Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
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/string.h>
14#include <linux/highmem.h>
15#include <linux/swap.h>
16#include <linux/quotaops.h>
17#include <linux/blkdev.h>
18#include <linux/sched/signal.h>
19
20#include <cluster/masklog.h>
21
22#include "ocfs2.h"
23
24#include "alloc.h"
25#include "aops.h"
26#include "blockcheck.h"
27#include "dlmglue.h"
28#include "extent_map.h"
29#include "inode.h"
30#include "journal.h"
31#include "localalloc.h"
32#include "suballoc.h"
33#include "sysfile.h"
34#include "file.h"
35#include "super.h"
36#include "uptodate.h"
37#include "xattr.h"
38#include "refcounttree.h"
39#include "ocfs2_trace.h"
40
41#include "buffer_head_io.h"
42
43enum ocfs2_contig_type {
44 CONTIG_NONE = 0,
45 CONTIG_LEFT,
46 CONTIG_RIGHT,
47 CONTIG_LEFTRIGHT,
48};
49
50static enum ocfs2_contig_type
51 ocfs2_extent_rec_contig(struct super_block *sb,
52 struct ocfs2_extent_rec *ext,
53 struct ocfs2_extent_rec *insert_rec);
54/*
55 * Operations for a specific extent tree type.
56 *
57 * To implement an on-disk btree (extent tree) type in ocfs2, add
58 * an ocfs2_extent_tree_operations structure and the matching
59 * ocfs2_init_<thingy>_extent_tree() function. That's pretty much it
60 * for the allocation portion of the extent tree.
61 */
62struct ocfs2_extent_tree_operations {
63 /*
64 * last_eb_blk is the block number of the right most leaf extent
65 * block. Most on-disk structures containing an extent tree store
66 * this value for fast access. The ->eo_set_last_eb_blk() and
67 * ->eo_get_last_eb_blk() operations access this value. They are
68 * both required.
69 */
70 void (*eo_set_last_eb_blk)(struct ocfs2_extent_tree *et,
71 u64 blkno);
72 u64 (*eo_get_last_eb_blk)(struct ocfs2_extent_tree *et);
73
74 /*
75 * The on-disk structure usually keeps track of how many total
76 * clusters are stored in this extent tree. This function updates
77 * that value. new_clusters is the delta, and must be
78 * added to the total. Required.
79 */
80 void (*eo_update_clusters)(struct ocfs2_extent_tree *et,
81 u32 new_clusters);
82
83 /*
84 * If this extent tree is supported by an extent map, insert
85 * a record into the map.
86 */
87 void (*eo_extent_map_insert)(struct ocfs2_extent_tree *et,
88 struct ocfs2_extent_rec *rec);
89
90 /*
91 * If this extent tree is supported by an extent map, truncate the
92 * map to clusters,
93 */
94 void (*eo_extent_map_truncate)(struct ocfs2_extent_tree *et,
95 u32 clusters);
96
97 /*
98 * If ->eo_insert_check() exists, it is called before rec is
99 * inserted into the extent tree. It is optional.
100 */
101 int (*eo_insert_check)(struct ocfs2_extent_tree *et,
102 struct ocfs2_extent_rec *rec);
103 int (*eo_sanity_check)(struct ocfs2_extent_tree *et);
104
105 /*
106 * --------------------------------------------------------------
107 * The remaining are internal to ocfs2_extent_tree and don't have
108 * accessor functions
109 */
110
111 /*
112 * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el.
113 * It is required.
114 */
115 void (*eo_fill_root_el)(struct ocfs2_extent_tree *et);
116
117 /*
118 * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if
119 * it exists. If it does not, et->et_max_leaf_clusters is set
120 * to 0 (unlimited). Optional.
121 */
122 void (*eo_fill_max_leaf_clusters)(struct ocfs2_extent_tree *et);
123
124 /*
125 * ->eo_extent_contig test whether the 2 ocfs2_extent_rec
126 * are contiguous or not. Optional. Don't need to set it if use
127 * ocfs2_extent_rec as the tree leaf.
128 */
129 enum ocfs2_contig_type
130 (*eo_extent_contig)(struct ocfs2_extent_tree *et,
131 struct ocfs2_extent_rec *ext,
132 struct ocfs2_extent_rec *insert_rec);
133};
134
135
136/*
137 * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check
138 * in the methods.
139 */
140static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et);
141static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
142 u64 blkno);
143static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
144 u32 clusters);
145static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
146 struct ocfs2_extent_rec *rec);
147static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
148 u32 clusters);
149static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
150 struct ocfs2_extent_rec *rec);
151static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et);
152static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et);
153
154static int ocfs2_reuse_blk_from_dealloc(handle_t *handle,
155 struct ocfs2_extent_tree *et,
156 struct buffer_head **new_eb_bh,
157 int blk_wanted, int *blk_given);
158static int ocfs2_is_dealloc_empty(struct ocfs2_extent_tree *et);
159
160static const struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops = {
161 .eo_set_last_eb_blk = ocfs2_dinode_set_last_eb_blk,
162 .eo_get_last_eb_blk = ocfs2_dinode_get_last_eb_blk,
163 .eo_update_clusters = ocfs2_dinode_update_clusters,
164 .eo_extent_map_insert = ocfs2_dinode_extent_map_insert,
165 .eo_extent_map_truncate = ocfs2_dinode_extent_map_truncate,
166 .eo_insert_check = ocfs2_dinode_insert_check,
167 .eo_sanity_check = ocfs2_dinode_sanity_check,
168 .eo_fill_root_el = ocfs2_dinode_fill_root_el,
169};
170
171static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
172 u64 blkno)
173{
174 struct ocfs2_dinode *di = et->et_object;
175
176 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
177 di->i_last_eb_blk = cpu_to_le64(blkno);
178}
179
180static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et)
181{
182 struct ocfs2_dinode *di = et->et_object;
183
184 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
185 return le64_to_cpu(di->i_last_eb_blk);
186}
187
188static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
189 u32 clusters)
190{
191 struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
192 struct ocfs2_dinode *di = et->et_object;
193
194 le32_add_cpu(&di->i_clusters, clusters);
195 spin_lock(&oi->ip_lock);
196 oi->ip_clusters = le32_to_cpu(di->i_clusters);
197 spin_unlock(&oi->ip_lock);
198}
199
200static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
201 struct ocfs2_extent_rec *rec)
202{
203 struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
204
205 ocfs2_extent_map_insert_rec(inode, rec);
206}
207
208static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
209 u32 clusters)
210{
211 struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
212
213 ocfs2_extent_map_trunc(inode, clusters);
214}
215
216static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
217 struct ocfs2_extent_rec *rec)
218{
219 struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
220 struct ocfs2_super *osb = OCFS2_SB(oi->vfs_inode.i_sb);
221
222 BUG_ON(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL);
223 mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) &&
224 (oi->ip_clusters != le32_to_cpu(rec->e_cpos)),
225 "Device %s, asking for sparse allocation: inode %llu, "
226 "cpos %u, clusters %u\n",
227 osb->dev_str,
228 (unsigned long long)oi->ip_blkno,
229 rec->e_cpos, oi->ip_clusters);
230
231 return 0;
232}
233
234static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et)
235{
236 struct ocfs2_dinode *di = et->et_object;
237
238 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
239 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
240
241 return 0;
242}
243
244static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et)
245{
246 struct ocfs2_dinode *di = et->et_object;
247
248 et->et_root_el = &di->id2.i_list;
249}
250
251
252static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree *et)
253{
254 struct ocfs2_xattr_value_buf *vb = et->et_object;
255
256 et->et_root_el = &vb->vb_xv->xr_list;
257}
258
259static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree *et,
260 u64 blkno)
261{
262 struct ocfs2_xattr_value_buf *vb = et->et_object;
263
264 vb->vb_xv->xr_last_eb_blk = cpu_to_le64(blkno);
265}
266
267static u64 ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree *et)
268{
269 struct ocfs2_xattr_value_buf *vb = et->et_object;
270
271 return le64_to_cpu(vb->vb_xv->xr_last_eb_blk);
272}
273
274static void ocfs2_xattr_value_update_clusters(struct ocfs2_extent_tree *et,
275 u32 clusters)
276{
277 struct ocfs2_xattr_value_buf *vb = et->et_object;
278
279 le32_add_cpu(&vb->vb_xv->xr_clusters, clusters);
280}
281
282static const struct ocfs2_extent_tree_operations ocfs2_xattr_value_et_ops = {
283 .eo_set_last_eb_blk = ocfs2_xattr_value_set_last_eb_blk,
284 .eo_get_last_eb_blk = ocfs2_xattr_value_get_last_eb_blk,
285 .eo_update_clusters = ocfs2_xattr_value_update_clusters,
286 .eo_fill_root_el = ocfs2_xattr_value_fill_root_el,
287};
288
289static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree *et)
290{
291 struct ocfs2_xattr_block *xb = et->et_object;
292
293 et->et_root_el = &xb->xb_attrs.xb_root.xt_list;
294}
295
296static void ocfs2_xattr_tree_fill_max_leaf_clusters(struct ocfs2_extent_tree *et)
297{
298 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
299 et->et_max_leaf_clusters =
300 ocfs2_clusters_for_bytes(sb, OCFS2_MAX_XATTR_TREE_LEAF_SIZE);
301}
302
303static void ocfs2_xattr_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
304 u64 blkno)
305{
306 struct ocfs2_xattr_block *xb = et->et_object;
307 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
308
309 xt->xt_last_eb_blk = cpu_to_le64(blkno);
310}
311
312static u64 ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
313{
314 struct ocfs2_xattr_block *xb = et->et_object;
315 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
316
317 return le64_to_cpu(xt->xt_last_eb_blk);
318}
319
320static void ocfs2_xattr_tree_update_clusters(struct ocfs2_extent_tree *et,
321 u32 clusters)
322{
323 struct ocfs2_xattr_block *xb = et->et_object;
324
325 le32_add_cpu(&xb->xb_attrs.xb_root.xt_clusters, clusters);
326}
327
328static const struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops = {
329 .eo_set_last_eb_blk = ocfs2_xattr_tree_set_last_eb_blk,
330 .eo_get_last_eb_blk = ocfs2_xattr_tree_get_last_eb_blk,
331 .eo_update_clusters = ocfs2_xattr_tree_update_clusters,
332 .eo_fill_root_el = ocfs2_xattr_tree_fill_root_el,
333 .eo_fill_max_leaf_clusters = ocfs2_xattr_tree_fill_max_leaf_clusters,
334};
335
336static void ocfs2_dx_root_set_last_eb_blk(struct ocfs2_extent_tree *et,
337 u64 blkno)
338{
339 struct ocfs2_dx_root_block *dx_root = et->et_object;
340
341 dx_root->dr_last_eb_blk = cpu_to_le64(blkno);
342}
343
344static u64 ocfs2_dx_root_get_last_eb_blk(struct ocfs2_extent_tree *et)
345{
346 struct ocfs2_dx_root_block *dx_root = et->et_object;
347
348 return le64_to_cpu(dx_root->dr_last_eb_blk);
349}
350
351static void ocfs2_dx_root_update_clusters(struct ocfs2_extent_tree *et,
352 u32 clusters)
353{
354 struct ocfs2_dx_root_block *dx_root = et->et_object;
355
356 le32_add_cpu(&dx_root->dr_clusters, clusters);
357}
358
359static int ocfs2_dx_root_sanity_check(struct ocfs2_extent_tree *et)
360{
361 struct ocfs2_dx_root_block *dx_root = et->et_object;
362
363 BUG_ON(!OCFS2_IS_VALID_DX_ROOT(dx_root));
364
365 return 0;
366}
367
368static void ocfs2_dx_root_fill_root_el(struct ocfs2_extent_tree *et)
369{
370 struct ocfs2_dx_root_block *dx_root = et->et_object;
371
372 et->et_root_el = &dx_root->dr_list;
373}
374
375static const struct ocfs2_extent_tree_operations ocfs2_dx_root_et_ops = {
376 .eo_set_last_eb_blk = ocfs2_dx_root_set_last_eb_blk,
377 .eo_get_last_eb_blk = ocfs2_dx_root_get_last_eb_blk,
378 .eo_update_clusters = ocfs2_dx_root_update_clusters,
379 .eo_sanity_check = ocfs2_dx_root_sanity_check,
380 .eo_fill_root_el = ocfs2_dx_root_fill_root_el,
381};
382
383static void ocfs2_refcount_tree_fill_root_el(struct ocfs2_extent_tree *et)
384{
385 struct ocfs2_refcount_block *rb = et->et_object;
386
387 et->et_root_el = &rb->rf_list;
388}
389
390static void ocfs2_refcount_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
391 u64 blkno)
392{
393 struct ocfs2_refcount_block *rb = et->et_object;
394
395 rb->rf_last_eb_blk = cpu_to_le64(blkno);
396}
397
398static u64 ocfs2_refcount_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
399{
400 struct ocfs2_refcount_block *rb = et->et_object;
401
402 return le64_to_cpu(rb->rf_last_eb_blk);
403}
404
405static void ocfs2_refcount_tree_update_clusters(struct ocfs2_extent_tree *et,
406 u32 clusters)
407{
408 struct ocfs2_refcount_block *rb = et->et_object;
409
410 le32_add_cpu(&rb->rf_clusters, clusters);
411}
412
413static enum ocfs2_contig_type
414ocfs2_refcount_tree_extent_contig(struct ocfs2_extent_tree *et,
415 struct ocfs2_extent_rec *ext,
416 struct ocfs2_extent_rec *insert_rec)
417{
418 return CONTIG_NONE;
419}
420
421static const struct ocfs2_extent_tree_operations ocfs2_refcount_tree_et_ops = {
422 .eo_set_last_eb_blk = ocfs2_refcount_tree_set_last_eb_blk,
423 .eo_get_last_eb_blk = ocfs2_refcount_tree_get_last_eb_blk,
424 .eo_update_clusters = ocfs2_refcount_tree_update_clusters,
425 .eo_fill_root_el = ocfs2_refcount_tree_fill_root_el,
426 .eo_extent_contig = ocfs2_refcount_tree_extent_contig,
427};
428
429static void __ocfs2_init_extent_tree(struct ocfs2_extent_tree *et,
430 struct ocfs2_caching_info *ci,
431 struct buffer_head *bh,
432 ocfs2_journal_access_func access,
433 void *obj,
434 const struct ocfs2_extent_tree_operations *ops)
435{
436 et->et_ops = ops;
437 et->et_root_bh = bh;
438 et->et_ci = ci;
439 et->et_root_journal_access = access;
440 if (!obj)
441 obj = (void *)bh->b_data;
442 et->et_object = obj;
443 et->et_dealloc = NULL;
444
445 et->et_ops->eo_fill_root_el(et);
446 if (!et->et_ops->eo_fill_max_leaf_clusters)
447 et->et_max_leaf_clusters = 0;
448 else
449 et->et_ops->eo_fill_max_leaf_clusters(et);
450}
451
452void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree *et,
453 struct ocfs2_caching_info *ci,
454 struct buffer_head *bh)
455{
456 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_di,
457 NULL, &ocfs2_dinode_et_ops);
458}
459
460void ocfs2_init_xattr_tree_extent_tree(struct ocfs2_extent_tree *et,
461 struct ocfs2_caching_info *ci,
462 struct buffer_head *bh)
463{
464 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_xb,
465 NULL, &ocfs2_xattr_tree_et_ops);
466}
467
468void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree *et,
469 struct ocfs2_caching_info *ci,
470 struct ocfs2_xattr_value_buf *vb)
471{
472 __ocfs2_init_extent_tree(et, ci, vb->vb_bh, vb->vb_access, vb,
473 &ocfs2_xattr_value_et_ops);
474}
475
476void ocfs2_init_dx_root_extent_tree(struct ocfs2_extent_tree *et,
477 struct ocfs2_caching_info *ci,
478 struct buffer_head *bh)
479{
480 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_dr,
481 NULL, &ocfs2_dx_root_et_ops);
482}
483
484void ocfs2_init_refcount_extent_tree(struct ocfs2_extent_tree *et,
485 struct ocfs2_caching_info *ci,
486 struct buffer_head *bh)
487{
488 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_rb,
489 NULL, &ocfs2_refcount_tree_et_ops);
490}
491
492static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree *et,
493 u64 new_last_eb_blk)
494{
495 et->et_ops->eo_set_last_eb_blk(et, new_last_eb_blk);
496}
497
498static inline u64 ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree *et)
499{
500 return et->et_ops->eo_get_last_eb_blk(et);
501}
502
503static inline void ocfs2_et_update_clusters(struct ocfs2_extent_tree *et,
504 u32 clusters)
505{
506 et->et_ops->eo_update_clusters(et, clusters);
507}
508
509static inline void ocfs2_et_extent_map_insert(struct ocfs2_extent_tree *et,
510 struct ocfs2_extent_rec *rec)
511{
512 if (et->et_ops->eo_extent_map_insert)
513 et->et_ops->eo_extent_map_insert(et, rec);
514}
515
516static inline void ocfs2_et_extent_map_truncate(struct ocfs2_extent_tree *et,
517 u32 clusters)
518{
519 if (et->et_ops->eo_extent_map_truncate)
520 et->et_ops->eo_extent_map_truncate(et, clusters);
521}
522
523static inline int ocfs2_et_root_journal_access(handle_t *handle,
524 struct ocfs2_extent_tree *et,
525 int type)
526{
527 return et->et_root_journal_access(handle, et->et_ci, et->et_root_bh,
528 type);
529}
530
531static inline enum ocfs2_contig_type
532 ocfs2_et_extent_contig(struct ocfs2_extent_tree *et,
533 struct ocfs2_extent_rec *rec,
534 struct ocfs2_extent_rec *insert_rec)
535{
536 if (et->et_ops->eo_extent_contig)
537 return et->et_ops->eo_extent_contig(et, rec, insert_rec);
538
539 return ocfs2_extent_rec_contig(
540 ocfs2_metadata_cache_get_super(et->et_ci),
541 rec, insert_rec);
542}
543
544static inline int ocfs2_et_insert_check(struct ocfs2_extent_tree *et,
545 struct ocfs2_extent_rec *rec)
546{
547 int ret = 0;
548
549 if (et->et_ops->eo_insert_check)
550 ret = et->et_ops->eo_insert_check(et, rec);
551 return ret;
552}
553
554static inline int ocfs2_et_sanity_check(struct ocfs2_extent_tree *et)
555{
556 int ret = 0;
557
558 if (et->et_ops->eo_sanity_check)
559 ret = et->et_ops->eo_sanity_check(et);
560 return ret;
561}
562
563static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
564 struct ocfs2_extent_block *eb);
565static void ocfs2_adjust_rightmost_records(handle_t *handle,
566 struct ocfs2_extent_tree *et,
567 struct ocfs2_path *path,
568 struct ocfs2_extent_rec *insert_rec);
569/*
570 * Reset the actual path elements so that we can reuse the structure
571 * to build another path. Generally, this involves freeing the buffer
572 * heads.
573 */
574void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root)
575{
576 int i, start = 0, depth = 0;
577 struct ocfs2_path_item *node;
578
579 if (keep_root)
580 start = 1;
581
582 for(i = start; i < path_num_items(path); i++) {
583 node = &path->p_node[i];
584
585 brelse(node->bh);
586 node->bh = NULL;
587 node->el = NULL;
588 }
589
590 /*
591 * Tree depth may change during truncate, or insert. If we're
592 * keeping the root extent list, then make sure that our path
593 * structure reflects the proper depth.
594 */
595 if (keep_root)
596 depth = le16_to_cpu(path_root_el(path)->l_tree_depth);
597 else
598 path_root_access(path) = NULL;
599
600 path->p_tree_depth = depth;
601}
602
603void ocfs2_free_path(struct ocfs2_path *path)
604{
605 if (path) {
606 ocfs2_reinit_path(path, 0);
607 kfree(path);
608 }
609}
610
611/*
612 * All the elements of src into dest. After this call, src could be freed
613 * without affecting dest.
614 *
615 * Both paths should have the same root. Any non-root elements of dest
616 * will be freed.
617 */
618static void ocfs2_cp_path(struct ocfs2_path *dest, struct ocfs2_path *src)
619{
620 int i;
621
622 BUG_ON(path_root_bh(dest) != path_root_bh(src));
623 BUG_ON(path_root_el(dest) != path_root_el(src));
624 BUG_ON(path_root_access(dest) != path_root_access(src));
625
626 ocfs2_reinit_path(dest, 1);
627
628 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
629 dest->p_node[i].bh = src->p_node[i].bh;
630 dest->p_node[i].el = src->p_node[i].el;
631
632 if (dest->p_node[i].bh)
633 get_bh(dest->p_node[i].bh);
634 }
635}
636
637/*
638 * Make the *dest path the same as src and re-initialize src path to
639 * have a root only.
640 */
641static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src)
642{
643 int i;
644
645 BUG_ON(path_root_bh(dest) != path_root_bh(src));
646 BUG_ON(path_root_access(dest) != path_root_access(src));
647
648 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
649 brelse(dest->p_node[i].bh);
650
651 dest->p_node[i].bh = src->p_node[i].bh;
652 dest->p_node[i].el = src->p_node[i].el;
653
654 src->p_node[i].bh = NULL;
655 src->p_node[i].el = NULL;
656 }
657}
658
659/*
660 * Insert an extent block at given index.
661 *
662 * This will not take an additional reference on eb_bh.
663 */
664static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index,
665 struct buffer_head *eb_bh)
666{
667 struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data;
668
669 /*
670 * Right now, no root bh is an extent block, so this helps
671 * catch code errors with dinode trees. The assertion can be
672 * safely removed if we ever need to insert extent block
673 * structures at the root.
674 */
675 BUG_ON(index == 0);
676
677 path->p_node[index].bh = eb_bh;
678 path->p_node[index].el = &eb->h_list;
679}
680
681static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh,
682 struct ocfs2_extent_list *root_el,
683 ocfs2_journal_access_func access)
684{
685 struct ocfs2_path *path;
686
687 BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH);
688
689 path = kzalloc_obj(*path, GFP_NOFS);
690 if (path) {
691 path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth);
692 get_bh(root_bh);
693 path_root_bh(path) = root_bh;
694 path_root_el(path) = root_el;
695 path_root_access(path) = access;
696 }
697
698 return path;
699}
700
701struct ocfs2_path *ocfs2_new_path_from_path(struct ocfs2_path *path)
702{
703 return ocfs2_new_path(path_root_bh(path), path_root_el(path),
704 path_root_access(path));
705}
706
707struct ocfs2_path *ocfs2_new_path_from_et(struct ocfs2_extent_tree *et)
708{
709 return ocfs2_new_path(et->et_root_bh, et->et_root_el,
710 et->et_root_journal_access);
711}
712
713/*
714 * Journal the buffer at depth idx. All idx>0 are extent_blocks,
715 * otherwise it's the root_access function.
716 *
717 * I don't like the way this function's name looks next to
718 * ocfs2_journal_access_path(), but I don't have a better one.
719 */
720int ocfs2_path_bh_journal_access(handle_t *handle,
721 struct ocfs2_caching_info *ci,
722 struct ocfs2_path *path,
723 int idx)
724{
725 ocfs2_journal_access_func access = path_root_access(path);
726
727 if (!access)
728 access = ocfs2_journal_access;
729
730 if (idx)
731 access = ocfs2_journal_access_eb;
732
733 return access(handle, ci, path->p_node[idx].bh,
734 OCFS2_JOURNAL_ACCESS_WRITE);
735}
736
737/*
738 * Convenience function to journal all components in a path.
739 */
740int ocfs2_journal_access_path(struct ocfs2_caching_info *ci,
741 handle_t *handle,
742 struct ocfs2_path *path)
743{
744 int i, ret = 0;
745
746 if (!path)
747 goto out;
748
749 for(i = 0; i < path_num_items(path); i++) {
750 ret = ocfs2_path_bh_journal_access(handle, ci, path, i);
751 if (ret < 0) {
752 mlog_errno(ret);
753 goto out;
754 }
755 }
756
757out:
758 return ret;
759}
760
761/*
762 * Return the index of the extent record which contains cluster #v_cluster.
763 * -1 is returned if it was not found.
764 *
765 * Should work fine on interior and exterior nodes.
766 */
767int ocfs2_search_extent_list(struct ocfs2_extent_list *el, u32 v_cluster)
768{
769 int ret = -1;
770 int i;
771 struct ocfs2_extent_rec *rec;
772 u32 rec_end, rec_start, clusters;
773
774 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
775 rec = &el->l_recs[i];
776
777 rec_start = le32_to_cpu(rec->e_cpos);
778 clusters = ocfs2_rec_clusters(el, rec);
779
780 rec_end = rec_start + clusters;
781
782 if (v_cluster >= rec_start && v_cluster < rec_end) {
783 ret = i;
784 break;
785 }
786 }
787
788 return ret;
789}
790
791/*
792 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
793 * ocfs2_extent_rec_contig only work properly against leaf nodes!
794 */
795static int ocfs2_block_extent_contig(struct super_block *sb,
796 struct ocfs2_extent_rec *ext,
797 u64 blkno)
798{
799 u64 blk_end = le64_to_cpu(ext->e_blkno);
800
801 blk_end += ocfs2_clusters_to_blocks(sb,
802 le16_to_cpu(ext->e_leaf_clusters));
803
804 return blkno == blk_end;
805}
806
807static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
808 struct ocfs2_extent_rec *right)
809{
810 u32 left_range;
811
812 left_range = le32_to_cpu(left->e_cpos) +
813 le16_to_cpu(left->e_leaf_clusters);
814
815 return (left_range == le32_to_cpu(right->e_cpos));
816}
817
818static enum ocfs2_contig_type
819 ocfs2_extent_rec_contig(struct super_block *sb,
820 struct ocfs2_extent_rec *ext,
821 struct ocfs2_extent_rec *insert_rec)
822{
823 u64 blkno = le64_to_cpu(insert_rec->e_blkno);
824
825 /*
826 * Refuse to coalesce extent records with different flag
827 * fields - we don't want to mix unwritten extents with user
828 * data.
829 */
830 if (ext->e_flags != insert_rec->e_flags)
831 return CONTIG_NONE;
832
833 if (ocfs2_extents_adjacent(ext, insert_rec) &&
834 ocfs2_block_extent_contig(sb, ext, blkno))
835 return CONTIG_RIGHT;
836
837 blkno = le64_to_cpu(ext->e_blkno);
838 if (ocfs2_extents_adjacent(insert_rec, ext) &&
839 ocfs2_block_extent_contig(sb, insert_rec, blkno))
840 return CONTIG_LEFT;
841
842 return CONTIG_NONE;
843}
844
845/*
846 * NOTE: We can have pretty much any combination of contiguousness and
847 * appending.
848 *
849 * The usefulness of APPEND_TAIL is more in that it lets us know that
850 * we'll have to update the path to that leaf.
851 */
852enum ocfs2_append_type {
853 APPEND_NONE = 0,
854 APPEND_TAIL,
855};
856
857enum ocfs2_split_type {
858 SPLIT_NONE = 0,
859 SPLIT_LEFT,
860 SPLIT_RIGHT,
861};
862
863struct ocfs2_insert_type {
864 enum ocfs2_split_type ins_split;
865 enum ocfs2_append_type ins_appending;
866 enum ocfs2_contig_type ins_contig;
867 int ins_contig_index;
868 int ins_tree_depth;
869};
870
871struct ocfs2_merge_ctxt {
872 enum ocfs2_contig_type c_contig_type;
873 int c_has_empty_extent;
874 int c_split_covers_rec;
875};
876
877static int ocfs2_validate_extent_block(struct super_block *sb,
878 struct buffer_head *bh)
879{
880 int rc;
881 struct ocfs2_extent_block *eb =
882 (struct ocfs2_extent_block *)bh->b_data;
883
884 trace_ocfs2_validate_extent_block((unsigned long long)bh->b_blocknr);
885
886 BUG_ON(!buffer_uptodate(bh));
887
888 /*
889 * If the ecc fails, we return the error but otherwise
890 * leave the filesystem running. We know any error is
891 * local to this block.
892 */
893 rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &eb->h_check);
894 if (rc) {
895 mlog(ML_ERROR, "Checksum failed for extent block %llu\n",
896 (unsigned long long)bh->b_blocknr);
897 return rc;
898 }
899
900 /*
901 * Errors after here are fatal.
902 */
903
904 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
905 rc = ocfs2_error(sb,
906 "Extent block #%llu has bad signature %.*s\n",
907 (unsigned long long)bh->b_blocknr, 7,
908 eb->h_signature);
909 goto bail;
910 }
911
912 if (le64_to_cpu(eb->h_blkno) != bh->b_blocknr) {
913 rc = ocfs2_error(sb,
914 "Extent block #%llu has an invalid h_blkno of %llu\n",
915 (unsigned long long)bh->b_blocknr,
916 (unsigned long long)le64_to_cpu(eb->h_blkno));
917 goto bail;
918 }
919
920 if (le32_to_cpu(eb->h_fs_generation) != OCFS2_SB(sb)->fs_generation) {
921 rc = ocfs2_error(sb,
922 "Extent block #%llu has an invalid h_fs_generation of #%u\n",
923 (unsigned long long)bh->b_blocknr,
924 le32_to_cpu(eb->h_fs_generation));
925 goto bail;
926 }
927
928 if (le16_to_cpu(eb->h_list.l_count) != ocfs2_extent_recs_per_eb(sb)) {
929 rc = ocfs2_error(sb,
930 "Extent block #%llu has invalid l_count %u (expected %u)\n",
931 (unsigned long long)bh->b_blocknr,
932 le16_to_cpu(eb->h_list.l_count),
933 ocfs2_extent_recs_per_eb(sb));
934 goto bail;
935 }
936
937 if (le16_to_cpu(eb->h_list.l_next_free_rec) > le16_to_cpu(eb->h_list.l_count)) {
938 rc = ocfs2_error(sb,
939 "Extent block #%llu has invalid l_next_free_rec %u (l_count %u)\n",
940 (unsigned long long)bh->b_blocknr,
941 le16_to_cpu(eb->h_list.l_next_free_rec),
942 le16_to_cpu(eb->h_list.l_count));
943 goto bail;
944 }
945
946bail:
947 return rc;
948}
949
950int ocfs2_read_extent_block(struct ocfs2_caching_info *ci, u64 eb_blkno,
951 struct buffer_head **bh)
952{
953 int rc;
954 struct buffer_head *tmp = *bh;
955
956 rc = ocfs2_read_block(ci, eb_blkno, &tmp,
957 ocfs2_validate_extent_block);
958
959 /* If ocfs2_read_block() got us a new bh, pass it up. */
960 if (!rc && !*bh)
961 *bh = tmp;
962
963 return rc;
964}
965
966
967/*
968 * How many free extents have we got before we need more meta data?
969 */
970int ocfs2_num_free_extents(struct ocfs2_extent_tree *et)
971{
972 int retval;
973 struct ocfs2_extent_list *el = NULL;
974 struct ocfs2_extent_block *eb;
975 struct buffer_head *eb_bh = NULL;
976 u64 last_eb_blk = 0;
977
978 el = et->et_root_el;
979 last_eb_blk = ocfs2_et_get_last_eb_blk(et);
980
981 if (last_eb_blk) {
982 retval = ocfs2_read_extent_block(et->et_ci, last_eb_blk,
983 &eb_bh);
984 if (retval < 0) {
985 mlog_errno(retval);
986 goto bail;
987 }
988 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
989 el = &eb->h_list;
990 }
991
992 if (el->l_tree_depth != 0) {
993 retval = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
994 "Owner %llu has leaf extent block %llu with an invalid l_tree_depth of %u\n",
995 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
996 (unsigned long long)last_eb_blk,
997 le16_to_cpu(el->l_tree_depth));
998 goto bail;
999 }
1000
1001 retval = le16_to_cpu(el->l_count) - le16_to_cpu(el->l_next_free_rec);
1002bail:
1003 brelse(eb_bh);
1004
1005 trace_ocfs2_num_free_extents(retval);
1006 return retval;
1007}
1008
1009/* expects array to already be allocated
1010 *
1011 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
1012 * l_count for you
1013 */
1014static int ocfs2_create_new_meta_bhs(handle_t *handle,
1015 struct ocfs2_extent_tree *et,
1016 int wanted,
1017 struct ocfs2_alloc_context *meta_ac,
1018 struct buffer_head *bhs[])
1019{
1020 int count, status, i;
1021 u16 suballoc_bit_start;
1022 u32 num_got;
1023 u64 suballoc_loc, first_blkno;
1024 struct ocfs2_super *osb =
1025 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
1026 struct ocfs2_extent_block *eb;
1027
1028 count = 0;
1029 while (count < wanted) {
1030 status = ocfs2_claim_metadata(handle,
1031 meta_ac,
1032 wanted - count,
1033 &suballoc_loc,
1034 &suballoc_bit_start,
1035 &num_got,
1036 &first_blkno);
1037 if (status < 0) {
1038 mlog_errno(status);
1039 goto bail;
1040 }
1041
1042 for(i = count; i < (num_got + count); i++) {
1043 bhs[i] = sb_getblk(osb->sb, first_blkno);
1044 if (bhs[i] == NULL) {
1045 status = -ENOMEM;
1046 mlog_errno(status);
1047 goto bail;
1048 }
1049 ocfs2_set_new_buffer_uptodate(et->et_ci, bhs[i]);
1050
1051 status = ocfs2_journal_access_eb(handle, et->et_ci,
1052 bhs[i],
1053 OCFS2_JOURNAL_ACCESS_CREATE);
1054 if (status < 0) {
1055 mlog_errno(status);
1056 goto bail;
1057 }
1058
1059 memset(bhs[i]->b_data, 0, osb->sb->s_blocksize);
1060 eb = (struct ocfs2_extent_block *) bhs[i]->b_data;
1061 /* Ok, setup the minimal stuff here. */
1062 strscpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
1063 eb->h_blkno = cpu_to_le64(first_blkno);
1064 eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
1065 eb->h_suballoc_slot =
1066 cpu_to_le16(meta_ac->ac_alloc_slot);
1067 eb->h_suballoc_loc = cpu_to_le64(suballoc_loc);
1068 eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1069 eb->h_list.l_count =
1070 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
1071
1072 suballoc_bit_start++;
1073 first_blkno++;
1074
1075 /* We'll also be dirtied by the caller, so
1076 * this isn't absolutely necessary. */
1077 ocfs2_journal_dirty(handle, bhs[i]);
1078 }
1079
1080 count += num_got;
1081 }
1082
1083 status = 0;
1084bail:
1085 if (status < 0) {
1086 for(i = 0; i < wanted; i++) {
1087 brelse(bhs[i]);
1088 bhs[i] = NULL;
1089 }
1090 }
1091 return status;
1092}
1093
1094/*
1095 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
1096 *
1097 * Returns the sum of the rightmost extent rec logical offset and
1098 * cluster count.
1099 *
1100 * ocfs2_add_branch() uses this to determine what logical cluster
1101 * value should be populated into the leftmost new branch records.
1102 *
1103 * ocfs2_shift_tree_depth() uses this to determine the # clusters
1104 * value for the new topmost tree record.
1105 */
1106static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list *el)
1107{
1108 int i;
1109
1110 i = le16_to_cpu(el->l_next_free_rec) - 1;
1111
1112 return le32_to_cpu(el->l_recs[i].e_cpos) +
1113 ocfs2_rec_clusters(el, &el->l_recs[i]);
1114}
1115
1116/*
1117 * Change range of the branches in the right most path according to the leaf
1118 * extent block's rightmost record.
1119 */
1120static int ocfs2_adjust_rightmost_branch(handle_t *handle,
1121 struct ocfs2_extent_tree *et)
1122{
1123 int status;
1124 struct ocfs2_path *path = NULL;
1125 struct ocfs2_extent_list *el;
1126 struct ocfs2_extent_rec *rec;
1127
1128 path = ocfs2_new_path_from_et(et);
1129 if (!path) {
1130 status = -ENOMEM;
1131 return status;
1132 }
1133
1134 status = ocfs2_find_path(et->et_ci, path, UINT_MAX);
1135 if (status < 0) {
1136 mlog_errno(status);
1137 goto out;
1138 }
1139
1140 status = ocfs2_extend_trans(handle, path_num_items(path));
1141 if (status < 0) {
1142 mlog_errno(status);
1143 goto out;
1144 }
1145
1146 status = ocfs2_journal_access_path(et->et_ci, handle, path);
1147 if (status < 0) {
1148 mlog_errno(status);
1149 goto out;
1150 }
1151
1152 el = path_leaf_el(path);
1153 rec = &el->l_recs[le16_to_cpu(el->l_next_free_rec) - 1];
1154
1155 ocfs2_adjust_rightmost_records(handle, et, path, rec);
1156
1157out:
1158 ocfs2_free_path(path);
1159 return status;
1160}
1161
1162/*
1163 * Add an entire tree branch to our inode. eb_bh is the extent block
1164 * to start at, if we don't want to start the branch at the root
1165 * structure.
1166 *
1167 * last_eb_bh is required as we have to update it's next_leaf pointer
1168 * for the new last extent block.
1169 *
1170 * the new branch will be 'empty' in the sense that every block will
1171 * contain a single record with cluster count == 0.
1172 */
1173static int ocfs2_add_branch(handle_t *handle,
1174 struct ocfs2_extent_tree *et,
1175 struct buffer_head *eb_bh,
1176 struct buffer_head **last_eb_bh,
1177 struct ocfs2_alloc_context *meta_ac)
1178{
1179 int status, new_blocks, i, block_given = 0;
1180 u64 next_blkno, new_last_eb_blk;
1181 struct buffer_head *bh;
1182 struct buffer_head **new_eb_bhs = NULL;
1183 struct ocfs2_extent_block *eb;
1184 struct ocfs2_extent_list *eb_el;
1185 struct ocfs2_extent_list *el;
1186 u32 new_cpos, root_end;
1187
1188 BUG_ON(!last_eb_bh || !*last_eb_bh);
1189
1190 if (eb_bh) {
1191 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1192 el = &eb->h_list;
1193 } else
1194 el = et->et_root_el;
1195
1196 /* we never add a branch to a leaf. */
1197 BUG_ON(!el->l_tree_depth);
1198
1199 new_blocks = le16_to_cpu(el->l_tree_depth);
1200
1201 eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
1202 new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
1203 root_end = ocfs2_sum_rightmost_rec(et->et_root_el);
1204
1205 /*
1206 * If there is a gap before the root end and the real end
1207 * of the rightmost leaf block, we need to remove the gap
1208 * between new_cpos and root_end first so that the tree
1209 * is consistent after we add a new branch(it will start
1210 * from new_cpos).
1211 */
1212 if (root_end > new_cpos) {
1213 trace_ocfs2_adjust_rightmost_branch(
1214 (unsigned long long)
1215 ocfs2_metadata_cache_owner(et->et_ci),
1216 root_end, new_cpos);
1217
1218 status = ocfs2_adjust_rightmost_branch(handle, et);
1219 if (status) {
1220 mlog_errno(status);
1221 goto bail;
1222 }
1223 }
1224
1225 /* allocate the number of new eb blocks we need */
1226 new_eb_bhs = kzalloc_objs(struct buffer_head *, new_blocks);
1227 if (!new_eb_bhs) {
1228 status = -ENOMEM;
1229 mlog_errno(status);
1230 goto bail;
1231 }
1232
1233 /* Firstyly, try to reuse dealloc since we have already estimated how
1234 * many extent blocks we may use.
1235 */
1236 if (!ocfs2_is_dealloc_empty(et)) {
1237 status = ocfs2_reuse_blk_from_dealloc(handle, et,
1238 new_eb_bhs, new_blocks,
1239 &block_given);
1240 if (status < 0) {
1241 mlog_errno(status);
1242 goto bail;
1243 }
1244 }
1245
1246 BUG_ON(block_given > new_blocks);
1247
1248 if (block_given < new_blocks) {
1249 BUG_ON(!meta_ac);
1250 status = ocfs2_create_new_meta_bhs(handle, et,
1251 new_blocks - block_given,
1252 meta_ac,
1253 &new_eb_bhs[block_given]);
1254 if (status < 0) {
1255 mlog_errno(status);
1256 goto bail;
1257 }
1258 }
1259
1260 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
1261 * linked with the rest of the tree.
1262 * conversely, new_eb_bhs[0] is the new bottommost leaf.
1263 *
1264 * when we leave the loop, new_last_eb_blk will point to the
1265 * newest leaf, and next_blkno will point to the topmost extent
1266 * block. */
1267 next_blkno = new_last_eb_blk = 0;
1268 for(i = 0; i < new_blocks; i++) {
1269 bh = new_eb_bhs[i];
1270 eb = (struct ocfs2_extent_block *) bh->b_data;
1271 /* ocfs2_create_new_meta_bhs() should create it right! */
1272 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1273 eb_el = &eb->h_list;
1274
1275 status = ocfs2_journal_access_eb(handle, et->et_ci, bh,
1276 OCFS2_JOURNAL_ACCESS_CREATE);
1277 if (status < 0) {
1278 mlog_errno(status);
1279 goto bail;
1280 }
1281
1282 eb->h_next_leaf_blk = 0;
1283 eb_el->l_tree_depth = cpu_to_le16(i);
1284 eb_el->l_next_free_rec = cpu_to_le16(1);
1285 /*
1286 * This actually counts as an empty extent as
1287 * c_clusters == 0
1288 */
1289 eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
1290 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
1291 /*
1292 * eb_el isn't always an interior node, but even leaf
1293 * nodes want a zero'd flags and reserved field so
1294 * this gets the whole 32 bits regardless of use.
1295 */
1296 eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
1297 if (!eb_el->l_tree_depth)
1298 new_last_eb_blk = le64_to_cpu(eb->h_blkno);
1299
1300 ocfs2_journal_dirty(handle, bh);
1301 next_blkno = le64_to_cpu(eb->h_blkno);
1302 }
1303
1304 /* This is a bit hairy. We want to update up to three blocks
1305 * here without leaving any of them in an inconsistent state
1306 * in case of error. We don't have to worry about
1307 * journal_dirty erroring as it won't unless we've aborted the
1308 * handle (in which case we would never be here) so reserving
1309 * the write with journal_access is all we need to do. */
1310 status = ocfs2_journal_access_eb(handle, et->et_ci, *last_eb_bh,
1311 OCFS2_JOURNAL_ACCESS_WRITE);
1312 if (status < 0) {
1313 mlog_errno(status);
1314 goto bail;
1315 }
1316 status = ocfs2_et_root_journal_access(handle, et,
1317 OCFS2_JOURNAL_ACCESS_WRITE);
1318 if (status < 0) {
1319 mlog_errno(status);
1320 goto bail;
1321 }
1322 if (eb_bh) {
1323 status = ocfs2_journal_access_eb(handle, et->et_ci, eb_bh,
1324 OCFS2_JOURNAL_ACCESS_WRITE);
1325 if (status < 0) {
1326 mlog_errno(status);
1327 goto bail;
1328 }
1329 }
1330
1331 /* Link the new branch into the rest of the tree (el will
1332 * either be on the root_bh, or the extent block passed in. */
1333 i = le16_to_cpu(el->l_next_free_rec);
1334 el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
1335 el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1336 el->l_recs[i].e_int_clusters = 0;
1337 le16_add_cpu(&el->l_next_free_rec, 1);
1338
1339 /* fe needs a new last extent block pointer, as does the
1340 * next_leaf on the previously last-extent-block. */
1341 ocfs2_et_set_last_eb_blk(et, new_last_eb_blk);
1342
1343 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
1344 eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);
1345
1346 ocfs2_journal_dirty(handle, *last_eb_bh);
1347 ocfs2_journal_dirty(handle, et->et_root_bh);
1348 if (eb_bh)
1349 ocfs2_journal_dirty(handle, eb_bh);
1350
1351 /*
1352 * Some callers want to track the rightmost leaf so pass it
1353 * back here.
1354 */
1355 brelse(*last_eb_bh);
1356 get_bh(new_eb_bhs[0]);
1357 *last_eb_bh = new_eb_bhs[0];
1358
1359 status = 0;
1360bail:
1361 if (new_eb_bhs) {
1362 for (i = 0; i < new_blocks; i++)
1363 brelse(new_eb_bhs[i]);
1364 kfree(new_eb_bhs);
1365 }
1366
1367 return status;
1368}
1369
1370/*
1371 * adds another level to the allocation tree.
1372 * returns back the new extent block so you can add a branch to it
1373 * after this call.
1374 */
1375static int ocfs2_shift_tree_depth(handle_t *handle,
1376 struct ocfs2_extent_tree *et,
1377 struct ocfs2_alloc_context *meta_ac,
1378 struct buffer_head **ret_new_eb_bh)
1379{
1380 int status, i, block_given = 0;
1381 u32 new_clusters;
1382 struct buffer_head *new_eb_bh = NULL;
1383 struct ocfs2_extent_block *eb;
1384 struct ocfs2_extent_list *root_el;
1385 struct ocfs2_extent_list *eb_el;
1386
1387 if (!ocfs2_is_dealloc_empty(et)) {
1388 status = ocfs2_reuse_blk_from_dealloc(handle, et,
1389 &new_eb_bh, 1,
1390 &block_given);
1391 } else if (meta_ac) {
1392 status = ocfs2_create_new_meta_bhs(handle, et, 1, meta_ac,
1393 &new_eb_bh);
1394
1395 } else {
1396 BUG();
1397 }
1398
1399 if (status < 0) {
1400 mlog_errno(status);
1401 goto bail;
1402 }
1403
1404 eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
1405 /* ocfs2_create_new_meta_bhs() should create it right! */
1406 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1407
1408 eb_el = &eb->h_list;
1409 root_el = et->et_root_el;
1410
1411 status = ocfs2_journal_access_eb(handle, et->et_ci, new_eb_bh,
1412 OCFS2_JOURNAL_ACCESS_CREATE);
1413 if (status < 0) {
1414 mlog_errno(status);
1415 goto bail;
1416 }
1417
1418 /* copy the root extent list data into the new extent block */
1419 eb_el->l_tree_depth = root_el->l_tree_depth;
1420 eb_el->l_next_free_rec = root_el->l_next_free_rec;
1421 for (i = 0; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1422 eb_el->l_recs[i] = root_el->l_recs[i];
1423
1424 ocfs2_journal_dirty(handle, new_eb_bh);
1425
1426 status = ocfs2_et_root_journal_access(handle, et,
1427 OCFS2_JOURNAL_ACCESS_WRITE);
1428 if (status < 0) {
1429 mlog_errno(status);
1430 goto bail;
1431 }
1432
1433 new_clusters = ocfs2_sum_rightmost_rec(eb_el);
1434
1435 /* update root_bh now */
1436 le16_add_cpu(&root_el->l_tree_depth, 1);
1437 root_el->l_recs[0].e_cpos = 0;
1438 root_el->l_recs[0].e_blkno = eb->h_blkno;
1439 root_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
1440 for (i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1441 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
1442 root_el->l_next_free_rec = cpu_to_le16(1);
1443
1444 /* If this is our 1st tree depth shift, then last_eb_blk
1445 * becomes the allocated extent block */
1446 if (root_el->l_tree_depth == cpu_to_le16(1))
1447 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
1448
1449 ocfs2_journal_dirty(handle, et->et_root_bh);
1450
1451 *ret_new_eb_bh = new_eb_bh;
1452 new_eb_bh = NULL;
1453 status = 0;
1454bail:
1455 brelse(new_eb_bh);
1456
1457 return status;
1458}
1459
1460/*
1461 * Should only be called when there is no space left in any of the
1462 * leaf nodes. What we want to do is find the lowest tree depth
1463 * non-leaf extent block with room for new records. There are three
1464 * valid results of this search:
1465 *
1466 * 1) a lowest extent block is found, then we pass it back in
1467 * *lowest_eb_bh and return '0'
1468 *
1469 * 2) the search fails to find anything, but the root_el has room. We
1470 * pass NULL back in *lowest_eb_bh, but still return '0'
1471 *
1472 * 3) the search fails to find anything AND the root_el is full, in
1473 * which case we return > 0
1474 *
1475 * return status < 0 indicates an error.
1476 */
1477static int ocfs2_find_branch_target(struct ocfs2_extent_tree *et,
1478 struct buffer_head **target_bh)
1479{
1480 int status = 0, i;
1481 u64 blkno;
1482 struct ocfs2_extent_block *eb;
1483 struct ocfs2_extent_list *el;
1484 struct buffer_head *bh = NULL;
1485 struct buffer_head *lowest_bh = NULL;
1486
1487 *target_bh = NULL;
1488
1489 el = et->et_root_el;
1490
1491 while(le16_to_cpu(el->l_tree_depth) > 1) {
1492 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1493 status = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1494 "Owner %llu has empty extent list (next_free_rec == 0)\n",
1495 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
1496 goto bail;
1497 }
1498 i = le16_to_cpu(el->l_next_free_rec) - 1;
1499 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1500 if (!blkno) {
1501 status = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1502 "Owner %llu has extent list where extent # %d has no physical block start\n",
1503 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), i);
1504 goto bail;
1505 }
1506
1507 brelse(bh);
1508 bh = NULL;
1509
1510 status = ocfs2_read_extent_block(et->et_ci, blkno, &bh);
1511 if (status < 0) {
1512 mlog_errno(status);
1513 goto bail;
1514 }
1515
1516 eb = (struct ocfs2_extent_block *) bh->b_data;
1517 el = &eb->h_list;
1518
1519 if (le16_to_cpu(el->l_next_free_rec) <
1520 le16_to_cpu(el->l_count)) {
1521 brelse(lowest_bh);
1522 lowest_bh = bh;
1523 get_bh(lowest_bh);
1524 }
1525 }
1526
1527 /* If we didn't find one and the fe doesn't have any room,
1528 * then return '1' */
1529 el = et->et_root_el;
1530 if (!lowest_bh && (el->l_next_free_rec == el->l_count))
1531 status = 1;
1532
1533 *target_bh = lowest_bh;
1534bail:
1535 brelse(bh);
1536
1537 return status;
1538}
1539
1540/*
1541 * Grow a b-tree so that it has more records.
1542 *
1543 * We might shift the tree depth in which case existing paths should
1544 * be considered invalid.
1545 *
1546 * Tree depth after the grow is returned via *final_depth.
1547 *
1548 * *last_eb_bh will be updated by ocfs2_add_branch().
1549 */
1550static int ocfs2_grow_tree(handle_t *handle, struct ocfs2_extent_tree *et,
1551 int *final_depth, struct buffer_head **last_eb_bh,
1552 struct ocfs2_alloc_context *meta_ac)
1553{
1554 int ret, shift;
1555 struct ocfs2_extent_list *el = et->et_root_el;
1556 int depth = le16_to_cpu(el->l_tree_depth);
1557 struct buffer_head *bh = NULL;
1558
1559 BUG_ON(meta_ac == NULL && ocfs2_is_dealloc_empty(et));
1560
1561 shift = ocfs2_find_branch_target(et, &bh);
1562 if (shift < 0) {
1563 ret = shift;
1564 mlog_errno(ret);
1565 goto out;
1566 }
1567
1568 /* We traveled all the way to the bottom of the allocation tree
1569 * and didn't find room for any more extents - we need to add
1570 * another tree level */
1571 if (shift) {
1572 BUG_ON(bh);
1573 trace_ocfs2_grow_tree(
1574 (unsigned long long)
1575 ocfs2_metadata_cache_owner(et->et_ci),
1576 depth);
1577
1578 /* ocfs2_shift_tree_depth will return us a buffer with
1579 * the new extent block (so we can pass that to
1580 * ocfs2_add_branch). */
1581 ret = ocfs2_shift_tree_depth(handle, et, meta_ac, &bh);
1582 if (ret < 0) {
1583 mlog_errno(ret);
1584 goto out;
1585 }
1586 depth++;
1587 if (depth == 1) {
1588 /*
1589 * Special case: we have room now if we shifted from
1590 * tree_depth 0, so no more work needs to be done.
1591 *
1592 * We won't be calling add_branch, so pass
1593 * back *last_eb_bh as the new leaf. At depth
1594 * zero, it should always be null so there's
1595 * no reason to brelse.
1596 */
1597 BUG_ON(*last_eb_bh);
1598 get_bh(bh);
1599 *last_eb_bh = bh;
1600 goto out;
1601 }
1602 }
1603
1604 /* call ocfs2_add_branch to add the final part of the tree with
1605 * the new data. */
1606 ret = ocfs2_add_branch(handle, et, bh, last_eb_bh,
1607 meta_ac);
1608 if (ret < 0)
1609 mlog_errno(ret);
1610
1611out:
1612 if (final_depth)
1613 *final_depth = depth;
1614 brelse(bh);
1615 return ret;
1616}
1617
1618/*
1619 * This function will discard the rightmost extent record.
1620 */
1621static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
1622{
1623 int next_free = le16_to_cpu(el->l_next_free_rec);
1624 int count = le16_to_cpu(el->l_count);
1625 unsigned int num_bytes;
1626
1627 BUG_ON(!next_free);
1628 /* This will cause us to go off the end of our extent list. */
1629 BUG_ON(next_free >= count);
1630
1631 num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
1632
1633 memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
1634}
1635
1636static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
1637 struct ocfs2_extent_rec *insert_rec)
1638{
1639 int i, insert_index, next_free, has_empty, num_bytes;
1640 u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
1641 struct ocfs2_extent_rec *rec;
1642
1643 next_free = le16_to_cpu(el->l_next_free_rec);
1644 has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
1645
1646 BUG_ON(!next_free);
1647
1648 /* The tree code before us didn't allow enough room in the leaf. */
1649 BUG_ON(el->l_next_free_rec == el->l_count && !has_empty);
1650
1651 /*
1652 * The easiest way to approach this is to just remove the
1653 * empty extent and temporarily decrement next_free.
1654 */
1655 if (has_empty) {
1656 /*
1657 * If next_free was 1 (only an empty extent), this
1658 * loop won't execute, which is fine. We still want
1659 * the decrement above to happen.
1660 */
1661 for(i = 0; i < (next_free - 1); i++)
1662 el->l_recs[i] = el->l_recs[i+1];
1663
1664 next_free--;
1665 }
1666
1667 /*
1668 * Figure out what the new record index should be.
1669 */
1670 for(i = 0; i < next_free; i++) {
1671 rec = &el->l_recs[i];
1672
1673 if (insert_cpos < le32_to_cpu(rec->e_cpos))
1674 break;
1675 }
1676 insert_index = i;
1677
1678 trace_ocfs2_rotate_leaf(insert_cpos, insert_index,
1679 has_empty, next_free,
1680 le16_to_cpu(el->l_count));
1681
1682 BUG_ON(insert_index < 0);
1683 BUG_ON(insert_index >= le16_to_cpu(el->l_count));
1684 BUG_ON(insert_index > next_free);
1685
1686 /*
1687 * No need to memmove if we're just adding to the tail.
1688 */
1689 if (insert_index != next_free) {
1690 BUG_ON(next_free >= le16_to_cpu(el->l_count));
1691
1692 num_bytes = next_free - insert_index;
1693 num_bytes *= sizeof(struct ocfs2_extent_rec);
1694 memmove(&el->l_recs[insert_index + 1],
1695 &el->l_recs[insert_index],
1696 num_bytes);
1697 }
1698
1699 /*
1700 * Either we had an empty extent, and need to re-increment or
1701 * there was no empty extent on a non full rightmost leaf node,
1702 * in which case we still need to increment.
1703 */
1704 next_free++;
1705 el->l_next_free_rec = cpu_to_le16(next_free);
1706 /*
1707 * Make sure none of the math above just messed up our tree.
1708 */
1709 BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
1710
1711 el->l_recs[insert_index] = *insert_rec;
1712
1713}
1714
1715static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
1716{
1717 int size, num_recs = le16_to_cpu(el->l_next_free_rec);
1718
1719 BUG_ON(num_recs == 0);
1720
1721 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
1722 num_recs--;
1723 size = num_recs * sizeof(struct ocfs2_extent_rec);
1724 memmove(&el->l_recs[0], &el->l_recs[1], size);
1725 memset(&el->l_recs[num_recs], 0,
1726 sizeof(struct ocfs2_extent_rec));
1727 el->l_next_free_rec = cpu_to_le16(num_recs);
1728 }
1729}
1730
1731/*
1732 * Create an empty extent record .
1733 *
1734 * l_next_free_rec may be updated.
1735 *
1736 * If an empty extent already exists do nothing.
1737 */
1738static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
1739{
1740 int next_free = le16_to_cpu(el->l_next_free_rec);
1741
1742 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1743
1744 if (next_free == 0)
1745 goto set_and_inc;
1746
1747 if (ocfs2_is_empty_extent(&el->l_recs[0]))
1748 return;
1749
1750 mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
1751 "Asked to create an empty extent in a full list:\n"
1752 "count = %u, tree depth = %u",
1753 le16_to_cpu(el->l_count),
1754 le16_to_cpu(el->l_tree_depth));
1755
1756 ocfs2_shift_records_right(el);
1757
1758set_and_inc:
1759 le16_add_cpu(&el->l_next_free_rec, 1);
1760 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1761}
1762
1763/*
1764 * For a rotation which involves two leaf nodes, the "root node" is
1765 * the lowest level tree node which contains a path to both leafs. This
1766 * resulting set of information can be used to form a complete "subtree"
1767 *
1768 * This function is passed two full paths from the dinode down to a
1769 * pair of adjacent leaves. It's task is to figure out which path
1770 * index contains the subtree root - this can be the root index itself
1771 * in a worst-case rotation.
1772 *
1773 * The array index of the subtree root is passed back.
1774 */
1775int ocfs2_find_subtree_root(struct ocfs2_extent_tree *et,
1776 struct ocfs2_path *left,
1777 struct ocfs2_path *right)
1778{
1779 int i = 0;
1780
1781 /*
1782 * Check that the caller passed in two paths from the same tree.
1783 */
1784 BUG_ON(path_root_bh(left) != path_root_bh(right));
1785
1786 do {
1787 i++;
1788
1789 /*
1790 * The caller didn't pass two adjacent paths.
1791 */
1792 mlog_bug_on_msg(i > left->p_tree_depth,
1793 "Owner %llu, left depth %u, right depth %u\n"
1794 "left leaf blk %llu, right leaf blk %llu\n",
1795 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
1796 left->p_tree_depth, right->p_tree_depth,
1797 (unsigned long long)path_leaf_bh(left)->b_blocknr,
1798 (unsigned long long)path_leaf_bh(right)->b_blocknr);
1799 } while (left->p_node[i].bh->b_blocknr ==
1800 right->p_node[i].bh->b_blocknr);
1801
1802 return i - 1;
1803}
1804
1805typedef void (path_insert_t)(void *, struct buffer_head *);
1806
1807/*
1808 * Traverse a btree path in search of cpos, starting at root_el.
1809 *
1810 * This code can be called with a cpos larger than the tree, in which
1811 * case it will return the rightmost path.
1812 */
1813static int __ocfs2_find_path(struct ocfs2_caching_info *ci,
1814 struct ocfs2_extent_list *root_el, u32 cpos,
1815 path_insert_t *func, void *data)
1816{
1817 int i, ret = 0;
1818 u32 range;
1819 u64 blkno;
1820 struct buffer_head *bh = NULL;
1821 struct ocfs2_extent_block *eb;
1822 struct ocfs2_extent_list *el;
1823 struct ocfs2_extent_rec *rec;
1824
1825 el = root_el;
1826 while (el->l_tree_depth) {
1827 if (unlikely(le16_to_cpu(el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH)) {
1828 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1829 "Owner %llu has invalid tree depth %u in extent list\n",
1830 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1831 le16_to_cpu(el->l_tree_depth));
1832 ret = -EROFS;
1833 goto out;
1834 }
1835 if (!el->l_next_free_rec || !el->l_count) {
1836 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1837 "Owner %llu has empty extent list at depth %u\n"
1838 "(next free=%u count=%u)\n",
1839 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1840 le16_to_cpu(el->l_tree_depth),
1841 le16_to_cpu(el->l_next_free_rec), le16_to_cpu(el->l_count));
1842 ret = -EROFS;
1843 goto out;
1844 }
1845
1846 for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
1847 rec = &el->l_recs[i];
1848
1849 /*
1850 * In the case that cpos is off the allocation
1851 * tree, this should just wind up returning the
1852 * rightmost record.
1853 */
1854 range = le32_to_cpu(rec->e_cpos) +
1855 ocfs2_rec_clusters(el, rec);
1856 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1857 break;
1858 }
1859
1860 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1861 if (blkno == 0) {
1862 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1863 "Owner %llu has bad blkno in extent list at depth %u (index %d)\n",
1864 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1865 le16_to_cpu(el->l_tree_depth), i);
1866 ret = -EROFS;
1867 goto out;
1868 }
1869
1870 brelse(bh);
1871 bh = NULL;
1872 ret = ocfs2_read_extent_block(ci, blkno, &bh);
1873 if (ret) {
1874 mlog_errno(ret);
1875 goto out;
1876 }
1877
1878 eb = (struct ocfs2_extent_block *) bh->b_data;
1879 el = &eb->h_list;
1880
1881 if (func)
1882 func(data, bh);
1883 }
1884
1885out:
1886 /*
1887 * Catch any trailing bh that the loop didn't handle.
1888 */
1889 brelse(bh);
1890
1891 return ret;
1892}
1893
1894/*
1895 * Given an initialized path (that is, it has a valid root extent
1896 * list), this function will traverse the btree in search of the path
1897 * which would contain cpos.
1898 *
1899 * The path traveled is recorded in the path structure.
1900 *
1901 * Note that this will not do any comparisons on leaf node extent
1902 * records, so it will work fine in the case that we just added a tree
1903 * branch.
1904 */
1905struct find_path_data {
1906 int index;
1907 struct ocfs2_path *path;
1908};
1909static void find_path_ins(void *data, struct buffer_head *bh)
1910{
1911 struct find_path_data *fp = data;
1912
1913 get_bh(bh);
1914 ocfs2_path_insert_eb(fp->path, fp->index, bh);
1915 fp->index++;
1916}
1917int ocfs2_find_path(struct ocfs2_caching_info *ci,
1918 struct ocfs2_path *path, u32 cpos)
1919{
1920 struct find_path_data data;
1921
1922 data.index = 1;
1923 data.path = path;
1924 return __ocfs2_find_path(ci, path_root_el(path), cpos,
1925 find_path_ins, &data);
1926}
1927
1928static void find_leaf_ins(void *data, struct buffer_head *bh)
1929{
1930 struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1931 struct ocfs2_extent_list *el = &eb->h_list;
1932 struct buffer_head **ret = data;
1933
1934 /* We want to retain only the leaf block. */
1935 if (le16_to_cpu(el->l_tree_depth) == 0) {
1936 get_bh(bh);
1937 *ret = bh;
1938 }
1939}
1940/*
1941 * Find the leaf block in the tree which would contain cpos. No
1942 * checking of the actual leaf is done.
1943 *
1944 * Some paths want to call this instead of allocating a path structure
1945 * and calling ocfs2_find_path().
1946 *
1947 * This function doesn't handle non btree extent lists.
1948 */
1949int ocfs2_find_leaf(struct ocfs2_caching_info *ci,
1950 struct ocfs2_extent_list *root_el, u32 cpos,
1951 struct buffer_head **leaf_bh)
1952{
1953 int ret;
1954 struct buffer_head *bh = NULL;
1955
1956 ret = __ocfs2_find_path(ci, root_el, cpos, find_leaf_ins, &bh);
1957 if (ret) {
1958 mlog_errno(ret);
1959 goto out;
1960 }
1961
1962 *leaf_bh = bh;
1963out:
1964 return ret;
1965}
1966
1967/*
1968 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1969 *
1970 * Basically, we've moved stuff around at the bottom of the tree and
1971 * we need to fix up the extent records above the changes to reflect
1972 * the new changes.
1973 *
1974 * left_rec: the record on the left.
1975 * right_rec: the record to the right of left_rec
1976 * right_child_el: is the child list pointed to by right_rec
1977 *
1978 * By definition, this only works on interior nodes.
1979 */
1980static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1981 struct ocfs2_extent_rec *right_rec,
1982 struct ocfs2_extent_list *right_child_el)
1983{
1984 u32 left_clusters, right_end;
1985
1986 /*
1987 * Interior nodes never have holes. Their cpos is the cpos of
1988 * the leftmost record in their child list. Their cluster
1989 * count covers the full theoretical range of their child list
1990 * - the range between their cpos and the cpos of the record
1991 * immediately to their right.
1992 */
1993 left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1994 if (!ocfs2_rec_clusters(right_child_el, &right_child_el->l_recs[0])) {
1995 BUG_ON(right_child_el->l_tree_depth);
1996 BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
1997 left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
1998 }
1999 left_clusters -= le32_to_cpu(left_rec->e_cpos);
2000 left_rec->e_int_clusters = cpu_to_le32(left_clusters);
2001
2002 /*
2003 * Calculate the rightmost cluster count boundary before
2004 * moving cpos - we will need to adjust clusters after
2005 * updating e_cpos to keep the same highest cluster count.
2006 */
2007 right_end = le32_to_cpu(right_rec->e_cpos);
2008 right_end += le32_to_cpu(right_rec->e_int_clusters);
2009
2010 right_rec->e_cpos = left_rec->e_cpos;
2011 le32_add_cpu(&right_rec->e_cpos, left_clusters);
2012
2013 right_end -= le32_to_cpu(right_rec->e_cpos);
2014 right_rec->e_int_clusters = cpu_to_le32(right_end);
2015}
2016
2017/*
2018 * Adjust the adjacent root node records involved in a
2019 * rotation. left_el_blkno is passed in as a key so that we can easily
2020 * find it's index in the root list.
2021 */
2022static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
2023 struct ocfs2_extent_list *left_el,
2024 struct ocfs2_extent_list *right_el,
2025 u64 left_el_blkno)
2026{
2027 int i;
2028
2029 BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
2030 le16_to_cpu(left_el->l_tree_depth));
2031
2032 for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
2033 if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
2034 break;
2035 }
2036
2037 /*
2038 * The path walking code should have never returned a root and
2039 * two paths which are not adjacent.
2040 */
2041 BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
2042
2043 ocfs2_adjust_adjacent_records(&root_el->l_recs[i],
2044 &root_el->l_recs[i + 1], right_el);
2045}
2046
2047/*
2048 * We've changed a leaf block (in right_path) and need to reflect that
2049 * change back up the subtree.
2050 *
2051 * This happens in multiple places:
2052 * - When we've moved an extent record from the left path leaf to the right
2053 * path leaf to make room for an empty extent in the left path leaf.
2054 * - When our insert into the right path leaf is at the leftmost edge
2055 * and requires an update of the path immediately to it's left. This
2056 * can occur at the end of some types of rotation and appending inserts.
2057 * - When we've adjusted the last extent record in the left path leaf and the
2058 * 1st extent record in the right path leaf during cross extent block merge.
2059 */
2060static void ocfs2_complete_edge_insert(handle_t *handle,
2061 struct ocfs2_path *left_path,
2062 struct ocfs2_path *right_path,
2063 int subtree_index)
2064{
2065 int i, idx;
2066 struct ocfs2_extent_list *el, *left_el, *right_el;
2067 struct ocfs2_extent_rec *left_rec, *right_rec;
2068 struct buffer_head *root_bh;
2069
2070 /*
2071 * Update the counts and position values within all the
2072 * interior nodes to reflect the leaf rotation we just did.
2073 *
2074 * The root node is handled below the loop.
2075 *
2076 * We begin the loop with right_el and left_el pointing to the
2077 * leaf lists and work our way up.
2078 *
2079 * NOTE: within this loop, left_el and right_el always refer
2080 * to the *child* lists.
2081 */
2082 left_el = path_leaf_el(left_path);
2083 right_el = path_leaf_el(right_path);
2084 for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
2085 trace_ocfs2_complete_edge_insert(i);
2086
2087 /*
2088 * One nice property of knowing that all of these
2089 * nodes are below the root is that we only deal with
2090 * the leftmost right node record and the rightmost
2091 * left node record.
2092 */
2093 el = left_path->p_node[i].el;
2094 idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
2095 left_rec = &el->l_recs[idx];
2096
2097 el = right_path->p_node[i].el;
2098 right_rec = &el->l_recs[0];
2099
2100 ocfs2_adjust_adjacent_records(left_rec, right_rec, right_el);
2101
2102 ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
2103 ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
2104
2105 /*
2106 * Setup our list pointers now so that the current
2107 * parents become children in the next iteration.
2108 */
2109 left_el = left_path->p_node[i].el;
2110 right_el = right_path->p_node[i].el;
2111 }
2112
2113 /*
2114 * At the root node, adjust the two adjacent records which
2115 * begin our path to the leaves.
2116 */
2117
2118 el = left_path->p_node[subtree_index].el;
2119 left_el = left_path->p_node[subtree_index + 1].el;
2120 right_el = right_path->p_node[subtree_index + 1].el;
2121
2122 ocfs2_adjust_root_records(el, left_el, right_el,
2123 left_path->p_node[subtree_index + 1].bh->b_blocknr);
2124
2125 root_bh = left_path->p_node[subtree_index].bh;
2126
2127 ocfs2_journal_dirty(handle, root_bh);
2128}
2129
2130static int ocfs2_rotate_subtree_right(handle_t *handle,
2131 struct ocfs2_extent_tree *et,
2132 struct ocfs2_path *left_path,
2133 struct ocfs2_path *right_path,
2134 int subtree_index)
2135{
2136 int ret, i;
2137 struct buffer_head *right_leaf_bh;
2138 struct buffer_head *left_leaf_bh = NULL;
2139 struct buffer_head *root_bh;
2140 struct ocfs2_extent_list *right_el, *left_el;
2141 struct ocfs2_extent_rec move_rec;
2142
2143 left_leaf_bh = path_leaf_bh(left_path);
2144 left_el = path_leaf_el(left_path);
2145
2146 if (left_el->l_next_free_rec != left_el->l_count) {
2147 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
2148 "Inode %llu has non-full interior leaf node %llu (next free = %u)\n",
2149 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2150 (unsigned long long)left_leaf_bh->b_blocknr,
2151 le16_to_cpu(left_el->l_next_free_rec));
2152 return -EROFS;
2153 }
2154
2155 /*
2156 * This extent block may already have an empty record, so we
2157 * return early if so.
2158 */
2159 if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
2160 return 0;
2161
2162 root_bh = left_path->p_node[subtree_index].bh;
2163 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2164
2165 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2166 subtree_index);
2167 if (ret) {
2168 mlog_errno(ret);
2169 goto out;
2170 }
2171
2172 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2173 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2174 right_path, i);
2175 if (ret) {
2176 mlog_errno(ret);
2177 goto out;
2178 }
2179
2180 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2181 left_path, i);
2182 if (ret) {
2183 mlog_errno(ret);
2184 goto out;
2185 }
2186 }
2187
2188 right_leaf_bh = path_leaf_bh(right_path);
2189 right_el = path_leaf_el(right_path);
2190
2191 /* This is a code error, not a disk corruption. */
2192 mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
2193 "because rightmost leaf block %llu is empty\n",
2194 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2195 (unsigned long long)right_leaf_bh->b_blocknr);
2196
2197 ocfs2_create_empty_extent(right_el);
2198
2199 ocfs2_journal_dirty(handle, right_leaf_bh);
2200
2201 /* Do the copy now. */
2202 i = le16_to_cpu(left_el->l_next_free_rec) - 1;
2203 move_rec = left_el->l_recs[i];
2204 right_el->l_recs[0] = move_rec;
2205
2206 /*
2207 * Clear out the record we just copied and shift everything
2208 * over, leaving an empty extent in the left leaf.
2209 *
2210 * We temporarily subtract from next_free_rec so that the
2211 * shift will lose the tail record (which is now defunct).
2212 */
2213 le16_add_cpu(&left_el->l_next_free_rec, -1);
2214 ocfs2_shift_records_right(left_el);
2215 memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2216 le16_add_cpu(&left_el->l_next_free_rec, 1);
2217
2218 ocfs2_journal_dirty(handle, left_leaf_bh);
2219
2220 ocfs2_complete_edge_insert(handle, left_path, right_path,
2221 subtree_index);
2222
2223out:
2224 return ret;
2225}
2226
2227/*
2228 * Given a full path, determine what cpos value would return us a path
2229 * containing the leaf immediately to the left of the current one.
2230 *
2231 * Will return zero if the path passed in is already the leftmost path.
2232 */
2233int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
2234 struct ocfs2_path *path, u32 *cpos)
2235{
2236 int i, j, ret = 0;
2237 u64 blkno;
2238 struct ocfs2_extent_list *el;
2239
2240 BUG_ON(path->p_tree_depth == 0);
2241
2242 *cpos = 0;
2243
2244 blkno = path_leaf_bh(path)->b_blocknr;
2245
2246 /* Start at the tree node just above the leaf and work our way up. */
2247 i = path->p_tree_depth - 1;
2248 while (i >= 0) {
2249 el = path->p_node[i].el;
2250
2251 /*
2252 * Find the extent record just before the one in our
2253 * path.
2254 */
2255 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2256 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2257 if (j == 0) {
2258 if (i == 0) {
2259 /*
2260 * We've determined that the
2261 * path specified is already
2262 * the leftmost one - return a
2263 * cpos of zero.
2264 */
2265 goto out;
2266 }
2267 /*
2268 * The leftmost record points to our
2269 * leaf - we need to travel up the
2270 * tree one level.
2271 */
2272 goto next_node;
2273 }
2274
2275 *cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
2276 *cpos = *cpos + ocfs2_rec_clusters(el,
2277 &el->l_recs[j - 1]);
2278 *cpos = *cpos - 1;
2279 goto out;
2280 }
2281 }
2282
2283 /*
2284 * If we got here, we never found a valid node where
2285 * the tree indicated one should be.
2286 */
2287 ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
2288 (unsigned long long)blkno);
2289 ret = -EROFS;
2290 goto out;
2291
2292next_node:
2293 blkno = path->p_node[i].bh->b_blocknr;
2294 i--;
2295 }
2296
2297out:
2298 return ret;
2299}
2300
2301/*
2302 * Extend the transaction by enough credits to complete the rotation,
2303 * and still leave at least the original number of credits allocated
2304 * to this transaction.
2305 */
2306static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
2307 int op_credits,
2308 struct ocfs2_path *path)
2309{
2310 int ret = 0;
2311 int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;
2312
2313 if (jbd2_handle_buffer_credits(handle) < credits)
2314 ret = ocfs2_extend_trans(handle,
2315 credits - jbd2_handle_buffer_credits(handle));
2316
2317 return ret;
2318}
2319
2320/*
2321 * Trap the case where we're inserting into the theoretical range past
2322 * the _actual_ left leaf range. Otherwise, we'll rotate a record
2323 * whose cpos is less than ours into the right leaf.
2324 *
2325 * It's only necessary to look at the rightmost record of the left
2326 * leaf because the logic that calls us should ensure that the
2327 * theoretical ranges in the path components above the leaves are
2328 * correct.
2329 */
2330static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
2331 u32 insert_cpos)
2332{
2333 struct ocfs2_extent_list *left_el;
2334 struct ocfs2_extent_rec *rec;
2335 int next_free;
2336
2337 left_el = path_leaf_el(left_path);
2338 next_free = le16_to_cpu(left_el->l_next_free_rec);
2339 rec = &left_el->l_recs[next_free - 1];
2340
2341 if (insert_cpos > le32_to_cpu(rec->e_cpos))
2342 return 1;
2343 return 0;
2344}
2345
2346static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
2347{
2348 int next_free = le16_to_cpu(el->l_next_free_rec);
2349 unsigned int range;
2350 struct ocfs2_extent_rec *rec;
2351
2352 if (next_free == 0)
2353 return 0;
2354
2355 rec = &el->l_recs[0];
2356 if (ocfs2_is_empty_extent(rec)) {
2357 /* Empty list. */
2358 if (next_free == 1)
2359 return 0;
2360 rec = &el->l_recs[1];
2361 }
2362
2363 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2364 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
2365 return 1;
2366 return 0;
2367}
2368
2369/*
2370 * Rotate all the records in a btree right one record, starting at insert_cpos.
2371 *
2372 * The path to the rightmost leaf should be passed in.
2373 *
2374 * The array is assumed to be large enough to hold an entire path (tree depth).
2375 *
2376 * Upon successful return from this function:
2377 *
2378 * - The 'right_path' array will contain a path to the leaf block
2379 * whose range contains e_cpos.
2380 * - That leaf block will have a single empty extent in list index 0.
2381 * - In the case that the rotation requires a post-insert update,
2382 * *ret_left_path will contain a valid path which can be passed to
2383 * ocfs2_insert_path().
2384 */
2385static int ocfs2_rotate_tree_right(handle_t *handle,
2386 struct ocfs2_extent_tree *et,
2387 enum ocfs2_split_type split,
2388 u32 insert_cpos,
2389 struct ocfs2_path *right_path,
2390 struct ocfs2_path **ret_left_path)
2391{
2392 int ret, start, orig_credits = jbd2_handle_buffer_credits(handle);
2393 u32 cpos;
2394 struct ocfs2_path *left_path = NULL;
2395 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2396
2397 *ret_left_path = NULL;
2398
2399 left_path = ocfs2_new_path_from_path(right_path);
2400 if (!left_path) {
2401 ret = -ENOMEM;
2402 mlog_errno(ret);
2403 goto out;
2404 }
2405
2406 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2407 if (ret) {
2408 mlog_errno(ret);
2409 goto out;
2410 }
2411
2412 trace_ocfs2_rotate_tree_right(
2413 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2414 insert_cpos, cpos);
2415
2416 /*
2417 * What we want to do here is:
2418 *
2419 * 1) Start with the rightmost path.
2420 *
2421 * 2) Determine a path to the leaf block directly to the left
2422 * of that leaf.
2423 *
2424 * 3) Determine the 'subtree root' - the lowest level tree node
2425 * which contains a path to both leaves.
2426 *
2427 * 4) Rotate the subtree.
2428 *
2429 * 5) Find the next subtree by considering the left path to be
2430 * the new right path.
2431 *
2432 * The check at the top of this while loop also accepts
2433 * insert_cpos == cpos because cpos is only a _theoretical_
2434 * value to get us the left path - insert_cpos might very well
2435 * be filling that hole.
2436 *
2437 * Stop at a cpos of '0' because we either started at the
2438 * leftmost branch (i.e., a tree with one branch and a
2439 * rotation inside of it), or we've gone as far as we can in
2440 * rotating subtrees.
2441 */
2442 while (cpos && insert_cpos <= cpos) {
2443 trace_ocfs2_rotate_tree_right(
2444 (unsigned long long)
2445 ocfs2_metadata_cache_owner(et->et_ci),
2446 insert_cpos, cpos);
2447
2448 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
2449 if (ret) {
2450 mlog_errno(ret);
2451 goto out;
2452 }
2453
2454 mlog_bug_on_msg(path_leaf_bh(left_path) ==
2455 path_leaf_bh(right_path),
2456 "Owner %llu: error during insert of %u "
2457 "(left path cpos %u) results in two identical "
2458 "paths ending at %llu\n",
2459 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2460 insert_cpos, cpos,
2461 (unsigned long long)
2462 path_leaf_bh(left_path)->b_blocknr);
2463
2464 if (split == SPLIT_NONE &&
2465 ocfs2_rotate_requires_path_adjustment(left_path,
2466 insert_cpos)) {
2467
2468 /*
2469 * We've rotated the tree as much as we
2470 * should. The rest is up to
2471 * ocfs2_insert_path() to complete, after the
2472 * record insertion. We indicate this
2473 * situation by returning the left path.
2474 *
2475 * The reason we don't adjust the records here
2476 * before the record insert is that an error
2477 * later might break the rule where a parent
2478 * record e_cpos will reflect the actual
2479 * e_cpos of the 1st nonempty record of the
2480 * child list.
2481 */
2482 *ret_left_path = left_path;
2483 goto out_ret_path;
2484 }
2485
2486 start = ocfs2_find_subtree_root(et, left_path, right_path);
2487
2488 trace_ocfs2_rotate_subtree(start,
2489 (unsigned long long)
2490 right_path->p_node[start].bh->b_blocknr,
2491 right_path->p_tree_depth);
2492
2493 ret = ocfs2_extend_rotate_transaction(handle, start,
2494 orig_credits, right_path);
2495 if (ret) {
2496 mlog_errno(ret);
2497 goto out;
2498 }
2499
2500 ret = ocfs2_rotate_subtree_right(handle, et, left_path,
2501 right_path, start);
2502 if (ret) {
2503 mlog_errno(ret);
2504 goto out;
2505 }
2506
2507 if (split != SPLIT_NONE &&
2508 ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
2509 insert_cpos)) {
2510 /*
2511 * A rotate moves the rightmost left leaf
2512 * record over to the leftmost right leaf
2513 * slot. If we're doing an extent split
2514 * instead of a real insert, then we have to
2515 * check that the extent to be split wasn't
2516 * just moved over. If it was, then we can
2517 * exit here, passing left_path back -
2518 * ocfs2_split_extent() is smart enough to
2519 * search both leaves.
2520 */
2521 *ret_left_path = left_path;
2522 goto out_ret_path;
2523 }
2524
2525 /*
2526 * There is no need to re-read the next right path
2527 * as we know that it'll be our current left
2528 * path. Optimize by copying values instead.
2529 */
2530 ocfs2_mv_path(right_path, left_path);
2531
2532 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2533 if (ret) {
2534 mlog_errno(ret);
2535 goto out;
2536 }
2537 }
2538
2539out:
2540 ocfs2_free_path(left_path);
2541
2542out_ret_path:
2543 return ret;
2544}
2545
2546static int ocfs2_update_edge_lengths(handle_t *handle,
2547 struct ocfs2_extent_tree *et,
2548 struct ocfs2_path *path)
2549{
2550 int i, idx, ret;
2551 struct ocfs2_extent_rec *rec;
2552 struct ocfs2_extent_list *el;
2553 struct ocfs2_extent_block *eb;
2554 u32 range;
2555
2556 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
2557 if (ret) {
2558 mlog_errno(ret);
2559 goto out;
2560 }
2561
2562 /* Path should always be rightmost. */
2563 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2564 BUG_ON(eb->h_next_leaf_blk != 0ULL);
2565
2566 el = &eb->h_list;
2567 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
2568 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2569 rec = &el->l_recs[idx];
2570 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2571
2572 for (i = 0; i < path->p_tree_depth; i++) {
2573 el = path->p_node[i].el;
2574 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2575 rec = &el->l_recs[idx];
2576
2577 rec->e_int_clusters = cpu_to_le32(range);
2578 le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));
2579
2580 ocfs2_journal_dirty(handle, path->p_node[i].bh);
2581 }
2582out:
2583 return ret;
2584}
2585
2586static void ocfs2_unlink_path(handle_t *handle,
2587 struct ocfs2_extent_tree *et,
2588 struct ocfs2_cached_dealloc_ctxt *dealloc,
2589 struct ocfs2_path *path, int unlink_start)
2590{
2591 int ret, i;
2592 struct ocfs2_extent_block *eb;
2593 struct ocfs2_extent_list *el;
2594 struct buffer_head *bh;
2595
2596 for(i = unlink_start; i < path_num_items(path); i++) {
2597 bh = path->p_node[i].bh;
2598
2599 eb = (struct ocfs2_extent_block *)bh->b_data;
2600 /*
2601 * Not all nodes might have had their final count
2602 * decremented by the caller - handle this here.
2603 */
2604 el = &eb->h_list;
2605 if (le16_to_cpu(el->l_next_free_rec) > 1) {
2606 mlog(ML_ERROR,
2607 "Inode %llu, attempted to remove extent block "
2608 "%llu with %u records\n",
2609 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2610 (unsigned long long)le64_to_cpu(eb->h_blkno),
2611 le16_to_cpu(el->l_next_free_rec));
2612
2613 ocfs2_journal_dirty(handle, bh);
2614 ocfs2_remove_from_cache(et->et_ci, bh);
2615 continue;
2616 }
2617
2618 el->l_next_free_rec = 0;
2619 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2620
2621 ocfs2_journal_dirty(handle, bh);
2622
2623 ret = ocfs2_cache_extent_block_free(dealloc, eb);
2624 if (ret)
2625 mlog_errno(ret);
2626
2627 ocfs2_remove_from_cache(et->et_ci, bh);
2628 }
2629}
2630
2631static void ocfs2_unlink_subtree(handle_t *handle,
2632 struct ocfs2_extent_tree *et,
2633 struct ocfs2_path *left_path,
2634 struct ocfs2_path *right_path,
2635 int subtree_index,
2636 struct ocfs2_cached_dealloc_ctxt *dealloc)
2637{
2638 int i;
2639 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2640 struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2641 struct ocfs2_extent_block *eb;
2642
2643 eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2644
2645 for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2646 if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2647 break;
2648
2649 BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2650
2651 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2652 le16_add_cpu(&root_el->l_next_free_rec, -1);
2653
2654 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2655 eb->h_next_leaf_blk = 0;
2656
2657 ocfs2_journal_dirty(handle, root_bh);
2658 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2659
2660 ocfs2_unlink_path(handle, et, dealloc, right_path,
2661 subtree_index + 1);
2662}
2663
2664static int ocfs2_rotate_subtree_left(handle_t *handle,
2665 struct ocfs2_extent_tree *et,
2666 struct ocfs2_path *left_path,
2667 struct ocfs2_path *right_path,
2668 int subtree_index,
2669 struct ocfs2_cached_dealloc_ctxt *dealloc,
2670 int *deleted)
2671{
2672 int ret, i, del_right_subtree = 0, right_has_empty = 0;
2673 struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path);
2674 struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2675 struct ocfs2_extent_block *eb;
2676
2677 *deleted = 0;
2678
2679 right_leaf_el = path_leaf_el(right_path);
2680 left_leaf_el = path_leaf_el(left_path);
2681 root_bh = left_path->p_node[subtree_index].bh;
2682 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2683
2684 if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2685 return 0;
2686
2687 eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2688 if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2689 /*
2690 * It's legal for us to proceed if the right leaf is
2691 * the rightmost one and it has an empty extent. There
2692 * are two cases to handle - whether the leaf will be
2693 * empty after removal or not. If the leaf isn't empty
2694 * then just remove the empty extent up front. The
2695 * next block will handle empty leaves by flagging
2696 * them for unlink.
2697 *
2698 * Non rightmost leaves will throw -EAGAIN and the
2699 * caller can manually move the subtree and retry.
2700 */
2701
2702 if (eb->h_next_leaf_blk != 0ULL)
2703 return -EAGAIN;
2704
2705 if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2706 ret = ocfs2_journal_access_eb(handle, et->et_ci,
2707 path_leaf_bh(right_path),
2708 OCFS2_JOURNAL_ACCESS_WRITE);
2709 if (ret) {
2710 mlog_errno(ret);
2711 goto out;
2712 }
2713
2714 ocfs2_remove_empty_extent(right_leaf_el);
2715 } else
2716 right_has_empty = 1;
2717 }
2718
2719 if (eb->h_next_leaf_blk == 0ULL &&
2720 le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2721 /*
2722 * We have to update i_last_eb_blk during the meta
2723 * data delete.
2724 */
2725 ret = ocfs2_et_root_journal_access(handle, et,
2726 OCFS2_JOURNAL_ACCESS_WRITE);
2727 if (ret) {
2728 mlog_errno(ret);
2729 goto out;
2730 }
2731
2732 del_right_subtree = 1;
2733 }
2734
2735 /*
2736 * Getting here with an empty extent in the right path implies
2737 * that it's the rightmost path and will be deleted.
2738 */
2739 BUG_ON(right_has_empty && !del_right_subtree);
2740
2741 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2742 subtree_index);
2743 if (ret) {
2744 mlog_errno(ret);
2745 goto out;
2746 }
2747
2748 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2749 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2750 right_path, i);
2751 if (ret) {
2752 mlog_errno(ret);
2753 goto out;
2754 }
2755
2756 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2757 left_path, i);
2758 if (ret) {
2759 mlog_errno(ret);
2760 goto out;
2761 }
2762 }
2763
2764 if (!right_has_empty) {
2765 /*
2766 * Only do this if we're moving a real
2767 * record. Otherwise, the action is delayed until
2768 * after removal of the right path in which case we
2769 * can do a simple shift to remove the empty extent.
2770 */
2771 ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2772 memset(&right_leaf_el->l_recs[0], 0,
2773 sizeof(struct ocfs2_extent_rec));
2774 }
2775 if (eb->h_next_leaf_blk == 0ULL) {
2776 /*
2777 * Move recs over to get rid of empty extent, decrease
2778 * next_free. This is allowed to remove the last
2779 * extent in our leaf (setting l_next_free_rec to
2780 * zero) - the delete code below won't care.
2781 */
2782 ocfs2_remove_empty_extent(right_leaf_el);
2783 }
2784
2785 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2786 ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2787
2788 if (del_right_subtree) {
2789 ocfs2_unlink_subtree(handle, et, left_path, right_path,
2790 subtree_index, dealloc);
2791 ret = ocfs2_update_edge_lengths(handle, et, left_path);
2792 if (ret) {
2793 mlog_errno(ret);
2794 goto out;
2795 }
2796
2797 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2798 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2799
2800 /*
2801 * Removal of the extent in the left leaf was skipped
2802 * above so we could delete the right path
2803 * 1st.
2804 */
2805 if (right_has_empty)
2806 ocfs2_remove_empty_extent(left_leaf_el);
2807
2808 ocfs2_journal_dirty(handle, et_root_bh);
2809
2810 *deleted = 1;
2811 } else
2812 ocfs2_complete_edge_insert(handle, left_path, right_path,
2813 subtree_index);
2814
2815out:
2816 return ret;
2817}
2818
2819/*
2820 * Given a full path, determine what cpos value would return us a path
2821 * containing the leaf immediately to the right of the current one.
2822 *
2823 * Will return zero if the path passed in is already the rightmost path.
2824 *
2825 * This looks similar, but is subtly different to
2826 * ocfs2_find_cpos_for_left_leaf().
2827 */
2828int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
2829 struct ocfs2_path *path, u32 *cpos)
2830{
2831 int i, j, ret = 0;
2832 u64 blkno;
2833 struct ocfs2_extent_list *el;
2834
2835 *cpos = 0;
2836
2837 if (path->p_tree_depth == 0)
2838 return 0;
2839
2840 blkno = path_leaf_bh(path)->b_blocknr;
2841
2842 /* Start at the tree node just above the leaf and work our way up. */
2843 i = path->p_tree_depth - 1;
2844 while (i >= 0) {
2845 int next_free;
2846
2847 el = path->p_node[i].el;
2848
2849 /*
2850 * Find the extent record just after the one in our
2851 * path.
2852 */
2853 next_free = le16_to_cpu(el->l_next_free_rec);
2854 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2855 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2856 if (j == (next_free - 1)) {
2857 if (i == 0) {
2858 /*
2859 * We've determined that the
2860 * path specified is already
2861 * the rightmost one - return a
2862 * cpos of zero.
2863 */
2864 goto out;
2865 }
2866 /*
2867 * The rightmost record points to our
2868 * leaf - we need to travel up the
2869 * tree one level.
2870 */
2871 goto next_node;
2872 }
2873
2874 *cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
2875 goto out;
2876 }
2877 }
2878
2879 /*
2880 * If we got here, we never found a valid node where
2881 * the tree indicated one should be.
2882 */
2883 ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
2884 (unsigned long long)blkno);
2885 ret = -EROFS;
2886 goto out;
2887
2888next_node:
2889 blkno = path->p_node[i].bh->b_blocknr;
2890 i--;
2891 }
2892
2893out:
2894 return ret;
2895}
2896
2897static int ocfs2_rotate_rightmost_leaf_left(handle_t *handle,
2898 struct ocfs2_extent_tree *et,
2899 struct ocfs2_path *path)
2900{
2901 int ret;
2902 struct buffer_head *bh = path_leaf_bh(path);
2903 struct ocfs2_extent_list *el = path_leaf_el(path);
2904
2905 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2906 return 0;
2907
2908 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
2909 path_num_items(path) - 1);
2910 if (ret) {
2911 mlog_errno(ret);
2912 goto out;
2913 }
2914
2915 ocfs2_remove_empty_extent(el);
2916 ocfs2_journal_dirty(handle, bh);
2917
2918out:
2919 return ret;
2920}
2921
2922static int __ocfs2_rotate_tree_left(handle_t *handle,
2923 struct ocfs2_extent_tree *et,
2924 int orig_credits,
2925 struct ocfs2_path *path,
2926 struct ocfs2_cached_dealloc_ctxt *dealloc,
2927 struct ocfs2_path **empty_extent_path)
2928{
2929 int ret, subtree_root, deleted;
2930 u32 right_cpos;
2931 struct ocfs2_path *left_path = NULL;
2932 struct ocfs2_path *right_path = NULL;
2933 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2934
2935 if (!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])))
2936 return 0;
2937
2938 *empty_extent_path = NULL;
2939
2940 ret = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
2941 if (ret) {
2942 mlog_errno(ret);
2943 goto out;
2944 }
2945
2946 left_path = ocfs2_new_path_from_path(path);
2947 if (!left_path) {
2948 ret = -ENOMEM;
2949 mlog_errno(ret);
2950 goto out;
2951 }
2952
2953 ocfs2_cp_path(left_path, path);
2954
2955 right_path = ocfs2_new_path_from_path(path);
2956 if (!right_path) {
2957 ret = -ENOMEM;
2958 mlog_errno(ret);
2959 goto out;
2960 }
2961
2962 while (right_cpos) {
2963 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
2964 if (ret) {
2965 mlog_errno(ret);
2966 goto out;
2967 }
2968
2969 subtree_root = ocfs2_find_subtree_root(et, left_path,
2970 right_path);
2971
2972 trace_ocfs2_rotate_subtree(subtree_root,
2973 (unsigned long long)
2974 right_path->p_node[subtree_root].bh->b_blocknr,
2975 right_path->p_tree_depth);
2976
2977 ret = ocfs2_extend_rotate_transaction(handle, 0,
2978 orig_credits, left_path);
2979 if (ret) {
2980 mlog_errno(ret);
2981 goto out;
2982 }
2983
2984 /*
2985 * Caller might still want to make changes to the
2986 * tree root, so re-add it to the journal here.
2987 */
2988 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2989 left_path, 0);
2990 if (ret) {
2991 mlog_errno(ret);
2992 goto out;
2993 }
2994
2995 ret = ocfs2_rotate_subtree_left(handle, et, left_path,
2996 right_path, subtree_root,
2997 dealloc, &deleted);
2998 if (ret == -EAGAIN) {
2999 /*
3000 * The rotation has to temporarily stop due to
3001 * the right subtree having an empty
3002 * extent. Pass it back to the caller for a
3003 * fixup.
3004 */
3005 *empty_extent_path = right_path;
3006 right_path = NULL;
3007 goto out;
3008 }
3009 if (ret) {
3010 mlog_errno(ret);
3011 goto out;
3012 }
3013
3014 /*
3015 * The subtree rotate might have removed records on
3016 * the rightmost edge. If so, then rotation is
3017 * complete.
3018 */
3019 if (deleted)
3020 break;
3021
3022 ocfs2_mv_path(left_path, right_path);
3023
3024 ret = ocfs2_find_cpos_for_right_leaf(sb, left_path,
3025 &right_cpos);
3026 if (ret) {
3027 mlog_errno(ret);
3028 goto out;
3029 }
3030 }
3031
3032out:
3033 ocfs2_free_path(right_path);
3034 ocfs2_free_path(left_path);
3035
3036 return ret;
3037}
3038
3039static int ocfs2_remove_rightmost_path(handle_t *handle,
3040 struct ocfs2_extent_tree *et,
3041 struct ocfs2_path *path,
3042 struct ocfs2_cached_dealloc_ctxt *dealloc)
3043{
3044 int ret, subtree_index;
3045 u32 cpos;
3046 struct ocfs2_path *left_path = NULL;
3047 struct ocfs2_extent_block *eb;
3048 struct ocfs2_extent_list *el;
3049
3050 ret = ocfs2_et_sanity_check(et);
3051 if (ret)
3052 goto out;
3053
3054 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
3055 if (ret) {
3056 mlog_errno(ret);
3057 goto out;
3058 }
3059
3060 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3061 path, &cpos);
3062 if (ret) {
3063 mlog_errno(ret);
3064 goto out;
3065 }
3066
3067 if (cpos) {
3068 /*
3069 * We have a path to the left of this one - it needs
3070 * an update too.
3071 */
3072 left_path = ocfs2_new_path_from_path(path);
3073 if (!left_path) {
3074 ret = -ENOMEM;
3075 mlog_errno(ret);
3076 goto out;
3077 }
3078
3079 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
3080 if (ret) {
3081 mlog_errno(ret);
3082 goto out;
3083 }
3084
3085 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
3086 if (ret) {
3087 mlog_errno(ret);
3088 goto out;
3089 }
3090
3091 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
3092
3093 ocfs2_unlink_subtree(handle, et, left_path, path,
3094 subtree_index, dealloc);
3095 ret = ocfs2_update_edge_lengths(handle, et, left_path);
3096 if (ret) {
3097 mlog_errno(ret);
3098 goto out;
3099 }
3100
3101 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
3102 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
3103 } else {
3104 /*
3105 * 'path' is also the leftmost path which
3106 * means it must be the only one. This gets
3107 * handled differently because we want to
3108 * revert the root back to having extents
3109 * in-line.
3110 */
3111 ocfs2_unlink_path(handle, et, dealloc, path, 1);
3112
3113 el = et->et_root_el;
3114 el->l_tree_depth = 0;
3115 el->l_next_free_rec = 0;
3116 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3117
3118 ocfs2_et_set_last_eb_blk(et, 0);
3119 }
3120
3121 ocfs2_journal_dirty(handle, path_root_bh(path));
3122
3123out:
3124 ocfs2_free_path(left_path);
3125 return ret;
3126}
3127
3128static int ocfs2_remove_rightmost_empty_extent(struct ocfs2_super *osb,
3129 struct ocfs2_extent_tree *et,
3130 struct ocfs2_path *path,
3131 struct ocfs2_cached_dealloc_ctxt *dealloc)
3132{
3133 handle_t *handle;
3134 int ret;
3135 int credits = path->p_tree_depth * 2 + 1;
3136
3137 handle = ocfs2_start_trans(osb, credits);
3138 if (IS_ERR(handle)) {
3139 ret = PTR_ERR(handle);
3140 mlog_errno(ret);
3141 return ret;
3142 }
3143
3144 ret = ocfs2_remove_rightmost_path(handle, et, path, dealloc);
3145 if (ret)
3146 mlog_errno(ret);
3147
3148 ocfs2_commit_trans(osb, handle);
3149 return ret;
3150}
3151
3152/*
3153 * Left rotation of btree records.
3154 *
3155 * In many ways, this is (unsurprisingly) the opposite of right
3156 * rotation. We start at some non-rightmost path containing an empty
3157 * extent in the leaf block. The code works its way to the rightmost
3158 * path by rotating records to the left in every subtree.
3159 *
3160 * This is used by any code which reduces the number of extent records
3161 * in a leaf. After removal, an empty record should be placed in the
3162 * leftmost list position.
3163 *
3164 * This won't handle a length update of the rightmost path records if
3165 * the rightmost tree leaf record is removed so the caller is
3166 * responsible for detecting and correcting that.
3167 */
3168static int ocfs2_rotate_tree_left(handle_t *handle,
3169 struct ocfs2_extent_tree *et,
3170 struct ocfs2_path *path,
3171 struct ocfs2_cached_dealloc_ctxt *dealloc)
3172{
3173 int ret, orig_credits = jbd2_handle_buffer_credits(handle);
3174 struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
3175 struct ocfs2_extent_block *eb;
3176 struct ocfs2_extent_list *el;
3177
3178 el = path_leaf_el(path);
3179 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
3180 return 0;
3181
3182 if (path->p_tree_depth == 0) {
3183rightmost_no_delete:
3184 /*
3185 * Inline extents. This is trivially handled, so do
3186 * it up front.
3187 */
3188 ret = ocfs2_rotate_rightmost_leaf_left(handle, et, path);
3189 if (ret)
3190 mlog_errno(ret);
3191 goto out;
3192 }
3193
3194 /*
3195 * Handle rightmost branch now. There's several cases:
3196 * 1) simple rotation leaving records in there. That's trivial.
3197 * 2) rotation requiring a branch delete - there's no more
3198 * records left. Two cases of this:
3199 * a) There are branches to the left.
3200 * b) This is also the leftmost (the only) branch.
3201 *
3202 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
3203 * 2a) we need the left branch so that we can update it with the unlink
3204 * 2b) we need to bring the root back to inline extents.
3205 */
3206
3207 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
3208 el = &eb->h_list;
3209 if (eb->h_next_leaf_blk == 0) {
3210 /*
3211 * This gets a bit tricky if we're going to delete the
3212 * rightmost path. Get the other cases out of the way
3213 * 1st.
3214 */
3215 if (le16_to_cpu(el->l_next_free_rec) > 1)
3216 goto rightmost_no_delete;
3217
3218 if (le16_to_cpu(el->l_next_free_rec) == 0) {
3219 ret = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3220 "Owner %llu has empty extent block at %llu\n",
3221 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
3222 (unsigned long long)le64_to_cpu(eb->h_blkno));
3223 goto out;
3224 }
3225
3226 /*
3227 * XXX: The caller can not trust "path" any more after
3228 * this as it will have been deleted. What do we do?
3229 *
3230 * In theory the rotate-for-merge code will never get
3231 * here because it'll always ask for a rotate in a
3232 * nonempty list.
3233 */
3234
3235 ret = ocfs2_remove_rightmost_path(handle, et, path,
3236 dealloc);
3237 if (ret)
3238 mlog_errno(ret);
3239 goto out;
3240 }
3241
3242 /*
3243 * Now we can loop, remembering the path we get from -EAGAIN
3244 * and restarting from there.
3245 */
3246try_rotate:
3247 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits, path,
3248 dealloc, &restart_path);
3249 if (ret && ret != -EAGAIN) {
3250 mlog_errno(ret);
3251 goto out;
3252 }
3253
3254 while (ret == -EAGAIN) {
3255 tmp_path = restart_path;
3256 restart_path = NULL;
3257
3258 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits,
3259 tmp_path, dealloc,
3260 &restart_path);
3261 if (ret && ret != -EAGAIN) {
3262 mlog_errno(ret);
3263 goto out;
3264 }
3265
3266 ocfs2_free_path(tmp_path);
3267 tmp_path = NULL;
3268
3269 if (ret == 0)
3270 goto try_rotate;
3271 }
3272
3273out:
3274 ocfs2_free_path(tmp_path);
3275 ocfs2_free_path(restart_path);
3276 return ret;
3277}
3278
3279static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
3280 int index)
3281{
3282 struct ocfs2_extent_rec *rec = &el->l_recs[index];
3283 unsigned int size;
3284
3285 if (rec->e_leaf_clusters == 0) {
3286 /*
3287 * We consumed all of the merged-from record. An empty
3288 * extent cannot exist anywhere but the 1st array
3289 * position, so move things over if the merged-from
3290 * record doesn't occupy that position.
3291 *
3292 * This creates a new empty extent so the caller
3293 * should be smart enough to have removed any existing
3294 * ones.
3295 */
3296 if (index > 0) {
3297 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3298 size = index * sizeof(struct ocfs2_extent_rec);
3299 memmove(&el->l_recs[1], &el->l_recs[0], size);
3300 }
3301
3302 /*
3303 * Always memset - the caller doesn't check whether it
3304 * created an empty extent, so there could be junk in
3305 * the other fields.
3306 */
3307 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3308 }
3309}
3310
3311static int ocfs2_get_right_path(struct ocfs2_extent_tree *et,
3312 struct ocfs2_path *left_path,
3313 struct ocfs2_path **ret_right_path)
3314{
3315 int ret;
3316 u32 right_cpos;
3317 struct ocfs2_path *right_path = NULL;
3318 struct ocfs2_extent_list *left_el;
3319
3320 *ret_right_path = NULL;
3321
3322 /* This function shouldn't be called for non-trees. */
3323 BUG_ON(left_path->p_tree_depth == 0);
3324
3325 left_el = path_leaf_el(left_path);
3326 BUG_ON(left_el->l_next_free_rec != left_el->l_count);
3327
3328 ret = ocfs2_find_cpos_for_right_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3329 left_path, &right_cpos);
3330 if (ret) {
3331 mlog_errno(ret);
3332 goto out;
3333 }
3334
3335 /* This function shouldn't be called for the rightmost leaf. */
3336 BUG_ON(right_cpos == 0);
3337
3338 right_path = ocfs2_new_path_from_path(left_path);
3339 if (!right_path) {
3340 ret = -ENOMEM;
3341 mlog_errno(ret);
3342 goto out;
3343 }
3344
3345 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
3346 if (ret) {
3347 mlog_errno(ret);
3348 goto out;
3349 }
3350
3351 *ret_right_path = right_path;
3352out:
3353 if (ret)
3354 ocfs2_free_path(right_path);
3355 return ret;
3356}
3357
3358/*
3359 * Remove split_rec clusters from the record at index and merge them
3360 * onto the beginning of the record "next" to it.
3361 * For index < l_count - 1, the next means the extent rec at index + 1.
3362 * For index == l_count - 1, the "next" means the 1st extent rec of the
3363 * next extent block.
3364 */
3365static int ocfs2_merge_rec_right(struct ocfs2_path *left_path,
3366 handle_t *handle,
3367 struct ocfs2_extent_tree *et,
3368 struct ocfs2_extent_rec *split_rec,
3369 int index)
3370{
3371 int ret, next_free, i;
3372 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3373 struct ocfs2_extent_rec *left_rec;
3374 struct ocfs2_extent_rec *right_rec;
3375 struct ocfs2_extent_list *right_el;
3376 struct ocfs2_path *right_path = NULL;
3377 int subtree_index = 0;
3378 struct ocfs2_extent_list *el = path_leaf_el(left_path);
3379 struct buffer_head *bh = path_leaf_bh(left_path);
3380 struct buffer_head *root_bh = NULL;
3381
3382 BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
3383 left_rec = &el->l_recs[index];
3384
3385 if (index == le16_to_cpu(el->l_next_free_rec) - 1 &&
3386 le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count)) {
3387 /* we meet with a cross extent block merge. */
3388 ret = ocfs2_get_right_path(et, left_path, &right_path);
3389 if (ret) {
3390 mlog_errno(ret);
3391 return ret;
3392 }
3393
3394 right_el = path_leaf_el(right_path);
3395 next_free = le16_to_cpu(right_el->l_next_free_rec);
3396 BUG_ON(next_free <= 0);
3397 right_rec = &right_el->l_recs[0];
3398 if (ocfs2_is_empty_extent(right_rec)) {
3399 BUG_ON(next_free <= 1);
3400 right_rec = &right_el->l_recs[1];
3401 }
3402
3403 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3404 le16_to_cpu(left_rec->e_leaf_clusters) !=
3405 le32_to_cpu(right_rec->e_cpos));
3406
3407 subtree_index = ocfs2_find_subtree_root(et, left_path,
3408 right_path);
3409
3410 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3411 jbd2_handle_buffer_credits(handle),
3412 right_path);
3413 if (ret) {
3414 mlog_errno(ret);
3415 goto out;
3416 }
3417
3418 root_bh = left_path->p_node[subtree_index].bh;
3419 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3420
3421 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3422 subtree_index);
3423 if (ret) {
3424 mlog_errno(ret);
3425 goto out;
3426 }
3427
3428 for (i = subtree_index + 1;
3429 i < path_num_items(right_path); i++) {
3430 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3431 right_path, i);
3432 if (ret) {
3433 mlog_errno(ret);
3434 goto out;
3435 }
3436
3437 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3438 left_path, i);
3439 if (ret) {
3440 mlog_errno(ret);
3441 goto out;
3442 }
3443 }
3444
3445 } else {
3446 BUG_ON(index == le16_to_cpu(el->l_next_free_rec) - 1);
3447 right_rec = &el->l_recs[index + 1];
3448 }
3449
3450 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, left_path,
3451 path_num_items(left_path) - 1);
3452 if (ret) {
3453 mlog_errno(ret);
3454 goto out;
3455 }
3456
3457 le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
3458
3459 le32_add_cpu(&right_rec->e_cpos, -split_clusters);
3460 le64_add_cpu(&right_rec->e_blkno,
3461 -ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3462 split_clusters));
3463 le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
3464
3465 ocfs2_cleanup_merge(el, index);
3466
3467 ocfs2_journal_dirty(handle, bh);
3468 if (right_path) {
3469 ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
3470 ocfs2_complete_edge_insert(handle, left_path, right_path,
3471 subtree_index);
3472 }
3473out:
3474 ocfs2_free_path(right_path);
3475 return ret;
3476}
3477
3478static int ocfs2_get_left_path(struct ocfs2_extent_tree *et,
3479 struct ocfs2_path *right_path,
3480 struct ocfs2_path **ret_left_path)
3481{
3482 int ret;
3483 u32 left_cpos;
3484 struct ocfs2_path *left_path = NULL;
3485
3486 *ret_left_path = NULL;
3487
3488 /* This function shouldn't be called for non-trees. */
3489 BUG_ON(right_path->p_tree_depth == 0);
3490
3491 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3492 right_path, &left_cpos);
3493 if (ret) {
3494 mlog_errno(ret);
3495 goto out;
3496 }
3497
3498 /* This function shouldn't be called for the leftmost leaf. */
3499 BUG_ON(left_cpos == 0);
3500
3501 left_path = ocfs2_new_path_from_path(right_path);
3502 if (!left_path) {
3503 ret = -ENOMEM;
3504 mlog_errno(ret);
3505 goto out;
3506 }
3507
3508 ret = ocfs2_find_path(et->et_ci, left_path, left_cpos);
3509 if (ret) {
3510 mlog_errno(ret);
3511 goto out;
3512 }
3513
3514 *ret_left_path = left_path;
3515out:
3516 if (ret)
3517 ocfs2_free_path(left_path);
3518 return ret;
3519}
3520
3521/*
3522 * Remove split_rec clusters from the record at index and merge them
3523 * onto the tail of the record "before" it.
3524 * For index > 0, the "before" means the extent rec at index - 1.
3525 *
3526 * For index == 0, the "before" means the last record of the previous
3527 * extent block. And there is also a situation that we may need to
3528 * remove the rightmost leaf extent block in the right_path and change
3529 * the right path to indicate the new rightmost path.
3530 */
3531static int ocfs2_merge_rec_left(struct ocfs2_path *right_path,
3532 handle_t *handle,
3533 struct ocfs2_extent_tree *et,
3534 struct ocfs2_extent_rec *split_rec,
3535 struct ocfs2_cached_dealloc_ctxt *dealloc,
3536 int index)
3537{
3538 int ret, i, subtree_index = 0, has_empty_extent = 0;
3539 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3540 struct ocfs2_extent_rec *left_rec;
3541 struct ocfs2_extent_rec *right_rec;
3542 struct ocfs2_extent_list *el = path_leaf_el(right_path);
3543 struct buffer_head *bh = path_leaf_bh(right_path);
3544 struct buffer_head *root_bh = NULL;
3545 struct ocfs2_path *left_path = NULL;
3546 struct ocfs2_extent_list *left_el;
3547
3548 BUG_ON(index < 0);
3549
3550 right_rec = &el->l_recs[index];
3551 if (index == 0) {
3552 /* we meet with a cross extent block merge. */
3553 ret = ocfs2_get_left_path(et, right_path, &left_path);
3554 if (ret) {
3555 mlog_errno(ret);
3556 return ret;
3557 }
3558
3559 left_el = path_leaf_el(left_path);
3560 BUG_ON(le16_to_cpu(left_el->l_next_free_rec) !=
3561 le16_to_cpu(left_el->l_count));
3562
3563 left_rec = &left_el->l_recs[
3564 le16_to_cpu(left_el->l_next_free_rec) - 1];
3565 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3566 le16_to_cpu(left_rec->e_leaf_clusters) !=
3567 le32_to_cpu(split_rec->e_cpos));
3568
3569 subtree_index = ocfs2_find_subtree_root(et, left_path,
3570 right_path);
3571
3572 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3573 jbd2_handle_buffer_credits(handle),
3574 left_path);
3575 if (ret) {
3576 mlog_errno(ret);
3577 goto out;
3578 }
3579
3580 root_bh = left_path->p_node[subtree_index].bh;
3581 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3582
3583 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3584 subtree_index);
3585 if (ret) {
3586 mlog_errno(ret);
3587 goto out;
3588 }
3589
3590 for (i = subtree_index + 1;
3591 i < path_num_items(right_path); i++) {
3592 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3593 right_path, i);
3594 if (ret) {
3595 mlog_errno(ret);
3596 goto out;
3597 }
3598
3599 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3600 left_path, i);
3601 if (ret) {
3602 mlog_errno(ret);
3603 goto out;
3604 }
3605 }
3606 } else {
3607 left_rec = &el->l_recs[index - 1];
3608 if (ocfs2_is_empty_extent(&el->l_recs[0]))
3609 has_empty_extent = 1;
3610 }
3611
3612 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3613 path_num_items(right_path) - 1);
3614 if (ret) {
3615 mlog_errno(ret);
3616 goto out;
3617 }
3618
3619 if (has_empty_extent && index == 1) {
3620 /*
3621 * The easy case - we can just plop the record right in.
3622 */
3623 *left_rec = *split_rec;
3624 } else
3625 le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
3626
3627 le32_add_cpu(&right_rec->e_cpos, split_clusters);
3628 le64_add_cpu(&right_rec->e_blkno,
3629 ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3630 split_clusters));
3631 le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
3632
3633 ocfs2_cleanup_merge(el, index);
3634
3635 ocfs2_journal_dirty(handle, bh);
3636 if (left_path) {
3637 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
3638
3639 /*
3640 * In the situation that the right_rec is empty and the extent
3641 * block is empty also, ocfs2_complete_edge_insert can't handle
3642 * it and we need to delete the right extent block.
3643 */
3644 if (le16_to_cpu(right_rec->e_leaf_clusters) == 0 &&
3645 le16_to_cpu(el->l_next_free_rec) == 1) {
3646 /* extend credit for ocfs2_remove_rightmost_path */
3647 ret = ocfs2_extend_rotate_transaction(handle, 0,
3648 jbd2_handle_buffer_credits(handle),
3649 right_path);
3650 if (ret) {
3651 mlog_errno(ret);
3652 goto out;
3653 }
3654
3655 ret = ocfs2_remove_rightmost_path(handle, et,
3656 right_path,
3657 dealloc);
3658 if (ret) {
3659 mlog_errno(ret);
3660 goto out;
3661 }
3662
3663 /* Now the rightmost extent block has been deleted.
3664 * So we use the new rightmost path.
3665 */
3666 ocfs2_mv_path(right_path, left_path);
3667 } else
3668 ocfs2_complete_edge_insert(handle, left_path,
3669 right_path, subtree_index);
3670 }
3671out:
3672 ocfs2_free_path(left_path);
3673 return ret;
3674}
3675
3676static int ocfs2_try_to_merge_extent(handle_t *handle,
3677 struct ocfs2_extent_tree *et,
3678 struct ocfs2_path *path,
3679 int split_index,
3680 struct ocfs2_extent_rec *split_rec,
3681 struct ocfs2_cached_dealloc_ctxt *dealloc,
3682 struct ocfs2_merge_ctxt *ctxt)
3683{
3684 int ret = 0;
3685 struct ocfs2_extent_list *el = path_leaf_el(path);
3686 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3687
3688 BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
3689
3690 if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) {
3691 /* extend credit for ocfs2_remove_rightmost_path */
3692 ret = ocfs2_extend_rotate_transaction(handle, 0,
3693 jbd2_handle_buffer_credits(handle),
3694 path);
3695 if (ret) {
3696 mlog_errno(ret);
3697 goto out;
3698 }
3699 /*
3700 * The merge code will need to create an empty
3701 * extent to take the place of the newly
3702 * emptied slot. Remove any pre-existing empty
3703 * extents - having more than one in a leaf is
3704 * illegal.
3705 */
3706 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3707 if (ret) {
3708 mlog_errno(ret);
3709 goto out;
3710 }
3711 split_index--;
3712 rec = &el->l_recs[split_index];
3713 }
3714
3715 if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
3716 /*
3717 * Left-right contig implies this.
3718 */
3719 BUG_ON(!ctxt->c_split_covers_rec);
3720
3721 /*
3722 * Since the leftright insert always covers the entire
3723 * extent, this call will delete the insert record
3724 * entirely, resulting in an empty extent record added to
3725 * the extent block.
3726 *
3727 * Since the adding of an empty extent shifts
3728 * everything back to the right, there's no need to
3729 * update split_index here.
3730 *
3731 * When the split_index is zero, we need to merge it to the
3732 * previous extent block. It is more efficient and easier
3733 * if we do merge_right first and merge_left later.
3734 */
3735 ret = ocfs2_merge_rec_right(path, handle, et, split_rec,
3736 split_index);
3737 if (ret) {
3738 mlog_errno(ret);
3739 goto out;
3740 }
3741
3742 /*
3743 * We can only get this from logic error above.
3744 */
3745 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
3746
3747 /* extend credit for ocfs2_remove_rightmost_path */
3748 ret = ocfs2_extend_rotate_transaction(handle, 0,
3749 jbd2_handle_buffer_credits(handle),
3750 path);
3751 if (ret) {
3752 mlog_errno(ret);
3753 goto out;
3754 }
3755
3756 /* The merge left us with an empty extent, remove it. */
3757 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3758 if (ret) {
3759 mlog_errno(ret);
3760 goto out;
3761 }
3762
3763 rec = &el->l_recs[split_index];
3764
3765 /*
3766 * Note that we don't pass split_rec here on purpose -
3767 * we've merged it into the rec already.
3768 */
3769 ret = ocfs2_merge_rec_left(path, handle, et, rec,
3770 dealloc, split_index);
3771
3772 if (ret) {
3773 mlog_errno(ret);
3774 goto out;
3775 }
3776
3777 /* extend credit for ocfs2_remove_rightmost_path */
3778 ret = ocfs2_extend_rotate_transaction(handle, 0,
3779 jbd2_handle_buffer_credits(handle),
3780 path);
3781 if (ret) {
3782 mlog_errno(ret);
3783 goto out;
3784 }
3785
3786 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3787 /*
3788 * Error from this last rotate is not critical, so
3789 * print but don't bubble it up.
3790 */
3791 if (ret)
3792 mlog_errno(ret);
3793 ret = 0;
3794 } else {
3795 /*
3796 * Merge a record to the left or right.
3797 *
3798 * 'contig_type' is relative to the existing record,
3799 * so for example, if we're "right contig", it's to
3800 * the record on the left (hence the left merge).
3801 */
3802 if (ctxt->c_contig_type == CONTIG_RIGHT) {
3803 ret = ocfs2_merge_rec_left(path, handle, et,
3804 split_rec, dealloc,
3805 split_index);
3806 if (ret) {
3807 mlog_errno(ret);
3808 goto out;
3809 }
3810 } else {
3811 ret = ocfs2_merge_rec_right(path, handle,
3812 et, split_rec,
3813 split_index);
3814 if (ret) {
3815 mlog_errno(ret);
3816 goto out;
3817 }
3818 }
3819
3820 if (ctxt->c_split_covers_rec) {
3821 /* extend credit for ocfs2_remove_rightmost_path */
3822 ret = ocfs2_extend_rotate_transaction(handle, 0,
3823 jbd2_handle_buffer_credits(handle),
3824 path);
3825 if (ret) {
3826 mlog_errno(ret);
3827 ret = 0;
3828 goto out;
3829 }
3830
3831 /*
3832 * The merge may have left an empty extent in
3833 * our leaf. Try to rotate it away.
3834 */
3835 ret = ocfs2_rotate_tree_left(handle, et, path,
3836 dealloc);
3837 if (ret)
3838 mlog_errno(ret);
3839 ret = 0;
3840 }
3841 }
3842
3843out:
3844 return ret;
3845}
3846
3847static void ocfs2_subtract_from_rec(struct super_block *sb,
3848 enum ocfs2_split_type split,
3849 struct ocfs2_extent_rec *rec,
3850 struct ocfs2_extent_rec *split_rec)
3851{
3852 u64 len_blocks;
3853
3854 len_blocks = ocfs2_clusters_to_blocks(sb,
3855 le16_to_cpu(split_rec->e_leaf_clusters));
3856
3857 if (split == SPLIT_LEFT) {
3858 /*
3859 * Region is on the left edge of the existing
3860 * record.
3861 */
3862 le32_add_cpu(&rec->e_cpos,
3863 le16_to_cpu(split_rec->e_leaf_clusters));
3864 le64_add_cpu(&rec->e_blkno, len_blocks);
3865 le16_add_cpu(&rec->e_leaf_clusters,
3866 -le16_to_cpu(split_rec->e_leaf_clusters));
3867 } else {
3868 /*
3869 * Region is on the right edge of the existing
3870 * record.
3871 */
3872 le16_add_cpu(&rec->e_leaf_clusters,
3873 -le16_to_cpu(split_rec->e_leaf_clusters));
3874 }
3875}
3876
3877/*
3878 * Do the final bits of extent record insertion at the target leaf
3879 * list. If this leaf is part of an allocation tree, it is assumed
3880 * that the tree above has been prepared.
3881 */
3882static void ocfs2_insert_at_leaf(struct ocfs2_extent_tree *et,
3883 struct ocfs2_extent_rec *insert_rec,
3884 struct ocfs2_extent_list *el,
3885 struct ocfs2_insert_type *insert)
3886{
3887 int i = insert->ins_contig_index;
3888 unsigned int range;
3889 struct ocfs2_extent_rec *rec;
3890
3891 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3892
3893 if (insert->ins_split != SPLIT_NONE) {
3894 i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
3895 BUG_ON(i == -1);
3896 rec = &el->l_recs[i];
3897 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
3898 insert->ins_split, rec,
3899 insert_rec);
3900 goto rotate;
3901 }
3902
3903 /*
3904 * Contiguous insert - either left or right.
3905 */
3906 if (insert->ins_contig != CONTIG_NONE) {
3907 rec = &el->l_recs[i];
3908 if (insert->ins_contig == CONTIG_LEFT) {
3909 rec->e_blkno = insert_rec->e_blkno;
3910 rec->e_cpos = insert_rec->e_cpos;
3911 }
3912 le16_add_cpu(&rec->e_leaf_clusters,
3913 le16_to_cpu(insert_rec->e_leaf_clusters));
3914 return;
3915 }
3916
3917 /*
3918 * Handle insert into an empty leaf.
3919 */
3920 if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3921 ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3922 ocfs2_is_empty_extent(&el->l_recs[0]))) {
3923 el->l_recs[0] = *insert_rec;
3924 el->l_next_free_rec = cpu_to_le16(1);
3925 return;
3926 }
3927
3928 /*
3929 * Appending insert.
3930 */
3931 if (insert->ins_appending == APPEND_TAIL) {
3932 i = le16_to_cpu(el->l_next_free_rec) - 1;
3933 rec = &el->l_recs[i];
3934 range = le32_to_cpu(rec->e_cpos)
3935 + le16_to_cpu(rec->e_leaf_clusters);
3936 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3937
3938 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3939 le16_to_cpu(el->l_count),
3940 "owner %llu, depth %u, count %u, next free %u, "
3941 "rec.cpos %u, rec.clusters %u, "
3942 "insert.cpos %u, insert.clusters %u\n",
3943 ocfs2_metadata_cache_owner(et->et_ci),
3944 le16_to_cpu(el->l_tree_depth),
3945 le16_to_cpu(el->l_count),
3946 le16_to_cpu(el->l_next_free_rec),
3947 le32_to_cpu(el->l_recs[i].e_cpos),
3948 le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3949 le32_to_cpu(insert_rec->e_cpos),
3950 le16_to_cpu(insert_rec->e_leaf_clusters));
3951 i++;
3952 el->l_recs[i] = *insert_rec;
3953 le16_add_cpu(&el->l_next_free_rec, 1);
3954 return;
3955 }
3956
3957rotate:
3958 /*
3959 * Ok, we have to rotate.
3960 *
3961 * At this point, it is safe to assume that inserting into an
3962 * empty leaf and appending to a leaf have both been handled
3963 * above.
3964 *
3965 * This leaf needs to have space, either by the empty 1st
3966 * extent record, or by virtue of an l_next_free_rec < l_count.
3967 */
3968 ocfs2_rotate_leaf(el, insert_rec);
3969}
3970
3971static void ocfs2_adjust_rightmost_records(handle_t *handle,
3972 struct ocfs2_extent_tree *et,
3973 struct ocfs2_path *path,
3974 struct ocfs2_extent_rec *insert_rec)
3975{
3976 int i, next_free;
3977 struct buffer_head *bh;
3978 struct ocfs2_extent_list *el;
3979 struct ocfs2_extent_rec *rec;
3980
3981 /*
3982 * Update everything except the leaf block.
3983 */
3984 for (i = 0; i < path->p_tree_depth; i++) {
3985 bh = path->p_node[i].bh;
3986 el = path->p_node[i].el;
3987
3988 next_free = le16_to_cpu(el->l_next_free_rec);
3989 if (next_free == 0) {
3990 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3991 "Owner %llu has a bad extent list\n",
3992 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
3993 return;
3994 }
3995
3996 rec = &el->l_recs[next_free - 1];
3997
3998 rec->e_int_clusters = insert_rec->e_cpos;
3999 le32_add_cpu(&rec->e_int_clusters,
4000 le16_to_cpu(insert_rec->e_leaf_clusters));
4001 le32_add_cpu(&rec->e_int_clusters,
4002 -le32_to_cpu(rec->e_cpos));
4003
4004 ocfs2_journal_dirty(handle, bh);
4005 }
4006}
4007
4008static int ocfs2_append_rec_to_path(handle_t *handle,
4009 struct ocfs2_extent_tree *et,
4010 struct ocfs2_extent_rec *insert_rec,
4011 struct ocfs2_path *right_path,
4012 struct ocfs2_path **ret_left_path)
4013{
4014 int ret, next_free;
4015 struct ocfs2_extent_list *el;
4016 struct ocfs2_path *left_path = NULL;
4017
4018 *ret_left_path = NULL;
4019
4020 /*
4021 * This shouldn't happen for non-trees. The extent rec cluster
4022 * count manipulation below only works for interior nodes.
4023 */
4024 BUG_ON(right_path->p_tree_depth == 0);
4025
4026 /*
4027 * If our appending insert is at the leftmost edge of a leaf,
4028 * then we might need to update the rightmost records of the
4029 * neighboring path.
4030 */
4031 el = path_leaf_el(right_path);
4032 next_free = le16_to_cpu(el->l_next_free_rec);
4033 if (next_free == 0 ||
4034 (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
4035 u32 left_cpos;
4036
4037 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
4038 right_path, &left_cpos);
4039 if (ret) {
4040 mlog_errno(ret);
4041 goto out;
4042 }
4043
4044 trace_ocfs2_append_rec_to_path(
4045 (unsigned long long)
4046 ocfs2_metadata_cache_owner(et->et_ci),
4047 le32_to_cpu(insert_rec->e_cpos),
4048 left_cpos);
4049
4050 /*
4051 * No need to worry if the append is already in the
4052 * leftmost leaf.
4053 */
4054 if (left_cpos) {
4055 left_path = ocfs2_new_path_from_path(right_path);
4056 if (!left_path) {
4057 ret = -ENOMEM;
4058 mlog_errno(ret);
4059 goto out;
4060 }
4061
4062 ret = ocfs2_find_path(et->et_ci, left_path,
4063 left_cpos);
4064 if (ret) {
4065 mlog_errno(ret);
4066 goto out;
4067 }
4068
4069 /*
4070 * ocfs2_insert_path() will pass the left_path to the
4071 * journal for us.
4072 */
4073 }
4074 }
4075
4076 ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4077 if (ret) {
4078 mlog_errno(ret);
4079 goto out;
4080 }
4081
4082 ocfs2_adjust_rightmost_records(handle, et, right_path, insert_rec);
4083
4084 *ret_left_path = left_path;
4085 ret = 0;
4086out:
4087 if (ret != 0)
4088 ocfs2_free_path(left_path);
4089
4090 return ret;
4091}
4092
4093static void ocfs2_split_record(struct ocfs2_extent_tree *et,
4094 struct ocfs2_path *left_path,
4095 struct ocfs2_path *right_path,
4096 struct ocfs2_extent_rec *split_rec,
4097 enum ocfs2_split_type split)
4098{
4099 int index;
4100 u32 cpos = le32_to_cpu(split_rec->e_cpos);
4101 struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
4102 struct ocfs2_extent_rec *rec, *tmprec;
4103
4104 right_el = path_leaf_el(right_path);
4105 if (left_path)
4106 left_el = path_leaf_el(left_path);
4107
4108 el = right_el;
4109 insert_el = right_el;
4110 index = ocfs2_search_extent_list(el, cpos);
4111 if (index != -1) {
4112 if (index == 0 && left_path) {
4113 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
4114
4115 /*
4116 * This typically means that the record
4117 * started in the left path but moved to the
4118 * right as a result of rotation. We either
4119 * move the existing record to the left, or we
4120 * do the later insert there.
4121 *
4122 * In this case, the left path should always
4123 * exist as the rotate code will have passed
4124 * it back for a post-insert update.
4125 */
4126
4127 if (split == SPLIT_LEFT) {
4128 /*
4129 * It's a left split. Since we know
4130 * that the rotate code gave us an
4131 * empty extent in the left path, we
4132 * can just do the insert there.
4133 */
4134 insert_el = left_el;
4135 } else {
4136 /*
4137 * Right split - we have to move the
4138 * existing record over to the left
4139 * leaf. The insert will be into the
4140 * newly created empty extent in the
4141 * right leaf.
4142 */
4143 tmprec = &right_el->l_recs[index];
4144 ocfs2_rotate_leaf(left_el, tmprec);
4145 el = left_el;
4146
4147 memset(tmprec, 0, sizeof(*tmprec));
4148 index = ocfs2_search_extent_list(left_el, cpos);
4149 BUG_ON(index == -1);
4150 }
4151 }
4152 } else {
4153 BUG_ON(!left_path);
4154 BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
4155 /*
4156 * Left path is easy - we can just allow the insert to
4157 * happen.
4158 */
4159 el = left_el;
4160 insert_el = left_el;
4161 index = ocfs2_search_extent_list(el, cpos);
4162 BUG_ON(index == -1);
4163 }
4164
4165 rec = &el->l_recs[index];
4166 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4167 split, rec, split_rec);
4168 ocfs2_rotate_leaf(insert_el, split_rec);
4169}
4170
4171/*
4172 * This function only does inserts on an allocation b-tree. For tree
4173 * depth = 0, ocfs2_insert_at_leaf() is called directly.
4174 *
4175 * right_path is the path we want to do the actual insert
4176 * in. left_path should only be passed in if we need to update that
4177 * portion of the tree after an edge insert.
4178 */
4179static int ocfs2_insert_path(handle_t *handle,
4180 struct ocfs2_extent_tree *et,
4181 struct ocfs2_path *left_path,
4182 struct ocfs2_path *right_path,
4183 struct ocfs2_extent_rec *insert_rec,
4184 struct ocfs2_insert_type *insert)
4185{
4186 int ret, subtree_index;
4187 struct buffer_head *leaf_bh = path_leaf_bh(right_path);
4188
4189 if (left_path) {
4190 /*
4191 * There's a chance that left_path got passed back to
4192 * us without being accounted for in the
4193 * journal. Extend our transaction here to be sure we
4194 * can change those blocks.
4195 */
4196 ret = ocfs2_extend_trans(handle, left_path->p_tree_depth);
4197 if (ret < 0) {
4198 mlog_errno(ret);
4199 goto out;
4200 }
4201
4202 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
4203 if (ret < 0) {
4204 mlog_errno(ret);
4205 goto out;
4206 }
4207 }
4208
4209 /*
4210 * Pass both paths to the journal. The majority of inserts
4211 * will be touching all components anyway.
4212 */
4213 ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4214 if (ret < 0) {
4215 mlog_errno(ret);
4216 goto out;
4217 }
4218
4219 if (insert->ins_split != SPLIT_NONE) {
4220 /*
4221 * We could call ocfs2_insert_at_leaf() for some types
4222 * of splits, but it's easier to just let one separate
4223 * function sort it all out.
4224 */
4225 ocfs2_split_record(et, left_path, right_path,
4226 insert_rec, insert->ins_split);
4227
4228 /*
4229 * Split might have modified either leaf and we don't
4230 * have a guarantee that the later edge insert will
4231 * dirty this for us.
4232 */
4233 if (left_path)
4234 ocfs2_journal_dirty(handle,
4235 path_leaf_bh(left_path));
4236 } else
4237 ocfs2_insert_at_leaf(et, insert_rec, path_leaf_el(right_path),
4238 insert);
4239
4240 ocfs2_journal_dirty(handle, leaf_bh);
4241
4242 if (left_path) {
4243 /*
4244 * The rotate code has indicated that we need to fix
4245 * up portions of the tree after the insert.
4246 *
4247 * XXX: Should we extend the transaction here?
4248 */
4249 subtree_index = ocfs2_find_subtree_root(et, left_path,
4250 right_path);
4251 ocfs2_complete_edge_insert(handle, left_path, right_path,
4252 subtree_index);
4253 }
4254
4255 ret = 0;
4256out:
4257 return ret;
4258}
4259
4260static int ocfs2_do_insert_extent(handle_t *handle,
4261 struct ocfs2_extent_tree *et,
4262 struct ocfs2_extent_rec *insert_rec,
4263 struct ocfs2_insert_type *type)
4264{
4265 int ret, rotate = 0;
4266 u32 cpos;
4267 struct ocfs2_path *right_path = NULL;
4268 struct ocfs2_path *left_path = NULL;
4269 struct ocfs2_extent_list *el;
4270
4271 el = et->et_root_el;
4272
4273 ret = ocfs2_et_root_journal_access(handle, et,
4274 OCFS2_JOURNAL_ACCESS_WRITE);
4275 if (ret) {
4276 mlog_errno(ret);
4277 goto out;
4278 }
4279
4280 if (le16_to_cpu(el->l_tree_depth) == 0) {
4281 ocfs2_insert_at_leaf(et, insert_rec, el, type);
4282 goto out_update_clusters;
4283 }
4284
4285 right_path = ocfs2_new_path_from_et(et);
4286 if (!right_path) {
4287 ret = -ENOMEM;
4288 mlog_errno(ret);
4289 goto out;
4290 }
4291
4292 /*
4293 * Determine the path to start with. Rotations need the
4294 * rightmost path, everything else can go directly to the
4295 * target leaf.
4296 */
4297 cpos = le32_to_cpu(insert_rec->e_cpos);
4298 if (type->ins_appending == APPEND_NONE &&
4299 type->ins_contig == CONTIG_NONE) {
4300 rotate = 1;
4301 cpos = UINT_MAX;
4302 }
4303
4304 ret = ocfs2_find_path(et->et_ci, right_path, cpos);
4305 if (ret) {
4306 mlog_errno(ret);
4307 goto out;
4308 }
4309
4310 /*
4311 * Rotations and appends need special treatment - they modify
4312 * parts of the tree's above them.
4313 *
4314 * Both might pass back a path immediate to the left of the
4315 * one being inserted to. This will be cause
4316 * ocfs2_insert_path() to modify the rightmost records of
4317 * left_path to account for an edge insert.
4318 *
4319 * XXX: When modifying this code, keep in mind that an insert
4320 * can wind up skipping both of these two special cases...
4321 */
4322 if (rotate) {
4323 ret = ocfs2_rotate_tree_right(handle, et, type->ins_split,
4324 le32_to_cpu(insert_rec->e_cpos),
4325 right_path, &left_path);
4326 if (ret) {
4327 mlog_errno(ret);
4328 goto out;
4329 }
4330
4331 /*
4332 * ocfs2_rotate_tree_right() might have extended the
4333 * transaction without re-journaling our tree root.
4334 */
4335 ret = ocfs2_et_root_journal_access(handle, et,
4336 OCFS2_JOURNAL_ACCESS_WRITE);
4337 if (ret) {
4338 mlog_errno(ret);
4339 goto out;
4340 }
4341 } else if (type->ins_appending == APPEND_TAIL
4342 && type->ins_contig != CONTIG_LEFT) {
4343 ret = ocfs2_append_rec_to_path(handle, et, insert_rec,
4344 right_path, &left_path);
4345 if (ret) {
4346 mlog_errno(ret);
4347 goto out;
4348 }
4349 }
4350
4351 ret = ocfs2_insert_path(handle, et, left_path, right_path,
4352 insert_rec, type);
4353 if (ret) {
4354 mlog_errno(ret);
4355 goto out;
4356 }
4357
4358out_update_clusters:
4359 if (type->ins_split == SPLIT_NONE)
4360 ocfs2_et_update_clusters(et,
4361 le16_to_cpu(insert_rec->e_leaf_clusters));
4362
4363 ocfs2_journal_dirty(handle, et->et_root_bh);
4364
4365out:
4366 ocfs2_free_path(left_path);
4367 ocfs2_free_path(right_path);
4368
4369 return ret;
4370}
4371
4372static int ocfs2_figure_merge_contig_type(struct ocfs2_extent_tree *et,
4373 struct ocfs2_path *path,
4374 struct ocfs2_extent_list *el, int index,
4375 struct ocfs2_extent_rec *split_rec,
4376 struct ocfs2_merge_ctxt *ctxt)
4377{
4378 int status = 0;
4379 enum ocfs2_contig_type ret = CONTIG_NONE;
4380 u32 left_cpos, right_cpos;
4381 struct ocfs2_extent_rec *rec = NULL;
4382 struct ocfs2_extent_list *new_el;
4383 struct ocfs2_path *left_path = NULL, *right_path = NULL;
4384 struct buffer_head *bh;
4385 struct ocfs2_extent_block *eb;
4386 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
4387
4388 if (index > 0) {
4389 rec = &el->l_recs[index - 1];
4390 } else if (path->p_tree_depth > 0) {
4391 status = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
4392 if (status)
4393 goto exit;
4394
4395 if (left_cpos != 0) {
4396 left_path = ocfs2_new_path_from_path(path);
4397 if (!left_path) {
4398 status = -ENOMEM;
4399 mlog_errno(status);
4400 goto exit;
4401 }
4402
4403 status = ocfs2_find_path(et->et_ci, left_path,
4404 left_cpos);
4405 if (status)
4406 goto free_left_path;
4407
4408 new_el = path_leaf_el(left_path);
4409
4410 if (le16_to_cpu(new_el->l_next_free_rec) !=
4411 le16_to_cpu(new_el->l_count)) {
4412 bh = path_leaf_bh(left_path);
4413 eb = (struct ocfs2_extent_block *)bh->b_data;
4414 status = ocfs2_error(sb,
4415 "Extent block #%llu has an invalid l_next_free_rec of %d. It should have matched the l_count of %d\n",
4416 (unsigned long long)le64_to_cpu(eb->h_blkno),
4417 le16_to_cpu(new_el->l_next_free_rec),
4418 le16_to_cpu(new_el->l_count));
4419 goto free_left_path;
4420 }
4421 rec = &new_el->l_recs[
4422 le16_to_cpu(new_el->l_next_free_rec) - 1];
4423 }
4424 }
4425
4426 /*
4427 * We're careful to check for an empty extent record here -
4428 * the merge code will know what to do if it sees one.
4429 */
4430 if (rec) {
4431 if (index == 1 && ocfs2_is_empty_extent(rec)) {
4432 if (split_rec->e_cpos == el->l_recs[index].e_cpos)
4433 ret = CONTIG_RIGHT;
4434 } else {
4435 ret = ocfs2_et_extent_contig(et, rec, split_rec);
4436 }
4437 }
4438
4439 rec = NULL;
4440 if (index < (le16_to_cpu(el->l_next_free_rec) - 1))
4441 rec = &el->l_recs[index + 1];
4442 else if (le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count) &&
4443 path->p_tree_depth > 0) {
4444 status = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
4445 if (status)
4446 goto free_left_path;
4447
4448 if (right_cpos == 0)
4449 goto free_left_path;
4450
4451 right_path = ocfs2_new_path_from_path(path);
4452 if (!right_path) {
4453 status = -ENOMEM;
4454 mlog_errno(status);
4455 goto free_left_path;
4456 }
4457
4458 status = ocfs2_find_path(et->et_ci, right_path, right_cpos);
4459 if (status)
4460 goto free_right_path;
4461
4462 new_el = path_leaf_el(right_path);
4463 rec = &new_el->l_recs[0];
4464 if (ocfs2_is_empty_extent(rec)) {
4465 if (le16_to_cpu(new_el->l_next_free_rec) <= 1) {
4466 bh = path_leaf_bh(right_path);
4467 eb = (struct ocfs2_extent_block *)bh->b_data;
4468 status = ocfs2_error(sb,
4469 "Extent block #%llu has an invalid l_next_free_rec of %d\n",
4470 (unsigned long long)le64_to_cpu(eb->h_blkno),
4471 le16_to_cpu(new_el->l_next_free_rec));
4472 goto free_right_path;
4473 }
4474 rec = &new_el->l_recs[1];
4475 }
4476 }
4477
4478 if (rec) {
4479 enum ocfs2_contig_type contig_type;
4480
4481 contig_type = ocfs2_et_extent_contig(et, rec, split_rec);
4482
4483 if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
4484 ret = CONTIG_LEFTRIGHT;
4485 else if (ret == CONTIG_NONE)
4486 ret = contig_type;
4487 }
4488
4489free_right_path:
4490 ocfs2_free_path(right_path);
4491free_left_path:
4492 ocfs2_free_path(left_path);
4493exit:
4494 if (status == 0)
4495 ctxt->c_contig_type = ret;
4496
4497 return status;
4498}
4499
4500static void ocfs2_figure_contig_type(struct ocfs2_extent_tree *et,
4501 struct ocfs2_insert_type *insert,
4502 struct ocfs2_extent_list *el,
4503 struct ocfs2_extent_rec *insert_rec)
4504{
4505 int i;
4506 enum ocfs2_contig_type contig_type = CONTIG_NONE;
4507
4508 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4509
4510 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
4511 contig_type = ocfs2_et_extent_contig(et, &el->l_recs[i],
4512 insert_rec);
4513 if (contig_type != CONTIG_NONE) {
4514 insert->ins_contig_index = i;
4515 break;
4516 }
4517 }
4518 insert->ins_contig = contig_type;
4519
4520 if (insert->ins_contig != CONTIG_NONE) {
4521 struct ocfs2_extent_rec *rec =
4522 &el->l_recs[insert->ins_contig_index];
4523 unsigned int len = le16_to_cpu(rec->e_leaf_clusters) +
4524 le16_to_cpu(insert_rec->e_leaf_clusters);
4525
4526 /*
4527 * Caller might want us to limit the size of extents, don't
4528 * calculate contiguousness if we might exceed that limit.
4529 */
4530 if (et->et_max_leaf_clusters &&
4531 (len > et->et_max_leaf_clusters))
4532 insert->ins_contig = CONTIG_NONE;
4533 }
4534}
4535
4536/*
4537 * This should only be called against the rightmost leaf extent list.
4538 *
4539 * ocfs2_figure_appending_type() will figure out whether we'll have to
4540 * insert at the tail of the rightmost leaf.
4541 *
4542 * This should also work against the root extent list for tree's with 0
4543 * depth. If we consider the root extent list to be the rightmost leaf node
4544 * then the logic here makes sense.
4545 */
4546static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
4547 struct ocfs2_extent_list *el,
4548 struct ocfs2_extent_rec *insert_rec)
4549{
4550 int i;
4551 u32 cpos = le32_to_cpu(insert_rec->e_cpos);
4552 struct ocfs2_extent_rec *rec;
4553
4554 insert->ins_appending = APPEND_NONE;
4555
4556 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4557
4558 if (!el->l_next_free_rec)
4559 goto set_tail_append;
4560
4561 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
4562 /* Were all records empty? */
4563 if (le16_to_cpu(el->l_next_free_rec) == 1)
4564 goto set_tail_append;
4565 }
4566
4567 i = le16_to_cpu(el->l_next_free_rec) - 1;
4568 rec = &el->l_recs[i];
4569
4570 if (cpos >=
4571 (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
4572 goto set_tail_append;
4573
4574 return;
4575
4576set_tail_append:
4577 insert->ins_appending = APPEND_TAIL;
4578}
4579
4580/*
4581 * Helper function called at the beginning of an insert.
4582 *
4583 * This computes a few things that are commonly used in the process of
4584 * inserting into the btree:
4585 * - Whether the new extent is contiguous with an existing one.
4586 * - The current tree depth.
4587 * - Whether the insert is an appending one.
4588 * - The total # of free records in the tree.
4589 *
4590 * All of the information is stored on the ocfs2_insert_type
4591 * structure.
4592 */
4593static int ocfs2_figure_insert_type(struct ocfs2_extent_tree *et,
4594 struct buffer_head **last_eb_bh,
4595 struct ocfs2_extent_rec *insert_rec,
4596 int *free_records,
4597 struct ocfs2_insert_type *insert)
4598{
4599 int ret;
4600 struct ocfs2_extent_block *eb;
4601 struct ocfs2_extent_list *el;
4602 struct ocfs2_path *path = NULL;
4603 struct buffer_head *bh = NULL;
4604
4605 insert->ins_split = SPLIT_NONE;
4606
4607 el = et->et_root_el;
4608 insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
4609
4610 if (el->l_tree_depth) {
4611 /*
4612 * If we have tree depth, we read in the
4613 * rightmost extent block ahead of time as
4614 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4615 * may want it later.
4616 */
4617 ret = ocfs2_read_extent_block(et->et_ci,
4618 ocfs2_et_get_last_eb_blk(et),
4619 &bh);
4620 if (ret) {
4621 mlog_errno(ret);
4622 goto out;
4623 }
4624 eb = (struct ocfs2_extent_block *) bh->b_data;
4625 el = &eb->h_list;
4626 }
4627
4628 /*
4629 * Unless we have a contiguous insert, we'll need to know if
4630 * there is room left in our allocation tree for another
4631 * extent record.
4632 *
4633 * XXX: This test is simplistic, we can search for empty
4634 * extent records too.
4635 */
4636 *free_records = le16_to_cpu(el->l_count) -
4637 le16_to_cpu(el->l_next_free_rec);
4638
4639 if (!insert->ins_tree_depth) {
4640 ocfs2_figure_contig_type(et, insert, el, insert_rec);
4641 ocfs2_figure_appending_type(insert, el, insert_rec);
4642 return 0;
4643 }
4644
4645 path = ocfs2_new_path_from_et(et);
4646 if (!path) {
4647 ret = -ENOMEM;
4648 mlog_errno(ret);
4649 goto out;
4650 }
4651
4652 /*
4653 * In the case that we're inserting past what the tree
4654 * currently accounts for, ocfs2_find_path() will return for
4655 * us the rightmost tree path. This is accounted for below in
4656 * the appending code.
4657 */
4658 ret = ocfs2_find_path(et->et_ci, path, le32_to_cpu(insert_rec->e_cpos));
4659 if (ret) {
4660 mlog_errno(ret);
4661 goto out;
4662 }
4663
4664 el = path_leaf_el(path);
4665
4666 /*
4667 * Now that we have the path, there's two things we want to determine:
4668 * 1) Contiguousness (also set contig_index if this is so)
4669 *
4670 * 2) Are we doing an append? We can trivially break this up
4671 * into two types of appends: simple record append, or a
4672 * rotate inside the tail leaf.
4673 */
4674 ocfs2_figure_contig_type(et, insert, el, insert_rec);
4675
4676 /*
4677 * The insert code isn't quite ready to deal with all cases of
4678 * left contiguousness. Specifically, if it's an insert into
4679 * the 1st record in a leaf, it will require the adjustment of
4680 * cluster count on the last record of the path directly to it's
4681 * left. For now, just catch that case and fool the layers
4682 * above us. This works just fine for tree_depth == 0, which
4683 * is why we allow that above.
4684 */
4685 if (insert->ins_contig == CONTIG_LEFT &&
4686 insert->ins_contig_index == 0)
4687 insert->ins_contig = CONTIG_NONE;
4688
4689 /*
4690 * Ok, so we can simply compare against last_eb to figure out
4691 * whether the path doesn't exist. This will only happen in
4692 * the case that we're doing a tail append, so maybe we can
4693 * take advantage of that information somehow.
4694 */
4695 if (ocfs2_et_get_last_eb_blk(et) ==
4696 path_leaf_bh(path)->b_blocknr) {
4697 /*
4698 * Ok, ocfs2_find_path() returned us the rightmost
4699 * tree path. This might be an appending insert. There are
4700 * two cases:
4701 * 1) We're doing a true append at the tail:
4702 * -This might even be off the end of the leaf
4703 * 2) We're "appending" by rotating in the tail
4704 */
4705 ocfs2_figure_appending_type(insert, el, insert_rec);
4706 }
4707
4708out:
4709 ocfs2_free_path(path);
4710
4711 if (ret == 0)
4712 *last_eb_bh = bh;
4713 else
4714 brelse(bh);
4715 return ret;
4716}
4717
4718/*
4719 * Insert an extent into a btree.
4720 *
4721 * The caller needs to update the owning btree's cluster count.
4722 */
4723int ocfs2_insert_extent(handle_t *handle,
4724 struct ocfs2_extent_tree *et,
4725 u32 cpos,
4726 u64 start_blk,
4727 u32 new_clusters,
4728 u8 flags,
4729 struct ocfs2_alloc_context *meta_ac)
4730{
4731 int status;
4732 int free_records;
4733 struct buffer_head *last_eb_bh = NULL;
4734 struct ocfs2_insert_type insert = {0, };
4735 struct ocfs2_extent_rec rec;
4736
4737 trace_ocfs2_insert_extent_start(
4738 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4739 cpos, new_clusters);
4740
4741 memset(&rec, 0, sizeof(rec));
4742 rec.e_cpos = cpu_to_le32(cpos);
4743 rec.e_blkno = cpu_to_le64(start_blk);
4744 rec.e_leaf_clusters = cpu_to_le16(new_clusters);
4745 rec.e_flags = flags;
4746 status = ocfs2_et_insert_check(et, &rec);
4747 if (status) {
4748 mlog_errno(status);
4749 goto bail;
4750 }
4751
4752 status = ocfs2_figure_insert_type(et, &last_eb_bh, &rec,
4753 &free_records, &insert);
4754 if (status < 0) {
4755 mlog_errno(status);
4756 goto bail;
4757 }
4758
4759 trace_ocfs2_insert_extent(insert.ins_appending, insert.ins_contig,
4760 insert.ins_contig_index, free_records,
4761 insert.ins_tree_depth);
4762
4763 if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
4764 status = ocfs2_grow_tree(handle, et,
4765 &insert.ins_tree_depth, &last_eb_bh,
4766 meta_ac);
4767 if (status) {
4768 mlog_errno(status);
4769 goto bail;
4770 }
4771 }
4772
4773 /* Finally, we can add clusters. This might rotate the tree for us. */
4774 status = ocfs2_do_insert_extent(handle, et, &rec, &insert);
4775 if (status < 0)
4776 mlog_errno(status);
4777 else
4778 ocfs2_et_extent_map_insert(et, &rec);
4779
4780bail:
4781 brelse(last_eb_bh);
4782
4783 return status;
4784}
4785
4786/*
4787 * Allocate and add clusters into the extent b-tree.
4788 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4789 * The extent b-tree's root is specified by et, and
4790 * it is not limited to the file storage. Any extent tree can use this
4791 * function if it implements the proper ocfs2_extent_tree.
4792 */
4793int ocfs2_add_clusters_in_btree(handle_t *handle,
4794 struct ocfs2_extent_tree *et,
4795 u32 *logical_offset,
4796 u32 clusters_to_add,
4797 int mark_unwritten,
4798 struct ocfs2_alloc_context *data_ac,
4799 struct ocfs2_alloc_context *meta_ac,
4800 enum ocfs2_alloc_restarted *reason_ret)
4801{
4802 int status = 0, err = 0;
4803 int need_free = 0;
4804 int free_extents;
4805 enum ocfs2_alloc_restarted reason = RESTART_NONE;
4806 u32 bit_off, num_bits;
4807 u64 block;
4808 u8 flags = 0;
4809 struct ocfs2_super *osb =
4810 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
4811
4812 BUG_ON(!clusters_to_add);
4813
4814 if (mark_unwritten)
4815 flags = OCFS2_EXT_UNWRITTEN;
4816
4817 free_extents = ocfs2_num_free_extents(et);
4818 if (free_extents < 0) {
4819 status = free_extents;
4820 mlog_errno(status);
4821 goto leave;
4822 }
4823
4824 /* there are two cases which could cause us to EAGAIN in the
4825 * we-need-more-metadata case:
4826 * 1) we haven't reserved *any*
4827 * 2) we are so fragmented, we've needed to add metadata too
4828 * many times. */
4829 if (!free_extents && !meta_ac) {
4830 err = -1;
4831 status = -EAGAIN;
4832 reason = RESTART_META;
4833 goto leave;
4834 } else if ((!free_extents)
4835 && (ocfs2_alloc_context_bits_left(meta_ac)
4836 < ocfs2_extend_meta_needed(et->et_root_el))) {
4837 err = -2;
4838 status = -EAGAIN;
4839 reason = RESTART_META;
4840 goto leave;
4841 }
4842
4843 status = __ocfs2_claim_clusters(handle, data_ac, 1,
4844 clusters_to_add, &bit_off, &num_bits);
4845 if (status < 0) {
4846 if (status != -ENOSPC)
4847 mlog_errno(status);
4848 goto leave;
4849 }
4850
4851 BUG_ON(num_bits > clusters_to_add);
4852
4853 /* reserve our write early -- insert_extent may update the tree root */
4854 status = ocfs2_et_root_journal_access(handle, et,
4855 OCFS2_JOURNAL_ACCESS_WRITE);
4856 if (status < 0) {
4857 mlog_errno(status);
4858 need_free = 1;
4859 goto bail;
4860 }
4861
4862 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
4863 trace_ocfs2_add_clusters_in_btree(
4864 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4865 bit_off, num_bits);
4866 status = ocfs2_insert_extent(handle, et, *logical_offset, block,
4867 num_bits, flags, meta_ac);
4868 if (status < 0) {
4869 mlog_errno(status);
4870 need_free = 1;
4871 goto bail;
4872 }
4873
4874 ocfs2_journal_dirty(handle, et->et_root_bh);
4875
4876 clusters_to_add -= num_bits;
4877 *logical_offset += num_bits;
4878
4879 if (clusters_to_add) {
4880 err = clusters_to_add;
4881 status = -EAGAIN;
4882 reason = RESTART_TRANS;
4883 }
4884
4885bail:
4886 if (need_free) {
4887 if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
4888 ocfs2_free_local_alloc_bits(osb, handle, data_ac,
4889 bit_off, num_bits);
4890 else
4891 ocfs2_free_clusters(handle,
4892 data_ac->ac_inode,
4893 data_ac->ac_bh,
4894 ocfs2_clusters_to_blocks(osb->sb, bit_off),
4895 num_bits);
4896 }
4897
4898leave:
4899 if (reason_ret)
4900 *reason_ret = reason;
4901 trace_ocfs2_add_clusters_in_btree_ret(status, reason, err);
4902 return status;
4903}
4904
4905static void ocfs2_make_right_split_rec(struct super_block *sb,
4906 struct ocfs2_extent_rec *split_rec,
4907 u32 cpos,
4908 struct ocfs2_extent_rec *rec)
4909{
4910 u32 rec_cpos = le32_to_cpu(rec->e_cpos);
4911 u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
4912
4913 memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
4914
4915 split_rec->e_cpos = cpu_to_le32(cpos);
4916 split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
4917
4918 split_rec->e_blkno = rec->e_blkno;
4919 le64_add_cpu(&split_rec->e_blkno,
4920 ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
4921
4922 split_rec->e_flags = rec->e_flags;
4923}
4924
4925static int ocfs2_split_and_insert(handle_t *handle,
4926 struct ocfs2_extent_tree *et,
4927 struct ocfs2_path *path,
4928 struct buffer_head **last_eb_bh,
4929 int split_index,
4930 struct ocfs2_extent_rec *orig_split_rec,
4931 struct ocfs2_alloc_context *meta_ac)
4932{
4933 int ret = 0, depth;
4934 unsigned int insert_range, rec_range, do_leftright = 0;
4935 struct ocfs2_extent_rec tmprec;
4936 struct ocfs2_extent_list *rightmost_el;
4937 struct ocfs2_extent_rec rec;
4938 struct ocfs2_extent_rec split_rec = *orig_split_rec;
4939 struct ocfs2_insert_type insert;
4940 struct ocfs2_extent_block *eb;
4941
4942leftright:
4943 /*
4944 * Store a copy of the record on the stack - it might move
4945 * around as the tree is manipulated below.
4946 */
4947 rec = path_leaf_el(path)->l_recs[split_index];
4948
4949 rightmost_el = et->et_root_el;
4950
4951 depth = le16_to_cpu(rightmost_el->l_tree_depth);
4952 if (depth) {
4953 BUG_ON(!(*last_eb_bh));
4954 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
4955 rightmost_el = &eb->h_list;
4956 }
4957
4958 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4959 le16_to_cpu(rightmost_el->l_count)) {
4960 ret = ocfs2_grow_tree(handle, et,
4961 &depth, last_eb_bh, meta_ac);
4962 if (ret) {
4963 mlog_errno(ret);
4964 goto out;
4965 }
4966 }
4967
4968 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4969 insert.ins_appending = APPEND_NONE;
4970 insert.ins_contig = CONTIG_NONE;
4971 insert.ins_tree_depth = depth;
4972
4973 insert_range = le32_to_cpu(split_rec.e_cpos) +
4974 le16_to_cpu(split_rec.e_leaf_clusters);
4975 rec_range = le32_to_cpu(rec.e_cpos) +
4976 le16_to_cpu(rec.e_leaf_clusters);
4977
4978 if (split_rec.e_cpos == rec.e_cpos) {
4979 insert.ins_split = SPLIT_LEFT;
4980 } else if (insert_range == rec_range) {
4981 insert.ins_split = SPLIT_RIGHT;
4982 } else {
4983 /*
4984 * Left/right split. We fake this as a right split
4985 * first and then make a second pass as a left split.
4986 */
4987 insert.ins_split = SPLIT_RIGHT;
4988
4989 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4990 &tmprec, insert_range, &rec);
4991
4992 split_rec = tmprec;
4993
4994 BUG_ON(do_leftright);
4995 do_leftright = 1;
4996 }
4997
4998 ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
4999 if (ret) {
5000 mlog_errno(ret);
5001 goto out;
5002 }
5003
5004 if (do_leftright == 1) {
5005 u32 cpos;
5006 struct ocfs2_extent_list *el;
5007
5008 do_leftright++;
5009 split_rec = *orig_split_rec;
5010
5011 ocfs2_reinit_path(path, 1);
5012
5013 cpos = le32_to_cpu(split_rec.e_cpos);
5014 ret = ocfs2_find_path(et->et_ci, path, cpos);
5015 if (ret) {
5016 mlog_errno(ret);
5017 goto out;
5018 }
5019
5020 el = path_leaf_el(path);
5021 split_index = ocfs2_search_extent_list(el, cpos);
5022 if (split_index == -1) {
5023 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5024 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5025 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5026 cpos);
5027 ret = -EROFS;
5028 goto out;
5029 }
5030 goto leftright;
5031 }
5032out:
5033
5034 return ret;
5035}
5036
5037static int ocfs2_replace_extent_rec(handle_t *handle,
5038 struct ocfs2_extent_tree *et,
5039 struct ocfs2_path *path,
5040 struct ocfs2_extent_list *el,
5041 int split_index,
5042 struct ocfs2_extent_rec *split_rec)
5043{
5044 int ret;
5045
5046 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
5047 path_num_items(path) - 1);
5048 if (ret) {
5049 mlog_errno(ret);
5050 goto out;
5051 }
5052
5053 el->l_recs[split_index] = *split_rec;
5054
5055 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5056out:
5057 return ret;
5058}
5059
5060/*
5061 * Split part or all of the extent record at split_index in the leaf
5062 * pointed to by path. Merge with the contiguous extent record if needed.
5063 *
5064 * Care is taken to handle contiguousness so as to not grow the tree.
5065 *
5066 * meta_ac is not strictly necessary - we only truly need it if growth
5067 * of the tree is required. All other cases will degrade into a less
5068 * optimal tree layout.
5069 *
5070 * last_eb_bh should be the rightmost leaf block for any extent
5071 * btree. Since a split may grow the tree or a merge might shrink it,
5072 * the caller cannot trust the contents of that buffer after this call.
5073 *
5074 * This code is optimized for readability - several passes might be
5075 * made over certain portions of the tree. All of those blocks will
5076 * have been brought into cache (and pinned via the journal), so the
5077 * extra overhead is not expressed in terms of disk reads.
5078 */
5079int ocfs2_split_extent(handle_t *handle,
5080 struct ocfs2_extent_tree *et,
5081 struct ocfs2_path *path,
5082 int split_index,
5083 struct ocfs2_extent_rec *split_rec,
5084 struct ocfs2_alloc_context *meta_ac,
5085 struct ocfs2_cached_dealloc_ctxt *dealloc)
5086{
5087 int ret = 0;
5088 struct ocfs2_extent_list *el = path_leaf_el(path);
5089 struct buffer_head *last_eb_bh = NULL;
5090 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
5091 struct ocfs2_merge_ctxt ctxt;
5092
5093 if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) ||
5094 ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) <
5095 (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) {
5096 ret = -EIO;
5097 mlog_errno(ret);
5098 goto out;
5099 }
5100
5101 ret = ocfs2_figure_merge_contig_type(et, path, el,
5102 split_index,
5103 split_rec,
5104 &ctxt);
5105 if (ret) {
5106 mlog_errno(ret);
5107 goto out;
5108 }
5109
5110 /*
5111 * The core merge / split code wants to know how much room is
5112 * left in this allocation tree, so we pass the
5113 * rightmost extent list.
5114 */
5115 if (path->p_tree_depth) {
5116 ret = ocfs2_read_extent_block(et->et_ci,
5117 ocfs2_et_get_last_eb_blk(et),
5118 &last_eb_bh);
5119 if (ret) {
5120 mlog_errno(ret);
5121 goto out;
5122 }
5123 }
5124
5125 if (rec->e_cpos == split_rec->e_cpos &&
5126 rec->e_leaf_clusters == split_rec->e_leaf_clusters)
5127 ctxt.c_split_covers_rec = 1;
5128 else
5129 ctxt.c_split_covers_rec = 0;
5130
5131 ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
5132
5133 trace_ocfs2_split_extent(split_index, ctxt.c_contig_type,
5134 ctxt.c_has_empty_extent,
5135 ctxt.c_split_covers_rec);
5136
5137 if (ctxt.c_contig_type == CONTIG_NONE) {
5138 if (ctxt.c_split_covers_rec)
5139 ret = ocfs2_replace_extent_rec(handle, et, path, el,
5140 split_index, split_rec);
5141 else
5142 ret = ocfs2_split_and_insert(handle, et, path,
5143 &last_eb_bh, split_index,
5144 split_rec, meta_ac);
5145 if (ret)
5146 mlog_errno(ret);
5147 } else {
5148 ret = ocfs2_try_to_merge_extent(handle, et, path,
5149 split_index, split_rec,
5150 dealloc, &ctxt);
5151 if (ret)
5152 mlog_errno(ret);
5153 }
5154
5155out:
5156 brelse(last_eb_bh);
5157 return ret;
5158}
5159
5160/*
5161 * Change the flags of the already-existing extent at cpos for len clusters.
5162 *
5163 * new_flags: the flags we want to set.
5164 * clear_flags: the flags we want to clear.
5165 * phys: the new physical offset we want this new extent starts from.
5166 *
5167 * If the existing extent is larger than the request, initiate a
5168 * split. An attempt will be made at merging with adjacent extents.
5169 *
5170 * The caller is responsible for passing down meta_ac if we'll need it.
5171 */
5172int ocfs2_change_extent_flag(handle_t *handle,
5173 struct ocfs2_extent_tree *et,
5174 u32 cpos, u32 len, u32 phys,
5175 struct ocfs2_alloc_context *meta_ac,
5176 struct ocfs2_cached_dealloc_ctxt *dealloc,
5177 int new_flags, int clear_flags)
5178{
5179 int ret, index;
5180 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5181 u64 start_blkno = ocfs2_clusters_to_blocks(sb, phys);
5182 struct ocfs2_extent_rec split_rec;
5183 struct ocfs2_path *left_path = NULL;
5184 struct ocfs2_extent_list *el;
5185 struct ocfs2_extent_rec *rec;
5186
5187 left_path = ocfs2_new_path_from_et(et);
5188 if (!left_path) {
5189 ret = -ENOMEM;
5190 mlog_errno(ret);
5191 goto out;
5192 }
5193
5194 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
5195 if (ret) {
5196 mlog_errno(ret);
5197 goto out;
5198 }
5199 el = path_leaf_el(left_path);
5200
5201 index = ocfs2_search_extent_list(el, cpos);
5202 if (index == -1) {
5203 ocfs2_error(sb,
5204 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5205 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5206 cpos);
5207 ret = -EROFS;
5208 goto out;
5209 }
5210
5211 ret = -EIO;
5212 rec = &el->l_recs[index];
5213 if (new_flags && (rec->e_flags & new_flags)) {
5214 mlog(ML_ERROR, "Owner %llu tried to set %d flags on an "
5215 "extent that already had them\n",
5216 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5217 new_flags);
5218 goto out;
5219 }
5220
5221 if (clear_flags && !(rec->e_flags & clear_flags)) {
5222 mlog(ML_ERROR, "Owner %llu tried to clear %d flags on an "
5223 "extent that didn't have them\n",
5224 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5225 clear_flags);
5226 goto out;
5227 }
5228
5229 memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
5230 split_rec.e_cpos = cpu_to_le32(cpos);
5231 split_rec.e_leaf_clusters = cpu_to_le16(len);
5232 split_rec.e_blkno = cpu_to_le64(start_blkno);
5233 split_rec.e_flags = rec->e_flags;
5234 if (new_flags)
5235 split_rec.e_flags |= new_flags;
5236 if (clear_flags)
5237 split_rec.e_flags &= ~clear_flags;
5238
5239 ret = ocfs2_split_extent(handle, et, left_path,
5240 index, &split_rec, meta_ac,
5241 dealloc);
5242 if (ret)
5243 mlog_errno(ret);
5244
5245out:
5246 ocfs2_free_path(left_path);
5247 return ret;
5248
5249}
5250
5251/*
5252 * Mark the already-existing extent at cpos as written for len clusters.
5253 * This removes the unwritten extent flag.
5254 *
5255 * If the existing extent is larger than the request, initiate a
5256 * split. An attempt will be made at merging with adjacent extents.
5257 *
5258 * The caller is responsible for passing down meta_ac if we'll need it.
5259 */
5260int ocfs2_mark_extent_written(struct inode *inode,
5261 struct ocfs2_extent_tree *et,
5262 handle_t *handle, u32 cpos, u32 len, u32 phys,
5263 struct ocfs2_alloc_context *meta_ac,
5264 struct ocfs2_cached_dealloc_ctxt *dealloc)
5265{
5266 int ret;
5267
5268 trace_ocfs2_mark_extent_written(
5269 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5270 cpos, len, phys);
5271
5272 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
5273 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",
5274 (unsigned long long)OCFS2_I(inode)->ip_blkno);
5275 ret = -EROFS;
5276 goto out;
5277 }
5278
5279 /*
5280 * XXX: This should be fixed up so that we just re-insert the
5281 * next extent records.
5282 */
5283 ocfs2_et_extent_map_truncate(et, 0);
5284
5285 ret = ocfs2_change_extent_flag(handle, et, cpos,
5286 len, phys, meta_ac, dealloc,
5287 0, OCFS2_EXT_UNWRITTEN);
5288 if (ret)
5289 mlog_errno(ret);
5290
5291out:
5292 return ret;
5293}
5294
5295static int ocfs2_split_tree(handle_t *handle, struct ocfs2_extent_tree *et,
5296 struct ocfs2_path *path,
5297 int index, u32 new_range,
5298 struct ocfs2_alloc_context *meta_ac)
5299{
5300 int ret, depth, credits;
5301 struct buffer_head *last_eb_bh = NULL;
5302 struct ocfs2_extent_block *eb;
5303 struct ocfs2_extent_list *rightmost_el, *el;
5304 struct ocfs2_extent_rec split_rec;
5305 struct ocfs2_extent_rec *rec;
5306 struct ocfs2_insert_type insert;
5307
5308 /*
5309 * Setup the record to split before we grow the tree.
5310 */
5311 el = path_leaf_el(path);
5312 rec = &el->l_recs[index];
5313 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
5314 &split_rec, new_range, rec);
5315
5316 depth = path->p_tree_depth;
5317 if (depth > 0) {
5318 ret = ocfs2_read_extent_block(et->et_ci,
5319 ocfs2_et_get_last_eb_blk(et),
5320 &last_eb_bh);
5321 if (ret < 0) {
5322 mlog_errno(ret);
5323 goto out;
5324 }
5325
5326 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5327 rightmost_el = &eb->h_list;
5328 } else
5329 rightmost_el = path_leaf_el(path);
5330
5331 credits = path->p_tree_depth +
5332 ocfs2_extend_meta_needed(et->et_root_el);
5333 ret = ocfs2_extend_trans(handle, credits);
5334 if (ret) {
5335 mlog_errno(ret);
5336 goto out;
5337 }
5338
5339 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
5340 le16_to_cpu(rightmost_el->l_count)) {
5341 ret = ocfs2_grow_tree(handle, et, &depth, &last_eb_bh,
5342 meta_ac);
5343 if (ret) {
5344 mlog_errno(ret);
5345 goto out;
5346 }
5347 }
5348
5349 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
5350 insert.ins_appending = APPEND_NONE;
5351 insert.ins_contig = CONTIG_NONE;
5352 insert.ins_split = SPLIT_RIGHT;
5353 insert.ins_tree_depth = depth;
5354
5355 ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
5356 if (ret)
5357 mlog_errno(ret);
5358
5359out:
5360 brelse(last_eb_bh);
5361 return ret;
5362}
5363
5364static int ocfs2_truncate_rec(handle_t *handle,
5365 struct ocfs2_extent_tree *et,
5366 struct ocfs2_path *path, int index,
5367 struct ocfs2_cached_dealloc_ctxt *dealloc,
5368 u32 cpos, u32 len)
5369{
5370 int ret;
5371 u32 left_cpos, rec_range, trunc_range;
5372 int is_rightmost_tree_rec = 0;
5373 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5374 struct ocfs2_path *left_path = NULL;
5375 struct ocfs2_extent_list *el = path_leaf_el(path);
5376 struct ocfs2_extent_rec *rec;
5377 struct ocfs2_extent_block *eb;
5378
5379 if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
5380 /* extend credit for ocfs2_remove_rightmost_path */
5381 ret = ocfs2_extend_rotate_transaction(handle, 0,
5382 jbd2_handle_buffer_credits(handle),
5383 path);
5384 if (ret) {
5385 mlog_errno(ret);
5386 goto out;
5387 }
5388
5389 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5390 if (ret) {
5391 mlog_errno(ret);
5392 goto out;
5393 }
5394
5395 index--;
5396 }
5397
5398 if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
5399 path->p_tree_depth) {
5400 /*
5401 * Check whether this is the rightmost tree record. If
5402 * we remove all of this record or part of its right
5403 * edge then an update of the record lengths above it
5404 * will be required.
5405 */
5406 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
5407 if (eb->h_next_leaf_blk == 0)
5408 is_rightmost_tree_rec = 1;
5409 }
5410
5411 rec = &el->l_recs[index];
5412 if (index == 0 && path->p_tree_depth &&
5413 le32_to_cpu(rec->e_cpos) == cpos) {
5414 /*
5415 * Changing the leftmost offset (via partial or whole
5416 * record truncate) of an interior (or rightmost) path
5417 * means we have to update the subtree that is formed
5418 * by this leaf and the one to it's left.
5419 *
5420 * There are two cases we can skip:
5421 * 1) Path is the leftmost one in our btree.
5422 * 2) The leaf is rightmost and will be empty after
5423 * we remove the extent record - the rotate code
5424 * knows how to update the newly formed edge.
5425 */
5426
5427 ret = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
5428 if (ret) {
5429 mlog_errno(ret);
5430 goto out;
5431 }
5432
5433 if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
5434 left_path = ocfs2_new_path_from_path(path);
5435 if (!left_path) {
5436 ret = -ENOMEM;
5437 mlog_errno(ret);
5438 goto out;
5439 }
5440
5441 ret = ocfs2_find_path(et->et_ci, left_path,
5442 left_cpos);
5443 if (ret) {
5444 mlog_errno(ret);
5445 goto out;
5446 }
5447 }
5448 }
5449
5450 ret = ocfs2_extend_rotate_transaction(handle, 0,
5451 jbd2_handle_buffer_credits(handle),
5452 path);
5453 if (ret) {
5454 mlog_errno(ret);
5455 goto out;
5456 }
5457
5458 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
5459 if (ret) {
5460 mlog_errno(ret);
5461 goto out;
5462 }
5463
5464 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
5465 if (ret) {
5466 mlog_errno(ret);
5467 goto out;
5468 }
5469
5470 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5471 trunc_range = cpos + len;
5472
5473 if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
5474 int next_free;
5475
5476 memset(rec, 0, sizeof(*rec));
5477 ocfs2_cleanup_merge(el, index);
5478
5479 next_free = le16_to_cpu(el->l_next_free_rec);
5480 if (is_rightmost_tree_rec && next_free > 1) {
5481 /*
5482 * We skip the edge update if this path will
5483 * be deleted by the rotate code.
5484 */
5485 rec = &el->l_recs[next_free - 1];
5486 ocfs2_adjust_rightmost_records(handle, et, path,
5487 rec);
5488 }
5489 } else if (le32_to_cpu(rec->e_cpos) == cpos) {
5490 /* Remove leftmost portion of the record. */
5491 le32_add_cpu(&rec->e_cpos, len);
5492 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
5493 le16_add_cpu(&rec->e_leaf_clusters, -len);
5494 } else if (rec_range == trunc_range) {
5495 /* Remove rightmost portion of the record */
5496 le16_add_cpu(&rec->e_leaf_clusters, -len);
5497 if (is_rightmost_tree_rec)
5498 ocfs2_adjust_rightmost_records(handle, et, path, rec);
5499 } else {
5500 /* Caller should have trapped this. */
5501 mlog(ML_ERROR, "Owner %llu: Invalid record truncate: (%u, %u) "
5502 "(%u, %u)\n",
5503 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5504 le32_to_cpu(rec->e_cpos),
5505 le16_to_cpu(rec->e_leaf_clusters), cpos, len);
5506 BUG();
5507 }
5508
5509 if (left_path) {
5510 int subtree_index;
5511
5512 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
5513 ocfs2_complete_edge_insert(handle, left_path, path,
5514 subtree_index);
5515 }
5516
5517 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5518
5519 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5520 if (ret)
5521 mlog_errno(ret);
5522
5523out:
5524 ocfs2_free_path(left_path);
5525 return ret;
5526}
5527
5528int ocfs2_remove_extent(handle_t *handle,
5529 struct ocfs2_extent_tree *et,
5530 u32 cpos, u32 len,
5531 struct ocfs2_alloc_context *meta_ac,
5532 struct ocfs2_cached_dealloc_ctxt *dealloc)
5533{
5534 int ret, index;
5535 u32 rec_range, trunc_range;
5536 struct ocfs2_extent_rec *rec;
5537 struct ocfs2_extent_list *el;
5538 struct ocfs2_path *path = NULL;
5539
5540 /*
5541 * XXX: Why are we truncating to 0 instead of wherever this
5542 * affects us?
5543 */
5544 ocfs2_et_extent_map_truncate(et, 0);
5545
5546 path = ocfs2_new_path_from_et(et);
5547 if (!path) {
5548 ret = -ENOMEM;
5549 mlog_errno(ret);
5550 goto out;
5551 }
5552
5553 ret = ocfs2_find_path(et->et_ci, path, cpos);
5554 if (ret) {
5555 mlog_errno(ret);
5556 goto out;
5557 }
5558
5559 el = path_leaf_el(path);
5560 index = ocfs2_search_extent_list(el, cpos);
5561 if (index == -1) {
5562 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5563 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5564 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5565 cpos);
5566 ret = -EROFS;
5567 goto out;
5568 }
5569
5570 /*
5571 * We have 3 cases of extent removal:
5572 * 1) Range covers the entire extent rec
5573 * 2) Range begins or ends on one edge of the extent rec
5574 * 3) Range is in the middle of the extent rec (no shared edges)
5575 *
5576 * For case 1 we remove the extent rec and left rotate to
5577 * fill the hole.
5578 *
5579 * For case 2 we just shrink the existing extent rec, with a
5580 * tree update if the shrinking edge is also the edge of an
5581 * extent block.
5582 *
5583 * For case 3 we do a right split to turn the extent rec into
5584 * something case 2 can handle.
5585 */
5586 rec = &el->l_recs[index];
5587 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5588 trunc_range = cpos + len;
5589
5590 BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
5591
5592 trace_ocfs2_remove_extent(
5593 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5594 cpos, len, index, le32_to_cpu(rec->e_cpos),
5595 ocfs2_rec_clusters(el, rec));
5596
5597 if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
5598 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5599 cpos, len);
5600 if (ret) {
5601 mlog_errno(ret);
5602 goto out;
5603 }
5604 } else {
5605 ret = ocfs2_split_tree(handle, et, path, index,
5606 trunc_range, meta_ac);
5607 if (ret) {
5608 mlog_errno(ret);
5609 goto out;
5610 }
5611
5612 /*
5613 * The split could have manipulated the tree enough to
5614 * move the record location, so we have to look for it again.
5615 */
5616 ocfs2_reinit_path(path, 1);
5617
5618 ret = ocfs2_find_path(et->et_ci, path, cpos);
5619 if (ret) {
5620 mlog_errno(ret);
5621 goto out;
5622 }
5623
5624 el = path_leaf_el(path);
5625 index = ocfs2_search_extent_list(el, cpos);
5626 if (index == -1) {
5627 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5628 "Owner %llu: split at cpos %u lost record\n",
5629 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5630 cpos);
5631 ret = -EROFS;
5632 goto out;
5633 }
5634
5635 /*
5636 * Double check our values here. If anything is fishy,
5637 * it's easier to catch it at the top level.
5638 */
5639 rec = &el->l_recs[index];
5640 rec_range = le32_to_cpu(rec->e_cpos) +
5641 ocfs2_rec_clusters(el, rec);
5642 if (rec_range != trunc_range) {
5643 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5644 "Owner %llu: error after split at cpos %u trunc len %u, existing record is (%u,%u)\n",
5645 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5646 cpos, len, le32_to_cpu(rec->e_cpos),
5647 ocfs2_rec_clusters(el, rec));
5648 ret = -EROFS;
5649 goto out;
5650 }
5651
5652 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5653 cpos, len);
5654 if (ret)
5655 mlog_errno(ret);
5656 }
5657
5658out:
5659 ocfs2_free_path(path);
5660 return ret;
5661}
5662
5663/*
5664 * ocfs2_reserve_blocks_for_rec_trunc() would look basically the
5665 * same as ocfs2_lock_alloctors(), except for it accepts a blocks
5666 * number to reserve some extra blocks, and it only handles meta
5667 * data allocations.
5668 *
5669 * Currently, only ocfs2_remove_btree_range() uses it for truncating
5670 * and punching holes.
5671 */
5672static int ocfs2_reserve_blocks_for_rec_trunc(struct inode *inode,
5673 struct ocfs2_extent_tree *et,
5674 u32 extents_to_split,
5675 struct ocfs2_alloc_context **ac,
5676 int extra_blocks)
5677{
5678 int ret = 0, num_free_extents;
5679 unsigned int max_recs_needed = 2 * extents_to_split;
5680 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5681
5682 *ac = NULL;
5683
5684 num_free_extents = ocfs2_num_free_extents(et);
5685 if (num_free_extents < 0) {
5686 ret = num_free_extents;
5687 mlog_errno(ret);
5688 goto out;
5689 }
5690
5691 if (!num_free_extents ||
5692 (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
5693 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
5694
5695 if (extra_blocks) {
5696 ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, ac);
5697 if (ret < 0) {
5698 if (ret != -ENOSPC)
5699 mlog_errno(ret);
5700 }
5701 }
5702
5703out:
5704 if (ret) {
5705 if (*ac) {
5706 ocfs2_free_alloc_context(*ac);
5707 *ac = NULL;
5708 }
5709 }
5710
5711 return ret;
5712}
5713
5714int ocfs2_remove_btree_range(struct inode *inode,
5715 struct ocfs2_extent_tree *et,
5716 u32 cpos, u32 phys_cpos, u32 len, int flags,
5717 struct ocfs2_cached_dealloc_ctxt *dealloc,
5718 u64 refcount_loc, bool refcount_tree_locked)
5719{
5720 int ret, credits = 0, extra_blocks = 0;
5721 u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
5722 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5723 struct inode *tl_inode = osb->osb_tl_inode;
5724 handle_t *handle;
5725 struct ocfs2_alloc_context *meta_ac = NULL;
5726 struct ocfs2_refcount_tree *ref_tree = NULL;
5727
5728 if ((flags & OCFS2_EXT_REFCOUNTED) && len) {
5729 BUG_ON(!ocfs2_is_refcount_inode(inode));
5730
5731 if (!refcount_tree_locked) {
5732 ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
5733 &ref_tree, NULL);
5734 if (ret) {
5735 mlog_errno(ret);
5736 goto bail;
5737 }
5738 }
5739
5740 ret = ocfs2_prepare_refcount_change_for_del(inode,
5741 refcount_loc,
5742 phys_blkno,
5743 len,
5744 &credits,
5745 &extra_blocks);
5746 if (ret < 0) {
5747 mlog_errno(ret);
5748 goto bail;
5749 }
5750 }
5751
5752 ret = ocfs2_reserve_blocks_for_rec_trunc(inode, et, 1, &meta_ac,
5753 extra_blocks);
5754 if (ret) {
5755 mlog_errno(ret);
5756 goto bail;
5757 }
5758
5759 inode_lock(tl_inode);
5760
5761 if (ocfs2_truncate_log_needs_flush(osb)) {
5762 ret = __ocfs2_flush_truncate_log(osb);
5763 if (ret < 0) {
5764 mlog_errno(ret);
5765 goto out;
5766 }
5767 }
5768
5769 handle = ocfs2_start_trans(osb,
5770 ocfs2_remove_extent_credits(osb->sb) + credits);
5771 if (IS_ERR(handle)) {
5772 ret = PTR_ERR(handle);
5773 mlog_errno(ret);
5774 goto out;
5775 }
5776
5777 ret = ocfs2_et_root_journal_access(handle, et,
5778 OCFS2_JOURNAL_ACCESS_WRITE);
5779 if (ret) {
5780 mlog_errno(ret);
5781 goto out_commit;
5782 }
5783
5784 dquot_free_space_nodirty(inode,
5785 ocfs2_clusters_to_bytes(inode->i_sb, len));
5786
5787 ret = ocfs2_remove_extent(handle, et, cpos, len, meta_ac, dealloc);
5788 if (ret) {
5789 mlog_errno(ret);
5790 goto out_commit;
5791 }
5792
5793 ocfs2_et_update_clusters(et, -len);
5794 ocfs2_update_inode_fsync_trans(handle, inode, 1);
5795
5796 ocfs2_journal_dirty(handle, et->et_root_bh);
5797
5798 if (phys_blkno) {
5799 if (flags & OCFS2_EXT_REFCOUNTED)
5800 ret = ocfs2_decrease_refcount(inode, handle,
5801 ocfs2_blocks_to_clusters(osb->sb,
5802 phys_blkno),
5803 len, meta_ac,
5804 dealloc, 1);
5805 else
5806 ret = ocfs2_truncate_log_append(osb, handle,
5807 phys_blkno, len);
5808 if (ret)
5809 mlog_errno(ret);
5810
5811 }
5812
5813out_commit:
5814 ocfs2_commit_trans(osb, handle);
5815out:
5816 inode_unlock(tl_inode);
5817bail:
5818 if (meta_ac)
5819 ocfs2_free_alloc_context(meta_ac);
5820
5821 if (ref_tree)
5822 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
5823
5824 return ret;
5825}
5826
5827int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
5828{
5829 struct buffer_head *tl_bh = osb->osb_tl_bh;
5830 struct ocfs2_dinode *di;
5831 struct ocfs2_truncate_log *tl;
5832
5833 di = (struct ocfs2_dinode *) tl_bh->b_data;
5834 tl = &di->id2.i_dealloc;
5835
5836 mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
5837 "slot %d, invalid truncate log parameters: used = "
5838 "%u, count = %u\n", osb->slot_num,
5839 le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
5840 return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
5841}
5842
5843static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
5844 unsigned int new_start)
5845{
5846 unsigned int tail_index;
5847 unsigned int current_tail;
5848
5849 /* No records, nothing to coalesce */
5850 if (!le16_to_cpu(tl->tl_used))
5851 return 0;
5852
5853 tail_index = le16_to_cpu(tl->tl_used) - 1;
5854 current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
5855 current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
5856
5857 return current_tail == new_start;
5858}
5859
5860int ocfs2_truncate_log_append(struct ocfs2_super *osb,
5861 handle_t *handle,
5862 u64 start_blk,
5863 unsigned int num_clusters)
5864{
5865 int status, index;
5866 unsigned int start_cluster, tl_count;
5867 struct inode *tl_inode = osb->osb_tl_inode;
5868 struct buffer_head *tl_bh = osb->osb_tl_bh;
5869 struct ocfs2_dinode *di;
5870 struct ocfs2_truncate_log *tl;
5871
5872 BUG_ON(inode_trylock(tl_inode));
5873
5874 start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
5875
5876 di = (struct ocfs2_dinode *) tl_bh->b_data;
5877
5878 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5879 * by the underlying call to ocfs2_read_inode_block(), so any
5880 * corruption is a code bug */
5881 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5882
5883 tl = &di->id2.i_dealloc;
5884 tl_count = le16_to_cpu(tl->tl_count);
5885 mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
5886 tl_count == 0,
5887 "Truncate record count on #%llu invalid "
5888 "wanted %u, actual %u\n",
5889 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5890 ocfs2_truncate_recs_per_inode(osb->sb),
5891 le16_to_cpu(tl->tl_count));
5892
5893 /* Caller should have known to flush before calling us. */
5894 index = le16_to_cpu(tl->tl_used);
5895 if (index >= tl_count) {
5896 status = -ENOSPC;
5897 mlog_errno(status);
5898 goto bail;
5899 }
5900
5901 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5902 OCFS2_JOURNAL_ACCESS_WRITE);
5903 if (status < 0) {
5904 mlog_errno(status);
5905 goto bail;
5906 }
5907
5908 trace_ocfs2_truncate_log_append(
5909 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index,
5910 start_cluster, num_clusters);
5911 if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
5912 /*
5913 * Move index back to the record we are coalescing with.
5914 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5915 */
5916 index--;
5917
5918 num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
5919 trace_ocfs2_truncate_log_append(
5920 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5921 index, le32_to_cpu(tl->tl_recs[index].t_start),
5922 num_clusters);
5923 } else {
5924 tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
5925 tl->tl_used = cpu_to_le16(index + 1);
5926 }
5927 tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
5928
5929 ocfs2_journal_dirty(handle, tl_bh);
5930
5931 osb->truncated_clusters += num_clusters;
5932bail:
5933 return status;
5934}
5935
5936static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
5937 struct inode *data_alloc_inode,
5938 struct buffer_head *data_alloc_bh)
5939{
5940 int status = 0;
5941 int i;
5942 unsigned int num_clusters;
5943 u64 start_blk;
5944 struct ocfs2_truncate_rec rec;
5945 struct ocfs2_dinode *di;
5946 struct ocfs2_truncate_log *tl;
5947 struct inode *tl_inode = osb->osb_tl_inode;
5948 struct buffer_head *tl_bh = osb->osb_tl_bh;
5949 handle_t *handle;
5950
5951 di = (struct ocfs2_dinode *) tl_bh->b_data;
5952 tl = &di->id2.i_dealloc;
5953 i = le16_to_cpu(tl->tl_used) - 1;
5954 while (i >= 0) {
5955 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
5956 if (IS_ERR(handle)) {
5957 status = PTR_ERR(handle);
5958 mlog_errno(status);
5959 goto bail;
5960 }
5961
5962 /* Caller has given us at least enough credits to
5963 * update the truncate log dinode */
5964 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5965 OCFS2_JOURNAL_ACCESS_WRITE);
5966 if (status < 0) {
5967 ocfs2_commit_trans(osb, handle);
5968 mlog_errno(status);
5969 goto bail;
5970 }
5971
5972 tl->tl_used = cpu_to_le16(i);
5973
5974 ocfs2_journal_dirty(handle, tl_bh);
5975
5976 rec = tl->tl_recs[i];
5977 start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
5978 le32_to_cpu(rec.t_start));
5979 num_clusters = le32_to_cpu(rec.t_clusters);
5980
5981 /* if start_blk is not set, we ignore the record as
5982 * invalid. */
5983 if (start_blk) {
5984 trace_ocfs2_replay_truncate_records(
5985 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5986 i, le32_to_cpu(rec.t_start), num_clusters);
5987
5988 status = ocfs2_free_clusters(handle, data_alloc_inode,
5989 data_alloc_bh, start_blk,
5990 num_clusters);
5991 if (status < 0) {
5992 ocfs2_commit_trans(osb, handle);
5993 mlog_errno(status);
5994 goto bail;
5995 }
5996 }
5997
5998 ocfs2_commit_trans(osb, handle);
5999 i--;
6000 }
6001
6002 osb->truncated_clusters = 0;
6003
6004bail:
6005 return status;
6006}
6007
6008/* Expects you to already be holding tl_inode->i_rwsem */
6009int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
6010{
6011 int status;
6012 unsigned int num_to_flush;
6013 struct inode *tl_inode = osb->osb_tl_inode;
6014 struct inode *data_alloc_inode = NULL;
6015 struct buffer_head *tl_bh = osb->osb_tl_bh;
6016 struct buffer_head *data_alloc_bh = NULL;
6017 struct ocfs2_dinode *di;
6018 struct ocfs2_truncate_log *tl;
6019 struct ocfs2_journal *journal = osb->journal;
6020
6021 BUG_ON(inode_trylock(tl_inode));
6022
6023 di = (struct ocfs2_dinode *) tl_bh->b_data;
6024
6025 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
6026 * by the underlying call to ocfs2_read_inode_block(), so any
6027 * corruption is a code bug */
6028 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6029
6030 tl = &di->id2.i_dealloc;
6031 num_to_flush = le16_to_cpu(tl->tl_used);
6032 trace_ocfs2_flush_truncate_log(
6033 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
6034 num_to_flush);
6035 if (!num_to_flush) {
6036 status = 0;
6037 goto out;
6038 }
6039
6040 /* Appending truncate log(TA) and flushing truncate log(TF) are
6041 * two separated transactions. They can be both committed but not
6042 * checkpointed. If crash occurs then, both two transaction will be
6043 * replayed with several already released to global bitmap clusters.
6044 * Then truncate log will be replayed resulting in cluster double free.
6045 */
6046 jbd2_journal_lock_updates(journal->j_journal);
6047 status = jbd2_journal_flush(journal->j_journal, 0);
6048 jbd2_journal_unlock_updates(journal->j_journal);
6049 if (status < 0) {
6050 mlog_errno(status);
6051 goto out;
6052 }
6053
6054 data_alloc_inode = ocfs2_get_system_file_inode(osb,
6055 GLOBAL_BITMAP_SYSTEM_INODE,
6056 OCFS2_INVALID_SLOT);
6057 if (!data_alloc_inode) {
6058 status = -EINVAL;
6059 mlog(ML_ERROR, "Could not get bitmap inode!\n");
6060 goto out;
6061 }
6062
6063 inode_lock(data_alloc_inode);
6064
6065 status = ocfs2_inode_lock(data_alloc_inode, &data_alloc_bh, 1);
6066 if (status < 0) {
6067 mlog_errno(status);
6068 goto out_mutex;
6069 }
6070
6071 status = ocfs2_replay_truncate_records(osb, data_alloc_inode,
6072 data_alloc_bh);
6073 if (status < 0)
6074 mlog_errno(status);
6075
6076 brelse(data_alloc_bh);
6077 ocfs2_inode_unlock(data_alloc_inode, 1);
6078
6079out_mutex:
6080 inode_unlock(data_alloc_inode);
6081 iput(data_alloc_inode);
6082
6083out:
6084 return status;
6085}
6086
6087int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
6088{
6089 int status;
6090 struct inode *tl_inode = osb->osb_tl_inode;
6091
6092 inode_lock(tl_inode);
6093 status = __ocfs2_flush_truncate_log(osb);
6094 inode_unlock(tl_inode);
6095
6096 return status;
6097}
6098
6099static void ocfs2_truncate_log_worker(struct work_struct *work)
6100{
6101 int status;
6102 struct ocfs2_super *osb =
6103 container_of(work, struct ocfs2_super,
6104 osb_truncate_log_wq.work);
6105
6106 status = ocfs2_flush_truncate_log(osb);
6107 if (status < 0)
6108 mlog_errno(status);
6109 else
6110 ocfs2_init_steal_slots(osb);
6111}
6112
6113#define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
6114void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
6115 int cancel)
6116{
6117 if (osb->osb_tl_inode &&
6118 atomic_read(&osb->osb_tl_disable) == 0) {
6119 /* We want to push off log flushes while truncates are
6120 * still running. */
6121 if (cancel)
6122 cancel_delayed_work(&osb->osb_truncate_log_wq);
6123
6124 queue_delayed_work(osb->ocfs2_wq, &osb->osb_truncate_log_wq,
6125 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
6126 }
6127}
6128
6129/*
6130 * Try to flush truncate logs if we can free enough clusters from it.
6131 * As for return value, "< 0" means error, "0" no space and "1" means
6132 * we have freed enough spaces and let the caller try to allocate again.
6133 */
6134int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
6135 unsigned int needed)
6136{
6137 tid_t target;
6138 int ret = 0;
6139 unsigned int truncated_clusters;
6140
6141 inode_lock(osb->osb_tl_inode);
6142 truncated_clusters = osb->truncated_clusters;
6143 inode_unlock(osb->osb_tl_inode);
6144
6145 /*
6146 * Check whether we can succeed in allocating if we free
6147 * the truncate log.
6148 */
6149 if (truncated_clusters < needed)
6150 goto out;
6151
6152 ret = ocfs2_flush_truncate_log(osb);
6153 if (ret) {
6154 mlog_errno(ret);
6155 goto out;
6156 }
6157
6158 if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
6159 jbd2_log_wait_commit(osb->journal->j_journal, target);
6160 ret = 1;
6161 }
6162out:
6163 return ret;
6164}
6165
6166static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
6167 int slot_num,
6168 struct inode **tl_inode,
6169 struct buffer_head **tl_bh)
6170{
6171 int status;
6172 struct inode *inode = NULL;
6173 struct buffer_head *bh = NULL;
6174 struct ocfs2_dinode *di;
6175 struct ocfs2_truncate_log *tl;
6176 unsigned int tl_count, tl_used;
6177
6178 inode = ocfs2_get_system_file_inode(osb,
6179 TRUNCATE_LOG_SYSTEM_INODE,
6180 slot_num);
6181 if (!inode) {
6182 status = -EINVAL;
6183 mlog(ML_ERROR, "Could not get load truncate log inode!\n");
6184 goto bail;
6185 }
6186
6187 status = ocfs2_read_inode_block(inode, &bh);
6188 if (status < 0) {
6189 iput(inode);
6190 mlog_errno(status);
6191 goto bail;
6192 }
6193
6194 di = (struct ocfs2_dinode *)bh->b_data;
6195 tl = &di->id2.i_dealloc;
6196 tl_count = le16_to_cpu(tl->tl_count);
6197 tl_used = le16_to_cpu(tl->tl_used);
6198 if (unlikely(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
6199 tl_count == 0 ||
6200 tl_used > tl_count)) {
6201 status = -EFSCORRUPTED;
6202 iput(inode);
6203 brelse(bh);
6204 mlog_errno(status);
6205 goto bail;
6206 }
6207
6208 *tl_inode = inode;
6209 *tl_bh = bh;
6210bail:
6211 return status;
6212}
6213
6214/* called during the 1st stage of node recovery. we stamp a clean
6215 * truncate log and pass back a copy for processing later. if the
6216 * truncate log does not require processing, a *tl_copy is set to
6217 * NULL. */
6218int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
6219 int slot_num,
6220 struct ocfs2_dinode **tl_copy)
6221{
6222 int status;
6223 struct inode *tl_inode = NULL;
6224 struct buffer_head *tl_bh = NULL;
6225 struct ocfs2_dinode *di;
6226 struct ocfs2_truncate_log *tl;
6227
6228 *tl_copy = NULL;
6229
6230 trace_ocfs2_begin_truncate_log_recovery(slot_num);
6231
6232 status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
6233 if (status < 0) {
6234 mlog_errno(status);
6235 goto bail;
6236 }
6237
6238 di = (struct ocfs2_dinode *) tl_bh->b_data;
6239
6240 /* tl_bh is loaded from ocfs2_get_truncate_log_info(). It's
6241 * validated by the underlying call to ocfs2_read_inode_block(),
6242 * so any corruption is a code bug */
6243 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6244
6245 tl = &di->id2.i_dealloc;
6246 if (le16_to_cpu(tl->tl_used)) {
6247 trace_ocfs2_truncate_log_recovery_num(le16_to_cpu(tl->tl_used));
6248
6249 /*
6250 * Assuming the write-out below goes well, this copy will be
6251 * passed back to recovery for processing.
6252 */
6253 *tl_copy = kmemdup(tl_bh->b_data, tl_bh->b_size, GFP_KERNEL);
6254 if (!(*tl_copy)) {
6255 status = -ENOMEM;
6256 mlog_errno(status);
6257 goto bail;
6258 }
6259
6260 /* All we need to do to clear the truncate log is set
6261 * tl_used. */
6262 tl->tl_used = 0;
6263
6264 ocfs2_compute_meta_ecc(osb->sb, tl_bh->b_data, &di->i_check);
6265 status = ocfs2_write_block(osb, tl_bh, INODE_CACHE(tl_inode));
6266 if (status < 0) {
6267 mlog_errno(status);
6268 goto bail;
6269 }
6270 }
6271
6272bail:
6273 iput(tl_inode);
6274 brelse(tl_bh);
6275
6276 if (status < 0) {
6277 kfree(*tl_copy);
6278 *tl_copy = NULL;
6279 mlog_errno(status);
6280 }
6281
6282 return status;
6283}
6284
6285int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
6286 struct ocfs2_dinode *tl_copy)
6287{
6288 int status = 0;
6289 int i;
6290 unsigned int clusters, num_recs, start_cluster;
6291 u64 start_blk;
6292 handle_t *handle;
6293 struct inode *tl_inode = osb->osb_tl_inode;
6294 struct ocfs2_truncate_log *tl;
6295
6296 if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
6297 mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
6298 return -EINVAL;
6299 }
6300
6301 tl = &tl_copy->id2.i_dealloc;
6302 num_recs = le16_to_cpu(tl->tl_used);
6303 trace_ocfs2_complete_truncate_log_recovery(
6304 (unsigned long long)le64_to_cpu(tl_copy->i_blkno),
6305 num_recs);
6306
6307 inode_lock(tl_inode);
6308 for(i = 0; i < num_recs; i++) {
6309 if (ocfs2_truncate_log_needs_flush(osb)) {
6310 status = __ocfs2_flush_truncate_log(osb);
6311 if (status < 0) {
6312 mlog_errno(status);
6313 goto bail_up;
6314 }
6315 }
6316
6317 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6318 if (IS_ERR(handle)) {
6319 status = PTR_ERR(handle);
6320 mlog_errno(status);
6321 goto bail_up;
6322 }
6323
6324 clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
6325 start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
6326 start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
6327
6328 status = ocfs2_truncate_log_append(osb, handle,
6329 start_blk, clusters);
6330 ocfs2_commit_trans(osb, handle);
6331 if (status < 0) {
6332 mlog_errno(status);
6333 goto bail_up;
6334 }
6335 }
6336
6337bail_up:
6338 inode_unlock(tl_inode);
6339
6340 return status;
6341}
6342
6343void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
6344{
6345 int status;
6346 struct inode *tl_inode = osb->osb_tl_inode;
6347
6348 atomic_set(&osb->osb_tl_disable, 1);
6349
6350 if (tl_inode) {
6351 cancel_delayed_work(&osb->osb_truncate_log_wq);
6352 flush_workqueue(osb->ocfs2_wq);
6353
6354 status = ocfs2_flush_truncate_log(osb);
6355 if (status < 0)
6356 mlog_errno(status);
6357
6358 brelse(osb->osb_tl_bh);
6359 iput(osb->osb_tl_inode);
6360 }
6361}
6362
6363int ocfs2_truncate_log_init(struct ocfs2_super *osb)
6364{
6365 int status;
6366 struct inode *tl_inode = NULL;
6367 struct buffer_head *tl_bh = NULL;
6368
6369 status = ocfs2_get_truncate_log_info(osb,
6370 osb->slot_num,
6371 &tl_inode,
6372 &tl_bh);
6373 if (status < 0)
6374 mlog_errno(status);
6375
6376 /* ocfs2_truncate_log_shutdown keys on the existence of
6377 * osb->osb_tl_inode so we don't set any of the osb variables
6378 * until we're sure all is well. */
6379 INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
6380 ocfs2_truncate_log_worker);
6381 atomic_set(&osb->osb_tl_disable, 0);
6382 osb->osb_tl_bh = tl_bh;
6383 osb->osb_tl_inode = tl_inode;
6384
6385 return status;
6386}
6387
6388/*
6389 * Delayed de-allocation of suballocator blocks.
6390 *
6391 * Some sets of block de-allocations might involve multiple suballocator inodes.
6392 *
6393 * The locking for this can get extremely complicated, especially when
6394 * the suballocator inodes to delete from aren't known until deep
6395 * within an unrelated codepath.
6396 *
6397 * ocfs2_extent_block structures are a good example of this - an inode
6398 * btree could have been grown by any number of nodes each allocating
6399 * out of their own suballoc inode.
6400 *
6401 * These structures allow the delay of block de-allocation until a
6402 * later time, when locking of multiple cluster inodes won't cause
6403 * deadlock.
6404 */
6405
6406/*
6407 * Describe a single bit freed from a suballocator. For the block
6408 * suballocators, it represents one block. For the global cluster
6409 * allocator, it represents some clusters and free_bit indicates
6410 * clusters number.
6411 */
6412struct ocfs2_cached_block_free {
6413 struct ocfs2_cached_block_free *free_next;
6414 u64 free_bg;
6415 u64 free_blk;
6416 unsigned int free_bit;
6417};
6418
6419struct ocfs2_per_slot_free_list {
6420 struct ocfs2_per_slot_free_list *f_next_suballocator;
6421 int f_inode_type;
6422 int f_slot;
6423 struct ocfs2_cached_block_free *f_first;
6424};
6425
6426static int ocfs2_free_cached_blocks(struct ocfs2_super *osb,
6427 int sysfile_type,
6428 int slot,
6429 struct ocfs2_cached_block_free *head)
6430{
6431 int ret;
6432 u64 bg_blkno;
6433 handle_t *handle;
6434 struct inode *inode;
6435 struct buffer_head *di_bh = NULL;
6436 struct ocfs2_cached_block_free *tmp;
6437
6438 inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
6439 if (!inode) {
6440 ret = -EINVAL;
6441 mlog_errno(ret);
6442 goto out;
6443 }
6444
6445 inode_lock(inode);
6446
6447 ret = ocfs2_inode_lock(inode, &di_bh, 1);
6448 if (ret) {
6449 mlog_errno(ret);
6450 goto out_mutex;
6451 }
6452
6453 while (head) {
6454 if (head->free_bg)
6455 bg_blkno = head->free_bg;
6456 else
6457 bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
6458 head->free_bit);
6459 handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
6460 if (IS_ERR(handle)) {
6461 ret = PTR_ERR(handle);
6462 mlog_errno(ret);
6463 goto out_unlock;
6464 }
6465
6466 trace_ocfs2_free_cached_blocks(
6467 (unsigned long long)head->free_blk, head->free_bit);
6468
6469 ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
6470 head->free_bit, bg_blkno, 1);
6471 if (ret)
6472 mlog_errno(ret);
6473
6474 ocfs2_commit_trans(osb, handle);
6475
6476 tmp = head;
6477 head = head->free_next;
6478 kfree(tmp);
6479 }
6480
6481out_unlock:
6482 ocfs2_inode_unlock(inode, 1);
6483 brelse(di_bh);
6484out_mutex:
6485 inode_unlock(inode);
6486 iput(inode);
6487out:
6488 while(head) {
6489 /* Premature exit may have left some dangling items. */
6490 tmp = head;
6491 head = head->free_next;
6492 kfree(tmp);
6493 }
6494
6495 return ret;
6496}
6497
6498int ocfs2_cache_cluster_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6499 u64 blkno, unsigned int bit)
6500{
6501 int ret = 0;
6502 struct ocfs2_cached_block_free *item;
6503
6504 item = kzalloc_obj(*item, GFP_NOFS);
6505 if (item == NULL) {
6506 ret = -ENOMEM;
6507 mlog_errno(ret);
6508 return ret;
6509 }
6510
6511 trace_ocfs2_cache_cluster_dealloc((unsigned long long)blkno, bit);
6512
6513 item->free_blk = blkno;
6514 item->free_bit = bit;
6515 item->free_next = ctxt->c_global_allocator;
6516
6517 ctxt->c_global_allocator = item;
6518 return ret;
6519}
6520
6521static int ocfs2_free_cached_clusters(struct ocfs2_super *osb,
6522 struct ocfs2_cached_block_free *head)
6523{
6524 struct ocfs2_cached_block_free *tmp;
6525 struct inode *tl_inode = osb->osb_tl_inode;
6526 handle_t *handle;
6527 int ret = 0;
6528
6529 inode_lock(tl_inode);
6530
6531 while (head) {
6532 if (ocfs2_truncate_log_needs_flush(osb)) {
6533 ret = __ocfs2_flush_truncate_log(osb);
6534 if (ret < 0) {
6535 mlog_errno(ret);
6536 break;
6537 }
6538 }
6539
6540 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6541 if (IS_ERR(handle)) {
6542 ret = PTR_ERR(handle);
6543 mlog_errno(ret);
6544 break;
6545 }
6546
6547 ret = ocfs2_truncate_log_append(osb, handle, head->free_blk,
6548 head->free_bit);
6549
6550 ocfs2_commit_trans(osb, handle);
6551 tmp = head;
6552 head = head->free_next;
6553 kfree(tmp);
6554
6555 if (ret < 0) {
6556 mlog_errno(ret);
6557 break;
6558 }
6559 }
6560
6561 inode_unlock(tl_inode);
6562
6563 while (head) {
6564 /* Premature exit may have left some dangling items. */
6565 tmp = head;
6566 head = head->free_next;
6567 kfree(tmp);
6568 }
6569
6570 return ret;
6571}
6572
6573int ocfs2_run_deallocs(struct ocfs2_super *osb,
6574 struct ocfs2_cached_dealloc_ctxt *ctxt)
6575{
6576 int ret = 0, ret2;
6577 struct ocfs2_per_slot_free_list *fl;
6578
6579 if (!ctxt)
6580 return 0;
6581
6582 while (ctxt->c_first_suballocator) {
6583 fl = ctxt->c_first_suballocator;
6584
6585 if (fl->f_first) {
6586 trace_ocfs2_run_deallocs(fl->f_inode_type,
6587 fl->f_slot);
6588 ret2 = ocfs2_free_cached_blocks(osb,
6589 fl->f_inode_type,
6590 fl->f_slot,
6591 fl->f_first);
6592 if (ret2)
6593 mlog_errno(ret2);
6594 if (!ret)
6595 ret = ret2;
6596 }
6597
6598 ctxt->c_first_suballocator = fl->f_next_suballocator;
6599 kfree(fl);
6600 }
6601
6602 if (ctxt->c_global_allocator) {
6603 ret2 = ocfs2_free_cached_clusters(osb,
6604 ctxt->c_global_allocator);
6605 if (ret2)
6606 mlog_errno(ret2);
6607 if (!ret)
6608 ret = ret2;
6609
6610 ctxt->c_global_allocator = NULL;
6611 }
6612
6613 return ret;
6614}
6615
6616static struct ocfs2_per_slot_free_list *
6617ocfs2_find_per_slot_free_list(int type,
6618 int slot,
6619 struct ocfs2_cached_dealloc_ctxt *ctxt)
6620{
6621 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6622
6623 while (fl) {
6624 if (fl->f_inode_type == type && fl->f_slot == slot)
6625 return fl;
6626
6627 fl = fl->f_next_suballocator;
6628 }
6629
6630 fl = kmalloc_obj(*fl, GFP_NOFS);
6631 if (fl) {
6632 fl->f_inode_type = type;
6633 fl->f_slot = slot;
6634 fl->f_first = NULL;
6635 fl->f_next_suballocator = ctxt->c_first_suballocator;
6636
6637 ctxt->c_first_suballocator = fl;
6638 }
6639 return fl;
6640}
6641
6642static struct ocfs2_per_slot_free_list *
6643ocfs2_find_preferred_free_list(int type,
6644 int preferred_slot,
6645 int *real_slot,
6646 struct ocfs2_cached_dealloc_ctxt *ctxt)
6647{
6648 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6649
6650 while (fl) {
6651 if (fl->f_inode_type == type && fl->f_slot == preferred_slot) {
6652 *real_slot = fl->f_slot;
6653 return fl;
6654 }
6655
6656 fl = fl->f_next_suballocator;
6657 }
6658
6659 /* If we can't find any free list matching preferred slot, just use
6660 * the first one.
6661 */
6662 fl = ctxt->c_first_suballocator;
6663 *real_slot = fl->f_slot;
6664
6665 return fl;
6666}
6667
6668/* Return Value 1 indicates empty */
6669static int ocfs2_is_dealloc_empty(struct ocfs2_extent_tree *et)
6670{
6671 struct ocfs2_per_slot_free_list *fl = NULL;
6672
6673 if (!et->et_dealloc)
6674 return 1;
6675
6676 fl = et->et_dealloc->c_first_suballocator;
6677 if (!fl)
6678 return 1;
6679
6680 if (!fl->f_first)
6681 return 1;
6682
6683 return 0;
6684}
6685
6686/* If extent was deleted from tree due to extent rotation and merging, and
6687 * no metadata is reserved ahead of time. Try to reuse some extents
6688 * just deleted. This is only used to reuse extent blocks.
6689 * It is supposed to find enough extent blocks in dealloc if our estimation
6690 * on metadata is accurate.
6691 */
6692static int ocfs2_reuse_blk_from_dealloc(handle_t *handle,
6693 struct ocfs2_extent_tree *et,
6694 struct buffer_head **new_eb_bh,
6695 int blk_wanted, int *blk_given)
6696{
6697 int i, status = 0, real_slot;
6698 struct ocfs2_cached_dealloc_ctxt *dealloc;
6699 struct ocfs2_per_slot_free_list *fl;
6700 struct ocfs2_cached_block_free *bf;
6701 struct ocfs2_extent_block *eb;
6702 struct ocfs2_super *osb =
6703 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
6704
6705 *blk_given = 0;
6706
6707 /* If extent tree doesn't have a dealloc, this is not faulty. Just
6708 * tell upper caller dealloc can't provide any block and it should
6709 * ask for alloc to claim more space.
6710 */
6711 dealloc = et->et_dealloc;
6712 if (!dealloc)
6713 goto bail;
6714
6715 for (i = 0; i < blk_wanted; i++) {
6716 /* Prefer to use local slot */
6717 fl = ocfs2_find_preferred_free_list(EXTENT_ALLOC_SYSTEM_INODE,
6718 osb->slot_num, &real_slot,
6719 dealloc);
6720 /* If no more block can be reused, we should claim more
6721 * from alloc. Just return here normally.
6722 */
6723 if (!fl) {
6724 status = 0;
6725 break;
6726 }
6727
6728 bf = fl->f_first;
6729 fl->f_first = bf->free_next;
6730
6731 new_eb_bh[i] = sb_getblk(osb->sb, bf->free_blk);
6732 if (new_eb_bh[i] == NULL) {
6733 status = -ENOMEM;
6734 mlog_errno(status);
6735 goto bail;
6736 }
6737
6738 mlog(0, "Reusing block(%llu) from "
6739 "dealloc(local slot:%d, real slot:%d)\n",
6740 bf->free_blk, osb->slot_num, real_slot);
6741
6742 ocfs2_set_new_buffer_uptodate(et->et_ci, new_eb_bh[i]);
6743
6744 status = ocfs2_journal_access_eb(handle, et->et_ci,
6745 new_eb_bh[i],
6746 OCFS2_JOURNAL_ACCESS_CREATE);
6747 if (status < 0) {
6748 mlog_errno(status);
6749 goto bail;
6750 }
6751
6752 memset(new_eb_bh[i]->b_data, 0, osb->sb->s_blocksize);
6753 eb = (struct ocfs2_extent_block *) new_eb_bh[i]->b_data;
6754
6755 /* We can't guarantee that buffer head is still cached, so
6756 * polutlate the extent block again.
6757 */
6758 strscpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
6759 eb->h_blkno = cpu_to_le64(bf->free_blk);
6760 eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
6761 eb->h_suballoc_slot = cpu_to_le16(real_slot);
6762 eb->h_suballoc_loc = cpu_to_le64(bf->free_bg);
6763 eb->h_suballoc_bit = cpu_to_le16(bf->free_bit);
6764 eb->h_list.l_count =
6765 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
6766
6767 /* We'll also be dirtied by the caller, so
6768 * this isn't absolutely necessary.
6769 */
6770 ocfs2_journal_dirty(handle, new_eb_bh[i]);
6771
6772 if (!fl->f_first) {
6773 dealloc->c_first_suballocator = fl->f_next_suballocator;
6774 kfree(fl);
6775 }
6776 kfree(bf);
6777 }
6778
6779 *blk_given = i;
6780
6781bail:
6782 if (unlikely(status < 0)) {
6783 for (i = 0; i < blk_wanted; i++)
6784 brelse(new_eb_bh[i]);
6785 }
6786
6787 return status;
6788}
6789
6790int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6791 int type, int slot, u64 suballoc,
6792 u64 blkno, unsigned int bit)
6793{
6794 int ret;
6795 struct ocfs2_per_slot_free_list *fl;
6796 struct ocfs2_cached_block_free *item;
6797
6798 fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
6799 if (fl == NULL) {
6800 ret = -ENOMEM;
6801 mlog_errno(ret);
6802 goto out;
6803 }
6804
6805 item = kzalloc_obj(*item, GFP_NOFS);
6806 if (item == NULL) {
6807 ret = -ENOMEM;
6808 mlog_errno(ret);
6809 goto out;
6810 }
6811
6812 trace_ocfs2_cache_block_dealloc(type, slot,
6813 (unsigned long long)suballoc,
6814 (unsigned long long)blkno, bit);
6815
6816 item->free_bg = suballoc;
6817 item->free_blk = blkno;
6818 item->free_bit = bit;
6819 item->free_next = fl->f_first;
6820
6821 fl->f_first = item;
6822
6823 ret = 0;
6824out:
6825 return ret;
6826}
6827
6828static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
6829 struct ocfs2_extent_block *eb)
6830{
6831 return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
6832 le16_to_cpu(eb->h_suballoc_slot),
6833 le64_to_cpu(eb->h_suballoc_loc),
6834 le64_to_cpu(eb->h_blkno),
6835 le16_to_cpu(eb->h_suballoc_bit));
6836}
6837
6838static int ocfs2_zero_func(handle_t *handle, struct buffer_head *bh)
6839{
6840 set_buffer_uptodate(bh);
6841 mark_buffer_dirty(bh);
6842 return 0;
6843}
6844
6845void ocfs2_map_and_dirty_folio(struct inode *inode, handle_t *handle,
6846 size_t from, size_t to, struct folio *folio, int zero,
6847 u64 *phys)
6848{
6849 int ret, partial = 0;
6850 loff_t start_byte = folio_pos(folio) + from;
6851 loff_t length = to - from;
6852
6853 ret = ocfs2_map_folio_blocks(folio, phys, inode, from, to, 0);
6854 if (ret)
6855 mlog_errno(ret);
6856
6857 if (zero)
6858 folio_zero_segment(folio, from, to);
6859
6860 /*
6861 * Need to set the buffers we zero'd into uptodate
6862 * here if they aren't - ocfs2_map_page_blocks()
6863 * might've skipped some
6864 */
6865 ret = walk_page_buffers(handle, folio_buffers(folio),
6866 from, to, &partial,
6867 ocfs2_zero_func);
6868 if (ret < 0)
6869 mlog_errno(ret);
6870 else if (ocfs2_should_order_data(inode)) {
6871 ret = ocfs2_jbd2_inode_add_write(handle, inode,
6872 start_byte, length);
6873 if (ret < 0)
6874 mlog_errno(ret);
6875 }
6876
6877 if (!partial)
6878 folio_mark_uptodate(folio);
6879
6880 flush_dcache_folio(folio);
6881}
6882
6883static void ocfs2_zero_cluster_folios(struct inode *inode, loff_t start,
6884 loff_t end, struct folio **folios, int numfolios,
6885 u64 phys, handle_t *handle)
6886{
6887 int i;
6888 struct super_block *sb = inode->i_sb;
6889
6890 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
6891
6892 if (numfolios == 0)
6893 goto out;
6894
6895 for (i = 0; i < numfolios; i++) {
6896 struct folio *folio = folios[i];
6897 size_t to = folio_size(folio);
6898 size_t from = offset_in_folio(folio, start);
6899
6900 if (to > end - folio_pos(folio))
6901 to = end - folio_pos(folio);
6902
6903 ocfs2_map_and_dirty_folio(inode, handle, from, to, folio, 1,
6904 &phys);
6905
6906 start = folio_next_pos(folio);
6907 }
6908out:
6909 if (folios)
6910 ocfs2_unlock_and_free_folios(folios, numfolios);
6911}
6912
6913static int ocfs2_grab_folios(struct inode *inode, loff_t start, loff_t end,
6914 struct folio **folios, int *num)
6915{
6916 int numfolios, ret = 0;
6917 struct address_space *mapping = inode->i_mapping;
6918 unsigned long index;
6919 loff_t last_page_bytes;
6920
6921 BUG_ON(start > end);
6922
6923 numfolios = 0;
6924 last_page_bytes = PAGE_ALIGN(end);
6925 index = start >> PAGE_SHIFT;
6926 do {
6927 folios[numfolios] = __filemap_get_folio(mapping, index,
6928 FGP_LOCK | FGP_ACCESSED | FGP_CREAT, GFP_NOFS);
6929 if (IS_ERR(folios[numfolios])) {
6930 ret = PTR_ERR(folios[numfolios]);
6931 mlog_errno(ret);
6932 folios[numfolios] = NULL;
6933 goto out;
6934 }
6935
6936 index = folio_next_index(folios[numfolios]);
6937 numfolios++;
6938 } while (index < (last_page_bytes >> PAGE_SHIFT));
6939
6940out:
6941 if (ret != 0) {
6942 ocfs2_unlock_and_free_folios(folios, numfolios);
6943 numfolios = 0;
6944 }
6945
6946 *num = numfolios;
6947
6948 return ret;
6949}
6950
6951static int ocfs2_grab_eof_folios(struct inode *inode, loff_t start, loff_t end,
6952 struct folio **folios, int *num)
6953{
6954 struct super_block *sb = inode->i_sb;
6955
6956 BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
6957 (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
6958
6959 return ocfs2_grab_folios(inode, start, end, folios, num);
6960}
6961
6962/*
6963 * Zero partial cluster for a hole punch or truncate. This avoids exposing
6964 * nonzero data on subsequent file extends.
6965 *
6966 * We need to call this before i_size is updated on the inode because
6967 * otherwise block_write_full_folio() will skip writeout of pages past
6968 * i_size.
6969 */
6970int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
6971 u64 range_start, u64 range_end)
6972{
6973 int ret = 0, numfolios;
6974 struct folio **folios = NULL;
6975 u64 phys;
6976 unsigned int ext_flags;
6977 struct super_block *sb = inode->i_sb;
6978
6979 /*
6980 * File systems which don't support sparse files zero on every
6981 * extend.
6982 */
6983 if (!ocfs2_sparse_alloc(OCFS2_SB(sb)))
6984 return 0;
6985
6986 /*
6987 * Avoid zeroing folios fully beyond current i_size. It is pointless as
6988 * underlying blocks of those folios should be already zeroed out and
6989 * page writeback will skip them anyway.
6990 */
6991 range_end = min_t(u64, range_end, i_size_read(inode));
6992 if (range_start >= range_end)
6993 return 0;
6994
6995 folios = kzalloc_objs(struct folio *, ocfs2_pages_per_cluster(sb),
6996 GFP_NOFS);
6997 if (folios == NULL) {
6998 ret = -ENOMEM;
6999 mlog_errno(ret);
7000 goto out;
7001 }
7002
7003 ret = ocfs2_extent_map_get_blocks(inode,
7004 range_start >> sb->s_blocksize_bits,
7005 &phys, NULL, &ext_flags);
7006 if (ret) {
7007 mlog_errno(ret);
7008 goto out;
7009 }
7010
7011 /*
7012 * Tail is a hole, or is marked unwritten. In either case, we
7013 * can count on read and write to return/push zero's.
7014 */
7015 if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN)
7016 goto out;
7017
7018 ret = ocfs2_grab_eof_folios(inode, range_start, range_end, folios,
7019 &numfolios);
7020 if (ret) {
7021 mlog_errno(ret);
7022 goto out;
7023 }
7024
7025 ocfs2_zero_cluster_folios(inode, range_start, range_end, folios,
7026 numfolios, phys, handle);
7027
7028 /*
7029 * Initiate writeout of the folios we zero'd here. We don't
7030 * wait on them - the truncate_inode_pages() call later will
7031 * do that for us.
7032 */
7033 ret = filemap_fdatawrite_range(inode->i_mapping, range_start,
7034 range_end - 1);
7035 if (ret)
7036 mlog_errno(ret);
7037
7038out:
7039 kfree(folios);
7040
7041 return ret;
7042}
7043
7044static void ocfs2_zero_dinode_id2_with_xattr(struct inode *inode,
7045 struct ocfs2_dinode *di)
7046{
7047 unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits;
7048 unsigned int xattrsize = le16_to_cpu(di->i_xattr_inline_size);
7049
7050 if (le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_XATTR_FL)
7051 memset(&di->id2, 0, blocksize -
7052 offsetof(struct ocfs2_dinode, id2) -
7053 xattrsize);
7054 else
7055 memset(&di->id2, 0, blocksize -
7056 offsetof(struct ocfs2_dinode, id2));
7057}
7058
7059void ocfs2_dinode_new_extent_list(struct inode *inode,
7060 struct ocfs2_dinode *di)
7061{
7062 ocfs2_zero_dinode_id2_with_xattr(inode, di);
7063 di->id2.i_list.l_tree_depth = 0;
7064 di->id2.i_list.l_next_free_rec = 0;
7065 di->id2.i_list.l_count = cpu_to_le16(
7066 ocfs2_extent_recs_per_inode_with_xattr(inode->i_sb, di));
7067}
7068
7069void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di)
7070{
7071 struct ocfs2_inode_info *oi = OCFS2_I(inode);
7072 struct ocfs2_inline_data *idata = &di->id2.i_data;
7073
7074 spin_lock(&oi->ip_lock);
7075 oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
7076 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7077 spin_unlock(&oi->ip_lock);
7078
7079 /*
7080 * We clear the entire i_data structure here so that all
7081 * fields can be properly initialized.
7082 */
7083 ocfs2_zero_dinode_id2_with_xattr(inode, di);
7084
7085 idata->id_count = cpu_to_le16(
7086 ocfs2_max_inline_data_with_xattr(inode->i_sb, di));
7087}
7088
7089int ocfs2_convert_inline_data_to_extents(struct inode *inode,
7090 struct buffer_head *di_bh)
7091{
7092 int ret, has_data, num_folios = 0;
7093 int need_free = 0;
7094 u32 bit_off, num;
7095 handle_t *handle;
7096 u64 block;
7097 struct ocfs2_inode_info *oi = OCFS2_I(inode);
7098 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7099 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7100 struct ocfs2_alloc_context *data_ac = NULL;
7101 struct folio *folio = NULL;
7102 struct ocfs2_extent_tree et;
7103 int did_quota = 0;
7104
7105 has_data = i_size_read(inode) ? 1 : 0;
7106
7107 if (has_data) {
7108 ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
7109 if (ret) {
7110 mlog_errno(ret);
7111 goto out;
7112 }
7113 }
7114
7115 handle = ocfs2_start_trans(osb,
7116 ocfs2_inline_to_extents_credits(osb->sb));
7117 if (IS_ERR(handle)) {
7118 ret = PTR_ERR(handle);
7119 mlog_errno(ret);
7120 goto out;
7121 }
7122
7123 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7124 OCFS2_JOURNAL_ACCESS_WRITE);
7125 if (ret) {
7126 mlog_errno(ret);
7127 goto out_commit;
7128 }
7129
7130 if (has_data) {
7131 unsigned int page_end = min_t(unsigned, PAGE_SIZE,
7132 osb->s_clustersize);
7133 u64 phys;
7134
7135 ret = dquot_alloc_space_nodirty(inode,
7136 ocfs2_clusters_to_bytes(osb->sb, 1));
7137 if (ret)
7138 goto out_commit;
7139 did_quota = 1;
7140
7141 data_ac->ac_resv = &oi->ip_la_data_resv;
7142
7143 ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off,
7144 &num);
7145 if (ret) {
7146 mlog_errno(ret);
7147 goto out_commit;
7148 }
7149
7150 /*
7151 * Save two copies, one for insert, and one that can
7152 * be changed by ocfs2_map_and_dirty_folio() below.
7153 */
7154 block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off);
7155
7156 ret = ocfs2_grab_eof_folios(inode, 0, page_end, &folio,
7157 &num_folios);
7158 if (ret) {
7159 mlog_errno(ret);
7160 need_free = 1;
7161 goto out_commit;
7162 }
7163
7164 /*
7165 * This should populate the 1st page for us and mark
7166 * it up to date.
7167 */
7168 ret = ocfs2_read_inline_data(inode, folio, di_bh);
7169 if (ret) {
7170 mlog_errno(ret);
7171 need_free = 1;
7172 goto out_unlock;
7173 }
7174
7175 ocfs2_map_and_dirty_folio(inode, handle, 0, page_end, folio, 0,
7176 &phys);
7177 }
7178
7179 spin_lock(&oi->ip_lock);
7180 oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
7181 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7182 spin_unlock(&oi->ip_lock);
7183
7184 ocfs2_update_inode_fsync_trans(handle, inode, 1);
7185 ocfs2_dinode_new_extent_list(inode, di);
7186
7187 ocfs2_journal_dirty(handle, di_bh);
7188
7189 if (has_data) {
7190 /*
7191 * An error at this point should be extremely rare. If
7192 * this proves to be false, we could always re-build
7193 * the in-inode data from our pages.
7194 */
7195 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7196 ret = ocfs2_insert_extent(handle, &et, 0, block, 1, 0, NULL);
7197 if (ret) {
7198 mlog_errno(ret);
7199 need_free = 1;
7200 goto out_unlock;
7201 }
7202
7203 inode->i_blocks = ocfs2_inode_sector_count(inode);
7204 }
7205
7206out_unlock:
7207 if (folio)
7208 ocfs2_unlock_and_free_folios(&folio, num_folios);
7209
7210out_commit:
7211 if (ret < 0 && did_quota)
7212 dquot_free_space_nodirty(inode,
7213 ocfs2_clusters_to_bytes(osb->sb, 1));
7214
7215 if (need_free) {
7216 if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
7217 ocfs2_free_local_alloc_bits(osb, handle, data_ac,
7218 bit_off, num);
7219 else
7220 ocfs2_free_clusters(handle,
7221 data_ac->ac_inode,
7222 data_ac->ac_bh,
7223 ocfs2_clusters_to_blocks(osb->sb, bit_off),
7224 num);
7225 }
7226
7227 ocfs2_commit_trans(osb, handle);
7228
7229out:
7230 if (data_ac)
7231 ocfs2_free_alloc_context(data_ac);
7232 return ret;
7233}
7234
7235/*
7236 * It is expected, that by the time you call this function,
7237 * inode->i_size and fe->i_size have been adjusted.
7238 *
7239 * WARNING: This will kfree the truncate context
7240 */
7241int ocfs2_commit_truncate(struct ocfs2_super *osb,
7242 struct inode *inode,
7243 struct buffer_head *di_bh)
7244{
7245 int status = 0, i, flags = 0;
7246 u32 new_highest_cpos, range, trunc_cpos, trunc_len, phys_cpos, coff;
7247 u64 blkno = 0;
7248 struct ocfs2_extent_list *el;
7249 struct ocfs2_extent_rec *rec;
7250 struct ocfs2_path *path = NULL;
7251 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7252 struct ocfs2_extent_list *root_el = &(di->id2.i_list);
7253 u64 refcount_loc = le64_to_cpu(di->i_refcount_loc);
7254 struct ocfs2_extent_tree et;
7255 struct ocfs2_cached_dealloc_ctxt dealloc;
7256 struct ocfs2_refcount_tree *ref_tree = NULL;
7257
7258 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7259 ocfs2_init_dealloc_ctxt(&dealloc);
7260
7261 new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
7262 i_size_read(inode));
7263
7264 path = ocfs2_new_path(di_bh, &di->id2.i_list,
7265 ocfs2_journal_access_di);
7266 if (!path) {
7267 status = -ENOMEM;
7268 mlog_errno(status);
7269 goto bail;
7270 }
7271
7272 ocfs2_extent_map_trunc(inode, new_highest_cpos);
7273
7274start:
7275 /*
7276 * Check that we still have allocation to delete.
7277 */
7278 if (OCFS2_I(inode)->ip_clusters == 0) {
7279 status = 0;
7280 goto bail;
7281 }
7282
7283 /*
7284 * Truncate always works against the rightmost tree branch.
7285 */
7286 status = ocfs2_find_path(INODE_CACHE(inode), path, UINT_MAX);
7287 if (status) {
7288 mlog_errno(status);
7289 goto bail;
7290 }
7291
7292 trace_ocfs2_commit_truncate(
7293 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7294 new_highest_cpos,
7295 OCFS2_I(inode)->ip_clusters,
7296 path->p_tree_depth);
7297
7298 /*
7299 * By now, el will point to the extent list on the bottom most
7300 * portion of this tree. Only the tail record is considered in
7301 * each pass.
7302 *
7303 * We handle the following cases, in order:
7304 * - empty extent: delete the remaining branch
7305 * - remove the entire record
7306 * - remove a partial record
7307 * - no record needs to be removed (truncate has completed)
7308 */
7309 el = path_leaf_el(path);
7310 if (le16_to_cpu(el->l_next_free_rec) == 0) {
7311 ocfs2_error(inode->i_sb,
7312 "Inode %llu has empty extent block at %llu\n",
7313 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7314 (unsigned long long)path_leaf_bh(path)->b_blocknr);
7315 status = -EROFS;
7316 goto bail;
7317 }
7318
7319 i = le16_to_cpu(el->l_next_free_rec) - 1;
7320 rec = &el->l_recs[i];
7321 flags = rec->e_flags;
7322 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
7323
7324 if (i == 0 && ocfs2_is_empty_extent(rec)) {
7325 /*
7326 * Lower levels depend on this never happening, but it's best
7327 * to check it up here before changing the tree.
7328 */
7329 if (root_el->l_tree_depth && rec->e_int_clusters == 0) {
7330 mlog(ML_ERROR, "Inode %llu has an empty "
7331 "extent record, depth %u\n", inode->i_ino,
7332 le16_to_cpu(root_el->l_tree_depth));
7333 status = ocfs2_remove_rightmost_empty_extent(osb,
7334 &et, path, &dealloc);
7335 if (status) {
7336 mlog_errno(status);
7337 goto bail;
7338 }
7339
7340 ocfs2_reinit_path(path, 1);
7341 goto start;
7342 } else {
7343 trunc_cpos = le32_to_cpu(rec->e_cpos);
7344 trunc_len = 0;
7345 blkno = 0;
7346 }
7347 } else if (le32_to_cpu(rec->e_cpos) >= new_highest_cpos) {
7348 /*
7349 * Truncate entire record.
7350 */
7351 trunc_cpos = le32_to_cpu(rec->e_cpos);
7352 trunc_len = ocfs2_rec_clusters(el, rec);
7353 blkno = le64_to_cpu(rec->e_blkno);
7354 } else if (range > new_highest_cpos) {
7355 /*
7356 * Partial truncate. it also should be
7357 * the last truncate we're doing.
7358 */
7359 trunc_cpos = new_highest_cpos;
7360 trunc_len = range - new_highest_cpos;
7361 coff = new_highest_cpos - le32_to_cpu(rec->e_cpos);
7362 blkno = le64_to_cpu(rec->e_blkno) +
7363 ocfs2_clusters_to_blocks(inode->i_sb, coff);
7364 } else {
7365 /*
7366 * Truncate completed, leave happily.
7367 */
7368 status = 0;
7369 goto bail;
7370 }
7371
7372 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
7373
7374 if ((flags & OCFS2_EXT_REFCOUNTED) && trunc_len && !ref_tree) {
7375 status = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
7376 &ref_tree, NULL);
7377 if (status) {
7378 mlog_errno(status);
7379 goto bail;
7380 }
7381 }
7382
7383 status = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
7384 phys_cpos, trunc_len, flags, &dealloc,
7385 refcount_loc, true);
7386 if (status < 0) {
7387 mlog_errno(status);
7388 goto bail;
7389 }
7390
7391 ocfs2_reinit_path(path, 1);
7392
7393 /*
7394 * The check above will catch the case where we've truncated
7395 * away all allocation.
7396 */
7397 goto start;
7398
7399bail:
7400 if (ref_tree)
7401 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
7402
7403 ocfs2_schedule_truncate_log_flush(osb, 1);
7404
7405 ocfs2_run_deallocs(osb, &dealloc);
7406
7407 ocfs2_free_path(path);
7408
7409 return status;
7410}
7411
7412/*
7413 * 'start' is inclusive, 'end' is not.
7414 */
7415int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh,
7416 unsigned int start, unsigned int end, int trunc)
7417{
7418 int ret;
7419 unsigned int numbytes;
7420 handle_t *handle;
7421 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7422 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7423 struct ocfs2_inline_data *idata = &di->id2.i_data;
7424
7425 /* No need to punch hole beyond i_size. */
7426 if (start >= i_size_read(inode))
7427 return 0;
7428
7429 if (end > i_size_read(inode))
7430 end = i_size_read(inode);
7431
7432 BUG_ON(start > end);
7433
7434 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
7435 !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) ||
7436 !ocfs2_supports_inline_data(osb)) {
7437 ocfs2_error(inode->i_sb,
7438 "Inline data flags for inode %llu don't agree! Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7439 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7440 le16_to_cpu(di->i_dyn_features),
7441 OCFS2_I(inode)->ip_dyn_features,
7442 osb->s_feature_incompat);
7443 ret = -EROFS;
7444 goto out;
7445 }
7446
7447 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
7448 if (IS_ERR(handle)) {
7449 ret = PTR_ERR(handle);
7450 mlog_errno(ret);
7451 goto out;
7452 }
7453
7454 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7455 OCFS2_JOURNAL_ACCESS_WRITE);
7456 if (ret) {
7457 mlog_errno(ret);
7458 goto out_commit;
7459 }
7460
7461 numbytes = end - start;
7462 memset(idata->id_data + start, 0, numbytes);
7463
7464 /*
7465 * No need to worry about the data page here - it's been
7466 * truncated already and inline data doesn't need it for
7467 * pushing zero's to disk, so we'll let read_folio pick it up
7468 * later.
7469 */
7470 if (trunc) {
7471 i_size_write(inode, start);
7472 di->i_size = cpu_to_le64(start);
7473 }
7474
7475 inode->i_blocks = ocfs2_inode_sector_count(inode);
7476 inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
7477
7478 di->i_ctime = di->i_mtime = cpu_to_le64(inode_get_ctime_sec(inode));
7479 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode_get_ctime_nsec(inode));
7480
7481 ocfs2_update_inode_fsync_trans(handle, inode, 1);
7482 ocfs2_journal_dirty(handle, di_bh);
7483
7484out_commit:
7485 ocfs2_commit_trans(osb, handle);
7486
7487out:
7488 return ret;
7489}
7490
7491static int ocfs2_trim_extent(struct super_block *sb,
7492 struct ocfs2_group_desc *gd,
7493 u64 group, u32 start, u32 count)
7494{
7495 u64 discard, bcount;
7496 struct ocfs2_super *osb = OCFS2_SB(sb);
7497
7498 bcount = ocfs2_clusters_to_blocks(sb, count);
7499 discard = ocfs2_clusters_to_blocks(sb, start);
7500
7501 /*
7502 * For the first cluster group, the gd->bg_blkno is not at the start
7503 * of the group, but at an offset from the start. If we add it while
7504 * calculating discard for first group, we will wrongly start fstrim a
7505 * few blocks after the desried start block and the range can cross
7506 * over into the next cluster group. So, add it only if this is not
7507 * the first cluster group.
7508 */
7509 if (group != osb->first_cluster_group_blkno)
7510 discard += le64_to_cpu(gd->bg_blkno);
7511
7512 trace_ocfs2_trim_extent(sb, (unsigned long long)discard, bcount);
7513
7514 return sb_issue_discard(sb, discard, bcount, GFP_NOFS, 0);
7515}
7516
7517static int ocfs2_trim_group(struct super_block *sb,
7518 struct ocfs2_group_desc *gd, u64 group,
7519 u32 start, u32 max, u32 minbits)
7520{
7521 int ret = 0, count = 0, next;
7522 void *bitmap = gd->bg_bitmap;
7523
7524 if (le16_to_cpu(gd->bg_free_bits_count) < minbits)
7525 return 0;
7526
7527 trace_ocfs2_trim_group((unsigned long long)le64_to_cpu(gd->bg_blkno),
7528 start, max, minbits);
7529
7530 while (start < max) {
7531 start = ocfs2_find_next_zero_bit(bitmap, max, start);
7532 if (start >= max)
7533 break;
7534 next = ocfs2_find_next_bit(bitmap, max, start);
7535
7536 if ((next - start) >= minbits) {
7537 ret = ocfs2_trim_extent(sb, gd, group,
7538 start, next - start);
7539 if (ret < 0) {
7540 mlog_errno(ret);
7541 break;
7542 }
7543 count += next - start;
7544 }
7545 start = next + 1;
7546
7547 if (fatal_signal_pending(current)) {
7548 count = -ERESTARTSYS;
7549 break;
7550 }
7551
7552 if ((le16_to_cpu(gd->bg_free_bits_count) - count) < minbits)
7553 break;
7554 }
7555
7556 if (ret < 0)
7557 count = ret;
7558
7559 return count;
7560}
7561
7562static
7563int ocfs2_trim_mainbm(struct super_block *sb, struct fstrim_range *range)
7564{
7565 struct ocfs2_super *osb = OCFS2_SB(sb);
7566 u64 start, len, trimmed = 0, first_group, last_group = 0, group = 0;
7567 int ret, cnt;
7568 u32 first_bit, last_bit, minlen;
7569 struct buffer_head *main_bm_bh = NULL;
7570 struct inode *main_bm_inode = NULL;
7571 struct buffer_head *gd_bh = NULL;
7572 struct ocfs2_dinode *main_bm;
7573 struct ocfs2_group_desc *gd = NULL;
7574
7575 start = range->start >> osb->s_clustersize_bits;
7576 len = range->len >> osb->s_clustersize_bits;
7577 minlen = range->minlen >> osb->s_clustersize_bits;
7578
7579 if (minlen >= osb->bitmap_cpg || range->len < sb->s_blocksize)
7580 return -EINVAL;
7581
7582 trace_ocfs2_trim_mainbm(start, len, minlen);
7583
7584next_group:
7585 main_bm_inode = ocfs2_get_system_file_inode(osb,
7586 GLOBAL_BITMAP_SYSTEM_INODE,
7587 OCFS2_INVALID_SLOT);
7588 if (!main_bm_inode) {
7589 ret = -EIO;
7590 mlog_errno(ret);
7591 goto out;
7592 }
7593
7594 inode_lock(main_bm_inode);
7595
7596 ret = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 0);
7597 if (ret < 0) {
7598 mlog_errno(ret);
7599 goto out_mutex;
7600 }
7601 main_bm = (struct ocfs2_dinode *)main_bm_bh->b_data;
7602
7603 /*
7604 * Do some check before trim the first group.
7605 */
7606 if (!group) {
7607 if (start >= le32_to_cpu(main_bm->i_clusters)) {
7608 ret = -EINVAL;
7609 goto out_unlock;
7610 }
7611
7612 if (start + len > le32_to_cpu(main_bm->i_clusters))
7613 len = le32_to_cpu(main_bm->i_clusters) - start;
7614
7615 /*
7616 * Determine first and last group to examine based on
7617 * start and len
7618 */
7619 first_group = ocfs2_which_cluster_group(main_bm_inode, start);
7620 if (first_group == osb->first_cluster_group_blkno)
7621 first_bit = start;
7622 else
7623 first_bit = start - ocfs2_blocks_to_clusters(sb,
7624 first_group);
7625 last_group = ocfs2_which_cluster_group(main_bm_inode,
7626 start + len - 1);
7627 group = first_group;
7628 }
7629
7630 do {
7631 if (first_bit + len >= osb->bitmap_cpg)
7632 last_bit = osb->bitmap_cpg;
7633 else
7634 last_bit = first_bit + len;
7635
7636 ret = ocfs2_read_group_descriptor(main_bm_inode,
7637 main_bm, group,
7638 &gd_bh);
7639 if (ret < 0) {
7640 mlog_errno(ret);
7641 break;
7642 }
7643
7644 gd = (struct ocfs2_group_desc *)gd_bh->b_data;
7645 cnt = ocfs2_trim_group(sb, gd, group,
7646 first_bit, last_bit, minlen);
7647 brelse(gd_bh);
7648 gd_bh = NULL;
7649 if (cnt < 0) {
7650 ret = cnt;
7651 mlog_errno(ret);
7652 break;
7653 }
7654
7655 trimmed += cnt;
7656 len -= osb->bitmap_cpg - first_bit;
7657 first_bit = 0;
7658 if (group == osb->first_cluster_group_blkno)
7659 group = ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7660 else
7661 group += ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7662 } while (0);
7663
7664out_unlock:
7665 ocfs2_inode_unlock(main_bm_inode, 0);
7666 brelse(main_bm_bh);
7667 main_bm_bh = NULL;
7668out_mutex:
7669 inode_unlock(main_bm_inode);
7670 iput(main_bm_inode);
7671
7672 /*
7673 * If all the groups trim are not done or failed, but we should release
7674 * main_bm related locks for avoiding the current IO starve, then go to
7675 * trim the next group
7676 */
7677 if (ret >= 0 && group <= last_group) {
7678 cond_resched();
7679 goto next_group;
7680 }
7681out:
7682 range->len = trimmed * osb->s_clustersize;
7683 return ret;
7684}
7685
7686int ocfs2_trim_fs(struct super_block *sb, struct fstrim_range *range)
7687{
7688 int ret;
7689 struct ocfs2_super *osb = OCFS2_SB(sb);
7690 struct ocfs2_trim_fs_info info, *pinfo = NULL;
7691
7692 ocfs2_trim_fs_lock_res_init(osb);
7693
7694 trace_ocfs2_trim_fs(range->start, range->len, range->minlen);
7695
7696 ret = ocfs2_trim_fs_lock(osb, NULL, 1);
7697 if (ret < 0) {
7698 if (ret != -EAGAIN) {
7699 mlog_errno(ret);
7700 ocfs2_trim_fs_lock_res_uninit(osb);
7701 return ret;
7702 }
7703
7704 mlog(ML_NOTICE, "Wait for trim on device (%s) to "
7705 "finish, which is running from another node.\n",
7706 osb->dev_str);
7707 ret = ocfs2_trim_fs_lock(osb, &info, 0);
7708 if (ret < 0) {
7709 mlog_errno(ret);
7710 ocfs2_trim_fs_lock_res_uninit(osb);
7711 return ret;
7712 }
7713
7714 if (info.tf_valid && info.tf_success &&
7715 info.tf_start == range->start &&
7716 info.tf_len == range->len &&
7717 info.tf_minlen == range->minlen) {
7718 /* Avoid sending duplicated trim to a shared device */
7719 mlog(ML_NOTICE, "The same trim on device (%s) was "
7720 "just done from node (%u), return.\n",
7721 osb->dev_str, info.tf_nodenum);
7722 range->len = info.tf_trimlen;
7723 goto out;
7724 }
7725 }
7726
7727 info.tf_nodenum = osb->node_num;
7728 info.tf_start = range->start;
7729 info.tf_len = range->len;
7730 info.tf_minlen = range->minlen;
7731
7732 ret = ocfs2_trim_mainbm(sb, range);
7733
7734 info.tf_trimlen = range->len;
7735 info.tf_success = (ret < 0 ? 0 : 1);
7736 pinfo = &info;
7737out:
7738 ocfs2_trim_fs_unlock(osb, pinfo);
7739 ocfs2_trim_fs_lock_res_uninit(osb);
7740 return ret;
7741}