_credits(handle),
3387 right_path);
3388 if (ret) {
3389 mlog_errno(ret);
3390 goto out;
3391 }
3392
3393 root_bh = left_path->p_node[subtree_index].bh;
3394 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3395
3396 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3397 subtree_index);
3398 if (ret) {
3399 mlog_errno(ret);
3400 goto out;
3401 }
3402
3403 for (i = subtree_index + 1;
3404 i < path_num_items(right_path); i++) {
3405 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3406 right_path, i);
3407 if (ret) {
3408 mlog_errno(ret);
3409 goto out;
3410 }
3411
3412 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3413 left_path, i);
3414 if (ret) {
3415 mlog_errno(ret);
3416 goto out;
3417 }
3418 }
3419
3420 } else {
3421 BUG_ON(index == le16_to_cpu(el->l_next_free_rec) - 1);
3422 right_rec = &el->l_recs[index + 1];
3423 }
3424
3425 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, left_path,
3426 path_num_items(left_path) - 1);
3427 if (ret) {
3428 mlog_errno(ret);
3429 goto out;
3430 }
3431
3432 le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
3433
3434 le32_add_cpu(&right_rec->e_cpos, -split_clusters);
3435 le64_add_cpu(&right_rec->e_blkno,
3436 -ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3437 split_clusters));
3438 le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
3439
3440 ocfs2_cleanup_merge(el, index);
3441
3442 ocfs2_journal_dirty(handle, bh);
3443 if (right_path) {
3444 ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
3445 ocfs2_complete_edge_insert(handle, left_path, right_path,
3446 subtree_index);
3447 }
3448out:
3449 ocfs2_free_path(right_path);
3450 return ret;
3451}
3452
3453static int ocfs2_get_left_path(struct ocfs2_extent_tree *et,
3454 struct ocfs2_path *right_path,
3455 struct ocfs2_path **ret_left_path)
3456{
3457 int ret;
3458 u32 left_cpos;
3459 struct ocfs2_path *left_path = NULL;
3460
3461 *ret_left_path = NULL;
3462
3463 /* This function shouldn't be called for non-trees. */
3464 BUG_ON(right_path->p_tree_depth == 0);
3465
3466 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3467 right_path, &left_cpos);
3468 if (ret) {
3469 mlog_errno(ret);
3470 goto out;
3471 }
3472
3473 /* This function shouldn't be called for the leftmost leaf. */
3474 BUG_ON(left_cpos == 0);
3475
3476 left_path = ocfs2_new_path_from_path(right_path);
3477 if (!left_path) {
3478 ret = -ENOMEM;
3479 mlog_errno(ret);
3480 goto out;
3481 }
3482
3483 ret = ocfs2_find_path(et->et_ci, left_path, left_cpos);
3484 if (ret) {
3485 mlog_errno(ret);
3486 goto out;
3487 }
3488
3489 *ret_left_path = left_path;
3490out:
3491 if (ret)
3492 ocfs2_free_path(left_path);
3493 return ret;
3494}
3495
3496/*
3497 * Remove split_rec clusters from the record at index and merge them
3498 * onto the tail of the record "before" it.
3499 * For index > 0, the "before" means the extent rec at index - 1.
3500 *
3501 * For index == 0, the "before" means the last record of the previous
3502 * extent block. And there is also a situation that we may need to
3503 * remove the rightmost leaf extent block in the right_path and change
3504 * the right path to indicate the new rightmost path.
3505 */
3506static int ocfs2_merge_rec_left(struct ocfs2_path *right_path,
3507 handle_t *handle,
3508 struct ocfs2_extent_tree *et,
3509 struct ocfs2_extent_rec *split_rec,
3510 struct ocfs2_cached_dealloc_ctxt *dealloc,
3511 int index)
3512{
3513 int ret, i, subtree_index = 0, has_empty_extent = 0;
3514 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3515 struct ocfs2_extent_rec *left_rec;
3516 struct ocfs2_extent_rec *right_rec;
3517 struct ocfs2_extent_list *el = path_leaf_el(right_path);
3518 struct buffer_head *bh = path_leaf_bh(right_path);
3519 struct buffer_head *root_bh = NULL;
3520 struct ocfs2_path *left_path = NULL;
3521 struct ocfs2_extent_list *left_el;
3522
3523 BUG_ON(index < 0);
3524
3525 right_rec = &el->l_recs[index];
3526 if (index == 0) {
3527 /* we meet with a cross extent block merge. */
3528 ret = ocfs2_get_left_path(et, right_path, &left_path);
3529 if (ret) {
3530 mlog_errno(ret);
3531 return ret;
3532 }
3533
3534 left_el = path_leaf_el(left_path);
3535 BUG_ON(le16_to_cpu(left_el->l_next_free_rec) !=
3536 le16_to_cpu(left_el->l_count));
3537
3538 left_rec = &left_el->l_recs[
3539 le16_to_cpu(left_el->l_next_free_rec) - 1];
3540 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3541 le16_to_cpu(left_rec->e_leaf_clusters) !=
3542 le32_to_cpu(split_rec->e_cpos));
3543
3544 subtree_index = ocfs2_find_subtree_root(et, left_path,
3545 right_path);
3546
3547 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3548 jbd2_handle_buffer_credits(handle),
3549 left_path);
3550 if (ret) {
3551 mlog_errno(ret);
3552 goto out;
3553 }
3554
3555 root_bh = left_path->p_node[subtree_index].bh;
3556 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3557
3558 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3559 subtree_index);
3560 if (ret) {
3561 mlog_errno(ret);
3562 goto out;
3563 }
3564
3565 for (i = subtree_index + 1;
3566 i < path_num_items(right_path); i++) {
3567 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3568 right_path, i);
3569 if (ret) {
3570 mlog_errno(ret);
3571 goto out;
3572 }
3573
3574 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3575 left_path, i);
3576 if (ret) {
3577 mlog_errno(ret);
3578 goto out;
3579 }
3580 }
3581 } else {
3582 left_rec = &el->l_recs[index - 1];
3583 if (ocfs2_is_empty_extent(&el->l_recs[0]))
3584 has_empty_extent = 1;
3585 }
3586
3587 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3588 path_num_items(right_path) - 1);
3589 if (ret) {
3590 mlog_errno(ret);
3591 goto out;
3592 }
3593
3594 if (has_empty_extent && index == 1) {
3595 /*
3596 * The easy case - we can just plop the record right in.
3597 */
3598 *left_rec = *split_rec;
3599 } else
3600 le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
3601
3602 le32_add_cpu(&right_rec->e_cpos, split_clusters);
3603 le64_add_cpu(&right_rec->e_blkno,
3604 ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3605 split_clusters));
3606 le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
3607
3608 ocfs2_cleanup_merge(el, index);
3609
3610 ocfs2_journal_dirty(handle, bh);
3611 if (left_path) {
3612 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
3613
3614 /*
3615 * In the situation that the right_rec is empty and the extent
3616 * block is empty also, ocfs2_complete_edge_insert can't handle
3617 * it and we need to delete the right extent block.
3618 */
3619 if (le16_to_cpu(right_rec->e_leaf_clusters) == 0 &&
3620 le16_to_cpu(el->l_next_free_rec) == 1) {
3621 /* extend credit for ocfs2_remove_rightmost_path */
3622 ret = ocfs2_extend_rotate_transaction(handle, 0,
3623 jbd2_handle_buffer_credits(handle),
3624 right_path);
3625 if (ret) {
3626 mlog_errno(ret);
3627 goto out;
3628 }
3629
3630 ret = ocfs2_remove_rightmost_path(handle, et,
3631 right_path,
3632 dealloc);
3633 if (ret) {
3634 mlog_errno(ret);
3635 goto out;
3636 }
3637
3638 /* Now the rightmost extent block has been deleted.
3639 * So we use the new rightmost path.
3640 */
3641 ocfs2_mv_path(right_path, left_path);
3642 left_path = NULL;
3643 } else
3644 ocfs2_complete_edge_insert(handle, left_path,
3645 right_path, subtree_index);
3646 }
3647out:
3648 ocfs2_free_path(left_path);
3649 return ret;
3650}
3651
3652static int ocfs2_try_to_merge_extent(handle_t *handle,
3653 struct ocfs2_extent_tree *et,
3654 struct ocfs2_path *path,
3655 int split_index,
3656 struct ocfs2_extent_rec *split_rec,
3657 struct ocfs2_cached_dealloc_ctxt *dealloc,
3658 struct ocfs2_merge_ctxt *ctxt)
3659{
3660 int ret = 0;
3661 struct ocfs2_extent_list *el = path_leaf_el(path);
3662 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3663
3664 BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
3665
3666 if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) {
3667 /* extend credit for ocfs2_remove_rightmost_path */
3668 ret = ocfs2_extend_rotate_transaction(handle, 0,
3669 jbd2_handle_buffer_credits(handle),
3670 path);
3671 if (ret) {
3672 mlog_errno(ret);
3673 goto out;
3674 }
3675 /*
3676 * The merge code will need to create an empty
3677 * extent to take the place of the newly
3678 * emptied slot. Remove any pre-existing empty
3679 * extents - having more than one in a leaf is
3680 * illegal.
3681 */
3682 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3683 if (ret) {
3684 mlog_errno(ret);
3685 goto out;
3686 }
3687 split_index--;
3688 rec = &el->l_recs[split_index];
3689 }
3690
3691 if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
3692 /*
3693 * Left-right contig implies this.
3694 */
3695 BUG_ON(!ctxt->c_split_covers_rec);
3696
3697 /*
3698 * Since the leftright insert always covers the entire
3699 * extent, this call will delete the insert record
3700 * entirely, resulting in an empty extent record added to
3701 * the extent block.
3702 *
3703 * Since the adding of an empty extent shifts
3704 * everything back to the right, there's no need to
3705 * update split_index here.
3706 *
3707 * When the split_index is zero, we need to merge it to the
3708 * prevoius extent block. It is more efficient and easier
3709 * if we do merge_right first and merge_left later.
3710 */
3711 ret = ocfs2_merge_rec_right(path, handle, et, split_rec,
3712 split_index);
3713 if (ret) {
3714 mlog_errno(ret);
3715 goto out;
3716 }
3717
3718 /*
3719 * We can only get this from logic error above.
3720 */
3721 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
3722
3723 /* extend credit for ocfs2_remove_rightmost_path */
3724 ret = ocfs2_extend_rotate_transaction(handle, 0,
3725 jbd2_handle_buffer_credits(handle),
3726 path);
3727 if (ret) {
3728 mlog_errno(ret);
3729 goto out;
3730 }
3731
3732 /* The merge left us with an empty extent, remove it. */
3733 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3734 if (ret) {
3735 mlog_errno(ret);
3736 goto out;
3737 }
3738
3739 rec = &el->l_recs[split_index];
3740
3741 /*
3742 * Note that we don't pass split_rec here on purpose -
3743 * we've merged it into the rec already.
3744 */
3745 ret = ocfs2_merge_rec_left(path, handle, et, rec,
3746 dealloc, split_index);
3747
3748 if (ret) {
3749 mlog_errno(ret);
3750 goto out;
3751 }
3752
3753 /* extend credit for ocfs2_remove_rightmost_path */
3754 ret = ocfs2_extend_rotate_transaction(handle, 0,
3755 jbd2_handle_buffer_credits(handle),
3756 path);
3757 if (ret) {
3758 mlog_errno(ret);
3759 goto out;
3760 }
3761
3762 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3763 /*
3764 * Error from this last rotate is not critical, so
3765 * print but don't bubble it up.
3766 */
3767 if (ret)
3768 mlog_errno(ret);
3769 ret = 0;
3770 } else {
3771 /*
3772 * Merge a record to the left or right.
3773 *
3774 * 'contig_type' is relative to the existing record,
3775 * so for example, if we're "right contig", it's to
3776 * the record on the left (hence the left merge).
3777 */
3778 if (ctxt->c_contig_type == CONTIG_RIGHT) {
3779 ret = ocfs2_merge_rec_left(path, handle, et,
3780 split_rec, dealloc,
3781 split_index);
3782 if (ret) {
3783 mlog_errno(ret);
3784 goto out;
3785 }
3786 } else {
3787 ret = ocfs2_merge_rec_right(path, handle,
3788 et, split_rec,
3789 split_index);
3790 if (ret) {
3791 mlog_errno(ret);
3792 goto out;
3793 }
3794 }
3795
3796 if (ctxt->c_split_covers_rec) {
3797 /* extend credit for ocfs2_remove_rightmost_path */
3798 ret = ocfs2_extend_rotate_transaction(handle, 0,
3799 jbd2_handle_buffer_credits(handle),
3800 path);
3801 if (ret) {
3802 mlog_errno(ret);
3803 ret = 0;
3804 goto out;
3805 }
3806
3807 /*
3808 * The merge may have left an empty extent in
3809 * our leaf. Try to rotate it away.
3810 */
3811 ret = ocfs2_rotate_tree_left(handle, et, path,
3812 dealloc);
3813 if (ret)
3814 mlog_errno(ret);
3815 ret = 0;
3816 }
3817 }
3818
3819out:
3820 return ret;
3821}
3822
3823static void ocfs2_subtract_from_rec(struct super_block *sb,
3824 enum ocfs2_split_type split,
3825 struct ocfs2_extent_rec *rec,
3826 struct ocfs2_extent_rec *split_rec)
3827{
3828 u64 len_blocks;
3829
3830 len_blocks = ocfs2_clusters_to_blocks(sb,
3831 le16_to_cpu(split_rec->e_leaf_clusters));
3832
3833 if (split == SPLIT_LEFT) {
3834 /*
3835 * Region is on the left edge of the existing
3836 * record.
3837 */
3838 le32_add_cpu(&rec->e_cpos,
3839 le16_to_cpu(split_rec->e_leaf_clusters));
3840 le64_add_cpu(&rec->e_blkno, len_blocks);
3841 le16_add_cpu(&rec->e_leaf_clusters,
3842 -le16_to_cpu(split_rec->e_leaf_clusters));
3843 } else {
3844 /*
3845 * Region is on the right edge of the existing
3846 * record.
3847 */
3848 le16_add_cpu(&rec->e_leaf_clusters,
3849 -le16_to_cpu(split_rec->e_leaf_clusters));
3850 }
3851}
3852
3853/*
3854 * Do the final bits of extent record insertion at the target leaf
3855 * list. If this leaf is part of an allocation tree, it is assumed
3856 * that the tree above has been prepared.
3857 */
3858static void ocfs2_insert_at_leaf(struct ocfs2_extent_tree *et,
3859 struct ocfs2_extent_rec *insert_rec,
3860 struct ocfs2_extent_list *el,
3861 struct ocfs2_insert_type *insert)
3862{
3863 int i = insert->ins_contig_index;
3864 unsigned int range;
3865 struct ocfs2_extent_rec *rec;
3866
3867 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3868
3869 if (insert->ins_split != SPLIT_NONE) {
3870 i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
3871 BUG_ON(i == -1);
3872 rec = &el->l_recs[i];
3873 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
3874 insert->ins_split, rec,
3875 insert_rec);
3876 goto rotate;
3877 }
3878
3879 /*
3880 * Contiguous insert - either left or right.
3881 */
3882 if (insert->ins_contig != CONTIG_NONE) {
3883 rec = &el->l_recs[i];
3884 if (insert->ins_contig == CONTIG_LEFT) {
3885 rec->e_blkno = insert_rec->e_blkno;
3886 rec->e_cpos = insert_rec->e_cpos;
3887 }
3888 le16_add_cpu(&rec->e_leaf_clusters,
3889 le16_to_cpu(insert_rec->e_leaf_clusters));
3890 return;
3891 }
3892
3893 /*
3894 * Handle insert into an empty leaf.
3895 */
3896 if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3897 ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3898 ocfs2_is_empty_extent(&el->l_recs[0]))) {
3899 el->l_recs[0] = *insert_rec;
3900 el->l_next_free_rec = cpu_to_le16(1);
3901 return;
3902 }
3903
3904 /*
3905 * Appending insert.
3906 */
3907 if (insert->ins_appending == APPEND_TAIL) {
3908 i = le16_to_cpu(el->l_next_free_rec) - 1;
3909 rec = &el->l_recs[i];
3910 range = le32_to_cpu(rec->e_cpos)
3911 + le16_to_cpu(rec->e_leaf_clusters);
3912 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3913
3914 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3915 le16_to_cpu(el->l_count),
3916 "owner %llu, depth %u, count %u, next free %u, "
3917 "rec.cpos %u, rec.clusters %u, "
3918 "insert.cpos %u, insert.clusters %u\n",
3919 ocfs2_metadata_cache_owner(et->et_ci),
3920 le16_to_cpu(el->l_tree_depth),
3921 le16_to_cpu(el->l_count),
3922 le16_to_cpu(el->l_next_free_rec),
3923 le32_to_cpu(el->l_recs[i].e_cpos),
3924 le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3925 le32_to_cpu(insert_rec->e_cpos),
3926 le16_to_cpu(insert_rec->e_leaf_clusters));
3927 i++;
3928 el->l_recs[i] = *insert_rec;
3929 le16_add_cpu(&el->l_next_free_rec, 1);
3930 return;
3931 }
3932
3933rotate:
3934 /*
3935 * Ok, we have to rotate.
3936 *
3937 * At this point, it is safe to assume that inserting into an
3938 * empty leaf and appending to a leaf have both been handled
3939 * above.
3940 *
3941 * This leaf needs to have space, either by the empty 1st
3942 * extent record, or by virtue of an l_next_free_rec < l_count.
3943 */
3944 ocfs2_rotate_leaf(el, insert_rec);
3945}
3946
3947static void ocfs2_adjust_rightmost_records(handle_t *handle,
3948 struct ocfs2_extent_tree *et,
3949 struct ocfs2_path *path,
3950 struct ocfs2_extent_rec *insert_rec)
3951{
3952 int i, next_free;
3953 struct buffer_head *bh;
3954 struct ocfs2_extent_list *el;
3955 struct ocfs2_extent_rec *rec;
3956
3957 /*
3958 * Update everything except the leaf block.
3959 */
3960 for (i = 0; i < path->p_tree_depth; i++) {
3961 bh = path->p_node[i].bh;
3962 el = path->p_node[i].el;
3963
3964 next_free = le16_to_cpu(el->l_next_free_rec);
3965 if (next_free == 0) {
3966 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3967 "Owner %llu has a bad extent list\n",
3968 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
3969 return;
3970 }
3971
3972 rec = &el->l_recs[next_free - 1];
3973
3974 rec->e_int_clusters = insert_rec->e_cpos;
3975 le32_add_cpu(&rec->e_int_clusters,
3976 le16_to_cpu(insert_rec->e_leaf_clusters));
3977 le32_add_cpu(&rec->e_int_clusters,
3978 -le32_to_cpu(rec->e_cpos));
3979
3980 ocfs2_journal_dirty(handle, bh);
3981 }
3982}
3983
3984static int ocfs2_append_rec_to_path(handle_t *handle,
3985 struct ocfs2_extent_tree *et,
3986 struct ocfs2_extent_rec *insert_rec,
3987 struct ocfs2_path *right_path,
3988 struct ocfs2_path **ret_left_path)
3989{
3990 int ret, next_free;
3991 struct ocfs2_extent_list *el;
3992 struct ocfs2_path *left_path = NULL;
3993
3994 *ret_left_path = NULL;
3995
3996 /*
3997 * This shouldn't happen for non-trees. The extent rec cluster
3998 * count manipulation below only works for interior nodes.
3999 */
4000 BUG_ON(right_path->p_tree_depth == 0);
4001
4002 /*
4003 * If our appending insert is at the leftmost edge of a leaf,
4004 * then we might need to update the rightmost records of the
4005 * neighboring path.
4006 */
4007 el = path_leaf_el(right_path);
4008 next_free = le16_to_cpu(el->l_next_free_rec);
4009 if (next_free == 0 ||
4010 (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
4011 u32 left_cpos;
4012
4013 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
4014 right_path, &left_cpos);
4015 if (ret) {
4016 mlog_errno(ret);
4017 goto out;
4018 }
4019
4020 trace_ocfs2_append_rec_to_path(
4021 (unsigned long long)
4022 ocfs2_metadata_cache_owner(et->et_ci),
4023 le32_to_cpu(insert_rec->e_cpos),
4024 left_cpos);
4025
4026 /*
4027 * No need to worry if the append is already in the
4028 * leftmost leaf.
4029 */
4030 if (left_cpos) {
4031 left_path = ocfs2_new_path_from_path(right_path);
4032 if (!left_path) {
4033 ret = -ENOMEM;
4034 mlog_errno(ret);
4035 goto out;
4036 }
4037
4038 ret = ocfs2_find_path(et->et_ci, left_path,
4039 left_cpos);
4040 if (ret) {
4041 mlog_errno(ret);
4042 goto out;
4043 }
4044
4045 /*
4046 * ocfs2_insert_path() will pass the left_path to the
4047 * journal for us.
4048 */
4049 }
4050 }
4051
4052 ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4053 if (ret) {
4054 mlog_errno(ret);
4055 goto out;
4056 }
4057
4058 ocfs2_adjust_rightmost_records(handle, et, right_path, insert_rec);
4059
4060 *ret_left_path = left_path;
4061 ret = 0;
4062out:
4063 if (ret != 0)
4064 ocfs2_free_path(left_path);
4065
4066 return ret;
4067}
4068
4069static void ocfs2_split_record(struct ocfs2_extent_tree *et,
4070 struct ocfs2_path *left_path,
4071 struct ocfs2_path *right_path,
4072 struct ocfs2_extent_rec *split_rec,
4073 enum ocfs2_split_type split)
4074{
4075 int index;
4076 u32 cpos = le32_to_cpu(split_rec->e_cpos);
4077 struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
4078 struct ocfs2_extent_rec *rec, *tmprec;
4079
4080 right_el = path_leaf_el(right_path);
4081 if (left_path)
4082 left_el = path_leaf_el(left_path);
4083
4084 el = right_el;
4085 insert_el = right_el;
4086 index = ocfs2_search_extent_list(el, cpos);
4087 if (index != -1) {
4088 if (index == 0 && left_path) {
4089 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
4090
4091 /*
4092 * This typically means that the record
4093 * started in the left path but moved to the
4094 * right as a result of rotation. We either
4095 * move the existing record to the left, or we
4096 * do the later insert there.
4097 *
4098 * In this case, the left path should always
4099 * exist as the rotate code will have passed
4100 * it back for a post-insert update.
4101 */
4102
4103 if (split == SPLIT_LEFT) {
4104 /*
4105 * It's a left split. Since we know
4106 * that the rotate code gave us an
4107 * empty extent in the left path, we
4108 * can just do the insert there.
4109 */
4110 insert_el = left_el;
4111 } else {
4112 /*
4113 * Right split - we have to move the
4114 * existing record over to the left
4115 * leaf. The insert will be into the
4116 * newly created empty extent in the
4117 * right leaf.
4118 */
4119 tmprec = &right_el->l_recs[index];
4120 ocfs2_rotate_leaf(left_el, tmprec);
4121 el = left_el;
4122
4123 memset(tmprec, 0, sizeof(*tmprec));
4124 index = ocfs2_search_extent_list(left_el, cpos);
4125 BUG_ON(index == -1);
4126 }
4127 }
4128 } else {
4129 BUG_ON(!left_path);
4130 BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
4131 /*
4132 * Left path is easy - we can just allow the insert to
4133 * happen.
4134 */
4135 el = left_el;
4136 insert_el = left_el;
4137 index = ocfs2_search_extent_list(el, cpos);
4138 BUG_ON(index == -1);
4139 }
4140
4141 rec = &el->l_recs[index];
4142 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4143 split, rec, split_rec);
4144 ocfs2_rotate_leaf(insert_el, split_rec);
4145}
4146
4147/*
4148 * This function only does inserts on an allocation b-tree. For tree
4149 * depth = 0, ocfs2_insert_at_leaf() is called directly.
4150 *
4151 * right_path is the path we want to do the actual insert
4152 * in. left_path should only be passed in if we need to update that
4153 * portion of the tree after an edge insert.
4154 */
4155static int ocfs2_insert_path(handle_t *handle,
4156 struct ocfs2_extent_tree *et,
4157 struct ocfs2_path *left_path,
4158 struct ocfs2_path *right_path,
4159 struct ocfs2_extent_rec *insert_rec,
4160 struct ocfs2_insert_type *insert)
4161{
4162 int ret, subtree_index;
4163 struct buffer_head *leaf_bh = path_leaf_bh(right_path);
4164
4165 if (left_path) {
4166 /*
4167 * There's a chance that left_path got passed back to
4168 * us without being accounted for in the
4169 * journal. Extend our transaction here to be sure we
4170 * can change those blocks.
4171 */
4172 ret = ocfs2_extend_trans(handle, left_path->p_tree_depth);
4173 if (ret < 0) {
4174 mlog_errno(ret);
4175 goto out;
4176 }
4177
4178 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
4179 if (ret < 0) {
4180 mlog_errno(ret);
4181 goto out;
4182 }
4183 }
4184
4185 /*
4186 * Pass both paths to the journal. The majority of inserts
4187 * will be touching all components anyway.
4188 */
4189 ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4190 if (ret < 0) {
4191 mlog_errno(ret);
4192 goto out;
4193 }
4194
4195 if (insert->ins_split != SPLIT_NONE) {
4196 /*
4197 * We could call ocfs2_insert_at_leaf() for some types
4198 * of splits, but it's easier to just let one separate
4199 * function sort it all out.
4200 */
4201 ocfs2_split_record(et, left_path, right_path,
4202 insert_rec, insert->ins_split);
4203
4204 /*
4205 * Split might have modified either leaf and we don't
4206 * have a guarantee that the later edge insert will
4207 * dirty this for us.
4208 */
4209 if (left_path)
4210 ocfs2_journal_dirty(handle,
4211 path_leaf_bh(left_path));
4212 } else
4213 ocfs2_insert_at_leaf(et, insert_rec, path_leaf_el(right_path),
4214 insert);
4215
4216 ocfs2_journal_dirty(handle, leaf_bh);
4217
4218 if (left_path) {
4219 /*
4220 * The rotate code has indicated that we need to fix
4221 * up portions of the tree after the insert.
4222 *
4223 * XXX: Should we extend the transaction here?
4224 */
4225 subtree_index = ocfs2_find_subtree_root(et, left_path,
4226 right_path);
4227 ocfs2_complete_edge_insert(handle, left_path, right_path,
4228 subtree_index);
4229 }
4230
4231 ret = 0;
4232out:
4233 return ret;
4234}
4235
4236static int ocfs2_do_insert_extent(handle_t *handle,
4237 struct ocfs2_extent_tree *et,
4238 struct ocfs2_extent_rec *insert_rec,
4239 struct ocfs2_insert_type *type)
4240{
4241 int ret, rotate = 0;
4242 u32 cpos;
4243 struct ocfs2_path *right_path = NULL;
4244 struct ocfs2_path *left_path = NULL;
4245 struct ocfs2_extent_list *el;
4246
4247 el = et->et_root_el;
4248
4249 ret = ocfs2_et_root_journal_access(handle, et,
4250 OCFS2_JOURNAL_ACCESS_WRITE);
4251 if (ret) {
4252 mlog_errno(ret);
4253 goto out;
4254 }
4255
4256 if (le16_to_cpu(el->l_tree_depth) == 0) {
4257 ocfs2_insert_at_leaf(et, insert_rec, el, type);
4258 goto out_update_clusters;
4259 }
4260
4261 right_path = ocfs2_new_path_from_et(et);
4262 if (!right_path) {
4263 ret = -ENOMEM;
4264 mlog_errno(ret);
4265 goto out;
4266 }
4267
4268 /*
4269 * Determine the path to start with. Rotations need the
4270 * rightmost path, everything else can go directly to the
4271 * target leaf.
4272 */
4273 cpos = le32_to_cpu(insert_rec->e_cpos);
4274 if (type->ins_appending == APPEND_NONE &&
4275 type->ins_contig == CONTIG_NONE) {
4276 rotate = 1;
4277 cpos = UINT_MAX;
4278 }
4279
4280 ret = ocfs2_find_path(et->et_ci, right_path, cpos);
4281 if (ret) {
4282 mlog_errno(ret);
4283 goto out;
4284 }
4285
4286 /*
4287 * Rotations and appends need special treatment - they modify
4288 * parts of the tree's above them.
4289 *
4290 * Both might pass back a path immediate to the left of the
4291 * one being inserted to. This will be cause
4292 * ocfs2_insert_path() to modify the rightmost records of
4293 * left_path to account for an edge insert.
4294 *
4295 * XXX: When modifying this code, keep in mind that an insert
4296 * can wind up skipping both of these two special cases...
4297 */
4298 if (rotate) {
4299 ret = ocfs2_rotate_tree_right(handle, et, type->ins_split,
4300 le32_to_cpu(insert_rec->e_cpos),
4301 right_path, &left_path);
4302 if (ret) {
4303 mlog_errno(ret);
4304 goto out;
4305 }
4306
4307 /*
4308 * ocfs2_rotate_tree_right() might have extended the
4309 * transaction without re-journaling our tree root.
4310 */
4311 ret = ocfs2_et_root_journal_access(handle, et,
4312 OCFS2_JOURNAL_ACCESS_WRITE);
4313 if (ret) {
4314 mlog_errno(ret);
4315 goto out;
4316 }
4317 } else if (type->ins_appending == APPEND_TAIL
4318 && type->ins_contig != CONTIG_LEFT) {
4319 ret = ocfs2_append_rec_to_path(handle, et, insert_rec,
4320 right_path, &left_path);
4321 if (ret) {
4322 mlog_errno(ret);
4323 goto out;
4324 }
4325 }
4326
4327 ret = ocfs2_insert_path(handle, et, left_path, right_path,
4328 insert_rec, type);
4329 if (ret) {
4330 mlog_errno(ret);
4331 goto out;
4332 }
4333
4334out_update_clusters:
4335 if (type->ins_split == SPLIT_NONE)
4336 ocfs2_et_update_clusters(et,
4337 le16_to_cpu(insert_rec->e_leaf_clusters));
4338
4339 ocfs2_journal_dirty(handle, et->et_root_bh);
4340
4341out:
4342 ocfs2_free_path(left_path);
4343 ocfs2_free_path(right_path);
4344
4345 return ret;
4346}
4347
4348static int ocfs2_figure_merge_contig_type(struct ocfs2_extent_tree *et,
4349 struct ocfs2_path *path,
4350 struct ocfs2_extent_list *el, int index,
4351 struct ocfs2_extent_rec *split_rec,
4352 struct ocfs2_merge_ctxt *ctxt)
4353{
4354 int status = 0;
4355 enum ocfs2_contig_type ret = CONTIG_NONE;
4356 u32 left_cpos, right_cpos;
4357 struct ocfs2_extent_rec *rec = NULL;
4358 struct ocfs2_extent_list *new_el;
4359 struct ocfs2_path *left_path = NULL, *right_path = NULL;
4360 struct buffer_head *bh;
4361 struct ocfs2_extent_block *eb;
4362 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
4363
4364 if (index > 0) {
4365 rec = &el->l_recs[index - 1];
4366 } else if (path->p_tree_depth > 0) {
4367 status = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
4368 if (status)
4369 goto exit;
4370
4371 if (left_cpos != 0) {
4372 left_path = ocfs2_new_path_from_path(path);
4373 if (!left_path) {
4374 status = -ENOMEM;
4375 mlog_errno(status);
4376 goto exit;
4377 }
4378
4379 status = ocfs2_find_path(et->et_ci, left_path,
4380 left_cpos);
4381 if (status)
4382 goto free_left_path;
4383
4384 new_el = path_leaf_el(left_path);
4385
4386 if (le16_to_cpu(new_el->l_next_free_rec) !=
4387 le16_to_cpu(new_el->l_count)) {
4388 bh = path_leaf_bh(left_path);
4389 eb = (struct ocfs2_extent_block *)bh->b_data;
4390 status = ocfs2_error(sb,
4391 "Extent block #%llu has an invalid l_next_free_rec of %d. It should have matched the l_count of %d\n",
4392 (unsigned long long)le64_to_cpu(eb->h_blkno),
4393 le16_to_cpu(new_el->l_next_free_rec),
4394 le16_to_cpu(new_el->l_count));
4395 goto free_left_path;
4396 }
4397 rec = &new_el->l_recs[
4398 le16_to_cpu(new_el->l_next_free_rec) - 1];
4399 }
4400 }
4401
4402 /*
4403 * We're careful to check for an empty extent record here -
4404 * the merge code will know what to do if it sees one.
4405 */
4406 if (rec) {
4407 if (index == 1 && ocfs2_is_empty_extent(rec)) {
4408 if (split_rec->e_cpos == el->l_recs[index].e_cpos)
4409 ret = CONTIG_RIGHT;
4410 } else {
4411 ret = ocfs2_et_extent_contig(et, rec, split_rec);
4412 }
4413 }
4414
4415 rec = NULL;
4416 if (index < (le16_to_cpu(el->l_next_free_rec) - 1))
4417 rec = &el->l_recs[index + 1];
4418 else if (le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count) &&
4419 path->p_tree_depth > 0) {
4420 status = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
4421 if (status)
4422 goto free_left_path;
4423
4424 if (right_cpos == 0)
4425 goto free_left_path;
4426
4427 right_path = ocfs2_new_path_from_path(path);
4428 if (!right_path) {
4429 status = -ENOMEM;
4430 mlog_errno(status);
4431 goto free_left_path;
4432 }
4433
4434 status = ocfs2_find_path(et->et_ci, right_path, right_cpos);
4435 if (status)
4436 goto free_right_path;
4437
4438 new_el = path_leaf_el(right_path);
4439 rec = &new_el->l_recs[0];
4440 if (ocfs2_is_empty_extent(rec)) {
4441 if (le16_to_cpu(new_el->l_next_free_rec) <= 1) {
4442 bh = path_leaf_bh(right_path);
4443 eb = (struct ocfs2_extent_block *)bh->b_data;
4444 status = ocfs2_error(sb,
4445 "Extent block #%llu has an invalid l_next_free_rec of %d\n",
4446 (unsigned long long)le64_to_cpu(eb->h_blkno),
4447 le16_to_cpu(new_el->l_next_free_rec));
4448 goto free_right_path;
4449 }
4450 rec = &new_el->l_recs[1];
4451 }
4452 }
4453
4454 if (rec) {
4455 enum ocfs2_contig_type contig_type;
4456
4457 contig_type = ocfs2_et_extent_contig(et, rec, split_rec);
4458
4459 if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
4460 ret = CONTIG_LEFTRIGHT;
4461 else if (ret == CONTIG_NONE)
4462 ret = contig_type;
4463 }
4464
4465free_right_path:
4466 ocfs2_free_path(right_path);
4467free_left_path:
4468 ocfs2_free_path(left_path);
4469exit:
4470 if (status == 0)
4471 ctxt->c_contig_type = ret;
4472
4473 return status;
4474}
4475
4476static void ocfs2_figure_contig_type(struct ocfs2_extent_tree *et,
4477 struct ocfs2_insert_type *insert,
4478 struct ocfs2_extent_list *el,
4479 struct ocfs2_extent_rec *insert_rec)
4480{
4481 int i;
4482 enum ocfs2_contig_type contig_type = CONTIG_NONE;
4483
4484 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4485
4486 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
4487 contig_type = ocfs2_et_extent_contig(et, &el->l_recs[i],
4488 insert_rec);
4489 if (contig_type != CONTIG_NONE) {
4490 insert->ins_contig_index = i;
4491 break;
4492 }
4493 }
4494 insert->ins_contig = contig_type;
4495
4496 if (insert->ins_contig != CONTIG_NONE) {
4497 struct ocfs2_extent_rec *rec =
4498 &el->l_recs[insert->ins_contig_index];
4499 unsigned int len = le16_to_cpu(rec->e_leaf_clusters) +
4500 le16_to_cpu(insert_rec->e_leaf_clusters);
4501
4502 /*
4503 * Caller might want us to limit the size of extents, don't
4504 * calculate contiguousness if we might exceed that limit.
4505 */
4506 if (et->et_max_leaf_clusters &&
4507 (len > et->et_max_leaf_clusters))
4508 insert->ins_contig = CONTIG_NONE;
4509 }
4510}
4511
4512/*
4513 * This should only be called against the righmost leaf extent list.
4514 *
4515 * ocfs2_figure_appending_type() will figure out whether we'll have to
4516 * insert at the tail of the rightmost leaf.
4517 *
4518 * This should also work against the root extent list for tree's with 0
4519 * depth. If we consider the root extent list to be the rightmost leaf node
4520 * then the logic here makes sense.
4521 */
4522static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
4523 struct ocfs2_extent_list *el,
4524 struct ocfs2_extent_rec *insert_rec)
4525{
4526 int i;
4527 u32 cpos = le32_to_cpu(insert_rec->e_cpos);
4528 struct ocfs2_extent_rec *rec;
4529
4530 insert->ins_appending = APPEND_NONE;
4531
4532 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4533
4534 if (!el->l_next_free_rec)
4535 goto set_tail_append;
4536
4537 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
4538 /* Were all records empty? */
4539 if (le16_to_cpu(el->l_next_free_rec) == 1)
4540 goto set_tail_append;
4541 }
4542
4543 i = le16_to_cpu(el->l_next_free_rec) - 1;
4544 rec = &el->l_recs[i];
4545
4546 if (cpos >=
4547 (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
4548 goto set_tail_append;
4549
4550 return;
4551
4552set_tail_append:
4553 insert->ins_appending = APPEND_TAIL;
4554}
4555
4556/*
4557 * Helper function called at the beginning of an insert.
4558 *
4559 * This computes a few things that are commonly used in the process of
4560 * inserting into the btree:
4561 * - Whether the new extent is contiguous with an existing one.
4562 * - The current tree depth.
4563 * - Whether the insert is an appending one.
4564 * - The total # of free records in the tree.
4565 *
4566 * All of the information is stored on the ocfs2_insert_type
4567 * structure.
4568 */
4569static int ocfs2_figure_insert_type(struct ocfs2_extent_tree *et,
4570 struct buffer_head **last_eb_bh,
4571 struct ocfs2_extent_rec *insert_rec,
4572 int *free_records,
4573 struct ocfs2_insert_type *insert)
4574{
4575 int ret;
4576 struct ocfs2_extent_block *eb;
4577 struct ocfs2_extent_list *el;
4578 struct ocfs2_path *path = NULL;
4579 struct buffer_head *bh = NULL;
4580
4581 insert->ins_split = SPLIT_NONE;
4582
4583 el = et->et_root_el;
4584 insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
4585
4586 if (el->l_tree_depth) {
4587 /*
4588 * If we have tree depth, we read in the
4589 * rightmost extent block ahead of time as
4590 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4591 * may want it later.
4592 */
4593 ret = ocfs2_read_extent_block(et->et_ci,
4594 ocfs2_et_get_last_eb_blk(et),
4595 &bh);
4596 if (ret) {
4597 mlog_errno(ret);
4598 goto out;
4599 }
4600 eb = (struct ocfs2_extent_block *) bh->b_data;
4601 el = &eb->h_list;
4602 }
4603
4604 /*
4605 * Unless we have a contiguous insert, we'll need to know if
4606 * there is room left in our allocation tree for another
4607 * extent record.
4608 *
4609 * XXX: This test is simplistic, we can search for empty
4610 * extent records too.
4611 */
4612 *free_records = le16_to_cpu(el->l_count) -
4613 le16_to_cpu(el->l_next_free_rec);
4614
4615 if (!insert->ins_tree_depth) {
4616 ocfs2_figure_contig_type(et, insert, el, insert_rec);
4617 ocfs2_figure_appending_type(insert, el, insert_rec);
4618 return 0;
4619 }
4620
4621 path = ocfs2_new_path_from_et(et);
4622 if (!path) {
4623 ret = -ENOMEM;
4624 mlog_errno(ret);
4625 goto out;
4626 }
4627
4628 /*
4629 * In the case that we're inserting past what the tree
4630 * currently accounts for, ocfs2_find_path() will return for
4631 * us the rightmost tree path. This is accounted for below in
4632 * the appending code.
4633 */
4634 ret = ocfs2_find_path(et->et_ci, path, le32_to_cpu(insert_rec->e_cpos));
4635 if (ret) {
4636 mlog_errno(ret);
4637 goto out;
4638 }
4639
4640 el = path_leaf_el(path);
4641
4642 /*
4643 * Now that we have the path, there's two things we want to determine:
4644 * 1) Contiguousness (also set contig_index if this is so)
4645 *
4646 * 2) Are we doing an append? We can trivially break this up
4647 * into two types of appends: simple record append, or a
4648 * rotate inside the tail leaf.
4649 */
4650 ocfs2_figure_contig_type(et, insert, el, insert_rec);
4651
4652 /*
4653 * The insert code isn't quite ready to deal with all cases of
4654 * left contiguousness. Specifically, if it's an insert into
4655 * the 1st record in a leaf, it will require the adjustment of
4656 * cluster count on the last record of the path directly to it's
4657 * left. For now, just catch that case and fool the layers
4658 * above us. This works just fine for tree_depth == 0, which
4659 * is why we allow that above.
4660 */
4661 if (insert->ins_contig == CONTIG_LEFT &&
4662 insert->ins_contig_index == 0)
4663 insert->ins_contig = CONTIG_NONE;
4664
4665 /*
4666 * Ok, so we can simply compare against last_eb to figure out
4667 * whether the path doesn't exist. This will only happen in
4668 * the case that we're doing a tail append, so maybe we can
4669 * take advantage of that information somehow.
4670 */
4671 if (ocfs2_et_get_last_eb_blk(et) ==
4672 path_leaf_bh(path)->b_blocknr) {
4673 /*
4674 * Ok, ocfs2_find_path() returned us the rightmost
4675 * tree path. This might be an appending insert. There are
4676 * two cases:
4677 * 1) We're doing a true append at the tail:
4678 * -This might even be off the end of the leaf
4679 * 2) We're "appending" by rotating in the tail
4680 */
4681 ocfs2_figure_appending_type(insert, el, insert_rec);
4682 }
4683
4684out:
4685 ocfs2_free_path(path);
4686
4687 if (ret == 0)
4688 *last_eb_bh = bh;
4689 else
4690 brelse(bh);
4691 return ret;
4692}
4693
4694/*
4695 * Insert an extent into a btree.
4696 *
4697 * The caller needs to update the owning btree's cluster count.
4698 */
4699int ocfs2_insert_extent(handle_t *handle,
4700 struct ocfs2_extent_tree *et,
4701 u32 cpos,
4702 u64 start_blk,
4703 u32 new_clusters,
4704 u8 flags,
4705 struct ocfs2_alloc_context *meta_ac)
4706{
4707 int status;
4708 int free_records;
4709 struct buffer_head *last_eb_bh = NULL;
4710 struct ocfs2_insert_type insert = {0, };
4711 struct ocfs2_extent_rec rec;
4712
4713 trace_ocfs2_insert_extent_start(
4714 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4715 cpos, new_clusters);
4716
4717 memset(&rec, 0, sizeof(rec));
4718 rec.e_cpos = cpu_to_le32(cpos);
4719 rec.e_blkno = cpu_to_le64(start_blk);
4720 rec.e_leaf_clusters = cpu_to_le16(new_clusters);
4721 rec.e_flags = flags;
4722 status = ocfs2_et_insert_check(et, &rec);
4723 if (status) {
4724 mlog_errno(status);
4725 goto bail;
4726 }
4727
4728 status = ocfs2_figure_insert_type(et, &last_eb_bh, &rec,
4729 &free_records, &insert);
4730 if (status < 0) {
4731 mlog_errno(status);
4732 goto bail;
4733 }
4734
4735 trace_ocfs2_insert_extent(insert.ins_appending, insert.ins_contig,
4736 insert.ins_contig_index, free_records,
4737 insert.ins_tree_depth);
4738
4739 if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
4740 status = ocfs2_grow_tree(handle, et,
4741 &insert.ins_tree_depth, &last_eb_bh,
4742 meta_ac);
4743 if (status) {
4744 mlog_errno(status);
4745 goto bail;
4746 }
4747 }
4748
4749 /* Finally, we can add clusters. This might rotate the tree for us. */
4750 status = ocfs2_do_insert_extent(handle, et, &rec, &insert);
4751 if (status < 0)
4752 mlog_errno(status);
4753 else
4754 ocfs2_et_extent_map_insert(et, &rec);
4755
4756bail:
4757 brelse(last_eb_bh);
4758
4759 return status;
4760}
4761
4762/*
4763 * Allcate and add clusters into the extent b-tree.
4764 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4765 * The extent b-tree's root is specified by et, and
4766 * it is not limited to the file storage. Any extent tree can use this
4767 * function if it implements the proper ocfs2_extent_tree.
4768 */
4769int ocfs2_add_clusters_in_btree(handle_t *handle,
4770 struct ocfs2_extent_tree *et,
4771 u32 *logical_offset,
4772 u32 clusters_to_add,
4773 int mark_unwritten,
4774 struct ocfs2_alloc_context *data_ac,
4775 struct ocfs2_alloc_context *meta_ac,
4776 enum ocfs2_alloc_restarted *reason_ret)
4777{
4778 int status = 0, err = 0;
4779 int need_free = 0;
4780 int free_extents;
4781 enum ocfs2_alloc_restarted reason = RESTART_NONE;
4782 u32 bit_off, num_bits;
4783 u64 block;
4784 u8 flags = 0;
4785 struct ocfs2_super *osb =
4786 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
4787
4788 BUG_ON(!clusters_to_add);
4789
4790 if (mark_unwritten)
4791 flags = OCFS2_EXT_UNWRITTEN;
4792
4793 free_extents = ocfs2_num_free_extents(et);
4794 if (free_extents < 0) {
4795 status = free_extents;
4796 mlog_errno(status);
4797 goto leave;
4798 }
4799
4800 /* there are two cases which could cause us to EAGAIN in the
4801 * we-need-more-metadata case:
4802 * 1) we haven't reserved *any*
4803 * 2) we are so fragmented, we've needed to add metadata too
4804 * many times. */
4805 if (!free_extents && !meta_ac) {
4806 err = -1;
4807 status = -EAGAIN;
4808 reason = RESTART_META;
4809 goto leave;
4810 } else if ((!free_extents)
4811 && (ocfs2_alloc_context_bits_left(meta_ac)
4812 < ocfs2_extend_meta_needed(et->et_root_el))) {
4813 err = -2;
4814 status = -EAGAIN;
4815 reason = RESTART_META;
4816 goto leave;
4817 }
4818
4819 status = __ocfs2_claim_clusters(handle, data_ac, 1,
4820 clusters_to_add, &bit_off, &num_bits);
4821 if (status < 0) {
4822 if (status != -ENOSPC)
4823 mlog_errno(status);
4824 goto leave;
4825 }
4826
4827 BUG_ON(num_bits > clusters_to_add);
4828
4829 /* reserve our write early -- insert_extent may update the tree root */
4830 status = ocfs2_et_root_journal_access(handle, et,
4831 OCFS2_JOURNAL_ACCESS_WRITE);
4832 if (status < 0) {
4833 mlog_errno(status);
4834 need_free = 1;
4835 goto bail;
4836 }
4837
4838 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
4839 trace_ocfs2_add_clusters_in_btree(
4840 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4841 bit_off, num_bits);
4842 status = ocfs2_insert_extent(handle, et, *logical_offset, block,
4843 num_bits, flags, meta_ac);
4844 if (status < 0) {
4845 mlog_errno(status);
4846 need_free = 1;
4847 goto bail;
4848 }
4849
4850 ocfs2_journal_dirty(handle, et->et_root_bh);
4851
4852 clusters_to_add -= num_bits;
4853 *logical_offset += num_bits;
4854
4855 if (clusters_to_add) {
4856 err = clusters_to_add;
4857 status = -EAGAIN;
4858 reason = RESTART_TRANS;
4859 }
4860
4861bail:
4862 if (need_free) {
4863 if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
4864 ocfs2_free_local_alloc_bits(osb, handle, data_ac,
4865 bit_off, num_bits);
4866 else
4867 ocfs2_free_clusters(handle,
4868 data_ac->ac_inode,
4869 data_ac->ac_bh,
4870 ocfs2_clusters_to_blocks(osb->sb, bit_off),
4871 num_bits);
4872 }
4873
4874leave:
4875 if (reason_ret)
4876 *reason_ret = reason;
4877 trace_ocfs2_add_clusters_in_btree_ret(status, reason, err);
4878 return status;
4879}
4880
4881static void ocfs2_make_right_split_rec(struct super_block *sb,
4882 struct ocfs2_extent_rec *split_rec,
4883 u32 cpos,
4884 struct ocfs2_extent_rec *rec)
4885{
4886 u32 rec_cpos = le32_to_cpu(rec->e_cpos);
4887 u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
4888
4889 memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
4890
4891 split_rec->e_cpos = cpu_to_le32(cpos);
4892 split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
4893
4894 split_rec->e_blkno = rec->e_blkno;
4895 le64_add_cpu(&split_rec->e_blkno,
4896 ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
4897
4898 split_rec->e_flags = rec->e_flags;
4899}
4900
4901static int ocfs2_split_and_insert(handle_t *handle,
4902 struct ocfs2_extent_tree *et,
4903 struct ocfs2_path *path,
4904 struct buffer_head **last_eb_bh,
4905 int split_index,
4906 struct ocfs2_extent_rec *orig_split_rec,
4907 struct ocfs2_alloc_context *meta_ac)
4908{
4909 int ret = 0, depth;
4910 unsigned int insert_range, rec_range, do_leftright = 0;
4911 struct ocfs2_extent_rec tmprec;
4912 struct ocfs2_extent_list *rightmost_el;
4913 struct ocfs2_extent_rec rec;
4914 struct ocfs2_extent_rec split_rec = *orig_split_rec;
4915 struct ocfs2_insert_type insert;
4916 struct ocfs2_extent_block *eb;
4917
4918leftright:
4919 /*
4920 * Store a copy of the record on the stack - it might move
4921 * around as the tree is manipulated below.
4922 */
4923 rec = path_leaf_el(path)->l_recs[split_index];
4924
4925 rightmost_el = et->et_root_el;
4926
4927 depth = le16_to_cpu(rightmost_el->l_tree_depth);
4928 if (depth) {
4929 BUG_ON(!(*last_eb_bh));
4930 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
4931 rightmost_el = &eb->h_list;
4932 }
4933
4934 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4935 le16_to_cpu(rightmost_el->l_count)) {
4936 ret = ocfs2_grow_tree(handle, et,
4937 &depth, last_eb_bh, meta_ac);
4938 if (ret) {
4939 mlog_errno(ret);
4940 goto out;
4941 }
4942 }
4943
4944 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4945 insert.ins_appending = APPEND_NONE;
4946 insert.ins_contig = CONTIG_NONE;
4947 insert.ins_tree_depth = depth;
4948
4949 insert_range = le32_to_cpu(split_rec.e_cpos) +
4950 le16_to_cpu(split_rec.e_leaf_clusters);
4951 rec_range = le32_to_cpu(rec.e_cpos) +
4952 le16_to_cpu(rec.e_leaf_clusters);
4953
4954 if (split_rec.e_cpos == rec.e_cpos) {
4955 insert.ins_split = SPLIT_LEFT;
4956 } else if (insert_range == rec_range) {
4957 insert.ins_split = SPLIT_RIGHT;
4958 } else {
4959 /*
4960 * Left/right split. We fake this as a right split
4961 * first and then make a second pass as a left split.
4962 */
4963 insert.ins_split = SPLIT_RIGHT;
4964
4965 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4966 &tmprec, insert_range, &rec);
4967
4968 split_rec = tmprec;
4969
4970 BUG_ON(do_leftright);
4971 do_leftright = 1;
4972 }
4973
4974 ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
4975 if (ret) {
4976 mlog_errno(ret);
4977 goto out;
4978 }
4979
4980 if (do_leftright == 1) {
4981 u32 cpos;
4982 struct ocfs2_extent_list *el;
4983
4984 do_leftright++;
4985 split_rec = *orig_split_rec;
4986
4987 ocfs2_reinit_path(path, 1);
4988
4989 cpos = le32_to_cpu(split_rec.e_cpos);
4990 ret = ocfs2_find_path(et->et_ci, path, cpos);
4991 if (ret) {
4992 mlog_errno(ret);
4993 goto out;
4994 }
4995
4996 el = path_leaf_el(path);
4997 split_index = ocfs2_search_extent_list(el, cpos);
4998 if (split_index == -1) {
4999 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5000 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5001 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5002 cpos);
5003 ret = -EROFS;
5004 goto out;
5005 }
5006 goto leftright;
5007 }
5008out:
5009
5010 return ret;
5011}
5012
5013static int ocfs2_replace_extent_rec(handle_t *handle,
5014 struct ocfs2_extent_tree *et,
5015 struct ocfs2_path *path,
5016 struct ocfs2_extent_list *el,
5017 int split_index,
5018 struct ocfs2_extent_rec *split_rec)
5019{
5020 int ret;
5021
5022 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
5023 path_num_items(path) - 1);
5024 if (ret) {
5025 mlog_errno(ret);
5026 goto out;
5027 }
5028
5029 el->l_recs[split_index] = *split_rec;
5030
5031 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5032out:
5033 return ret;
5034}
5035
5036/*
5037 * Split part or all of the extent record at split_index in the leaf
5038 * pointed to by path. Merge with the contiguous extent record if needed.
5039 *
5040 * Care is taken to handle contiguousness so as to not grow the tree.
5041 *
5042 * meta_ac is not strictly necessary - we only truly need it if growth
5043 * of the tree is required. All other cases will degrade into a less
5044 * optimal tree layout.
5045 *
5046 * last_eb_bh should be the rightmost leaf block for any extent
5047 * btree. Since a split may grow the tree or a merge might shrink it,
5048 * the caller cannot trust the contents of that buffer after this call.
5049 *
5050 * This code is optimized for readability - several passes might be
5051 * made over certain portions of the tree. All of those blocks will
5052 * have been brought into cache (and pinned via the journal), so the
5053 * extra overhead is not expressed in terms of disk reads.
5054 */
5055int ocfs2_split_extent(handle_t *handle,
5056 struct ocfs2_extent_tree *et,
5057 struct ocfs2_path *path,
5058 int split_index,
5059 struct ocfs2_extent_rec *split_rec,
5060 struct ocfs2_alloc_context *meta_ac,
5061 struct ocfs2_cached_dealloc_ctxt *dealloc)
5062{
5063 int ret = 0;
5064 struct ocfs2_extent_list *el = path_leaf_el(path);
5065 struct buffer_head *last_eb_bh = NULL;
5066 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
5067 struct ocfs2_merge_ctxt ctxt;
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 ret = ocfs2_read_extent_block(et->et_ci,
5093 ocfs2_et_get_last_eb_blk(et),
5094 &last_eb_bh);
5095 if (ret) {
5096 mlog_errno(ret);
5097 goto out;
5098 }
5099 }
5100
5101 if (rec->e_cpos == split_rec->e_cpos &&
5102 rec->e_leaf_clusters == split_rec->e_leaf_clusters)
5103 ctxt.c_split_covers_rec = 1;
5104 else
5105 ctxt.c_split_covers_rec = 0;
5106
5107 ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
5108
5109 trace_ocfs2_split_extent(split_index, ctxt.c_contig_type,
5110 ctxt.c_has_empty_extent,
5111 ctxt.c_split_covers_rec);
5112
5113 if (ctxt.c_contig_type == CONTIG_NONE) {
5114 if (ctxt.c_split_covers_rec)
5115 ret = ocfs2_replace_extent_rec(handle, et, path, el,
5116 split_index, split_rec);
5117 else
5118 ret = ocfs2_split_and_insert(handle, et, path,
5119 &last_eb_bh, split_index,
5120 split_rec, meta_ac);
5121 if (ret)
5122 mlog_errno(ret);
5123 } else {
5124 ret = ocfs2_try_to_merge_extent(handle, et, path,
5125 split_index, split_rec,
5126 dealloc, &ctxt);
5127 if (ret)
5128 mlog_errno(ret);
5129 }
5130
5131out:
5132 brelse(last_eb_bh);
5133 return ret;
5134}
5135
5136/*
5137 * Change the flags of the already-existing extent at cpos for len clusters.
5138 *
5139 * new_flags: the flags we want to set.
5140 * clear_flags: the flags we want to clear.
5141 * phys: the new physical offset we want this new extent starts from.
5142 *
5143 * If the existing extent is larger than the request, initiate a
5144 * split. An attempt will be made at merging with adjacent extents.
5145 *
5146 * The caller is responsible for passing down meta_ac if we'll need it.
5147 */
5148int ocfs2_change_extent_flag(handle_t *handle,
5149 struct ocfs2_extent_tree *et,
5150 u32 cpos, u32 len, u32 phys,
5151 struct ocfs2_alloc_context *meta_ac,
5152 struct ocfs2_cached_dealloc_ctxt *dealloc,
5153 int new_flags, int clear_flags)
5154{
5155 int ret, index;
5156 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5157 u64 start_blkno = ocfs2_clusters_to_blocks(sb, phys);
5158 struct ocfs2_extent_rec split_rec;
5159 struct ocfs2_path *left_path = NULL;
5160 struct ocfs2_extent_list *el;
5161 struct ocfs2_extent_rec *rec;
5162
5163 left_path = ocfs2_new_path_from_et(et);
5164 if (!left_path) {
5165 ret = -ENOMEM;
5166 mlog_errno(ret);
5167 goto out;
5168 }
5169
5170 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
5171 if (ret) {
5172 mlog_errno(ret);
5173 goto out;
5174 }
5175 el = path_leaf_el(left_path);
5176
5177 index = ocfs2_search_extent_list(el, cpos);
5178 if (index == -1) {
5179 ocfs2_error(sb,
5180 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5181 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5182 cpos);
5183 ret = -EROFS;
5184 goto out;
5185 }
5186
5187 ret = -EIO;
5188 rec = &el->l_recs[index];
5189 if (new_flags && (rec->e_flags & new_flags)) {
5190 mlog(ML_ERROR, "Owner %llu tried to set %d flags on an "
5191 "extent that already had them\n",
5192 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5193 new_flags);
5194 goto out;
5195 }
5196
5197 if (clear_flags && !(rec->e_flags & clear_flags)) {
5198 mlog(ML_ERROR, "Owner %llu tried to clear %d flags on an "
5199 "extent that didn't have them\n",
5200 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5201 clear_flags);
5202 goto out;
5203 }
5204
5205 memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
5206 split_rec.e_cpos = cpu_to_le32(cpos);
5207 split_rec.e_leaf_clusters = cpu_to_le16(len);
5208 split_rec.e_blkno = cpu_to_le64(start_blkno);
5209 split_rec.e_flags = rec->e_flags;
5210 if (new_flags)
5211 split_rec.e_flags |= new_flags;
5212 if (clear_flags)
5213 split_rec.e_flags &= ~clear_flags;
5214
5215 ret = ocfs2_split_extent(handle, et, left_path,
5216 index, &split_rec, meta_ac,
5217 dealloc);
5218 if (ret)
5219 mlog_errno(ret);
5220
5221out:
5222 ocfs2_free_path(left_path);
5223 return ret;
5224
5225}
5226
5227/*
5228 * Mark the already-existing extent at cpos as written for len clusters.
5229 * This removes the unwritten extent flag.
5230 *
5231 * If the existing extent is larger than the request, initiate a
5232 * split. An attempt will be made at merging with adjacent extents.
5233 *
5234 * The caller is responsible for passing down meta_ac if we'll need it.
5235 */
5236int ocfs2_mark_extent_written(struct inode *inode,
5237 struct ocfs2_extent_tree *et,
5238 handle_t *handle, u32 cpos, u32 len, u32 phys,
5239 struct ocfs2_alloc_context *meta_ac,
5240 struct ocfs2_cached_dealloc_ctxt *dealloc)
5241{
5242 int ret;
5243
5244 trace_ocfs2_mark_extent_written(
5245 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5246 cpos, len, phys);
5247
5248 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
5249 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",
5250 (unsigned long long)OCFS2_I(inode)->ip_blkno);
5251 ret = -EROFS;
5252 goto out;
5253 }
5254
5255 /*
5256 * XXX: This should be fixed up so that we just re-insert the
5257 * next extent records.
5258 */
5259 ocfs2_et_extent_map_truncate(et, 0);
5260
5261 ret = ocfs2_change_extent_flag(handle, et, cpos,
5262 len, phys, meta_ac, dealloc,
5263 0, OCFS2_EXT_UNWRITTEN);
5264 if (ret)
5265 mlog_errno(ret);
5266
5267out:
5268 return ret;
5269}
5270
5271static int ocfs2_split_tree(handle_t *handle, struct ocfs2_extent_tree *et,
5272 struct ocfs2_path *path,
5273 int index, u32 new_range,
5274 struct ocfs2_alloc_context *meta_ac)
5275{
5276 int ret, depth, credits;
5277 struct buffer_head *last_eb_bh = NULL;
5278 struct ocfs2_extent_block *eb;
5279 struct ocfs2_extent_list *rightmost_el, *el;
5280 struct ocfs2_extent_rec split_rec;
5281 struct ocfs2_extent_rec *rec;
5282 struct ocfs2_insert_type insert;
5283
5284 /*
5285 * Setup the record to split before we grow the tree.
5286 */
5287 el = path_leaf_el(path);
5288 rec = &el->l_recs[index];
5289 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
5290 &split_rec, new_range, rec);
5291
5292 depth = path->p_tree_depth;
5293 if (depth > 0) {
5294 ret = ocfs2_read_extent_block(et->et_ci,
5295 ocfs2_et_get_last_eb_blk(et),
5296 &last_eb_bh);
5297 if (ret < 0) {
5298 mlog_errno(ret);
5299 goto out;
5300 }
5301
5302 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5303 rightmost_el = &eb->h_list;
5304 } else
5305 rightmost_el = path_leaf_el(path);
5306
5307 credits = path->p_tree_depth +
5308 ocfs2_extend_meta_needed(et->et_root_el);
5309 ret = ocfs2_extend_trans(handle, credits);
5310 if (ret) {
5311 mlog_errno(ret);
5312 goto out;
5313 }
5314
5315 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
5316 le16_to_cpu(rightmost_el->l_count)) {
5317 ret = ocfs2_grow_tree(handle, et, &depth, &last_eb_bh,
5318 meta_ac);
5319 if (ret) {
5320 mlog_errno(ret);
5321 goto out;
5322 }
5323 }
5324
5325 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
5326 insert.ins_appending = APPEND_NONE;
5327 insert.ins_contig = CONTIG_NONE;
5328 insert.ins_split = SPLIT_RIGHT;
5329 insert.ins_tree_depth = depth;
5330
5331 ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
5332 if (ret)
5333 mlog_errno(ret);
5334
5335out:
5336 brelse(last_eb_bh);
5337 return ret;
5338}
5339
5340static int ocfs2_truncate_rec(handle_t *handle,
5341 struct ocfs2_extent_tree *et,
5342 struct ocfs2_path *path, int index,
5343 struct ocfs2_cached_dealloc_ctxt *dealloc,
5344 u32 cpos, u32 len)
5345{
5346 int ret;
5347 u32 left_cpos, rec_range, trunc_range;
5348 int is_rightmost_tree_rec = 0;
5349 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5350 struct ocfs2_path *left_path = NULL;
5351 struct ocfs2_extent_list *el = path_leaf_el(path);
5352 struct ocfs2_extent_rec *rec;
5353 struct ocfs2_extent_block *eb;
5354
5355 if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
5356 /* extend credit for ocfs2_remove_rightmost_path */
5357 ret = ocfs2_extend_rotate_transaction(handle, 0,
5358 jbd2_handle_buffer_credits(handle),
5359 path);
5360 if (ret) {
5361 mlog_errno(ret);
5362 goto out;
5363 }
5364
5365 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5366 if (ret) {
5367 mlog_errno(ret);
5368 goto out;
5369 }
5370
5371 index--;
5372 }
5373
5374 if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
5375 path->p_tree_depth) {
5376 /*
5377 * Check whether this is the rightmost tree record. If
5378 * we remove all of this record or part of its right
5379 * edge then an update of the record lengths above it
5380 * will be required.
5381 */
5382 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
5383 if (eb->h_next_leaf_blk == 0)
5384 is_rightmost_tree_rec = 1;
5385 }
5386
5387 rec = &el->l_recs[index];
5388 if (index == 0 && path->p_tree_depth &&
5389 le32_to_cpu(rec->e_cpos) == cpos) {
5390 /*
5391 * Changing the leftmost offset (via partial or whole
5392 * record truncate) of an interior (or rightmost) path
5393 * means we have to update the subtree that is formed
5394 * by this leaf and the one to it's left.
5395 *
5396 * There are two cases we can skip:
5397 * 1) Path is the leftmost one in our btree.
5398 * 2) The leaf is rightmost and will be empty after
5399 * we remove the extent record - the rotate code
5400 * knows how to update the newly formed edge.
5401 */
5402
5403 ret = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
5404 if (ret) {
5405 mlog_errno(ret);
5406 goto out;
5407 }
5408
5409 if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
5410 left_path = ocfs2_new_path_from_path(path);
5411 if (!left_path) {
5412 ret = -ENOMEM;
5413 mlog_errno(ret);
5414 goto out;
5415 }
5416
5417 ret = ocfs2_find_path(et->et_ci, left_path,
5418 left_cpos);
5419 if (ret) {
5420 mlog_errno(ret);
5421 goto out;
5422 }
5423 }
5424 }
5425
5426 ret = ocfs2_extend_rotate_transaction(handle, 0,
5427 jbd2_handle_buffer_credits(handle),
5428 path);
5429 if (ret) {
5430 mlog_errno(ret);
5431 goto out;
5432 }
5433
5434 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
5435 if (ret) {
5436 mlog_errno(ret);
5437 goto out;
5438 }
5439
5440 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
5441 if (ret) {
5442 mlog_errno(ret);
5443 goto out;
5444 }
5445
5446 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5447 trunc_range = cpos + len;
5448
5449 if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
5450 int next_free;
5451
5452 memset(rec, 0, sizeof(*rec));
5453 ocfs2_cleanup_merge(el, index);
5454
5455 next_free = le16_to_cpu(el->l_next_free_rec);
5456 if (is_rightmost_tree_rec && next_free > 1) {
5457 /*
5458 * We skip the edge update if this path will
5459 * be deleted by the rotate code.
5460 */
5461 rec = &el->l_recs[next_free - 1];
5462 ocfs2_adjust_rightmost_records(handle, et, path,
5463 rec);
5464 }
5465 } else if (le32_to_cpu(rec->e_cpos) == cpos) {
5466 /* Remove leftmost portion of the record. */
5467 le32_add_cpu(&rec->e_cpos, len);
5468 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
5469 le16_add_cpu(&rec->e_leaf_clusters, -len);
5470 } else if (rec_range == trunc_range) {
5471 /* Remove rightmost portion of the record */
5472 le16_add_cpu(&rec->e_leaf_clusters, -len);
5473 if (is_rightmost_tree_rec)
5474 ocfs2_adjust_rightmost_records(handle, et, path, rec);
5475 } else {
5476 /* Caller should have trapped this. */
5477 mlog(ML_ERROR, "Owner %llu: Invalid record truncate: (%u, %u) "
5478 "(%u, %u)\n",
5479 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5480 le32_to_cpu(rec->e_cpos),
5481 le16_to_cpu(rec->e_leaf_clusters), cpos, len);
5482 BUG();
5483 }
5484
5485 if (left_path) {
5486 int subtree_index;
5487
5488 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
5489 ocfs2_complete_edge_insert(handle, left_path, path,
5490 subtree_index);
5491 }
5492
5493 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5494
5495 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5496 if (ret)
5497 mlog_errno(ret);
5498
5499out:
5500 ocfs2_free_path(left_path);
5501 return ret;
5502}
5503
5504int ocfs2_remove_extent(handle_t *handle,
5505 struct ocfs2_extent_tree *et,
5506 u32 cpos, u32 len,
5507 struct ocfs2_alloc_context *meta_ac,
5508 struct ocfs2_cached_dealloc_ctxt *dealloc)
5509{
5510 int ret, index;
5511 u32 rec_range, trunc_range;
5512 struct ocfs2_extent_rec *rec;
5513 struct ocfs2_extent_list *el;
5514 struct ocfs2_path *path = NULL;
5515
5516 /*
5517 * XXX: Why are we truncating to 0 instead of wherever this
5518 * affects us?
5519 */
5520 ocfs2_et_extent_map_truncate(et, 0);
5521
5522 path = ocfs2_new_path_from_et(et);
5523 if (!path) {
5524 ret = -ENOMEM;
5525 mlog_errno(ret);
5526 goto out;
5527 }
5528
5529 ret = ocfs2_find_path(et->et_ci, path, cpos);
5530 if (ret) {
5531 mlog_errno(ret);
5532 goto out;
5533 }
5534
5535 el = path_leaf_el(path);
5536 index = ocfs2_search_extent_list(el, cpos);
5537 if (index == -1) {
5538 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5539 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5540 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5541 cpos);
5542 ret = -EROFS;
5543 goto out;
5544 }
5545
5546 /*
5547 * We have 3 cases of extent removal:
5548 * 1) Range covers the entire extent rec
5549 * 2) Range begins or ends on one edge of the extent rec
5550 * 3) Range is in the middle of the extent rec (no shared edges)
5551 *
5552 * For case 1 we remove the extent rec and left rotate to
5553 * fill the hole.
5554 *
5555 * For case 2 we just shrink the existing extent rec, with a
5556 * tree update if the shrinking edge is also the edge of an
5557 * extent block.
5558 *
5559 * For case 3 we do a right split to turn the extent rec into
5560 * something case 2 can handle.
5561 */
5562 rec = &el->l_recs[index];
5563 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5564 trunc_range = cpos + len;
5565
5566 BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
5567
5568 trace_ocfs2_remove_extent(
5569 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5570 cpos, len, index, le32_to_cpu(rec->e_cpos),
5571 ocfs2_rec_clusters(el, rec));
5572
5573 if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
5574 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5575 cpos, len);
5576 if (ret) {
5577 mlog_errno(ret);
5578 goto out;
5579 }
5580 } else {
5581 ret = ocfs2_split_tree(handle, et, path, index,
5582 trunc_range, meta_ac);
5583 if (ret) {
5584 mlog_errno(ret);
5585 goto out;
5586 }
5587
5588 /*
5589 * The split could have manipulated the tree enough to
5590 * move the record location, so we have to look for it again.
5591 */
5592 ocfs2_reinit_path(path, 1);
5593
5594 ret = ocfs2_find_path(et->et_ci, path, cpos);
5595 if (ret) {
5596 mlog_errno(ret);
5597 goto out;
5598 }
5599
5600 el = path_leaf_el(path);
5601 index = ocfs2_search_extent_list(el, cpos);
5602 if (index == -1) {
5603 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5604 "Owner %llu: split at cpos %u lost record\n",
5605 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5606 cpos);
5607 ret = -EROFS;
5608 goto out;
5609 }
5610
5611 /*
5612 * Double check our values here. If anything is fishy,
5613 * it's easier to catch it at the top level.
5614 */
5615 rec = &el->l_recs[index];
5616 rec_range = le32_to_cpu(rec->e_cpos) +
5617 ocfs2_rec_clusters(el, rec);
5618 if (rec_range != trunc_range) {
5619 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5620 "Owner %llu: error after split at cpos %u trunc len %u, existing record is (%u,%u)\n",
5621 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5622 cpos, len, le32_to_cpu(rec->e_cpos),
5623 ocfs2_rec_clusters(el, rec));
5624 ret = -EROFS;
5625 goto out;
5626 }
5627
5628 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5629 cpos, len);
5630 if (ret)
5631 mlog_errno(ret);
5632 }
5633
5634out:
5635 ocfs2_free_path(path);
5636 return ret;
5637}
5638
5639/*
5640 * ocfs2_reserve_blocks_for_rec_trunc() would look basically the
5641 * same as ocfs2_lock_alloctors(), except for it accepts a blocks
5642 * number to reserve some extra blocks, and it only handles meta
5643 * data allocations.
5644 *
5645 * Currently, only ocfs2_remove_btree_range() uses it for truncating
5646 * and punching holes.
5647 */
5648static int ocfs2_reserve_blocks_for_rec_trunc(struct inode *inode,
5649 struct ocfs2_extent_tree *et,
5650 u32 extents_to_split,
5651 struct ocfs2_alloc_context **ac,
5652 int extra_blocks)
5653{
5654 int ret = 0, num_free_extents;
5655 unsigned int max_recs_needed = 2 * extents_to_split;
5656 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5657
5658 *ac = NULL;
5659
5660 num_free_extents = ocfs2_num_free_extents(et);
5661 if (num_free_extents < 0) {
5662 ret = num_free_extents;
5663 mlog_errno(ret);
5664 goto out;
5665 }
5666
5667 if (!num_free_extents ||
5668 (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
5669 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
5670
5671 if (extra_blocks) {
5672 ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, ac);
5673 if (ret < 0) {
5674 if (ret != -ENOSPC)
5675 mlog_errno(ret);
5676 }
5677 }
5678
5679out:
5680 if (ret) {
5681 if (*ac) {
5682 ocfs2_free_alloc_context(*ac);
5683 *ac = NULL;
5684 }
5685 }
5686
5687 return ret;
5688}
5689
5690int ocfs2_remove_btree_range(struct inode *inode,
5691 struct ocfs2_extent_tree *et,
5692 u32 cpos, u32 phys_cpos, u32 len, int flags,
5693 struct ocfs2_cached_dealloc_ctxt *dealloc,
5694 u64 refcount_loc, bool refcount_tree_locked)
5695{
5696 int ret, credits = 0, extra_blocks = 0;
5697 u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
5698 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5699 struct inode *tl_inode = osb->osb_tl_inode;
5700 handle_t *handle;
5701 struct ocfs2_alloc_context *meta_ac = NULL;
5702 struct ocfs2_refcount_tree *ref_tree = NULL;
5703
5704 if ((flags & OCFS2_EXT_REFCOUNTED) && len) {
5705 BUG_ON(!ocfs2_is_refcount_inode(inode));
5706
5707 if (!refcount_tree_locked) {
5708 ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
5709 &ref_tree, NULL);
5710 if (ret) {
5711 mlog_errno(ret);
5712 goto bail;
5713 }
5714 }
5715
5716 ret = ocfs2_prepare_refcount_change_for_del(inode,
5717 refcount_loc,
5718 phys_blkno,
5719 len,
5720 &credits,
5721 &extra_blocks);
5722 if (ret < 0) {
5723 mlog_errno(ret);
5724 goto bail;
5725 }
5726 }
5727
5728 ret = ocfs2_reserve_blocks_for_rec_trunc(inode, et, 1, &meta_ac,
5729 extra_blocks);
5730 if (ret) {
5731 mlog_errno(ret);
5732 goto bail;
5733 }
5734
5735 inode_lock(tl_inode);
5736
5737 if (ocfs2_truncate_log_needs_flush(osb)) {
5738 ret = __ocfs2_flush_truncate_log(osb);
5739 if (ret < 0) {
5740 mlog_errno(ret);
5741 goto out;
5742 }
5743 }
5744
5745 handle = ocfs2_start_trans(osb,
5746 ocfs2_remove_extent_credits(osb->sb) + credits);
5747 if (IS_ERR(handle)) {
5748 ret = PTR_ERR(handle);
5749 mlog_errno(ret);
5750 goto out;
5751 }
5752
5753 ret = ocfs2_et_root_journal_access(handle, et,
5754 OCFS2_JOURNAL_ACCESS_WRITE);
5755 if (ret) {
5756 mlog_errno(ret);
5757 goto out_commit;
5758 }
5759
5760 dquot_free_space_nodirty(inode,
5761 ocfs2_clusters_to_bytes(inode->i_sb, len));
5762
5763 ret = ocfs2_remove_extent(handle, et, cpos, len, meta_ac, dealloc);
5764 if (ret) {
5765 mlog_errno(ret);
5766 goto out_commit;
5767 }
5768
5769 ocfs2_et_update_clusters(et, -len);
5770 ocfs2_update_inode_fsync_trans(handle, inode, 1);
5771
5772 ocfs2_journal_dirty(handle, et->et_root_bh);
5773
5774 if (phys_blkno) {
5775 if (flags & OCFS2_EXT_REFCOUNTED)
5776 ret = ocfs2_decrease_refcount(inode, handle,
5777 ocfs2_blocks_to_clusters(osb->sb,
5778 phys_blkno),
5779 len, meta_ac,
5780 dealloc, 1);
5781 else
5782 ret = ocfs2_truncate_log_append(osb, handle,
5783 phys_blkno, len);
5784 if (ret)
5785 mlog_errno(ret);
5786
5787 }
5788
5789out_commit:
5790 ocfs2_commit_trans(osb, handle);
5791out:
5792 inode_unlock(tl_inode);
5793bail:
5794 if (meta_ac)
5795 ocfs2_free_alloc_context(meta_ac);
5796
5797 if (ref_tree)
5798 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
5799
5800 return ret;
5801}
5802
5803int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
5804{
5805 struct buffer_head *tl_bh = osb->osb_tl_bh;
5806 struct ocfs2_dinode *di;
5807 struct ocfs2_truncate_log *tl;
5808
5809 di = (struct ocfs2_dinode *) tl_bh->b_data;
5810 tl = &di->id2.i_dealloc;
5811
5812 mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
5813 "slot %d, invalid truncate log parameters: used = "
5814 "%u, count = %u\n", osb->slot_num,
5815 le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
5816 return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
5817}
5818
5819static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
5820 unsigned int new_start)
5821{
5822 unsigned int tail_index;
5823 unsigned int current_tail;
5824
5825 /* No records, nothing to coalesce */
5826 if (!le16_to_cpu(tl->tl_used))
5827 return 0;
5828
5829 tail_index = le16_to_cpu(tl->tl_used) - 1;
5830 current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
5831 current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
5832
5833 return current_tail == new_start;
5834}
5835
5836int ocfs2_truncate_log_append(struct ocfs2_super *osb,
5837 handle_t *handle,
5838 u64 start_blk,
5839 unsigned int num_clusters)
5840{
5841 int status, index;
5842 unsigned int start_cluster, tl_count;
5843 struct inode *tl_inode = osb->osb_tl_inode;
5844 struct buffer_head *tl_bh = osb->osb_tl_bh;
5845 struct ocfs2_dinode *di;
5846 struct ocfs2_truncate_log *tl;
5847
5848 BUG_ON(inode_trylock(tl_inode));
5849
5850 start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
5851
5852 di = (struct ocfs2_dinode *) tl_bh->b_data;
5853
5854 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5855 * by the underlying call to ocfs2_read_inode_block(), so any
5856 * corruption is a code bug */
5857 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5858
5859 tl = &di->id2.i_dealloc;
5860 tl_count = le16_to_cpu(tl->tl_count);
5861 mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
5862 tl_count == 0,
5863 "Truncate record count on #%llu invalid "
5864 "wanted %u, actual %u\n",
5865 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5866 ocfs2_truncate_recs_per_inode(osb->sb),
5867 le16_to_cpu(tl->tl_count));
5868
5869 /* Caller should have known to flush before calling us. */
5870 index = le16_to_cpu(tl->tl_used);
5871 if (index >= tl_count) {
5872 status = -ENOSPC;
5873 mlog_errno(status);
5874 goto bail;
5875 }
5876
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 trace_ocfs2_truncate_log_append(
5885 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index,
5886 start_cluster, num_clusters);
5887 if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
5888 /*
5889 * Move index back to the record we are coalescing with.
5890 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5891 */
5892 index--;
5893
5894 num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
5895 trace_ocfs2_truncate_log_append(
5896 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5897 index, le32_to_cpu(tl->tl_recs[index].t_start),
5898 num_clusters);
5899 } else {
5900 tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
5901 tl->tl_used = cpu_to_le16(index + 1);
5902 }
5903 tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
5904
5905 ocfs2_journal_dirty(handle, tl_bh);
5906
5907 osb->truncated_clusters += num_clusters;
5908bail:
5909 return status;
5910}
5911
5912static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
5913 struct inode *data_alloc_inode,
5914 struct buffer_head *data_alloc_bh)
5915{
5916 int status = 0;
5917 int i;
5918 unsigned int num_clusters;
5919 u64 start_blk;
5920 struct ocfs2_truncate_rec rec;
5921 struct ocfs2_dinode *di;
5922 struct ocfs2_truncate_log *tl;
5923 struct inode *tl_inode = osb->osb_tl_inode;
5924 struct buffer_head *tl_bh = osb->osb_tl_bh;
5925 handle_t *handle;
5926
5927 di = (struct ocfs2_dinode *) tl_bh->b_data;
5928 tl = &di->id2.i_dealloc;
5929 i = le16_to_cpu(tl->tl_used) - 1;
5930 while (i >= 0) {
5931 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
5932 if (IS_ERR(handle)) {
5933 status = PTR_ERR(handle);
5934 mlog_errno(status);
5935 goto bail;
5936 }
5937
5938 /* Caller has given us at least enough credits to
5939 * update the truncate log dinode */
5940 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5941 OCFS2_JOURNAL_ACCESS_WRITE);
5942 if (status < 0) {
5943 mlog_errno(status);
5944 goto bail;
5945 }
5946
5947 tl->tl_used = cpu_to_le16(i);
5948
5949 ocfs2_journal_dirty(handle, tl_bh);
5950
5951 rec = tl->tl_recs[i];
5952 start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
5953 le32_to_cpu(rec.t_start));
5954 num_clusters = le32_to_cpu(rec.t_clusters);
5955
5956 /* if start_blk is not set, we ignore the record as
5957 * invalid. */
5958 if (start_blk) {
5959 trace_ocfs2_replay_truncate_records(
5960 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5961 i, le32_to_cpu(rec.t_start), num_clusters);
5962
5963 status = ocfs2_free_clusters(handle, data_alloc_inode,
5964 data_alloc_bh, start_blk,
5965 num_clusters);
5966 if (status < 0) {
5967 mlog_errno(status);
5968 goto bail;
5969 }
5970 }
5971
5972 ocfs2_commit_trans(osb, handle);
5973 i--;
5974 }
5975
5976 osb->truncated_clusters = 0;
5977
5978bail:
5979 return status;
5980}
5981
5982/* Expects you to already be holding tl_inode->i_mutex */
5983int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5984{
5985 int status;
5986 unsigned int num_to_flush;
5987 struct inode *tl_inode = osb->osb_tl_inode;
5988 struct inode *data_alloc_inode = NULL;
5989 struct buffer_head *tl_bh = osb->osb_tl_bh;
5990 struct buffer_head *data_alloc_bh = NULL;
5991 struct ocfs2_dinode *di;
5992 struct ocfs2_truncate_log *tl;
5993 struct ocfs2_journal *journal = osb->journal;
5994
5995 BUG_ON(inode_trylock(tl_inode));
5996
5997 di = (struct ocfs2_dinode *) tl_bh->b_data;
5998
5999 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
6000 * by the underlying call to ocfs2_read_inode_block(), so any
6001 * corruption is a code bug */
6002 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6003
6004 tl = &di->id2.i_dealloc;
6005 num_to_flush = le16_to_cpu(tl->tl_used);
6006 trace_ocfs2_flush_truncate_log(
6007 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
6008 num_to_flush);
6009 if (!num_to_flush) {
6010 status = 0;
6011 goto out;
6012 }
6013
6014 /* Appending truncate log(TA) and flushing truncate log(TF) are
6015 * two separated transactions. They can be both committed but not
6016 * checkpointed. If crash occurs then, both two transaction will be
6017 * replayed with several already released to global bitmap clusters.
6018 * Then truncate log will be replayed resulting in cluster double free.
6019 */
6020 jbd2_journal_lock_updates(journal->j_journal);
6021 status = jbd2_journal_flush(journal->j_journal, 0);
6022 jbd2_journal_unlock_updates(journal->j_journal);
6023 if (status < 0) {
6024 mlog_errno(status);
6025 goto out;
6026 }
6027
6028 data_alloc_inode = ocfs2_get_system_file_inode(osb,
6029 GLOBAL_BITMAP_SYSTEM_INODE,
6030 OCFS2_INVALID_SLOT);
6031 if (!data_alloc_inode) {
6032 status = -EINVAL;
6033 mlog(ML_ERROR, "Could not get bitmap inode!\n");
6034 goto out;
6035 }
6036
6037 inode_lock(data_alloc_inode);
6038
6039 status = ocfs2_inode_lock(data_alloc_inode, &data_alloc_bh, 1);
6040 if (status < 0) {
6041 mlog_errno(status);
6042 goto out_mutex;
6043 }
6044
6045 status = ocfs2_replay_truncate_records(osb, data_alloc_inode,
6046 data_alloc_bh);
6047 if (status < 0)
6048 mlog_errno(status);
6049
6050 brelse(data_alloc_bh);
6051 ocfs2_inode_unlock(data_alloc_inode, 1);
6052
6053out_mutex:
6054 inode_unlock(data_alloc_inode);
6055 iput(data_alloc_inode);
6056
6057out:
6058 return status;
6059}
6060
6061int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
6062{
6063 int status;
6064 struct inode *tl_inode = osb->osb_tl_inode;
6065
6066 inode_lock(tl_inode);
6067 status = __ocfs2_flush_truncate_log(osb);
6068 inode_unlock(tl_inode);
6069
6070 return status;
6071}
6072
6073static void ocfs2_truncate_log_worker(struct work_struct *work)
6074{
6075 int status;
6076 struct ocfs2_super *osb =
6077 container_of(work, struct ocfs2_super,
6078 osb_truncate_log_wq.work);
6079
6080 status = ocfs2_flush_truncate_log(osb);
6081 if (status < 0)
6082 mlog_errno(status);
6083 else
6084 ocfs2_init_steal_slots(osb);
6085}
6086
6087#define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
6088void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
6089 int cancel)
6090{
6091 if (osb->osb_tl_inode &&
6092 atomic_read(&osb->osb_tl_disable) == 0) {
6093 /* We want to push off log flushes while truncates are
6094 * still running. */
6095 if (cancel)
6096 cancel_delayed_work(&osb->osb_truncate_log_wq);
6097
6098 queue_delayed_work(osb->ocfs2_wq, &osb->osb_truncate_log_wq,
6099 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
6100 }
6101}
6102
6103/*
6104 * Try to flush truncate logs if we can free enough clusters from it.
6105 * As for return value, "< 0" means error, "0" no space and "1" means
6106 * we have freed enough spaces and let the caller try to allocate again.
6107 */
6108int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
6109 unsigned int needed)
6110{
6111 tid_t target;
6112 int ret = 0;
6113 unsigned int truncated_clusters;
6114
6115 inode_lock(osb->osb_tl_inode);
6116 truncated_clusters = osb->truncated_clusters;
6117 inode_unlock(osb->osb_tl_inode);
6118
6119 /*
6120 * Check whether we can succeed in allocating if we free
6121 * the truncate log.
6122 */
6123 if (truncated_clusters < needed)
6124 goto out;
6125
6126 ret = ocfs2_flush_truncate_log(osb);
6127 if (ret) {
6128 mlog_errno(ret);
6129 goto out;
6130 }
6131
6132 if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
6133 jbd2_log_wait_commit(osb->journal->j_journal, target);
6134 ret = 1;
6135 }
6136out:
6137 return ret;
6138}
6139
6140static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
6141 int slot_num,
6142 struct inode **tl_inode,
6143 struct buffer_head **tl_bh)
6144{
6145 int status;
6146 struct inode *inode = NULL;
6147 struct buffer_head *bh = NULL;
6148
6149 inode = ocfs2_get_system_file_inode(osb,
6150 TRUNCATE_LOG_SYSTEM_INODE,
6151 slot_num);
6152 if (!inode) {
6153 status = -EINVAL;
6154 mlog(ML_ERROR, "Could not get load truncate log inode!\n");
6155 goto bail;
6156 }
6157
6158 status = ocfs2_read_inode_block(inode, &bh);
6159 if (status < 0) {
6160 iput(inode);
6161 mlog_errno(status);
6162 goto bail;
6163 }
6164
6165 *tl_inode = inode;
6166 *tl_bh = bh;
6167bail:
6168 return status;
6169}
6170
6171/* called during the 1st stage of node recovery. we stamp a clean
6172 * truncate log and pass back a copy for processing later. if the
6173 * truncate log does not require processing, a *tl_copy is set to
6174 * NULL. */
6175int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
6176 int slot_num,
6177 struct ocfs2_dinode **tl_copy)
6178{
6179 int status;
6180 struct inode *tl_inode = NULL;
6181 struct buffer_head *tl_bh = NULL;
6182 struct ocfs2_dinode *di;
6183 struct ocfs2_truncate_log *tl;
6184
6185 *tl_copy = NULL;
6186
6187 trace_ocfs2_begin_truncate_log_recovery(slot_num);
6188
6189 status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
6190 if (status < 0) {
6191 mlog_errno(status);
6192 goto bail;
6193 }
6194
6195 di = (struct ocfs2_dinode *) tl_bh->b_data;
6196
6197 /* tl_bh is loaded from ocfs2_get_truncate_log_info(). It's
6198 * validated by the underlying call to ocfs2_read_inode_block(),
6199 * so any corruption is a code bug */
6200 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6201
6202 tl = &di->id2.i_dealloc;
6203 if (le16_to_cpu(tl->tl_used)) {
6204 trace_ocfs2_truncate_log_recovery_num(le16_to_cpu(tl->tl_used));
6205
6206 /*
6207 * Assuming the write-out below goes well, this copy will be
6208 * passed back to recovery for processing.
6209 */
6210 *tl_copy = kmemdup(tl_bh->b_data, tl_bh->b_size, GFP_KERNEL);
6211 if (!(*tl_copy)) {
6212 status = -ENOMEM;
6213 mlog_errno(status);
6214 goto bail;
6215 }
6216
6217 /* All we need to do to clear the truncate log is set
6218 * tl_used. */
6219 tl->tl_used = 0;
6220
6221 ocfs2_compute_meta_ecc(osb->sb, tl_bh->b_data, &di->i_check);
6222 status = ocfs2_write_block(osb, tl_bh, INODE_CACHE(tl_inode));
6223 if (status < 0) {
6224 mlog_errno(status);
6225 goto bail;
6226 }
6227 }
6228
6229bail:
6230 iput(tl_inode);
6231 brelse(tl_bh);
6232
6233 if (status < 0) {
6234 kfree(*tl_copy);
6235 *tl_copy = NULL;
6236 mlog_errno(status);
6237 }
6238
6239 return status;
6240}
6241
6242int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
6243 struct ocfs2_dinode *tl_copy)
6244{
6245 int status = 0;
6246 int i;
6247 unsigned int clusters, num_recs, start_cluster;
6248 u64 start_blk;
6249 handle_t *handle;
6250 struct inode *tl_inode = osb->osb_tl_inode;
6251 struct ocfs2_truncate_log *tl;
6252
6253 if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
6254 mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
6255 return -EINVAL;
6256 }
6257
6258 tl = &tl_copy->id2.i_dealloc;
6259 num_recs = le16_to_cpu(tl->tl_used);
6260 trace_ocfs2_complete_truncate_log_recovery(
6261 (unsigned long long)le64_to_cpu(tl_copy->i_blkno),
6262 num_recs);
6263
6264 inode_lock(tl_inode);
6265 for(i = 0; i < num_recs; i++) {
6266 if (ocfs2_truncate_log_needs_flush(osb)) {
6267 status = __ocfs2_flush_truncate_log(osb);
6268 if (status < 0) {
6269 mlog_errno(status);
6270 goto bail_up;
6271 }
6272 }
6273
6274 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6275 if (IS_ERR(handle)) {
6276 status = PTR_ERR(handle);
6277 mlog_errno(status);
6278 goto bail_up;
6279 }
6280
6281 clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
6282 start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
6283 start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
6284
6285 status = ocfs2_truncate_log_append(osb, handle,
6286 start_blk, clusters);
6287 ocfs2_commit_trans(osb, handle);
6288 if (status < 0) {
6289 mlog_errno(status);
6290 goto bail_up;
6291 }
6292 }
6293
6294bail_up:
6295 inode_unlock(tl_inode);
6296
6297 return status;
6298}
6299
6300void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
6301{
6302 int status;
6303 struct inode *tl_inode = osb->osb_tl_inode;
6304
6305 atomic_set(&osb->osb_tl_disable, 1);
6306
6307 if (tl_inode) {
6308 cancel_delayed_work(&osb->osb_truncate_log_wq);
6309 flush_workqueue(osb->ocfs2_wq);
6310
6311 status = ocfs2_flush_truncate_log(osb);
6312 if (status < 0)
6313 mlog_errno(status);
6314
6315 brelse(osb->osb_tl_bh);
6316 iput(osb->osb_tl_inode);
6317 }
6318}
6319
6320int ocfs2_truncate_log_init(struct ocfs2_super *osb)
6321{
6322 int status;
6323 struct inode *tl_inode = NULL;
6324 struct buffer_head *tl_bh = NULL;
6325
6326 status = ocfs2_get_truncate_log_info(osb,
6327 osb->slot_num,
6328 &tl_inode,
6329 &tl_bh);
6330 if (status < 0)
6331 mlog_errno(status);
6332
6333 /* ocfs2_truncate_log_shutdown keys on the existence of
6334 * osb->osb_tl_inode so we don't set any of the osb variables
6335 * until we're sure all is well. */
6336 INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
6337 ocfs2_truncate_log_worker);
6338 atomic_set(&osb->osb_tl_disable, 0);
6339 osb->osb_tl_bh = tl_bh;
6340 osb->osb_tl_inode = tl_inode;
6341
6342 return status;
6343}
6344
6345/*
6346 * Delayed de-allocation of suballocator blocks.
6347 *
6348 * Some sets of block de-allocations might involve multiple suballocator inodes.
6349 *
6350 * The locking for this can get extremely complicated, especially when
6351 * the suballocator inodes to delete from aren't known until deep
6352 * within an unrelated codepath.
6353 *
6354 * ocfs2_extent_block structures are a good example of this - an inode
6355 * btree could have been grown by any number of nodes each allocating
6356 * out of their own suballoc inode.
6357 *
6358 * These structures allow the delay of block de-allocation until a
6359 * later time, when locking of multiple cluster inodes won't cause
6360 * deadlock.
6361 */
6362
6363/*
6364 * Describe a single bit freed from a suballocator. For the block
6365 * suballocators, it represents one block. For the global cluster
6366 * allocator, it represents some clusters and free_bit indicates
6367 * clusters number.
6368 */
6369struct ocfs2_cached_block_free {
6370 struct ocfs2_cached_block_free *free_next;
6371 u64 free_bg;
6372 u64 free_blk;
6373 unsigned int free_bit;
6374};
6375
6376struct ocfs2_per_slot_free_list {
6377 struct ocfs2_per_slot_free_list *f_next_suballocator;
6378 int f_inode_type;
6379 int f_slot;
6380 struct ocfs2_cached_block_free *f_first;
6381};
6382
6383static int ocfs2_free_cached_blocks(struct ocfs2_super *osb,
6384 int sysfile_type,
6385 int slot,
6386 struct ocfs2_cached_block_free *head)
6387{
6388 int ret;
6389 u64 bg_blkno;
6390 handle_t *handle;
6391 struct inode *inode;
6392 struct buffer_head *di_bh = NULL;
6393 struct ocfs2_cached_block_free *tmp;
6394
6395 inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
6396 if (!inode) {
6397 ret = -EINVAL;
6398 mlog_errno(ret);
6399 goto out;
6400 }
6401
6402 inode_lock(inode);
6403
6404 ret = ocfs2_inode_lock(inode, &di_bh, 1);
6405 if (ret) {
6406 mlog_errno(ret);
6407 goto out_mutex;
6408 }
6409
6410 while (head) {
6411 if (head->free_bg)
6412 bg_blkno = head->free_bg;
6413 else
6414 bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
6415 head->free_bit);
6416 handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
6417 if (IS_ERR(handle)) {
6418 ret = PTR_ERR(handle);
6419 mlog_errno(ret);
6420 goto out_unlock;
6421 }
6422
6423 trace_ocfs2_free_cached_blocks(
6424 (unsigned long long)head->free_blk, head->free_bit);
6425
6426 ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
6427 head->free_bit, bg_blkno, 1);
6428 if (ret)
6429 mlog_errno(ret);
6430
6431 ocfs2_commit_trans(osb, handle);
6432
6433 tmp = head;
6434 head = head->free_next;
6435 kfree(tmp);
6436 }
6437
6438out_unlock:
6439 ocfs2_inode_unlock(inode, 1);
6440 brelse(di_bh);
6441out_mutex:
6442 inode_unlock(inode);
6443 iput(inode);
6444out:
6445 while(head) {
6446 /* Premature exit may have left some dangling items. */
6447 tmp = head;
6448 head = head->free_next;
6449 kfree(tmp);
6450 }
6451
6452 return ret;
6453}
6454
6455int ocfs2_cache_cluster_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6456 u64 blkno, unsigned int bit)
6457{
6458 int ret = 0;
6459 struct ocfs2_cached_block_free *item;
6460
6461 item = kzalloc(sizeof(*item), GFP_NOFS);
6462 if (item == NULL) {
6463 ret = -ENOMEM;
6464 mlog_errno(ret);
6465 return ret;
6466 }
6467
6468 trace_ocfs2_cache_cluster_dealloc((unsigned long long)blkno, bit);
6469
6470 item->free_blk = blkno;
6471 item->free_bit = bit;
6472 item->free_next = ctxt->c_global_allocator;
6473
6474 ctxt->c_global_allocator = item;
6475 return ret;
6476}
6477
6478static int ocfs2_free_cached_clusters(struct ocfs2_super *osb,
6479 struct ocfs2_cached_block_free *head)
6480{
6481 struct ocfs2_cached_block_free *tmp;
6482 struct inode *tl_inode = osb->osb_tl_inode;
6483 handle_t *handle;
6484 int ret = 0;
6485
6486 inode_lock(tl_inode);
6487
6488 while (head) {
6489 if (ocfs2_truncate_log_needs_flush(osb)) {
6490 ret = __ocfs2_flush_truncate_log(osb);
6491 if (ret < 0) {
6492 mlog_errno(ret);
6493 break;
6494 }
6495 }
6496
6497 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6498 if (IS_ERR(handle)) {
6499 ret = PTR_ERR(handle);
6500 mlog_errno(ret);
6501 break;
6502 }
6503
6504 ret = ocfs2_truncate_log_append(osb, handle, head->free_blk,
6505 head->free_bit);
6506
6507 ocfs2_commit_trans(osb, handle);
6508 tmp = head;
6509 head = head->free_next;
6510 kfree(tmp);
6511
6512 if (ret < 0) {
6513 mlog_errno(ret);
6514 break;
6515 }
6516 }
6517
6518 inode_unlock(tl_inode);
6519
6520 while (head) {
6521 /* Premature exit may have left some dangling items. */
6522 tmp = head;
6523 head = head->free_next;
6524 kfree(tmp);
6525 }
6526
6527 return ret;
6528}
6529
6530int ocfs2_run_deallocs(struct ocfs2_super *osb,
6531 struct ocfs2_cached_dealloc_ctxt *ctxt)
6532{
6533 int ret = 0, ret2;
6534 struct ocfs2_per_slot_free_list *fl;
6535
6536 if (!ctxt)
6537 return 0;
6538
6539 while (ctxt->c_first_suballocator) {
6540 fl = ctxt->c_first_suballocator;
6541
6542 if (fl->f_first) {
6543 trace_ocfs2_run_deallocs(fl->f_inode_type,
6544 fl->f_slot);
6545 ret2 = ocfs2_free_cached_blocks(osb,
6546 fl->f_inode_type,
6547 fl->f_slot,
6548 fl->f_first);
6549 if (ret2)
6550 mlog_errno(ret2);
6551 if (!ret)
6552 ret = ret2;
6553 }
6554
6555 ctxt->c_first_suballocator = fl->f_next_suballocator;
6556 kfree(fl);
6557 }
6558
6559 if (ctxt->c_global_allocator) {
6560 ret2 = ocfs2_free_cached_clusters(osb,
6561 ctxt->c_global_allocator);
6562 if (ret2)
6563 mlog_errno(ret2);
6564 if (!ret)
6565 ret = ret2;
6566
6567 ctxt->c_global_allocator = NULL;
6568 }
6569
6570 return ret;
6571}
6572
6573static struct ocfs2_per_slot_free_list *
6574ocfs2_find_per_slot_free_list(int type,
6575 int slot,
6576 struct ocfs2_cached_dealloc_ctxt *ctxt)
6577{
6578 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6579
6580 while (fl) {
6581 if (fl->f_inode_type == type && fl->f_slot == slot)
6582 return fl;
6583
6584 fl = fl->f_next_suballocator;
6585 }
6586
6587 fl = kmalloc(sizeof(*fl), GFP_NOFS);
6588 if (fl) {
6589 fl->f_inode_type = type;
6590 fl->f_slot = slot;
6591 fl->f_first = NULL;
6592 fl->f_next_suballocator = ctxt->c_first_suballocator;
6593
6594 ctxt->c_first_suballocator = fl;
6595 }
6596 return fl;
6597}
6598
6599static struct ocfs2_per_slot_free_list *
6600ocfs2_find_preferred_free_list(int type,
6601 int preferred_slot,
6602 int *real_slot,
6603 struct ocfs2_cached_dealloc_ctxt *ctxt)
6604{
6605 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6606
6607 while (fl) {
6608 if (fl->f_inode_type == type && fl->f_slot == preferred_slot) {
6609 *real_slot = fl->f_slot;
6610 return fl;
6611 }
6612
6613 fl = fl->f_next_suballocator;
6614 }
6615
6616 /* If we can't find any free list matching preferred slot, just use
6617 * the first one.
6618 */
6619 fl = ctxt->c_first_suballocator;
6620 *real_slot = fl->f_slot;
6621
6622 return fl;
6623}
6624
6625/* Return Value 1 indicates empty */
6626static int ocfs2_is_dealloc_empty(struct ocfs2_extent_tree *et)
6627{
6628 struct ocfs2_per_slot_free_list *fl = NULL;
6629
6630 if (!et->et_dealloc)
6631 return 1;
6632
6633 fl = et->et_dealloc->c_first_suballocator;
6634 if (!fl)
6635 return 1;
6636
6637 if (!fl->f_first)
6638 return 1;
6639
6640 return 0;
6641}
6642
6643/* If extent was deleted from tree due to extent rotation and merging, and
6644 * no metadata is reserved ahead of time. Try to reuse some extents
6645 * just deleted. This is only used to reuse extent blocks.
6646 * It is supposed to find enough extent blocks in dealloc if our estimation
6647 * on metadata is accurate.
6648 */
6649static int ocfs2_reuse_blk_from_dealloc(handle_t *handle,
6650 struct ocfs2_extent_tree *et,
6651 struct buffer_head **new_eb_bh,
6652 int blk_wanted, int *blk_given)
6653{
6654 int i, status = 0, real_slot;
6655 struct ocfs2_cached_dealloc_ctxt *dealloc;
6656 struct ocfs2_per_slot_free_list *fl;
6657 struct ocfs2_cached_block_free *bf;
6658 struct ocfs2_extent_block *eb;
6659 struct ocfs2_super *osb =
6660 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
6661
6662 *blk_given = 0;
6663
6664 /* If extent tree doesn't have a dealloc, this is not faulty. Just
6665 * tell upper caller dealloc can't provide any block and it should
6666 * ask for alloc to claim more space.
6667 */
6668 dealloc = et->et_dealloc;
6669 if (!dealloc)
6670 goto bail;
6671
6672 for (i = 0; i < blk_wanted; i++) {
6673 /* Prefer to use local slot */
6674 fl = ocfs2_find_preferred_free_list(EXTENT_ALLOC_SYSTEM_INODE,
6675 osb->slot_num, &real_slot,
6676 dealloc);
6677 /* If no more block can be reused, we should claim more
6678 * from alloc. Just return here normally.
6679 */
6680 if (!fl) {
6681 status = 0;
6682 break;
6683 }
6684
6685 bf = fl->f_first;
6686 fl->f_first = bf->free_next;
6687
6688 new_eb_bh[i] = sb_getblk(osb->sb, bf->free_blk);
6689 if (new_eb_bh[i] == NULL) {
6690 status = -ENOMEM;
6691 mlog_errno(status);
6692 goto bail;
6693 }
6694
6695 mlog(0, "Reusing block(%llu) from "
6696 "dealloc(local slot:%d, real slot:%d)\n",
6697 bf->free_blk, osb->slot_num, real_slot);
6698
6699 ocfs2_set_new_buffer_uptodate(et->et_ci, new_eb_bh[i]);
6700
6701 status = ocfs2_journal_access_eb(handle, et->et_ci,
6702 new_eb_bh[i],
6703 OCFS2_JOURNAL_ACCESS_CREATE);
6704 if (status < 0) {
6705 mlog_errno(status);
6706 goto bail;
6707 }
6708
6709 memset(new_eb_bh[i]->b_data, 0, osb->sb->s_blocksize);
6710 eb = (struct ocfs2_extent_block *) new_eb_bh[i]->b_data;
6711
6712 /* We can't guarantee that buffer head is still cached, so
6713 * polutlate the extent block again.
6714 */
6715 strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
6716 eb->h_blkno = cpu_to_le64(bf->free_blk);
6717 eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
6718 eb->h_suballoc_slot = cpu_to_le16(real_slot);
6719 eb->h_suballoc_loc = cpu_to_le64(bf->free_bg);
6720 eb->h_suballoc_bit = cpu_to_le16(bf->free_bit);
6721 eb->h_list.l_count =
6722 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
6723
6724 /* We'll also be dirtied by the caller, so
6725 * this isn't absolutely necessary.
6726 */
6727 ocfs2_journal_dirty(handle, new_eb_bh[i]);
6728
6729 if (!fl->f_first) {
6730 dealloc->c_first_suballocator = fl->f_next_suballocator;
6731 kfree(fl);
6732 }
6733 kfree(bf);
6734 }
6735
6736 *blk_given = i;
6737
6738bail:
6739 if (unlikely(status < 0)) {
6740 for (i = 0; i < blk_wanted; i++)
6741 brelse(new_eb_bh[i]);
6742 }
6743
6744 return status;
6745}
6746
6747int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6748 int type, int slot, u64 suballoc,
6749 u64 blkno, unsigned int bit)
6750{
6751 int ret;
6752 struct ocfs2_per_slot_free_list *fl;
6753 struct ocfs2_cached_block_free *item;
6754
6755 fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
6756 if (fl == NULL) {
6757 ret = -ENOMEM;
6758 mlog_errno(ret);
6759 goto out;
6760 }
6761
6762 item = kzalloc(sizeof(*item), GFP_NOFS);
6763 if (item == NULL) {
6764 ret = -ENOMEM;
6765 mlog_errno(ret);
6766 goto out;
6767 }
6768
6769 trace_ocfs2_cache_block_dealloc(type, slot,
6770 (unsigned long long)suballoc,
6771 (unsigned long long)blkno, bit);
6772
6773 item->free_bg = suballoc;
6774 item->free_blk = blkno;
6775 item->free_bit = bit;
6776 item->free_next = fl->f_first;
6777
6778 fl->f_first = item;
6779
6780 ret = 0;
6781out:
6782 return ret;
6783}
6784
6785static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
6786 struct ocfs2_extent_block *eb)
6787{
6788 return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
6789 le16_to_cpu(eb->h_suballoc_slot),
6790 le64_to_cpu(eb->h_suballoc_loc),
6791 le64_to_cpu(eb->h_blkno),
6792 le16_to_cpu(eb->h_suballoc_bit));
6793}
6794
6795static int ocfs2_zero_func(handle_t *handle, struct buffer_head *bh)
6796{
6797 set_buffer_uptodate(bh);
6798 mark_buffer_dirty(bh);
6799 return 0;
6800}
6801
6802void ocfs2_map_and_dirty_page(struct inode *inode, handle_t *handle,
6803 unsigned int from, unsigned int to,
6804 struct page *page, int zero, u64 *phys)
6805{
6806 int ret, partial = 0;
6807 loff_t start_byte = ((loff_t)page->index << PAGE_SHIFT) + from;
6808 loff_t length = to - from;
6809
6810 ret = ocfs2_map_page_blocks(page, phys, inode, from, to, 0);
6811 if (ret)
6812 mlog_errno(ret);
6813
6814 if (zero)
6815 zero_user_segment(page, from, to);
6816
6817 /*
6818 * Need to set the buffers we zero'd into uptodate
6819 * here if they aren't - ocfs2_map_page_blocks()
6820 * might've skipped some
6821 */
6822 ret = walk_page_buffers(handle, page_buffers(page),
6823 from, to, &partial,
6824 ocfs2_zero_func);
6825 if (ret < 0)
6826 mlog_errno(ret);
6827 else if (ocfs2_should_order_data(inode)) {
6828 ret = ocfs2_jbd2_inode_add_write(handle, inode,
6829 start_byte, length);
6830 if (ret < 0)
6831 mlog_errno(ret);
6832 }
6833
6834 if (!partial)
6835 SetPageUptodate(page);
6836
6837 flush_dcache_page(page);
6838}
6839
6840static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start,
6841 loff_t end, struct page **pages,
6842 int numpages, u64 phys, handle_t *handle)
6843{
6844 int i;
6845 struct page *page;
6846 unsigned int from, to = PAGE_SIZE;
6847 struct super_block *sb = inode->i_sb;
6848
6849 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
6850
6851 if (numpages == 0)
6852 goto out;
6853
6854 to = PAGE_SIZE;
6855 for(i = 0; i < numpages; i++) {
6856 page = pages[i];
6857
6858 from = start & (PAGE_SIZE - 1);
6859 if ((end >> PAGE_SHIFT) == page->index)
6860 to = end & (PAGE_SIZE - 1);
6861
6862 BUG_ON(from > PAGE_SIZE);
6863 BUG_ON(to > PAGE_SIZE);
6864
6865 ocfs2_map_and_dirty_page(inode, handle, from, to, page, 1,
6866 &phys);
6867
6868 start = (page->index + 1) << PAGE_SHIFT;
6869 }
6870out:
6871 if (pages)
6872 ocfs2_unlock_and_free_pages(pages, numpages);
6873}
6874
6875int ocfs2_grab_pages(struct inode *inode, loff_t start, loff_t end,
6876 struct page **pages, int *num)
6877{
6878 int numpages, ret = 0;
6879 struct address_space *mapping = inode->i_mapping;
6880 unsigned long index;
6881 loff_t last_page_bytes;
6882
6883 BUG_ON(start > end);
6884
6885 numpages = 0;
6886 last_page_bytes = PAGE_ALIGN(end);
6887 index = start >> PAGE_SHIFT;
6888 do {
6889 pages[numpages] = find_or_create_page(mapping, index, GFP_NOFS);
6890 if (!pages[numpages]) {
6891 ret = -ENOMEM;
6892 mlog_errno(ret);
6893 goto out;
6894 }
6895
6896 numpages++;
6897 index++;
6898 } while (index < (last_page_bytes >> PAGE_SHIFT));
6899
6900out:
6901 if (ret != 0) {
6902 if (pages)
6903 ocfs2_unlock_and_free_pages(pages, numpages);
6904 numpages = 0;
6905 }
6906
6907 *num = numpages;
6908
6909 return ret;
6910}
6911
6912static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end,
6913 struct page **pages, int *num)
6914{
6915 struct super_block *sb = inode->i_sb;
6916
6917 BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
6918 (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
6919
6920 return ocfs2_grab_pages(inode, start, end, pages, num);
6921}
6922
6923/*
6924 * Zero the area past i_size but still within an allocated
6925 * cluster. This avoids exposing nonzero data on subsequent file
6926 * extends.
6927 *
6928 * We need to call this before i_size is updated on the inode because
6929 * otherwise block_write_full_page() will skip writeout of pages past
6930 * i_size. The new_i_size parameter is passed for this reason.
6931 */
6932int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
6933 u64 range_start, u64 range_end)
6934{
6935 int ret = 0, numpages;
6936 struct page **pages = NULL;
6937 u64 phys;
6938 unsigned int ext_flags;
6939 struct super_block *sb = inode->i_sb;
6940
6941 /*
6942 * File systems which don't support sparse files zero on every
6943 * extend.
6944 */
6945 if (!ocfs2_sparse_alloc(OCFS2_SB(sb)))
6946 return 0;
6947
6948 pages = kcalloc(ocfs2_pages_per_cluster(sb),
6949 sizeof(struct page *), GFP_NOFS);
6950 if (pages == NULL) {
6951 ret = -ENOMEM;
6952 mlog_errno(ret);
6953 goto out;
6954 }
6955
6956 if (range_start == range_end)
6957 goto out;
6958
6959 ret = ocfs2_extent_map_get_blocks(inode,
6960 range_start >> sb->s_blocksize_bits,
6961 &phys, NULL, &ext_flags);
6962 if (ret) {
6963 mlog_errno(ret);
6964 goto out;
6965 }
6966
6967 /*
6968 * Tail is a hole, or is marked unwritten. In either case, we
6969 * can count on read and write to return/push zero's.
6970 */
6971 if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN)
6972 goto out;
6973
6974 ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages,
6975 &numpages);
6976 if (ret) {
6977 mlog_errno(ret);
6978 goto out;
6979 }
6980
6981 ocfs2_zero_cluster_pages(inode, range_start, range_end, pages,
6982 numpages, phys, handle);
6983
6984 /*
6985 * Initiate writeout of the pages we zero'd here. We don't
6986 * wait on them - the truncate_inode_pages() call later will
6987 * do that for us.
6988 */
6989 ret = filemap_fdatawrite_range(inode->i_mapping, range_start,
6990 range_end - 1);
6991 if (ret)
6992 mlog_errno(ret);
6993
6994out:
6995 kfree(pages);
6996
6997 return ret;
6998}
6999
7000static void ocfs2_zero_dinode_id2_with_xattr(struct inode *inode,
7001 struct ocfs2_dinode *di)
7002{
7003 unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits;
7004 unsigned int xattrsize = le16_to_cpu(di->i_xattr_inline_size);
7005
7006 if (le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_XATTR_FL)
7007 memset(&di->id2, 0, blocksize -
7008 offsetof(struct ocfs2_dinode, id2) -
7009 xattrsize);
7010 else
7011 memset(&di->id2, 0, blocksize -
7012 offsetof(struct ocfs2_dinode, id2));
7013}
7014
7015void ocfs2_dinode_new_extent_list(struct inode *inode,
7016 struct ocfs2_dinode *di)
7017{
7018 ocfs2_zero_dinode_id2_with_xattr(inode, di);
7019 di->id2.i_list.l_tree_depth = 0;
7020 di->id2.i_list.l_next_free_rec = 0;
7021 di->id2.i_list.l_count = cpu_to_le16(
7022 ocfs2_extent_recs_per_inode_with_xattr(inode->i_sb, di));
7023}
7024
7025void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di)
7026{
7027 struct ocfs2_inode_info *oi = OCFS2_I(inode);
7028 struct ocfs2_inline_data *idata = &di->id2.i_data;
7029
7030 spin_lock(&oi->ip_lock);
7031 oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
7032 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7033 spin_unlock(&oi->ip_lock);
7034
7035 /*
7036 * We clear the entire i_data structure here so that all
7037 * fields can be properly initialized.
7038 */
7039 ocfs2_zero_dinode_id2_with_xattr(inode, di);
7040
7041 idata->id_count = cpu_to_le16(
7042 ocfs2_max_inline_data_with_xattr(inode->i_sb, di));
7043}
7044
7045int ocfs2_convert_inline_data_to_extents(struct inode *inode,
7046 struct buffer_head *di_bh)
7047{
7048 int ret, has_data, num_pages = 0;
7049 int need_free = 0;
7050 u32 bit_off, num;
7051 handle_t *handle;
7052 u64 block;
7053 struct ocfs2_inode_info *oi = OCFS2_I(inode);
7054 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7055 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7056 struct ocfs2_alloc_context *data_ac = NULL;
7057 struct page *page = NULL;
7058 struct ocfs2_extent_tree et;
7059 int did_quota = 0;
7060
7061 has_data = i_size_read(inode) ? 1 : 0;
7062
7063 if (has_data) {
7064 ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
7065 if (ret) {
7066 mlog_errno(ret);
7067 goto out;
7068 }
7069 }
7070
7071 handle = ocfs2_start_trans(osb,
7072 ocfs2_inline_to_extents_credits(osb->sb));
7073 if (IS_ERR(handle)) {
7074 ret = PTR_ERR(handle);
7075 mlog_errno(ret);
7076 goto out;
7077 }
7078
7079 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7080 OCFS2_JOURNAL_ACCESS_WRITE);
7081 if (ret) {
7082 mlog_errno(ret);
7083 goto out_commit;
7084 }
7085
7086 if (has_data) {
7087 unsigned int page_end = min_t(unsigned, PAGE_SIZE,
7088 osb->s_clustersize);
7089 u64 phys;
7090
7091 ret = dquot_alloc_space_nodirty(inode,
7092 ocfs2_clusters_to_bytes(osb->sb, 1));
7093 if (ret)
7094 goto out_commit;
7095 did_quota = 1;
7096
7097 data_ac->ac_resv = &oi->ip_la_data_resv;
7098
7099 ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off,
7100 &num);
7101 if (ret) {
7102 mlog_errno(ret);
7103 goto out_commit;
7104 }
7105
7106 /*
7107 * Save two copies, one for insert, and one that can
7108 * be changed by ocfs2_map_and_dirty_page() below.
7109 */
7110 block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off);
7111
7112 ret = ocfs2_grab_eof_pages(inode, 0, page_end, &page,
7113 &num_pages);
7114 if (ret) {
7115 mlog_errno(ret);
7116 need_free = 1;
7117 goto out_commit;
7118 }
7119
7120 /*
7121 * This should populate the 1st page for us and mark
7122 * it up to date.
7123 */
7124 ret = ocfs2_read_inline_data(inode, page, di_bh);
7125 if (ret) {
7126 mlog_errno(ret);
7127 need_free = 1;
7128 goto out_unlock;
7129 }
7130
7131 ocfs2_map_and_dirty_page(inode, handle, 0, page_end, page, 0,
7132 &phys);
7133 }
7134
7135 spin_lock(&oi->ip_lock);
7136 oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
7137 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7138 spin_unlock(&oi->ip_lock);
7139
7140 ocfs2_update_inode_fsync_trans(handle, inode, 1);
7141 ocfs2_dinode_new_extent_list(inode, di);
7142
7143 ocfs2_journal_dirty(handle, di_bh);
7144
7145 if (has_data) {
7146 /*
7147 * An error at this point should be extremely rare. If
7148 * this proves to be false, we could always re-build
7149 * the in-inode data from our pages.
7150 */
7151 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7152 ret = ocfs2_insert_extent(handle, &et, 0, block, 1, 0, NULL);
7153 if (ret) {
7154 mlog_errno(ret);
7155 need_free = 1;
7156 goto out_unlock;
7157 }
7158
7159 inode->i_blocks = ocfs2_inode_sector_count(inode);
7160 }
7161
7162out_unlock:
7163 if (page)
7164 ocfs2_unlock_and_free_pages(&page, num_pages);
7165
7166out_commit:
7167 if (ret < 0 && did_quota)
7168 dquot_free_space_nodirty(inode,
7169 ocfs2_clusters_to_bytes(osb->sb, 1));
7170
7171 if (need_free) {
7172 if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
7173 ocfs2_free_local_alloc_bits(osb, handle, data_ac,
7174 bit_off, num);
7175 else
7176 ocfs2_free_clusters(handle,
7177 data_ac->ac_inode,
7178 data_ac->ac_bh,
7179 ocfs2_clusters_to_blocks(osb->sb, bit_off),
7180 num);
7181 }
7182
7183 ocfs2_commit_trans(osb, handle);
7184
7185out:
7186 if (data_ac)
7187 ocfs2_free_alloc_context(data_ac);
7188 return ret;
7189}
7190
7191/*
7192 * It is expected, that by the time you call this function,
7193 * inode->i_size and fe->i_size have been adjusted.
7194 *
7195 * WARNING: This will kfree the truncate context
7196 */
7197int ocfs2_commit_truncate(struct ocfs2_super *osb,
7198 struct inode *inode,
7199 struct buffer_head *di_bh)
7200{
7201 int status = 0, i, flags = 0;
7202 u32 new_highest_cpos, range, trunc_cpos, trunc_len, phys_cpos, coff;
7203 u64 blkno = 0;
7204 struct ocfs2_extent_list *el;
7205 struct ocfs2_extent_rec *rec;
7206 struct ocfs2_path *path = NULL;
7207 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7208 struct ocfs2_extent_list *root_el = &(di->id2.i_list);
7209 u64 refcount_loc = le64_to_cpu(di->i_refcount_loc);
7210 struct ocfs2_extent_tree et;
7211 struct ocfs2_cached_dealloc_ctxt dealloc;
7212 struct ocfs2_refcount_tree *ref_tree = NULL;
7213
7214 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7215 ocfs2_init_dealloc_ctxt(&dealloc);
7216
7217 new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
7218 i_size_read(inode));
7219
7220 path = ocfs2_new_path(di_bh, &di->id2.i_list,
7221 ocfs2_journal_access_di);
7222 if (!path) {
7223 status = -ENOMEM;
7224 mlog_errno(status);
7225 goto bail;
7226 }
7227
7228 ocfs2_extent_map_trunc(inode, new_highest_cpos);
7229
7230start:
7231 /*
7232 * Check that we still have allocation to delete.
7233 */
7234 if (OCFS2_I(inode)->ip_clusters == 0) {
7235 status = 0;
7236 goto bail;
7237 }
7238
7239 /*
7240 * Truncate always works against the rightmost tree branch.
7241 */
7242 status = ocfs2_find_path(INODE_CACHE(inode), path, UINT_MAX);
7243 if (status) {
7244 mlog_errno(status);
7245 goto bail;
7246 }
7247
7248 trace_ocfs2_commit_truncate(
7249 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7250 new_highest_cpos,
7251 OCFS2_I(inode)->ip_clusters,
7252 path->p_tree_depth);
7253
7254 /*
7255 * By now, el will point to the extent list on the bottom most
7256 * portion of this tree. Only the tail record is considered in
7257 * each pass.
7258 *
7259 * We handle the following cases, in order:
7260 * - empty extent: delete the remaining branch
7261 * - remove the entire record
7262 * - remove a partial record
7263 * - no record needs to be removed (truncate has completed)
7264 */
7265 el = path_leaf_el(path);
7266 if (le16_to_cpu(el->l_next_free_rec) == 0) {
7267 ocfs2_error(inode->i_sb,
7268 "Inode %llu has empty extent block at %llu\n",
7269 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7270 (unsigned long long)path_leaf_bh(path)->b_blocknr);
7271 status = -EROFS;
7272 goto bail;
7273 }
7274
7275 i = le16_to_cpu(el->l_next_free_rec) - 1;
7276 rec = &el->l_recs[i];
7277 flags = rec->e_flags;
7278 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
7279
7280 if (i == 0 && ocfs2_is_empty_extent(rec)) {
7281 /*
7282 * Lower levels depend on this never happening, but it's best
7283 * to check it up here before changing the tree.
7284 */
7285 if (root_el->l_tree_depth && rec->e_int_clusters == 0) {
7286 mlog(ML_ERROR, "Inode %lu has an empty "
7287 "extent record, depth %u\n", inode->i_ino,
7288 le16_to_cpu(root_el->l_tree_depth));
7289 status = ocfs2_remove_rightmost_empty_extent(osb,
7290 &et, path, &dealloc);
7291 if (status) {
7292 mlog_errno(status);
7293 goto bail;
7294 }
7295
7296 ocfs2_reinit_path(path, 1);
7297 goto start;
7298 } else {
7299 trunc_cpos = le32_to_cpu(rec->e_cpos);
7300 trunc_len = 0;
7301 blkno = 0;
7302 }
7303 } else if (le32_to_cpu(rec->e_cpos) >= new_highest_cpos) {
7304 /*
7305 * Truncate entire record.
7306 */
7307 trunc_cpos = le32_to_cpu(rec->e_cpos);
7308 trunc_len = ocfs2_rec_clusters(el, rec);
7309 blkno = le64_to_cpu(rec->e_blkno);
7310 } else if (range > new_highest_cpos) {
7311 /*
7312 * Partial truncate. it also should be
7313 * the last truncate we're doing.
7314 */
7315 trunc_cpos = new_highest_cpos;
7316 trunc_len = range - new_highest_cpos;
7317 coff = new_highest_cpos - le32_to_cpu(rec->e_cpos);
7318 blkno = le64_to_cpu(rec->e_blkno) +
7319 ocfs2_clusters_to_blocks(inode->i_sb, coff);
7320 } else {
7321 /*
7322 * Truncate completed, leave happily.
7323 */
7324 status = 0;
7325 goto bail;
7326 }
7327
7328 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
7329
7330 if ((flags & OCFS2_EXT_REFCOUNTED) && trunc_len && !ref_tree) {
7331 status = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
7332 &ref_tree, NULL);
7333 if (status) {
7334 mlog_errno(status);
7335 goto bail;
7336 }
7337 }
7338
7339 status = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
7340 phys_cpos, trunc_len, flags, &dealloc,
7341 refcount_loc, true);
7342 if (status < 0) {
7343 mlog_errno(status);
7344 goto bail;
7345 }
7346
7347 ocfs2_reinit_path(path, 1);
7348
7349 /*
7350 * The check above will catch the case where we've truncated
7351 * away all allocation.
7352 */
7353 goto start;
7354
7355bail:
7356 if (ref_tree)
7357 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
7358
7359 ocfs2_schedule_truncate_log_flush(osb, 1);
7360
7361 ocfs2_run_deallocs(osb, &dealloc);
7362
7363 ocfs2_free_path(path);
7364
7365 return status;
7366}
7367
7368/*
7369 * 'start' is inclusive, 'end' is not.
7370 */
7371int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh,
7372 unsigned int start, unsigned int end, int trunc)
7373{
7374 int ret;
7375 unsigned int numbytes;
7376 handle_t *handle;
7377 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7378 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7379 struct ocfs2_inline_data *idata = &di->id2.i_data;
7380
7381 /* No need to punch hole beyond i_size. */
7382 if (start >= i_size_read(inode))
7383 return 0;
7384
7385 if (end > i_size_read(inode))
7386 end = i_size_read(inode);
7387
7388 BUG_ON(start > end);
7389
7390 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
7391 !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) ||
7392 !ocfs2_supports_inline_data(osb)) {
7393 ocfs2_error(inode->i_sb,
7394 "Inline data flags for inode %llu don't agree! Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7395 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7396 le16_to_cpu(di->i_dyn_features),
7397 OCFS2_I(inode)->ip_dyn_features,
7398 osb->s_feature_incompat);
7399 ret = -EROFS;
7400 goto out;
7401 }
7402
7403 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
7404 if (IS_ERR(handle)) {
7405 ret = PTR_ERR(handle);
7406 mlog_errno(ret);
7407 goto out;
7408 }
7409
7410 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7411 OCFS2_JOURNAL_ACCESS_WRITE);
7412 if (ret) {
7413 mlog_errno(ret);
7414 goto out_commit;
7415 }
7416
7417 numbytes = end - start;
7418 memset(idata->id_data + start, 0, numbytes);
7419
7420 /*
7421 * No need to worry about the data page here - it's been
7422 * truncated already and inline data doesn't need it for
7423 * pushing zero's to disk, so we'll let readpage pick it up
7424 * later.
7425 */
7426 if (trunc) {
7427 i_size_write(inode, start);
7428 di->i_size = cpu_to_le64(start);
7429 }
7430
7431 inode->i_blocks = ocfs2_inode_sector_count(inode);
7432 inode->i_ctime = inode->i_mtime = current_time(inode);
7433
7434 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
7435 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
7436
7437 ocfs2_update_inode_fsync_trans(handle, inode, 1);
7438 ocfs2_journal_dirty(handle, di_bh);
7439
7440out_commit:
7441 ocfs2_commit_trans(osb, handle);
7442
7443out:
7444 return ret;
7445}
7446
7447static int ocfs2_trim_extent(struct super_block *sb,
7448 struct ocfs2_group_desc *gd,
7449 u64 group, u32 start, u32 count)
7450{
7451 u64 discard, bcount;
7452 struct ocfs2_super *osb = OCFS2_SB(sb);
7453
7454 bcount = ocfs2_clusters_to_blocks(sb, count);
7455 discard = ocfs2_clusters_to_blocks(sb, start);
7456
7457 /*
7458 * For the first cluster group, the gd->bg_blkno is not at the start
7459 * of the group, but at an offset from the start. If we add it while
7460 * calculating discard for first group, we will wrongly start fstrim a
7461 * few blocks after the desried start block and the range can cross
7462 * over into the next cluster group. So, add it only if this is not
7463 * the first cluster group.
7464 */
7465 if (group != osb->first_cluster_group_blkno)
7466 discard += le64_to_cpu(gd->bg_blkno);
7467
7468 trace_ocfs2_trim_extent(sb, (unsigned long long)discard, bcount);
7469
7470 return sb_issue_discard(sb, discard, bcount, GFP_NOFS, 0);
7471}
7472
7473static int ocfs2_trim_group(struct super_block *sb,
7474 struct ocfs2_group_desc *gd, u64 group,
7475 u32 start, u32 max, u32 minbits)
7476{
7477 int ret = 0, count = 0, next;
7478 void *bitmap = gd->bg_bitmap;
7479
7480 if (le16_to_cpu(gd->bg_free_bits_count) < minbits)
7481 return 0;
7482
7483 trace_ocfs2_trim_group((unsigned long long)le64_to_cpu(gd->bg_blkno),
7484 start, max, minbits);
7485
7486 while (start < max) {
7487 start = ocfs2_find_next_zero_bit(bitmap, max, start);
7488 if (start >= max)
7489 break;
7490 next = ocfs2_find_next_bit(bitmap, max, start);
7491
7492 if ((next - start) >= minbits) {
7493 ret = ocfs2_trim_extent(sb, gd, group,
7494 start, next - start);
7495 if (ret < 0) {
7496 mlog_errno(ret);
7497 break;
7498 }
7499 count += next - start;
7500 }
7501 start = next + 1;
7502
7503 if (fatal_signal_pending(current)) {
7504 count = -ERESTARTSYS;
7505 break;
7506 }
7507
7508 if ((le16_to_cpu(gd->bg_free_bits_count) - count) < minbits)
7509 break;
7510 }
7511
7512 if (ret < 0)
7513 count = ret;
7514
7515 return count;
7516}
7517
7518static
7519int ocfs2_trim_mainbm(struct super_block *sb, struct fstrim_range *range)
7520{
7521 struct ocfs2_super *osb = OCFS2_SB(sb);
7522 u64 start, len, trimmed = 0, first_group, last_group = 0, group = 0;
7523 int ret, cnt;
7524 u32 first_bit, last_bit, minlen;
7525 struct buffer_head *main_bm_bh = NULL;
7526 struct inode *main_bm_inode = NULL;
7527 struct buffer_head *gd_bh = NULL;
7528 struct ocfs2_dinode *main_bm;
7529 struct ocfs2_group_desc *gd = NULL;
7530
7531 start = range->start >> osb->s_clustersize_bits;
7532 len = range->len >> osb->s_clustersize_bits;
7533 minlen = range->minlen >> osb->s_clustersize_bits;
7534
7535 if (minlen >= osb->bitmap_cpg || range->len < sb->s_blocksize)
7536 return -EINVAL;
7537
7538 trace_ocfs2_trim_mainbm(start, len, minlen);
7539
7540next_group:
7541 main_bm_inode = ocfs2_get_system_file_inode(osb,
7542 GLOBAL_BITMAP_SYSTEM_INODE,
7543 OCFS2_INVALID_SLOT);
7544 if (!main_bm_inode) {
7545 ret = -EIO;
7546 mlog_errno(ret);
7547 goto out;
7548 }
7549
7550 inode_lock(main_bm_inode);
7551
7552 ret = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 0);
7553 if (ret < 0) {
7554 mlog_errno(ret);
7555 goto out_mutex;
7556 }
7557 main_bm = (struct ocfs2_dinode *)main_bm_bh->b_data;
7558
7559 /*
7560 * Do some check before trim the first group.
7561 */
7562 if (!group) {
7563 if (start >= le32_to_cpu(main_bm->i_clusters)) {
7564 ret = -EINVAL;
7565 goto out_unlock;
7566 }
7567
7568 if (start + len > le32_to_cpu(main_bm->i_clusters))
7569 len = le32_to_cpu(main_bm->i_clusters) - start;
7570
7571 /*
7572 * Determine first and last group to examine based on
7573 * start and len
7574 */
7575 first_group = ocfs2_which_cluster_group(main_bm_inode, start);
7576 if (first_group == osb->first_cluster_group_blkno)
7577 first_bit = start;
7578 else
7579 first_bit = start - ocfs2_blocks_to_clusters(sb,
7580 first_group);
7581 last_group = ocfs2_which_cluster_group(main_bm_inode,
7582 start + len - 1);
7583 group = first_group;
7584 }
7585
7586 do {
7587 if (first_bit + len >= osb->bitmap_cpg)
7588 last_bit = osb->bitmap_cpg;
7589 else
7590 last_bit = first_bit + len;
7591
7592 ret = ocfs2_read_group_descriptor(main_bm_inode,
7593 main_bm, group,
7594 &gd_bh);
7595 if (ret < 0) {
7596 mlog_errno(ret);
7597 break;
7598 }
7599
7600 gd = (struct ocfs2_group_desc *)gd_bh->b_data;
7601 cnt = ocfs2_trim_group(sb, gd, group,
7602 first_bit, last_bit, minlen);
7603 brelse(gd_bh);
7604 gd_bh = NULL;
7605 if (cnt < 0) {
7606 ret = cnt;
7607 mlog_errno(ret);
7608 break;
7609 }
7610
7611 trimmed += cnt;
7612 len -= osb->bitmap_cpg - first_bit;
7613 first_bit = 0;
7614 if (group == osb->first_cluster_group_blkno)
7615 group = ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7616 else
7617 group += ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7618 } while (0);
7619
7620out_unlock:
7621 ocfs2_inode_unlock(main_bm_inode, 0);
7622 brelse(main_bm_bh);
7623 main_bm_bh = NULL;
7624out_mutex:
7625 inode_unlock(main_bm_inode);
7626 iput(main_bm_inode);
7627
7628 /*
7629 * If all the groups trim are not done or failed, but we should release
7630 * main_bm related locks for avoiding the current IO starve, then go to
7631 * trim the next group
7632 */
7633 if (ret >= 0 && group <= last_group) {
7634 cond_resched();
7635 goto next_group;
7636 }
7637out:
7638 range->len = trimmed * sb->s_blocksize;
7639 return ret;
7640}
7641
7642int ocfs2_trim_fs(struct super_block *sb, struct fstrim_range *range)
7643{
7644 int ret;
7645 struct ocfs2_super *osb = OCFS2_SB(sb);
7646 struct ocfs2_trim_fs_info info, *pinfo = NULL;
7647
7648 ocfs2_trim_fs_lock_res_init(osb);
7649
7650 trace_ocfs2_trim_fs(range->start, range->len, range->minlen);
7651
7652 ret = ocfs2_trim_fs_lock(osb, NULL, 1);
7653 if (ret < 0) {
7654 if (ret != -EAGAIN) {
7655 mlog_errno(ret);
7656 ocfs2_trim_fs_lock_res_uninit(osb);
7657 return ret;
7658 }
7659
7660 mlog(ML_NOTICE, "Wait for trim on device (%s) to "
7661 "finish, which is running from another node.\n",
7662 osb->dev_str);
7663 ret = ocfs2_trim_fs_lock(osb, &info, 0);
7664 if (ret < 0) {
7665 mlog_errno(ret);
7666 ocfs2_trim_fs_lock_res_uninit(osb);
7667 return ret;
7668 }
7669
7670 if (info.tf_valid && info.tf_success &&
7671 info.tf_start == range->start &&
7672 info.tf_len == range->len &&
7673 info.tf_minlen == range->minlen) {
7674 /* Avoid sending duplicated trim to a shared device */
7675 mlog(ML_NOTICE, "The same trim on device (%s) was "
7676 "just done from node (%u), return.\n",
7677 osb->dev_str, info.tf_nodenum);
7678 range->len = info.tf_trimlen;
7679 goto out;
7680 }
7681 }
7682
7683 info.tf_nodenum = osb->node_num;
7684 info.tf_start = range->start;
7685 info.tf_len = range->len;
7686 info.tf_minlen = range->minlen;
7687
7688 ret = ocfs2_trim_mainbm(sb, range);
7689
7690 info.tf_trimlen = range->len;
7691 info.tf_success = (ret < 0 ? 0 : 1);
7692 pinfo = &info;
7693out:
7694 ocfs2_trim_fs_unlock(osb, pinfo);
7695 ocfs2_trim_fs_lock_res_uninit(osb);
7696 return ret;
7697}