Btrfs: fix space leak when we fail to reserve metadata space
[linux-2.6.git] / fs / ocfs2 / alloc.c
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * alloc.c
5  *
6  * Extent allocs and frees
7  *
8  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
9  *
10  * This program is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU General Public
12  * License as published by the Free Software Foundation; either
13  * version 2 of the License, or (at your option) any later version.
14  *
15  * This program is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * General Public License for more details.
19  *
20  * You should have received a copy of the GNU General Public
21  * License along with this program; if not, write to the
22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23  * Boston, MA 021110-1307, USA.
24  */
25
26 #include <linux/fs.h>
27 #include <linux/types.h>
28 #include <linux/slab.h>
29 #include <linux/highmem.h>
30 #include <linux/swap.h>
31 #include <linux/quotaops.h>
32 #include <linux/blkdev.h>
33
34 #include <cluster/masklog.h>
35
36 #include "ocfs2.h"
37
38 #include "alloc.h"
39 #include "aops.h"
40 #include "blockcheck.h"
41 #include "dlmglue.h"
42 #include "extent_map.h"
43 #include "inode.h"
44 #include "journal.h"
45 #include "localalloc.h"
46 #include "suballoc.h"
47 #include "sysfile.h"
48 #include "file.h"
49 #include "super.h"
50 #include "uptodate.h"
51 #include "xattr.h"
52 #include "refcounttree.h"
53 #include "ocfs2_trace.h"
54
55 #include "buffer_head_io.h"
56
57 enum ocfs2_contig_type {
58         CONTIG_NONE = 0,
59         CONTIG_LEFT,
60         CONTIG_RIGHT,
61         CONTIG_LEFTRIGHT,
62 };
63
64 static enum ocfs2_contig_type
65         ocfs2_extent_rec_contig(struct super_block *sb,
66                                 struct ocfs2_extent_rec *ext,
67                                 struct ocfs2_extent_rec *insert_rec);
68 /*
69  * Operations for a specific extent tree type.
70  *
71  * To implement an on-disk btree (extent tree) type in ocfs2, add
72  * an ocfs2_extent_tree_operations structure and the matching
73  * ocfs2_init_<thingy>_extent_tree() function.  That's pretty much it
74  * for the allocation portion of the extent tree.
75  */
76 struct ocfs2_extent_tree_operations {
77         /*
78          * last_eb_blk is the block number of the right most leaf extent
79          * block.  Most on-disk structures containing an extent tree store
80          * this value for fast access.  The ->eo_set_last_eb_blk() and
81          * ->eo_get_last_eb_blk() operations access this value.  They are
82          *  both required.
83          */
84         void (*eo_set_last_eb_blk)(struct ocfs2_extent_tree *et,
85                                    u64 blkno);
86         u64 (*eo_get_last_eb_blk)(struct ocfs2_extent_tree *et);
87
88         /*
89          * The on-disk structure usually keeps track of how many total
90          * clusters are stored in this extent tree.  This function updates
91          * that value.  new_clusters is the delta, and must be
92          * added to the total.  Required.
93          */
94         void (*eo_update_clusters)(struct ocfs2_extent_tree *et,
95                                    u32 new_clusters);
96
97         /*
98          * If this extent tree is supported by an extent map, insert
99          * a record into the map.
100          */
101         void (*eo_extent_map_insert)(struct ocfs2_extent_tree *et,
102                                      struct ocfs2_extent_rec *rec);
103
104         /*
105          * If this extent tree is supported by an extent map, truncate the
106          * map to clusters,
107          */
108         void (*eo_extent_map_truncate)(struct ocfs2_extent_tree *et,
109                                        u32 clusters);
110
111         /*
112          * If ->eo_insert_check() exists, it is called before rec is
113          * inserted into the extent tree.  It is optional.
114          */
115         int (*eo_insert_check)(struct ocfs2_extent_tree *et,
116                                struct ocfs2_extent_rec *rec);
117         int (*eo_sanity_check)(struct ocfs2_extent_tree *et);
118
119         /*
120          * --------------------------------------------------------------
121          * The remaining are internal to ocfs2_extent_tree and don't have
122          * accessor functions
123          */
124
125         /*
126          * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el.
127          * It is required.
128          */
129         void (*eo_fill_root_el)(struct ocfs2_extent_tree *et);
130
131         /*
132          * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if
133          * it exists.  If it does not, et->et_max_leaf_clusters is set
134          * to 0 (unlimited).  Optional.
135          */
136         void (*eo_fill_max_leaf_clusters)(struct ocfs2_extent_tree *et);
137
138         /*
139          * ->eo_extent_contig test whether the 2 ocfs2_extent_rec
140          * are contiguous or not. Optional. Don't need to set it if use
141          * ocfs2_extent_rec as the tree leaf.
142          */
143         enum ocfs2_contig_type
144                 (*eo_extent_contig)(struct ocfs2_extent_tree *et,
145                                     struct ocfs2_extent_rec *ext,
146                                     struct ocfs2_extent_rec *insert_rec);
147 };
148
149
150 /*
151  * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check
152  * in the methods.
153  */
154 static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et);
155 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
156                                          u64 blkno);
157 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
158                                          u32 clusters);
159 static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
160                                            struct ocfs2_extent_rec *rec);
161 static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
162                                              u32 clusters);
163 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
164                                      struct ocfs2_extent_rec *rec);
165 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et);
166 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et);
167 static struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops = {
168         .eo_set_last_eb_blk     = ocfs2_dinode_set_last_eb_blk,
169         .eo_get_last_eb_blk     = ocfs2_dinode_get_last_eb_blk,
170         .eo_update_clusters     = ocfs2_dinode_update_clusters,
171         .eo_extent_map_insert   = ocfs2_dinode_extent_map_insert,
172         .eo_extent_map_truncate = ocfs2_dinode_extent_map_truncate,
173         .eo_insert_check        = ocfs2_dinode_insert_check,
174         .eo_sanity_check        = ocfs2_dinode_sanity_check,
175         .eo_fill_root_el        = ocfs2_dinode_fill_root_el,
176 };
177
178 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
179                                          u64 blkno)
180 {
181         struct ocfs2_dinode *di = et->et_object;
182
183         BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
184         di->i_last_eb_blk = cpu_to_le64(blkno);
185 }
186
187 static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et)
188 {
189         struct ocfs2_dinode *di = et->et_object;
190
191         BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
192         return le64_to_cpu(di->i_last_eb_blk);
193 }
194
195 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
196                                          u32 clusters)
197 {
198         struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
199         struct ocfs2_dinode *di = et->et_object;
200
201         le32_add_cpu(&di->i_clusters, clusters);
202         spin_lock(&oi->ip_lock);
203         oi->ip_clusters = le32_to_cpu(di->i_clusters);
204         spin_unlock(&oi->ip_lock);
205 }
206
207 static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
208                                            struct ocfs2_extent_rec *rec)
209 {
210         struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
211
212         ocfs2_extent_map_insert_rec(inode, rec);
213 }
214
215 static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
216                                              u32 clusters)
217 {
218         struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
219
220         ocfs2_extent_map_trunc(inode, clusters);
221 }
222
223 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
224                                      struct ocfs2_extent_rec *rec)
225 {
226         struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
227         struct ocfs2_super *osb = OCFS2_SB(oi->vfs_inode.i_sb);
228
229         BUG_ON(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL);
230         mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) &&
231                         (oi->ip_clusters != le32_to_cpu(rec->e_cpos)),
232                         "Device %s, asking for sparse allocation: inode %llu, "
233                         "cpos %u, clusters %u\n",
234                         osb->dev_str,
235                         (unsigned long long)oi->ip_blkno,
236                         rec->e_cpos, oi->ip_clusters);
237
238         return 0;
239 }
240
241 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et)
242 {
243         struct ocfs2_dinode *di = et->et_object;
244
245         BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
246         BUG_ON(!OCFS2_IS_VALID_DINODE(di));
247
248         return 0;
249 }
250
251 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et)
252 {
253         struct ocfs2_dinode *di = et->et_object;
254
255         et->et_root_el = &di->id2.i_list;
256 }
257
258
259 static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree *et)
260 {
261         struct ocfs2_xattr_value_buf *vb = et->et_object;
262
263         et->et_root_el = &vb->vb_xv->xr_list;
264 }
265
266 static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree *et,
267                                               u64 blkno)
268 {
269         struct ocfs2_xattr_value_buf *vb = et->et_object;
270
271         vb->vb_xv->xr_last_eb_blk = cpu_to_le64(blkno);
272 }
273
274 static u64 ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree *et)
275 {
276         struct ocfs2_xattr_value_buf *vb = et->et_object;
277
278         return le64_to_cpu(vb->vb_xv->xr_last_eb_blk);
279 }
280
281 static void ocfs2_xattr_value_update_clusters(struct ocfs2_extent_tree *et,
282                                               u32 clusters)
283 {
284         struct ocfs2_xattr_value_buf *vb = et->et_object;
285
286         le32_add_cpu(&vb->vb_xv->xr_clusters, clusters);
287 }
288
289 static struct ocfs2_extent_tree_operations ocfs2_xattr_value_et_ops = {
290         .eo_set_last_eb_blk     = ocfs2_xattr_value_set_last_eb_blk,
291         .eo_get_last_eb_blk     = ocfs2_xattr_value_get_last_eb_blk,
292         .eo_update_clusters     = ocfs2_xattr_value_update_clusters,
293         .eo_fill_root_el        = ocfs2_xattr_value_fill_root_el,
294 };
295
296 static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree *et)
297 {
298         struct ocfs2_xattr_block *xb = et->et_object;
299
300         et->et_root_el = &xb->xb_attrs.xb_root.xt_list;
301 }
302
303 static void ocfs2_xattr_tree_fill_max_leaf_clusters(struct ocfs2_extent_tree *et)
304 {
305         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
306         et->et_max_leaf_clusters =
307                 ocfs2_clusters_for_bytes(sb, OCFS2_MAX_XATTR_TREE_LEAF_SIZE);
308 }
309
310 static void ocfs2_xattr_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
311                                              u64 blkno)
312 {
313         struct ocfs2_xattr_block *xb = et->et_object;
314         struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
315
316         xt->xt_last_eb_blk = cpu_to_le64(blkno);
317 }
318
319 static u64 ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
320 {
321         struct ocfs2_xattr_block *xb = et->et_object;
322         struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
323
324         return le64_to_cpu(xt->xt_last_eb_blk);
325 }
326
327 static void ocfs2_xattr_tree_update_clusters(struct ocfs2_extent_tree *et,
328                                              u32 clusters)
329 {
330         struct ocfs2_xattr_block *xb = et->et_object;
331
332         le32_add_cpu(&xb->xb_attrs.xb_root.xt_clusters, clusters);
333 }
334
335 static struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops = {
336         .eo_set_last_eb_blk     = ocfs2_xattr_tree_set_last_eb_blk,
337         .eo_get_last_eb_blk     = ocfs2_xattr_tree_get_last_eb_blk,
338         .eo_update_clusters     = ocfs2_xattr_tree_update_clusters,
339         .eo_fill_root_el        = ocfs2_xattr_tree_fill_root_el,
340         .eo_fill_max_leaf_clusters = ocfs2_xattr_tree_fill_max_leaf_clusters,
341 };
342
343 static void ocfs2_dx_root_set_last_eb_blk(struct ocfs2_extent_tree *et,
344                                           u64 blkno)
345 {
346         struct ocfs2_dx_root_block *dx_root = et->et_object;
347
348         dx_root->dr_last_eb_blk = cpu_to_le64(blkno);
349 }
350
351 static u64 ocfs2_dx_root_get_last_eb_blk(struct ocfs2_extent_tree *et)
352 {
353         struct ocfs2_dx_root_block *dx_root = et->et_object;
354
355         return le64_to_cpu(dx_root->dr_last_eb_blk);
356 }
357
358 static void ocfs2_dx_root_update_clusters(struct ocfs2_extent_tree *et,
359                                           u32 clusters)
360 {
361         struct ocfs2_dx_root_block *dx_root = et->et_object;
362
363         le32_add_cpu(&dx_root->dr_clusters, clusters);
364 }
365
366 static int ocfs2_dx_root_sanity_check(struct ocfs2_extent_tree *et)
367 {
368         struct ocfs2_dx_root_block *dx_root = et->et_object;
369
370         BUG_ON(!OCFS2_IS_VALID_DX_ROOT(dx_root));
371
372         return 0;
373 }
374
375 static void ocfs2_dx_root_fill_root_el(struct ocfs2_extent_tree *et)
376 {
377         struct ocfs2_dx_root_block *dx_root = et->et_object;
378
379         et->et_root_el = &dx_root->dr_list;
380 }
381
382 static struct ocfs2_extent_tree_operations ocfs2_dx_root_et_ops = {
383         .eo_set_last_eb_blk     = ocfs2_dx_root_set_last_eb_blk,
384         .eo_get_last_eb_blk     = ocfs2_dx_root_get_last_eb_blk,
385         .eo_update_clusters     = ocfs2_dx_root_update_clusters,
386         .eo_sanity_check        = ocfs2_dx_root_sanity_check,
387         .eo_fill_root_el        = ocfs2_dx_root_fill_root_el,
388 };
389
390 static void ocfs2_refcount_tree_fill_root_el(struct ocfs2_extent_tree *et)
391 {
392         struct ocfs2_refcount_block *rb = et->et_object;
393
394         et->et_root_el = &rb->rf_list;
395 }
396
397 static void ocfs2_refcount_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
398                                                 u64 blkno)
399 {
400         struct ocfs2_refcount_block *rb = et->et_object;
401
402         rb->rf_last_eb_blk = cpu_to_le64(blkno);
403 }
404
405 static u64 ocfs2_refcount_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
406 {
407         struct ocfs2_refcount_block *rb = et->et_object;
408
409         return le64_to_cpu(rb->rf_last_eb_blk);
410 }
411
412 static void ocfs2_refcount_tree_update_clusters(struct ocfs2_extent_tree *et,
413                                                 u32 clusters)
414 {
415         struct ocfs2_refcount_block *rb = et->et_object;
416
417         le32_add_cpu(&rb->rf_clusters, clusters);
418 }
419
420 static enum ocfs2_contig_type
421 ocfs2_refcount_tree_extent_contig(struct ocfs2_extent_tree *et,
422                                   struct ocfs2_extent_rec *ext,
423                                   struct ocfs2_extent_rec *insert_rec)
424 {
425         return CONTIG_NONE;
426 }
427
428 static struct ocfs2_extent_tree_operations ocfs2_refcount_tree_et_ops = {
429         .eo_set_last_eb_blk     = ocfs2_refcount_tree_set_last_eb_blk,
430         .eo_get_last_eb_blk     = ocfs2_refcount_tree_get_last_eb_blk,
431         .eo_update_clusters     = ocfs2_refcount_tree_update_clusters,
432         .eo_fill_root_el        = ocfs2_refcount_tree_fill_root_el,
433         .eo_extent_contig       = ocfs2_refcount_tree_extent_contig,
434 };
435
436 static void __ocfs2_init_extent_tree(struct ocfs2_extent_tree *et,
437                                      struct ocfs2_caching_info *ci,
438                                      struct buffer_head *bh,
439                                      ocfs2_journal_access_func access,
440                                      void *obj,
441                                      struct ocfs2_extent_tree_operations *ops)
442 {
443         et->et_ops = ops;
444         et->et_root_bh = bh;
445         et->et_ci = ci;
446         et->et_root_journal_access = access;
447         if (!obj)
448                 obj = (void *)bh->b_data;
449         et->et_object = obj;
450
451         et->et_ops->eo_fill_root_el(et);
452         if (!et->et_ops->eo_fill_max_leaf_clusters)
453                 et->et_max_leaf_clusters = 0;
454         else
455                 et->et_ops->eo_fill_max_leaf_clusters(et);
456 }
457
458 void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree *et,
459                                    struct ocfs2_caching_info *ci,
460                                    struct buffer_head *bh)
461 {
462         __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_di,
463                                  NULL, &ocfs2_dinode_et_ops);
464 }
465
466 void ocfs2_init_xattr_tree_extent_tree(struct ocfs2_extent_tree *et,
467                                        struct ocfs2_caching_info *ci,
468                                        struct buffer_head *bh)
469 {
470         __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_xb,
471                                  NULL, &ocfs2_xattr_tree_et_ops);
472 }
473
474 void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree *et,
475                                         struct ocfs2_caching_info *ci,
476                                         struct ocfs2_xattr_value_buf *vb)
477 {
478         __ocfs2_init_extent_tree(et, ci, vb->vb_bh, vb->vb_access, vb,
479                                  &ocfs2_xattr_value_et_ops);
480 }
481
482 void ocfs2_init_dx_root_extent_tree(struct ocfs2_extent_tree *et,
483                                     struct ocfs2_caching_info *ci,
484                                     struct buffer_head *bh)
485 {
486         __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_dr,
487                                  NULL, &ocfs2_dx_root_et_ops);
488 }
489
490 void ocfs2_init_refcount_extent_tree(struct ocfs2_extent_tree *et,
491                                      struct ocfs2_caching_info *ci,
492                                      struct buffer_head *bh)
493 {
494         __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_rb,
495                                  NULL, &ocfs2_refcount_tree_et_ops);
496 }
497
498 static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree *et,
499                                             u64 new_last_eb_blk)
500 {
501         et->et_ops->eo_set_last_eb_blk(et, new_last_eb_blk);
502 }
503
504 static inline u64 ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree *et)
505 {
506         return et->et_ops->eo_get_last_eb_blk(et);
507 }
508
509 static inline void ocfs2_et_update_clusters(struct ocfs2_extent_tree *et,
510                                             u32 clusters)
511 {
512         et->et_ops->eo_update_clusters(et, clusters);
513 }
514
515 static inline void ocfs2_et_extent_map_insert(struct ocfs2_extent_tree *et,
516                                               struct ocfs2_extent_rec *rec)
517 {
518         if (et->et_ops->eo_extent_map_insert)
519                 et->et_ops->eo_extent_map_insert(et, rec);
520 }
521
522 static inline void ocfs2_et_extent_map_truncate(struct ocfs2_extent_tree *et,
523                                                 u32 clusters)
524 {
525         if (et->et_ops->eo_extent_map_truncate)
526                 et->et_ops->eo_extent_map_truncate(et, clusters);
527 }
528
529 static inline int ocfs2_et_root_journal_access(handle_t *handle,
530                                                struct ocfs2_extent_tree *et,
531                                                int type)
532 {
533         return et->et_root_journal_access(handle, et->et_ci, et->et_root_bh,
534                                           type);
535 }
536
537 static inline enum ocfs2_contig_type
538         ocfs2_et_extent_contig(struct ocfs2_extent_tree *et,
539                                struct ocfs2_extent_rec *rec,
540                                struct ocfs2_extent_rec *insert_rec)
541 {
542         if (et->et_ops->eo_extent_contig)
543                 return et->et_ops->eo_extent_contig(et, rec, insert_rec);
544
545         return ocfs2_extent_rec_contig(
546                                 ocfs2_metadata_cache_get_super(et->et_ci),
547                                 rec, insert_rec);
548 }
549
550 static inline int ocfs2_et_insert_check(struct ocfs2_extent_tree *et,
551                                         struct ocfs2_extent_rec *rec)
552 {
553         int ret = 0;
554
555         if (et->et_ops->eo_insert_check)
556                 ret = et->et_ops->eo_insert_check(et, rec);
557         return ret;
558 }
559
560 static inline int ocfs2_et_sanity_check(struct ocfs2_extent_tree *et)
561 {
562         int ret = 0;
563
564         if (et->et_ops->eo_sanity_check)
565                 ret = et->et_ops->eo_sanity_check(et);
566         return ret;
567 }
568
569 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
570                                          struct ocfs2_extent_block *eb);
571 static void ocfs2_adjust_rightmost_records(handle_t *handle,
572                                            struct ocfs2_extent_tree *et,
573                                            struct ocfs2_path *path,
574                                            struct ocfs2_extent_rec *insert_rec);
575 /*
576  * Reset the actual path elements so that we can re-use the structure
577  * to build another path. Generally, this involves freeing the buffer
578  * heads.
579  */
580 void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root)
581 {
582         int i, start = 0, depth = 0;
583         struct ocfs2_path_item *node;
584
585         if (keep_root)
586                 start = 1;
587
588         for(i = start; i < path_num_items(path); i++) {
589                 node = &path->p_node[i];
590
591                 brelse(node->bh);
592                 node->bh = NULL;
593                 node->el = NULL;
594         }
595
596         /*
597          * Tree depth may change during truncate, or insert. If we're
598          * keeping the root extent list, then make sure that our path
599          * structure reflects the proper depth.
600          */
601         if (keep_root)
602                 depth = le16_to_cpu(path_root_el(path)->l_tree_depth);
603         else
604                 path_root_access(path) = NULL;
605
606         path->p_tree_depth = depth;
607 }
608
609 void ocfs2_free_path(struct ocfs2_path *path)
610 {
611         if (path) {
612                 ocfs2_reinit_path(path, 0);
613                 kfree(path);
614         }
615 }
616
617 /*
618  * All the elements of src into dest. After this call, src could be freed
619  * without affecting dest.
620  *
621  * Both paths should have the same root. Any non-root elements of dest
622  * will be freed.
623  */
624 static void ocfs2_cp_path(struct ocfs2_path *dest, struct ocfs2_path *src)
625 {
626         int i;
627
628         BUG_ON(path_root_bh(dest) != path_root_bh(src));
629         BUG_ON(path_root_el(dest) != path_root_el(src));
630         BUG_ON(path_root_access(dest) != path_root_access(src));
631
632         ocfs2_reinit_path(dest, 1);
633
634         for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
635                 dest->p_node[i].bh = src->p_node[i].bh;
636                 dest->p_node[i].el = src->p_node[i].el;
637
638                 if (dest->p_node[i].bh)
639                         get_bh(dest->p_node[i].bh);
640         }
641 }
642
643 /*
644  * Make the *dest path the same as src and re-initialize src path to
645  * have a root only.
646  */
647 static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src)
648 {
649         int i;
650
651         BUG_ON(path_root_bh(dest) != path_root_bh(src));
652         BUG_ON(path_root_access(dest) != path_root_access(src));
653
654         for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
655                 brelse(dest->p_node[i].bh);
656
657                 dest->p_node[i].bh = src->p_node[i].bh;
658                 dest->p_node[i].el = src->p_node[i].el;
659
660                 src->p_node[i].bh = NULL;
661                 src->p_node[i].el = NULL;
662         }
663 }
664
665 /*
666  * Insert an extent block at given index.
667  *
668  * This will not take an additional reference on eb_bh.
669  */
670 static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index,
671                                         struct buffer_head *eb_bh)
672 {
673         struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data;
674
675         /*
676          * Right now, no root bh is an extent block, so this helps
677          * catch code errors with dinode trees. The assertion can be
678          * safely removed if we ever need to insert extent block
679          * structures at the root.
680          */
681         BUG_ON(index == 0);
682
683         path->p_node[index].bh = eb_bh;
684         path->p_node[index].el = &eb->h_list;
685 }
686
687 static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh,
688                                          struct ocfs2_extent_list *root_el,
689                                          ocfs2_journal_access_func access)
690 {
691         struct ocfs2_path *path;
692
693         BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH);
694
695         path = kzalloc(sizeof(*path), GFP_NOFS);
696         if (path) {
697                 path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth);
698                 get_bh(root_bh);
699                 path_root_bh(path) = root_bh;
700                 path_root_el(path) = root_el;
701                 path_root_access(path) = access;
702         }
703
704         return path;
705 }
706
707 struct ocfs2_path *ocfs2_new_path_from_path(struct ocfs2_path *path)
708 {
709         return ocfs2_new_path(path_root_bh(path), path_root_el(path),
710                               path_root_access(path));
711 }
712
713 struct ocfs2_path *ocfs2_new_path_from_et(struct ocfs2_extent_tree *et)
714 {
715         return ocfs2_new_path(et->et_root_bh, et->et_root_el,
716                               et->et_root_journal_access);
717 }
718
719 /*
720  * Journal the buffer at depth idx.  All idx>0 are extent_blocks,
721  * otherwise it's the root_access function.
722  *
723  * I don't like the way this function's name looks next to
724  * ocfs2_journal_access_path(), but I don't have a better one.
725  */
726 int ocfs2_path_bh_journal_access(handle_t *handle,
727                                  struct ocfs2_caching_info *ci,
728                                  struct ocfs2_path *path,
729                                  int idx)
730 {
731         ocfs2_journal_access_func access = path_root_access(path);
732
733         if (!access)
734                 access = ocfs2_journal_access;
735
736         if (idx)
737                 access = ocfs2_journal_access_eb;
738
739         return access(handle, ci, path->p_node[idx].bh,
740                       OCFS2_JOURNAL_ACCESS_WRITE);
741 }
742
743 /*
744  * Convenience function to journal all components in a path.
745  */
746 int ocfs2_journal_access_path(struct ocfs2_caching_info *ci,
747                               handle_t *handle,
748                               struct ocfs2_path *path)
749 {
750         int i, ret = 0;
751
752         if (!path)
753                 goto out;
754
755         for(i = 0; i < path_num_items(path); i++) {
756                 ret = ocfs2_path_bh_journal_access(handle, ci, path, i);
757                 if (ret < 0) {
758                         mlog_errno(ret);
759                         goto out;
760                 }
761         }
762
763 out:
764         return ret;
765 }
766
767 /*
768  * Return the index of the extent record which contains cluster #v_cluster.
769  * -1 is returned if it was not found.
770  *
771  * Should work fine on interior and exterior nodes.
772  */
773 int ocfs2_search_extent_list(struct ocfs2_extent_list *el, u32 v_cluster)
774 {
775         int ret = -1;
776         int i;
777         struct ocfs2_extent_rec *rec;
778         u32 rec_end, rec_start, clusters;
779
780         for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
781                 rec = &el->l_recs[i];
782
783                 rec_start = le32_to_cpu(rec->e_cpos);
784                 clusters = ocfs2_rec_clusters(el, rec);
785
786                 rec_end = rec_start + clusters;
787
788                 if (v_cluster >= rec_start && v_cluster < rec_end) {
789                         ret = i;
790                         break;
791                 }
792         }
793
794         return ret;
795 }
796
797 /*
798  * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
799  * ocfs2_extent_rec_contig only work properly against leaf nodes!
800  */
801 static int ocfs2_block_extent_contig(struct super_block *sb,
802                                      struct ocfs2_extent_rec *ext,
803                                      u64 blkno)
804 {
805         u64 blk_end = le64_to_cpu(ext->e_blkno);
806
807         blk_end += ocfs2_clusters_to_blocks(sb,
808                                     le16_to_cpu(ext->e_leaf_clusters));
809
810         return blkno == blk_end;
811 }
812
813 static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
814                                   struct ocfs2_extent_rec *right)
815 {
816         u32 left_range;
817
818         left_range = le32_to_cpu(left->e_cpos) +
819                 le16_to_cpu(left->e_leaf_clusters);
820
821         return (left_range == le32_to_cpu(right->e_cpos));
822 }
823
824 static enum ocfs2_contig_type
825         ocfs2_extent_rec_contig(struct super_block *sb,
826                                 struct ocfs2_extent_rec *ext,
827                                 struct ocfs2_extent_rec *insert_rec)
828 {
829         u64 blkno = le64_to_cpu(insert_rec->e_blkno);
830
831         /*
832          * Refuse to coalesce extent records with different flag
833          * fields - we don't want to mix unwritten extents with user
834          * data.
835          */
836         if (ext->e_flags != insert_rec->e_flags)
837                 return CONTIG_NONE;
838
839         if (ocfs2_extents_adjacent(ext, insert_rec) &&
840             ocfs2_block_extent_contig(sb, ext, blkno))
841                         return CONTIG_RIGHT;
842
843         blkno = le64_to_cpu(ext->e_blkno);
844         if (ocfs2_extents_adjacent(insert_rec, ext) &&
845             ocfs2_block_extent_contig(sb, insert_rec, blkno))
846                 return CONTIG_LEFT;
847
848         return CONTIG_NONE;
849 }
850
851 /*
852  * NOTE: We can have pretty much any combination of contiguousness and
853  * appending.
854  *
855  * The usefulness of APPEND_TAIL is more in that it lets us know that
856  * we'll have to update the path to that leaf.
857  */
858 enum ocfs2_append_type {
859         APPEND_NONE = 0,
860         APPEND_TAIL,
861 };
862
863 enum ocfs2_split_type {
864         SPLIT_NONE = 0,
865         SPLIT_LEFT,
866         SPLIT_RIGHT,
867 };
868
869 struct ocfs2_insert_type {
870         enum ocfs2_split_type   ins_split;
871         enum ocfs2_append_type  ins_appending;
872         enum ocfs2_contig_type  ins_contig;
873         int                     ins_contig_index;
874         int                     ins_tree_depth;
875 };
876
877 struct ocfs2_merge_ctxt {
878         enum ocfs2_contig_type  c_contig_type;
879         int                     c_has_empty_extent;
880         int                     c_split_covers_rec;
881 };
882
883 static int ocfs2_validate_extent_block(struct super_block *sb,
884                                        struct buffer_head *bh)
885 {
886         int rc;
887         struct ocfs2_extent_block *eb =
888                 (struct ocfs2_extent_block *)bh->b_data;
889
890         trace_ocfs2_validate_extent_block((unsigned long long)bh->b_blocknr);
891
892         BUG_ON(!buffer_uptodate(bh));
893
894         /*
895          * If the ecc fails, we return the error but otherwise
896          * leave the filesystem running.  We know any error is
897          * local to this block.
898          */
899         rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &eb->h_check);
900         if (rc) {
901                 mlog(ML_ERROR, "Checksum failed for extent block %llu\n",
902                      (unsigned long long)bh->b_blocknr);
903                 return rc;
904         }
905
906         /*
907          * Errors after here are fatal.
908          */
909
910         if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
911                 ocfs2_error(sb,
912                             "Extent block #%llu has bad signature %.*s",
913                             (unsigned long long)bh->b_blocknr, 7,
914                             eb->h_signature);
915                 return -EINVAL;
916         }
917
918         if (le64_to_cpu(eb->h_blkno) != bh->b_blocknr) {
919                 ocfs2_error(sb,
920                             "Extent block #%llu has an invalid h_blkno "
921                             "of %llu",
922                             (unsigned long long)bh->b_blocknr,
923                             (unsigned long long)le64_to_cpu(eb->h_blkno));
924                 return -EINVAL;
925         }
926
927         if (le32_to_cpu(eb->h_fs_generation) != OCFS2_SB(sb)->fs_generation) {
928                 ocfs2_error(sb,
929                             "Extent block #%llu has an invalid "
930                             "h_fs_generation of #%u",
931                             (unsigned long long)bh->b_blocknr,
932                             le32_to_cpu(eb->h_fs_generation));
933                 return -EINVAL;
934         }
935
936         return 0;
937 }
938
939 int ocfs2_read_extent_block(struct ocfs2_caching_info *ci, u64 eb_blkno,
940                             struct buffer_head **bh)
941 {
942         int rc;
943         struct buffer_head *tmp = *bh;
944
945         rc = ocfs2_read_block(ci, eb_blkno, &tmp,
946                               ocfs2_validate_extent_block);
947
948         /* If ocfs2_read_block() got us a new bh, pass it up. */
949         if (!rc && !*bh)
950                 *bh = tmp;
951
952         return rc;
953 }
954
955
956 /*
957  * How many free extents have we got before we need more meta data?
958  */
959 int ocfs2_num_free_extents(struct ocfs2_super *osb,
960                            struct ocfs2_extent_tree *et)
961 {
962         int retval;
963         struct ocfs2_extent_list *el = NULL;
964         struct ocfs2_extent_block *eb;
965         struct buffer_head *eb_bh = NULL;
966         u64 last_eb_blk = 0;
967
968         el = et->et_root_el;
969         last_eb_blk = ocfs2_et_get_last_eb_blk(et);
970
971         if (last_eb_blk) {
972                 retval = ocfs2_read_extent_block(et->et_ci, last_eb_blk,
973                                                  &eb_bh);
974                 if (retval < 0) {
975                         mlog_errno(retval);
976                         goto bail;
977                 }
978                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
979                 el = &eb->h_list;
980         }
981
982         BUG_ON(el->l_tree_depth != 0);
983
984         retval = le16_to_cpu(el->l_count) - le16_to_cpu(el->l_next_free_rec);
985 bail:
986         brelse(eb_bh);
987
988         trace_ocfs2_num_free_extents(retval);
989         return retval;
990 }
991
992 /* expects array to already be allocated
993  *
994  * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
995  * l_count for you
996  */
997 static int ocfs2_create_new_meta_bhs(handle_t *handle,
998                                      struct ocfs2_extent_tree *et,
999                                      int wanted,
1000                                      struct ocfs2_alloc_context *meta_ac,
1001                                      struct buffer_head *bhs[])
1002 {
1003         int count, status, i;
1004         u16 suballoc_bit_start;
1005         u32 num_got;
1006         u64 suballoc_loc, first_blkno;
1007         struct ocfs2_super *osb =
1008                 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
1009         struct ocfs2_extent_block *eb;
1010
1011         count = 0;
1012         while (count < wanted) {
1013                 status = ocfs2_claim_metadata(handle,
1014                                               meta_ac,
1015                                               wanted - count,
1016                                               &suballoc_loc,
1017                                               &suballoc_bit_start,
1018                                               &num_got,
1019                                               &first_blkno);
1020                 if (status < 0) {
1021                         mlog_errno(status);
1022                         goto bail;
1023                 }
1024
1025                 for(i = count;  i < (num_got + count); i++) {
1026                         bhs[i] = sb_getblk(osb->sb, first_blkno);
1027                         if (bhs[i] == NULL) {
1028                                 status = -EIO;
1029                                 mlog_errno(status);
1030                                 goto bail;
1031                         }
1032                         ocfs2_set_new_buffer_uptodate(et->et_ci, bhs[i]);
1033
1034                         status = ocfs2_journal_access_eb(handle, et->et_ci,
1035                                                          bhs[i],
1036                                                          OCFS2_JOURNAL_ACCESS_CREATE);
1037                         if (status < 0) {
1038                                 mlog_errno(status);
1039                                 goto bail;
1040                         }
1041
1042                         memset(bhs[i]->b_data, 0, osb->sb->s_blocksize);
1043                         eb = (struct ocfs2_extent_block *) bhs[i]->b_data;
1044                         /* Ok, setup the minimal stuff here. */
1045                         strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
1046                         eb->h_blkno = cpu_to_le64(first_blkno);
1047                         eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
1048                         eb->h_suballoc_slot =
1049                                 cpu_to_le16(meta_ac->ac_alloc_slot);
1050                         eb->h_suballoc_loc = cpu_to_le64(suballoc_loc);
1051                         eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1052                         eb->h_list.l_count =
1053                                 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
1054
1055                         suballoc_bit_start++;
1056                         first_blkno++;
1057
1058                         /* We'll also be dirtied by the caller, so
1059                          * this isn't absolutely necessary. */
1060                         ocfs2_journal_dirty(handle, bhs[i]);
1061                 }
1062
1063                 count += num_got;
1064         }
1065
1066         status = 0;
1067 bail:
1068         if (status < 0) {
1069                 for(i = 0; i < wanted; i++) {
1070                         brelse(bhs[i]);
1071                         bhs[i] = NULL;
1072                 }
1073                 mlog_errno(status);
1074         }
1075         return status;
1076 }
1077
1078 /*
1079  * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
1080  *
1081  * Returns the sum of the rightmost extent rec logical offset and
1082  * cluster count.
1083  *
1084  * ocfs2_add_branch() uses this to determine what logical cluster
1085  * value should be populated into the leftmost new branch records.
1086  *
1087  * ocfs2_shift_tree_depth() uses this to determine the # clusters
1088  * value for the new topmost tree record.
1089  */
1090 static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list  *el)
1091 {
1092         int i;
1093
1094         i = le16_to_cpu(el->l_next_free_rec) - 1;
1095
1096         return le32_to_cpu(el->l_recs[i].e_cpos) +
1097                 ocfs2_rec_clusters(el, &el->l_recs[i]);
1098 }
1099
1100 /*
1101  * Change range of the branches in the right most path according to the leaf
1102  * extent block's rightmost record.
1103  */
1104 static int ocfs2_adjust_rightmost_branch(handle_t *handle,
1105                                          struct ocfs2_extent_tree *et)
1106 {
1107         int status;
1108         struct ocfs2_path *path = NULL;
1109         struct ocfs2_extent_list *el;
1110         struct ocfs2_extent_rec *rec;
1111
1112         path = ocfs2_new_path_from_et(et);
1113         if (!path) {
1114                 status = -ENOMEM;
1115                 return status;
1116         }
1117
1118         status = ocfs2_find_path(et->et_ci, path, UINT_MAX);
1119         if (status < 0) {
1120                 mlog_errno(status);
1121                 goto out;
1122         }
1123
1124         status = ocfs2_extend_trans(handle, path_num_items(path));
1125         if (status < 0) {
1126                 mlog_errno(status);
1127                 goto out;
1128         }
1129
1130         status = ocfs2_journal_access_path(et->et_ci, handle, path);
1131         if (status < 0) {
1132                 mlog_errno(status);
1133                 goto out;
1134         }
1135
1136         el = path_leaf_el(path);
1137         rec = &el->l_recs[le16_to_cpu(el->l_next_free_rec) - 1];
1138
1139         ocfs2_adjust_rightmost_records(handle, et, path, rec);
1140
1141 out:
1142         ocfs2_free_path(path);
1143         return status;
1144 }
1145
1146 /*
1147  * Add an entire tree branch to our inode. eb_bh is the extent block
1148  * to start at, if we don't want to start the branch at the root
1149  * structure.
1150  *
1151  * last_eb_bh is required as we have to update it's next_leaf pointer
1152  * for the new last extent block.
1153  *
1154  * the new branch will be 'empty' in the sense that every block will
1155  * contain a single record with cluster count == 0.
1156  */
1157 static int ocfs2_add_branch(handle_t *handle,
1158                             struct ocfs2_extent_tree *et,
1159                             struct buffer_head *eb_bh,
1160                             struct buffer_head **last_eb_bh,
1161                             struct ocfs2_alloc_context *meta_ac)
1162 {
1163         int status, new_blocks, i;
1164         u64 next_blkno, new_last_eb_blk;
1165         struct buffer_head *bh;
1166         struct buffer_head **new_eb_bhs = NULL;
1167         struct ocfs2_extent_block *eb;
1168         struct ocfs2_extent_list  *eb_el;
1169         struct ocfs2_extent_list  *el;
1170         u32 new_cpos, root_end;
1171
1172         BUG_ON(!last_eb_bh || !*last_eb_bh);
1173
1174         if (eb_bh) {
1175                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1176                 el = &eb->h_list;
1177         } else
1178                 el = et->et_root_el;
1179
1180         /* we never add a branch to a leaf. */
1181         BUG_ON(!el->l_tree_depth);
1182
1183         new_blocks = le16_to_cpu(el->l_tree_depth);
1184
1185         eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
1186         new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
1187         root_end = ocfs2_sum_rightmost_rec(et->et_root_el);
1188
1189         /*
1190          * If there is a gap before the root end and the real end
1191          * of the righmost leaf block, we need to remove the gap
1192          * between new_cpos and root_end first so that the tree
1193          * is consistent after we add a new branch(it will start
1194          * from new_cpos).
1195          */
1196         if (root_end > new_cpos) {
1197                 trace_ocfs2_adjust_rightmost_branch(
1198                         (unsigned long long)
1199                         ocfs2_metadata_cache_owner(et->et_ci),
1200                         root_end, new_cpos);
1201
1202                 status = ocfs2_adjust_rightmost_branch(handle, et);
1203                 if (status) {
1204                         mlog_errno(status);
1205                         goto bail;
1206                 }
1207         }
1208
1209         /* allocate the number of new eb blocks we need */
1210         new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
1211                              GFP_KERNEL);
1212         if (!new_eb_bhs) {
1213                 status = -ENOMEM;
1214                 mlog_errno(status);
1215                 goto bail;
1216         }
1217
1218         status = ocfs2_create_new_meta_bhs(handle, et, new_blocks,
1219                                            meta_ac, new_eb_bhs);
1220         if (status < 0) {
1221                 mlog_errno(status);
1222                 goto bail;
1223         }
1224
1225         /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
1226          * linked with the rest of the tree.
1227          * conversly, new_eb_bhs[0] is the new bottommost leaf.
1228          *
1229          * when we leave the loop, new_last_eb_blk will point to the
1230          * newest leaf, and next_blkno will point to the topmost extent
1231          * block. */
1232         next_blkno = new_last_eb_blk = 0;
1233         for(i = 0; i < new_blocks; i++) {
1234                 bh = new_eb_bhs[i];
1235                 eb = (struct ocfs2_extent_block *) bh->b_data;
1236                 /* ocfs2_create_new_meta_bhs() should create it right! */
1237                 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1238                 eb_el = &eb->h_list;
1239
1240                 status = ocfs2_journal_access_eb(handle, et->et_ci, bh,
1241                                                  OCFS2_JOURNAL_ACCESS_CREATE);
1242                 if (status < 0) {
1243                         mlog_errno(status);
1244                         goto bail;
1245                 }
1246
1247                 eb->h_next_leaf_blk = 0;
1248                 eb_el->l_tree_depth = cpu_to_le16(i);
1249                 eb_el->l_next_free_rec = cpu_to_le16(1);
1250                 /*
1251                  * This actually counts as an empty extent as
1252                  * c_clusters == 0
1253                  */
1254                 eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
1255                 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
1256                 /*
1257                  * eb_el isn't always an interior node, but even leaf
1258                  * nodes want a zero'd flags and reserved field so
1259                  * this gets the whole 32 bits regardless of use.
1260                  */
1261                 eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
1262                 if (!eb_el->l_tree_depth)
1263                         new_last_eb_blk = le64_to_cpu(eb->h_blkno);
1264
1265                 ocfs2_journal_dirty(handle, bh);
1266                 next_blkno = le64_to_cpu(eb->h_blkno);
1267         }
1268
1269         /* This is a bit hairy. We want to update up to three blocks
1270          * here without leaving any of them in an inconsistent state
1271          * in case of error. We don't have to worry about
1272          * journal_dirty erroring as it won't unless we've aborted the
1273          * handle (in which case we would never be here) so reserving
1274          * the write with journal_access is all we need to do. */
1275         status = ocfs2_journal_access_eb(handle, et->et_ci, *last_eb_bh,
1276                                          OCFS2_JOURNAL_ACCESS_WRITE);
1277         if (status < 0) {
1278                 mlog_errno(status);
1279                 goto bail;
1280         }
1281         status = ocfs2_et_root_journal_access(handle, et,
1282                                               OCFS2_JOURNAL_ACCESS_WRITE);
1283         if (status < 0) {
1284                 mlog_errno(status);
1285                 goto bail;
1286         }
1287         if (eb_bh) {
1288                 status = ocfs2_journal_access_eb(handle, et->et_ci, eb_bh,
1289                                                  OCFS2_JOURNAL_ACCESS_WRITE);
1290                 if (status < 0) {
1291                         mlog_errno(status);
1292                         goto bail;
1293                 }
1294         }
1295
1296         /* Link the new branch into the rest of the tree (el will
1297          * either be on the root_bh, or the extent block passed in. */
1298         i = le16_to_cpu(el->l_next_free_rec);
1299         el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
1300         el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1301         el->l_recs[i].e_int_clusters = 0;
1302         le16_add_cpu(&el->l_next_free_rec, 1);
1303
1304         /* fe needs a new last extent block pointer, as does the
1305          * next_leaf on the previously last-extent-block. */
1306         ocfs2_et_set_last_eb_blk(et, new_last_eb_blk);
1307
1308         eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
1309         eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);
1310
1311         ocfs2_journal_dirty(handle, *last_eb_bh);
1312         ocfs2_journal_dirty(handle, et->et_root_bh);
1313         if (eb_bh)
1314                 ocfs2_journal_dirty(handle, eb_bh);
1315
1316         /*
1317          * Some callers want to track the rightmost leaf so pass it
1318          * back here.
1319          */
1320         brelse(*last_eb_bh);
1321         get_bh(new_eb_bhs[0]);
1322         *last_eb_bh = new_eb_bhs[0];
1323
1324         status = 0;
1325 bail:
1326         if (new_eb_bhs) {
1327                 for (i = 0; i < new_blocks; i++)
1328                         brelse(new_eb_bhs[i]);
1329                 kfree(new_eb_bhs);
1330         }
1331
1332         return status;
1333 }
1334
1335 /*
1336  * adds another level to the allocation tree.
1337  * returns back the new extent block so you can add a branch to it
1338  * after this call.
1339  */
1340 static int ocfs2_shift_tree_depth(handle_t *handle,
1341                                   struct ocfs2_extent_tree *et,
1342                                   struct ocfs2_alloc_context *meta_ac,
1343                                   struct buffer_head **ret_new_eb_bh)
1344 {
1345         int status, i;
1346         u32 new_clusters;
1347         struct buffer_head *new_eb_bh = NULL;
1348         struct ocfs2_extent_block *eb;
1349         struct ocfs2_extent_list  *root_el;
1350         struct ocfs2_extent_list  *eb_el;
1351
1352         status = ocfs2_create_new_meta_bhs(handle, et, 1, meta_ac,
1353                                            &new_eb_bh);
1354         if (status < 0) {
1355                 mlog_errno(status);
1356                 goto bail;
1357         }
1358
1359         eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
1360         /* ocfs2_create_new_meta_bhs() should create it right! */
1361         BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1362
1363         eb_el = &eb->h_list;
1364         root_el = et->et_root_el;
1365
1366         status = ocfs2_journal_access_eb(handle, et->et_ci, new_eb_bh,
1367                                          OCFS2_JOURNAL_ACCESS_CREATE);
1368         if (status < 0) {
1369                 mlog_errno(status);
1370                 goto bail;
1371         }
1372
1373         /* copy the root extent list data into the new extent block */
1374         eb_el->l_tree_depth = root_el->l_tree_depth;
1375         eb_el->l_next_free_rec = root_el->l_next_free_rec;
1376         for (i = 0; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1377                 eb_el->l_recs[i] = root_el->l_recs[i];
1378
1379         ocfs2_journal_dirty(handle, new_eb_bh);
1380
1381         status = ocfs2_et_root_journal_access(handle, et,
1382                                               OCFS2_JOURNAL_ACCESS_WRITE);
1383         if (status < 0) {
1384                 mlog_errno(status);
1385                 goto bail;
1386         }
1387
1388         new_clusters = ocfs2_sum_rightmost_rec(eb_el);
1389
1390         /* update root_bh now */
1391         le16_add_cpu(&root_el->l_tree_depth, 1);
1392         root_el->l_recs[0].e_cpos = 0;
1393         root_el->l_recs[0].e_blkno = eb->h_blkno;
1394         root_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
1395         for (i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1396                 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
1397         root_el->l_next_free_rec = cpu_to_le16(1);
1398
1399         /* If this is our 1st tree depth shift, then last_eb_blk
1400          * becomes the allocated extent block */
1401         if (root_el->l_tree_depth == cpu_to_le16(1))
1402                 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
1403
1404         ocfs2_journal_dirty(handle, et->et_root_bh);
1405
1406         *ret_new_eb_bh = new_eb_bh;
1407         new_eb_bh = NULL;
1408         status = 0;
1409 bail:
1410         brelse(new_eb_bh);
1411
1412         return status;
1413 }
1414
1415 /*
1416  * Should only be called when there is no space left in any of the
1417  * leaf nodes. What we want to do is find the lowest tree depth
1418  * non-leaf extent block with room for new records. There are three
1419  * valid results of this search:
1420  *
1421  * 1) a lowest extent block is found, then we pass it back in
1422  *    *lowest_eb_bh and return '0'
1423  *
1424  * 2) the search fails to find anything, but the root_el has room. We
1425  *    pass NULL back in *lowest_eb_bh, but still return '0'
1426  *
1427  * 3) the search fails to find anything AND the root_el is full, in
1428  *    which case we return > 0
1429  *
1430  * return status < 0 indicates an error.
1431  */
1432 static int ocfs2_find_branch_target(struct ocfs2_extent_tree *et,
1433                                     struct buffer_head **target_bh)
1434 {
1435         int status = 0, i;
1436         u64 blkno;
1437         struct ocfs2_extent_block *eb;
1438         struct ocfs2_extent_list  *el;
1439         struct buffer_head *bh = NULL;
1440         struct buffer_head *lowest_bh = NULL;
1441
1442         *target_bh = NULL;
1443
1444         el = et->et_root_el;
1445
1446         while(le16_to_cpu(el->l_tree_depth) > 1) {
1447                 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1448                         ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1449                                     "Owner %llu has empty "
1450                                     "extent list (next_free_rec == 0)",
1451                                     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
1452                         status = -EIO;
1453                         goto bail;
1454                 }
1455                 i = le16_to_cpu(el->l_next_free_rec) - 1;
1456                 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1457                 if (!blkno) {
1458                         ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1459                                     "Owner %llu has extent "
1460                                     "list where extent # %d has no physical "
1461                                     "block start",
1462                                     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), i);
1463                         status = -EIO;
1464                         goto bail;
1465                 }
1466
1467                 brelse(bh);
1468                 bh = NULL;
1469
1470                 status = ocfs2_read_extent_block(et->et_ci, blkno, &bh);
1471                 if (status < 0) {
1472                         mlog_errno(status);
1473                         goto bail;
1474                 }
1475
1476                 eb = (struct ocfs2_extent_block *) bh->b_data;
1477                 el = &eb->h_list;
1478
1479                 if (le16_to_cpu(el->l_next_free_rec) <
1480                     le16_to_cpu(el->l_count)) {
1481                         brelse(lowest_bh);
1482                         lowest_bh = bh;
1483                         get_bh(lowest_bh);
1484                 }
1485         }
1486
1487         /* If we didn't find one and the fe doesn't have any room,
1488          * then return '1' */
1489         el = et->et_root_el;
1490         if (!lowest_bh && (el->l_next_free_rec == el->l_count))
1491                 status = 1;
1492
1493         *target_bh = lowest_bh;
1494 bail:
1495         brelse(bh);
1496
1497         return status;
1498 }
1499
1500 /*
1501  * Grow a b-tree so that it has more records.
1502  *
1503  * We might shift the tree depth in which case existing paths should
1504  * be considered invalid.
1505  *
1506  * Tree depth after the grow is returned via *final_depth.
1507  *
1508  * *last_eb_bh will be updated by ocfs2_add_branch().
1509  */
1510 static int ocfs2_grow_tree(handle_t *handle, struct ocfs2_extent_tree *et,
1511                            int *final_depth, struct buffer_head **last_eb_bh,
1512                            struct ocfs2_alloc_context *meta_ac)
1513 {
1514         int ret, shift;
1515         struct ocfs2_extent_list *el = et->et_root_el;
1516         int depth = le16_to_cpu(el->l_tree_depth);
1517         struct buffer_head *bh = NULL;
1518
1519         BUG_ON(meta_ac == NULL);
1520
1521         shift = ocfs2_find_branch_target(et, &bh);
1522         if (shift < 0) {
1523                 ret = shift;
1524                 mlog_errno(ret);
1525                 goto out;
1526         }
1527
1528         /* We traveled all the way to the bottom of the allocation tree
1529          * and didn't find room for any more extents - we need to add
1530          * another tree level */
1531         if (shift) {
1532                 BUG_ON(bh);
1533                 trace_ocfs2_grow_tree(
1534                         (unsigned long long)
1535                         ocfs2_metadata_cache_owner(et->et_ci),
1536                         depth);
1537
1538                 /* ocfs2_shift_tree_depth will return us a buffer with
1539                  * the new extent block (so we can pass that to
1540                  * ocfs2_add_branch). */
1541                 ret = ocfs2_shift_tree_depth(handle, et, meta_ac, &bh);
1542                 if (ret < 0) {
1543                         mlog_errno(ret);
1544                         goto out;
1545                 }
1546                 depth++;
1547                 if (depth == 1) {
1548                         /*
1549                          * Special case: we have room now if we shifted from
1550                          * tree_depth 0, so no more work needs to be done.
1551                          *
1552                          * We won't be calling add_branch, so pass
1553                          * back *last_eb_bh as the new leaf. At depth
1554                          * zero, it should always be null so there's
1555                          * no reason to brelse.
1556                          */
1557                         BUG_ON(*last_eb_bh);
1558                         get_bh(bh);
1559                         *last_eb_bh = bh;
1560                         goto out;
1561                 }
1562         }
1563
1564         /* call ocfs2_add_branch to add the final part of the tree with
1565          * the new data. */
1566         ret = ocfs2_add_branch(handle, et, bh, last_eb_bh,
1567                                meta_ac);
1568         if (ret < 0) {
1569                 mlog_errno(ret);
1570                 goto out;
1571         }
1572
1573 out:
1574         if (final_depth)
1575                 *final_depth = depth;
1576         brelse(bh);
1577         return ret;
1578 }
1579
1580 /*
1581  * This function will discard the rightmost extent record.
1582  */
1583 static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
1584 {
1585         int next_free = le16_to_cpu(el->l_next_free_rec);
1586         int count = le16_to_cpu(el->l_count);
1587         unsigned int num_bytes;
1588
1589         BUG_ON(!next_free);
1590         /* This will cause us to go off the end of our extent list. */
1591         BUG_ON(next_free >= count);
1592
1593         num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
1594
1595         memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
1596 }
1597
1598 static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
1599                               struct ocfs2_extent_rec *insert_rec)
1600 {
1601         int i, insert_index, next_free, has_empty, num_bytes;
1602         u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
1603         struct ocfs2_extent_rec *rec;
1604
1605         next_free = le16_to_cpu(el->l_next_free_rec);
1606         has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
1607
1608         BUG_ON(!next_free);
1609
1610         /* The tree code before us didn't allow enough room in the leaf. */
1611         BUG_ON(el->l_next_free_rec == el->l_count && !has_empty);
1612
1613         /*
1614          * The easiest way to approach this is to just remove the
1615          * empty extent and temporarily decrement next_free.
1616          */
1617         if (has_empty) {
1618                 /*
1619                  * If next_free was 1 (only an empty extent), this
1620                  * loop won't execute, which is fine. We still want
1621                  * the decrement above to happen.
1622                  */
1623                 for(i = 0; i < (next_free - 1); i++)
1624                         el->l_recs[i] = el->l_recs[i+1];
1625
1626                 next_free--;
1627         }
1628
1629         /*
1630          * Figure out what the new record index should be.
1631          */
1632         for(i = 0; i < next_free; i++) {
1633                 rec = &el->l_recs[i];
1634
1635                 if (insert_cpos < le32_to_cpu(rec->e_cpos))
1636                         break;
1637         }
1638         insert_index = i;
1639
1640         trace_ocfs2_rotate_leaf(insert_cpos, insert_index,
1641                                 has_empty, next_free,
1642                                 le16_to_cpu(el->l_count));
1643
1644         BUG_ON(insert_index < 0);
1645         BUG_ON(insert_index >= le16_to_cpu(el->l_count));
1646         BUG_ON(insert_index > next_free);
1647
1648         /*
1649          * No need to memmove if we're just adding to the tail.
1650          */
1651         if (insert_index != next_free) {
1652                 BUG_ON(next_free >= le16_to_cpu(el->l_count));
1653
1654                 num_bytes = next_free - insert_index;
1655                 num_bytes *= sizeof(struct ocfs2_extent_rec);
1656                 memmove(&el->l_recs[insert_index + 1],
1657                         &el->l_recs[insert_index],
1658                         num_bytes);
1659         }
1660
1661         /*
1662          * Either we had an empty extent, and need to re-increment or
1663          * there was no empty extent on a non full rightmost leaf node,
1664          * in which case we still need to increment.
1665          */
1666         next_free++;
1667         el->l_next_free_rec = cpu_to_le16(next_free);
1668         /*
1669          * Make sure none of the math above just messed up our tree.
1670          */
1671         BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
1672
1673         el->l_recs[insert_index] = *insert_rec;
1674
1675 }
1676
1677 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
1678 {
1679         int size, num_recs = le16_to_cpu(el->l_next_free_rec);
1680
1681         BUG_ON(num_recs == 0);
1682
1683         if (ocfs2_is_empty_extent(&el->l_recs[0])) {
1684                 num_recs--;
1685                 size = num_recs * sizeof(struct ocfs2_extent_rec);
1686                 memmove(&el->l_recs[0], &el->l_recs[1], size);
1687                 memset(&el->l_recs[num_recs], 0,
1688                        sizeof(struct ocfs2_extent_rec));
1689                 el->l_next_free_rec = cpu_to_le16(num_recs);
1690         }
1691 }
1692
1693 /*
1694  * Create an empty extent record .
1695  *
1696  * l_next_free_rec may be updated.
1697  *
1698  * If an empty extent already exists do nothing.
1699  */
1700 static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
1701 {
1702         int next_free = le16_to_cpu(el->l_next_free_rec);
1703
1704         BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1705
1706         if (next_free == 0)
1707                 goto set_and_inc;
1708
1709         if (ocfs2_is_empty_extent(&el->l_recs[0]))
1710                 return;
1711
1712         mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
1713                         "Asked to create an empty extent in a full list:\n"
1714                         "count = %u, tree depth = %u",
1715                         le16_to_cpu(el->l_count),
1716                         le16_to_cpu(el->l_tree_depth));
1717
1718         ocfs2_shift_records_right(el);
1719
1720 set_and_inc:
1721         le16_add_cpu(&el->l_next_free_rec, 1);
1722         memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1723 }
1724
1725 /*
1726  * For a rotation which involves two leaf nodes, the "root node" is
1727  * the lowest level tree node which contains a path to both leafs. This
1728  * resulting set of information can be used to form a complete "subtree"
1729  *
1730  * This function is passed two full paths from the dinode down to a
1731  * pair of adjacent leaves. It's task is to figure out which path
1732  * index contains the subtree root - this can be the root index itself
1733  * in a worst-case rotation.
1734  *
1735  * The array index of the subtree root is passed back.
1736  */
1737 int ocfs2_find_subtree_root(struct ocfs2_extent_tree *et,
1738                             struct ocfs2_path *left,
1739                             struct ocfs2_path *right)
1740 {
1741         int i = 0;
1742
1743         /*
1744          * Check that the caller passed in two paths from the same tree.
1745          */
1746         BUG_ON(path_root_bh(left) != path_root_bh(right));
1747
1748         do {
1749                 i++;
1750
1751                 /*
1752                  * The caller didn't pass two adjacent paths.
1753                  */
1754                 mlog_bug_on_msg(i > left->p_tree_depth,
1755                                 "Owner %llu, left depth %u, right depth %u\n"
1756                                 "left leaf blk %llu, right leaf blk %llu\n",
1757                                 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
1758                                 left->p_tree_depth, right->p_tree_depth,
1759                                 (unsigned long long)path_leaf_bh(left)->b_blocknr,
1760                                 (unsigned long long)path_leaf_bh(right)->b_blocknr);
1761         } while (left->p_node[i].bh->b_blocknr ==
1762                  right->p_node[i].bh->b_blocknr);
1763
1764         return i - 1;
1765 }
1766
1767 typedef void (path_insert_t)(void *, struct buffer_head *);
1768
1769 /*
1770  * Traverse a btree path in search of cpos, starting at root_el.
1771  *
1772  * This code can be called with a cpos larger than the tree, in which
1773  * case it will return the rightmost path.
1774  */
1775 static int __ocfs2_find_path(struct ocfs2_caching_info *ci,
1776                              struct ocfs2_extent_list *root_el, u32 cpos,
1777                              path_insert_t *func, void *data)
1778 {
1779         int i, ret = 0;
1780         u32 range;
1781         u64 blkno;
1782         struct buffer_head *bh = NULL;
1783         struct ocfs2_extent_block *eb;
1784         struct ocfs2_extent_list *el;
1785         struct ocfs2_extent_rec *rec;
1786
1787         el = root_el;
1788         while (el->l_tree_depth) {
1789                 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1790                         ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1791                                     "Owner %llu has empty extent list at "
1792                                     "depth %u\n",
1793                                     (unsigned long long)ocfs2_metadata_cache_owner(ci),
1794                                     le16_to_cpu(el->l_tree_depth));
1795                         ret = -EROFS;
1796                         goto out;
1797
1798                 }
1799
1800                 for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
1801                         rec = &el->l_recs[i];
1802
1803                         /*
1804                          * In the case that cpos is off the allocation
1805                          * tree, this should just wind up returning the
1806                          * rightmost record.
1807                          */
1808                         range = le32_to_cpu(rec->e_cpos) +
1809                                 ocfs2_rec_clusters(el, rec);
1810                         if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1811                             break;
1812                 }
1813
1814                 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1815                 if (blkno == 0) {
1816                         ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1817                                     "Owner %llu has bad blkno in extent list "
1818                                     "at depth %u (index %d)\n",
1819                                     (unsigned long long)ocfs2_metadata_cache_owner(ci),
1820                                     le16_to_cpu(el->l_tree_depth), i);
1821                         ret = -EROFS;
1822                         goto out;
1823                 }
1824
1825                 brelse(bh);
1826                 bh = NULL;
1827                 ret = ocfs2_read_extent_block(ci, blkno, &bh);
1828                 if (ret) {
1829                         mlog_errno(ret);
1830                         goto out;
1831                 }
1832
1833                 eb = (struct ocfs2_extent_block *) bh->b_data;
1834                 el = &eb->h_list;
1835
1836                 if (le16_to_cpu(el->l_next_free_rec) >
1837                     le16_to_cpu(el->l_count)) {
1838                         ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1839                                     "Owner %llu has bad count in extent list "
1840                                     "at block %llu (next free=%u, count=%u)\n",
1841                                     (unsigned long long)ocfs2_metadata_cache_owner(ci),
1842                                     (unsigned long long)bh->b_blocknr,
1843                                     le16_to_cpu(el->l_next_free_rec),
1844                                     le16_to_cpu(el->l_count));
1845                         ret = -EROFS;
1846                         goto out;
1847                 }
1848
1849                 if (func)
1850                         func(data, bh);
1851         }
1852
1853 out:
1854         /*
1855          * Catch any trailing bh that the loop didn't handle.
1856          */
1857         brelse(bh);
1858
1859         return ret;
1860 }
1861
1862 /*
1863  * Given an initialized path (that is, it has a valid root extent
1864  * list), this function will traverse the btree in search of the path
1865  * which would contain cpos.
1866  *
1867  * The path traveled is recorded in the path structure.
1868  *
1869  * Note that this will not do any comparisons on leaf node extent
1870  * records, so it will work fine in the case that we just added a tree
1871  * branch.
1872  */
1873 struct find_path_data {
1874         int index;
1875         struct ocfs2_path *path;
1876 };
1877 static void find_path_ins(void *data, struct buffer_head *bh)
1878 {
1879         struct find_path_data *fp = data;
1880
1881         get_bh(bh);
1882         ocfs2_path_insert_eb(fp->path, fp->index, bh);
1883         fp->index++;
1884 }
1885 int ocfs2_find_path(struct ocfs2_caching_info *ci,
1886                     struct ocfs2_path *path, u32 cpos)
1887 {
1888         struct find_path_data data;
1889
1890         data.index = 1;
1891         data.path = path;
1892         return __ocfs2_find_path(ci, path_root_el(path), cpos,
1893                                  find_path_ins, &data);
1894 }
1895
1896 static void find_leaf_ins(void *data, struct buffer_head *bh)
1897 {
1898         struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1899         struct ocfs2_extent_list *el = &eb->h_list;
1900         struct buffer_head **ret = data;
1901
1902         /* We want to retain only the leaf block. */
1903         if (le16_to_cpu(el->l_tree_depth) == 0) {
1904                 get_bh(bh);
1905                 *ret = bh;
1906         }
1907 }
1908 /*
1909  * Find the leaf block in the tree which would contain cpos. No
1910  * checking of the actual leaf is done.
1911  *
1912  * Some paths want to call this instead of allocating a path structure
1913  * and calling ocfs2_find_path().
1914  *
1915  * This function doesn't handle non btree extent lists.
1916  */
1917 int ocfs2_find_leaf(struct ocfs2_caching_info *ci,
1918                     struct ocfs2_extent_list *root_el, u32 cpos,
1919                     struct buffer_head **leaf_bh)
1920 {
1921         int ret;
1922         struct buffer_head *bh = NULL;
1923
1924         ret = __ocfs2_find_path(ci, root_el, cpos, find_leaf_ins, &bh);
1925         if (ret) {
1926                 mlog_errno(ret);
1927                 goto out;
1928         }
1929
1930         *leaf_bh = bh;
1931 out:
1932         return ret;
1933 }
1934
1935 /*
1936  * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1937  *
1938  * Basically, we've moved stuff around at the bottom of the tree and
1939  * we need to fix up the extent records above the changes to reflect
1940  * the new changes.
1941  *
1942  * left_rec: the record on the left.
1943  * left_child_el: is the child list pointed to by left_rec
1944  * right_rec: the record to the right of left_rec
1945  * right_child_el: is the child list pointed to by right_rec
1946  *
1947  * By definition, this only works on interior nodes.
1948  */
1949 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1950                                   struct ocfs2_extent_list *left_child_el,
1951                                   struct ocfs2_extent_rec *right_rec,
1952                                   struct ocfs2_extent_list *right_child_el)
1953 {
1954         u32 left_clusters, right_end;
1955
1956         /*
1957          * Interior nodes never have holes. Their cpos is the cpos of
1958          * the leftmost record in their child list. Their cluster
1959          * count covers the full theoretical range of their child list
1960          * - the range between their cpos and the cpos of the record
1961          * immediately to their right.
1962          */
1963         left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1964         if (!ocfs2_rec_clusters(right_child_el, &right_child_el->l_recs[0])) {
1965                 BUG_ON(right_child_el->l_tree_depth);
1966                 BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
1967                 left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
1968         }
1969         left_clusters -= le32_to_cpu(left_rec->e_cpos);
1970         left_rec->e_int_clusters = cpu_to_le32(left_clusters);
1971
1972         /*
1973          * Calculate the rightmost cluster count boundary before
1974          * moving cpos - we will need to adjust clusters after
1975          * updating e_cpos to keep the same highest cluster count.
1976          */
1977         right_end = le32_to_cpu(right_rec->e_cpos);
1978         right_end += le32_to_cpu(right_rec->e_int_clusters);
1979
1980         right_rec->e_cpos = left_rec->e_cpos;
1981         le32_add_cpu(&right_rec->e_cpos, left_clusters);
1982
1983         right_end -= le32_to_cpu(right_rec->e_cpos);
1984         right_rec->e_int_clusters = cpu_to_le32(right_end);
1985 }
1986
1987 /*
1988  * Adjust the adjacent root node records involved in a
1989  * rotation. left_el_blkno is passed in as a key so that we can easily
1990  * find it's index in the root list.
1991  */
1992 static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
1993                                       struct ocfs2_extent_list *left_el,
1994                                       struct ocfs2_extent_list *right_el,
1995                                       u64 left_el_blkno)
1996 {
1997         int i;
1998
1999         BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
2000                le16_to_cpu(left_el->l_tree_depth));
2001
2002         for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
2003                 if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
2004                         break;
2005         }
2006
2007         /*
2008          * The path walking code should have never returned a root and
2009          * two paths which are not adjacent.
2010          */
2011         BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
2012
2013         ocfs2_adjust_adjacent_records(&root_el->l_recs[i], left_el,
2014                                       &root_el->l_recs[i + 1], right_el);
2015 }
2016
2017 /*
2018  * We've changed a leaf block (in right_path) and need to reflect that
2019  * change back up the subtree.
2020  *
2021  * This happens in multiple places:
2022  *   - When we've moved an extent record from the left path leaf to the right
2023  *     path leaf to make room for an empty extent in the left path leaf.
2024  *   - When our insert into the right path leaf is at the leftmost edge
2025  *     and requires an update of the path immediately to it's left. This
2026  *     can occur at the end of some types of rotation and appending inserts.
2027  *   - When we've adjusted the last extent record in the left path leaf and the
2028  *     1st extent record in the right path leaf during cross extent block merge.
2029  */
2030 static void ocfs2_complete_edge_insert(handle_t *handle,
2031                                        struct ocfs2_path *left_path,
2032                                        struct ocfs2_path *right_path,
2033                                        int subtree_index)
2034 {
2035         int i, idx;
2036         struct ocfs2_extent_list *el, *left_el, *right_el;
2037         struct ocfs2_extent_rec *left_rec, *right_rec;
2038         struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2039
2040         /*
2041          * Update the counts and position values within all the
2042          * interior nodes to reflect the leaf rotation we just did.
2043          *
2044          * The root node is handled below the loop.
2045          *
2046          * We begin the loop with right_el and left_el pointing to the
2047          * leaf lists and work our way up.
2048          *
2049          * NOTE: within this loop, left_el and right_el always refer
2050          * to the *child* lists.
2051          */
2052         left_el = path_leaf_el(left_path);
2053         right_el = path_leaf_el(right_path);
2054         for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
2055                 trace_ocfs2_complete_edge_insert(i);
2056
2057                 /*
2058                  * One nice property of knowing that all of these
2059                  * nodes are below the root is that we only deal with
2060                  * the leftmost right node record and the rightmost
2061                  * left node record.
2062                  */
2063                 el = left_path->p_node[i].el;
2064                 idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
2065                 left_rec = &el->l_recs[idx];
2066
2067                 el = right_path->p_node[i].el;
2068                 right_rec = &el->l_recs[0];
2069
2070                 ocfs2_adjust_adjacent_records(left_rec, left_el, right_rec,
2071                                               right_el);
2072
2073                 ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
2074                 ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
2075
2076                 /*
2077                  * Setup our list pointers now so that the current
2078                  * parents become children in the next iteration.
2079                  */
2080                 left_el = left_path->p_node[i].el;
2081                 right_el = right_path->p_node[i].el;
2082         }
2083
2084         /*
2085          * At the root node, adjust the two adjacent records which
2086          * begin our path to the leaves.
2087          */
2088
2089         el = left_path->p_node[subtree_index].el;
2090         left_el = left_path->p_node[subtree_index + 1].el;
2091         right_el = right_path->p_node[subtree_index + 1].el;
2092
2093         ocfs2_adjust_root_records(el, left_el, right_el,
2094                                   left_path->p_node[subtree_index + 1].bh->b_blocknr);
2095
2096         root_bh = left_path->p_node[subtree_index].bh;
2097
2098         ocfs2_journal_dirty(handle, root_bh);
2099 }
2100
2101 static int ocfs2_rotate_subtree_right(handle_t *handle,
2102                                       struct ocfs2_extent_tree *et,
2103                                       struct ocfs2_path *left_path,
2104                                       struct ocfs2_path *right_path,
2105                                       int subtree_index)
2106 {
2107         int ret, i;
2108         struct buffer_head *right_leaf_bh;
2109         struct buffer_head *left_leaf_bh = NULL;
2110         struct buffer_head *root_bh;
2111         struct ocfs2_extent_list *right_el, *left_el;
2112         struct ocfs2_extent_rec move_rec;
2113
2114         left_leaf_bh = path_leaf_bh(left_path);
2115         left_el = path_leaf_el(left_path);
2116
2117         if (left_el->l_next_free_rec != left_el->l_count) {
2118                 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
2119                             "Inode %llu has non-full interior leaf node %llu"
2120                             "(next free = %u)",
2121                             (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2122                             (unsigned long long)left_leaf_bh->b_blocknr,
2123                             le16_to_cpu(left_el->l_next_free_rec));
2124                 return -EROFS;
2125         }
2126
2127         /*
2128          * This extent block may already have an empty record, so we
2129          * return early if so.
2130          */
2131         if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
2132                 return 0;
2133
2134         root_bh = left_path->p_node[subtree_index].bh;
2135         BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2136
2137         ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2138                                            subtree_index);
2139         if (ret) {
2140                 mlog_errno(ret);
2141                 goto out;
2142         }
2143
2144         for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2145                 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2146                                                    right_path, i);
2147                 if (ret) {
2148                         mlog_errno(ret);
2149                         goto out;
2150                 }
2151
2152                 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2153                                                    left_path, i);
2154                 if (ret) {
2155                         mlog_errno(ret);
2156                         goto out;
2157                 }
2158         }
2159
2160         right_leaf_bh = path_leaf_bh(right_path);
2161         right_el = path_leaf_el(right_path);
2162
2163         /* This is a code error, not a disk corruption. */
2164         mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
2165                         "because rightmost leaf block %llu is empty\n",
2166                         (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2167                         (unsigned long long)right_leaf_bh->b_blocknr);
2168
2169         ocfs2_create_empty_extent(right_el);
2170
2171         ocfs2_journal_dirty(handle, right_leaf_bh);
2172
2173         /* Do the copy now. */
2174         i = le16_to_cpu(left_el->l_next_free_rec) - 1;
2175         move_rec = left_el->l_recs[i];
2176         right_el->l_recs[0] = move_rec;
2177
2178         /*
2179          * Clear out the record we just copied and shift everything
2180          * over, leaving an empty extent in the left leaf.
2181          *
2182          * We temporarily subtract from next_free_rec so that the
2183          * shift will lose the tail record (which is now defunct).
2184          */
2185         le16_add_cpu(&left_el->l_next_free_rec, -1);
2186         ocfs2_shift_records_right(left_el);
2187         memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2188         le16_add_cpu(&left_el->l_next_free_rec, 1);
2189
2190         ocfs2_journal_dirty(handle, left_leaf_bh);
2191
2192         ocfs2_complete_edge_insert(handle, left_path, right_path,
2193                                    subtree_index);
2194
2195 out:
2196         return ret;
2197 }
2198
2199 /*
2200  * Given a full path, determine what cpos value would return us a path
2201  * containing the leaf immediately to the left of the current one.
2202  *
2203  * Will return zero if the path passed in is already the leftmost path.
2204  */
2205 int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
2206                                   struct ocfs2_path *path, u32 *cpos)
2207 {
2208         int i, j, ret = 0;
2209         u64 blkno;
2210         struct ocfs2_extent_list *el;
2211
2212         BUG_ON(path->p_tree_depth == 0);
2213
2214         *cpos = 0;
2215
2216         blkno = path_leaf_bh(path)->b_blocknr;
2217
2218         /* Start at the tree node just above the leaf and work our way up. */
2219         i = path->p_tree_depth - 1;
2220         while (i >= 0) {
2221                 el = path->p_node[i].el;
2222
2223                 /*
2224                  * Find the extent record just before the one in our
2225                  * path.
2226                  */
2227                 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2228                         if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2229                                 if (j == 0) {
2230                                         if (i == 0) {
2231                                                 /*
2232                                                  * We've determined that the
2233                                                  * path specified is already
2234                                                  * the leftmost one - return a
2235                                                  * cpos of zero.
2236                                                  */
2237                                                 goto out;
2238                                         }
2239                                         /*
2240                                          * The leftmost record points to our
2241                                          * leaf - we need to travel up the
2242                                          * tree one level.
2243                                          */
2244                                         goto next_node;
2245                                 }
2246
2247                                 *cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
2248                                 *cpos = *cpos + ocfs2_rec_clusters(el,
2249                                                            &el->l_recs[j - 1]);
2250                                 *cpos = *cpos - 1;
2251                                 goto out;
2252                         }
2253                 }
2254
2255                 /*
2256                  * If we got here, we never found a valid node where
2257                  * the tree indicated one should be.
2258                  */
2259                 ocfs2_error(sb,
2260                             "Invalid extent tree at extent block %llu\n",
2261                             (unsigned long long)blkno);
2262                 ret = -EROFS;
2263                 goto out;
2264
2265 next_node:
2266                 blkno = path->p_node[i].bh->b_blocknr;
2267                 i--;
2268         }
2269
2270 out:
2271         return ret;
2272 }
2273
2274 /*
2275  * Extend the transaction by enough credits to complete the rotation,
2276  * and still leave at least the original number of credits allocated
2277  * to this transaction.
2278  */
2279 static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
2280                                            int op_credits,
2281                                            struct ocfs2_path *path)
2282 {
2283         int ret = 0;
2284         int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;
2285
2286         if (handle->h_buffer_credits < credits)
2287                 ret = ocfs2_extend_trans(handle,
2288                                          credits - handle->h_buffer_credits);
2289
2290         return ret;
2291 }
2292
2293 /*
2294  * Trap the case where we're inserting into the theoretical range past
2295  * the _actual_ left leaf range. Otherwise, we'll rotate a record
2296  * whose cpos is less than ours into the right leaf.
2297  *
2298  * It's only necessary to look at the rightmost record of the left
2299  * leaf because the logic that calls us should ensure that the
2300  * theoretical ranges in the path components above the leaves are
2301  * correct.
2302  */
2303 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
2304                                                  u32 insert_cpos)
2305 {
2306         struct ocfs2_extent_list *left_el;
2307         struct ocfs2_extent_rec *rec;
2308         int next_free;
2309
2310         left_el = path_leaf_el(left_path);
2311         next_free = le16_to_cpu(left_el->l_next_free_rec);
2312         rec = &left_el->l_recs[next_free - 1];
2313
2314         if (insert_cpos > le32_to_cpu(rec->e_cpos))
2315                 return 1;
2316         return 0;
2317 }
2318
2319 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
2320 {
2321         int next_free = le16_to_cpu(el->l_next_free_rec);
2322         unsigned int range;
2323         struct ocfs2_extent_rec *rec;
2324
2325         if (next_free == 0)
2326                 return 0;
2327
2328         rec = &el->l_recs[0];
2329         if (ocfs2_is_empty_extent(rec)) {
2330                 /* Empty list. */
2331                 if (next_free == 1)
2332                         return 0;
2333                 rec = &el->l_recs[1];
2334         }
2335
2336         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2337         if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
2338                 return 1;
2339         return 0;
2340 }
2341
2342 /*
2343  * Rotate all the records in a btree right one record, starting at insert_cpos.
2344  *
2345  * The path to the rightmost leaf should be passed in.
2346  *
2347  * The array is assumed to be large enough to hold an entire path (tree depth).
2348  *
2349  * Upon successful return from this function:
2350  *
2351  * - The 'right_path' array will contain a path to the leaf block
2352  *   whose range contains e_cpos.
2353  * - That leaf block will have a single empty extent in list index 0.
2354  * - In the case that the rotation requires a post-insert update,
2355  *   *ret_left_path will contain a valid path which can be passed to
2356  *   ocfs2_insert_path().
2357  */
2358 static int ocfs2_rotate_tree_right(handle_t *handle,
2359                                    struct ocfs2_extent_tree *et,
2360                                    enum ocfs2_split_type split,
2361                                    u32 insert_cpos,
2362                                    struct ocfs2_path *right_path,
2363                                    struct ocfs2_path **ret_left_path)
2364 {
2365         int ret, start, orig_credits = handle->h_buffer_credits;
2366         u32 cpos;
2367         struct ocfs2_path *left_path = NULL;
2368         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2369
2370         *ret_left_path = NULL;
2371
2372         left_path = ocfs2_new_path_from_path(right_path);
2373         if (!left_path) {
2374                 ret = -ENOMEM;
2375                 mlog_errno(ret);
2376                 goto out;
2377         }
2378
2379         ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2380         if (ret) {
2381                 mlog_errno(ret);
2382                 goto out;
2383         }
2384
2385         trace_ocfs2_rotate_tree_right(
2386                 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2387                 insert_cpos, cpos);
2388
2389         /*
2390          * What we want to do here is:
2391          *
2392          * 1) Start with the rightmost path.
2393          *
2394          * 2) Determine a path to the leaf block directly to the left
2395          *    of that leaf.
2396          *
2397          * 3) Determine the 'subtree root' - the lowest level tree node
2398          *    which contains a path to both leaves.
2399          *
2400          * 4) Rotate the subtree.
2401          *
2402          * 5) Find the next subtree by considering the left path to be
2403          *    the new right path.
2404          *
2405          * The check at the top of this while loop also accepts
2406          * insert_cpos == cpos because cpos is only a _theoretical_
2407          * value to get us the left path - insert_cpos might very well
2408          * be filling that hole.
2409          *
2410          * Stop at a cpos of '0' because we either started at the
2411          * leftmost branch (i.e., a tree with one branch and a
2412          * rotation inside of it), or we've gone as far as we can in
2413          * rotating subtrees.
2414          */
2415         while (cpos && insert_cpos <= cpos) {
2416                 trace_ocfs2_rotate_tree_right(
2417                         (unsigned long long)
2418                         ocfs2_metadata_cache_owner(et->et_ci),
2419                         insert_cpos, cpos);
2420
2421                 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
2422                 if (ret) {
2423                         mlog_errno(ret);
2424                         goto out;
2425                 }
2426
2427                 mlog_bug_on_msg(path_leaf_bh(left_path) ==
2428                                 path_leaf_bh(right_path),
2429                                 "Owner %llu: error during insert of %u "
2430                                 "(left path cpos %u) results in two identical "
2431                                 "paths ending at %llu\n",
2432                                 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2433                                 insert_cpos, cpos,
2434                                 (unsigned long long)
2435                                 path_leaf_bh(left_path)->b_blocknr);
2436
2437                 if (split == SPLIT_NONE &&
2438                     ocfs2_rotate_requires_path_adjustment(left_path,
2439                                                           insert_cpos)) {
2440
2441                         /*
2442                          * We've rotated the tree as much as we
2443                          * should. The rest is up to
2444                          * ocfs2_insert_path() to complete, after the
2445                          * record insertion. We indicate this
2446                          * situation by returning the left path.
2447                          *
2448                          * The reason we don't adjust the records here
2449                          * before the record insert is that an error
2450                          * later might break the rule where a parent
2451                          * record e_cpos will reflect the actual
2452                          * e_cpos of the 1st nonempty record of the
2453                          * child list.
2454                          */
2455                         *ret_left_path = left_path;
2456                         goto out_ret_path;
2457                 }
2458
2459                 start = ocfs2_find_subtree_root(et, left_path, right_path);
2460
2461                 trace_ocfs2_rotate_subtree(start,
2462                         (unsigned long long)
2463                         right_path->p_node[start].bh->b_blocknr,
2464                         right_path->p_tree_depth);
2465
2466                 ret = ocfs2_extend_rotate_transaction(handle, start,
2467                                                       orig_credits, right_path);
2468                 if (ret) {
2469                         mlog_errno(ret);
2470                         goto out;
2471                 }
2472
2473                 ret = ocfs2_rotate_subtree_right(handle, et, left_path,
2474                                                  right_path, start);
2475                 if (ret) {
2476                         mlog_errno(ret);
2477                         goto out;
2478                 }
2479
2480                 if (split != SPLIT_NONE &&
2481                     ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
2482                                                 insert_cpos)) {
2483                         /*
2484                          * A rotate moves the rightmost left leaf
2485                          * record over to the leftmost right leaf
2486                          * slot. If we're doing an extent split
2487                          * instead of a real insert, then we have to
2488                          * check that the extent to be split wasn't
2489                          * just moved over. If it was, then we can
2490                          * exit here, passing left_path back -
2491                          * ocfs2_split_extent() is smart enough to
2492                          * search both leaves.
2493                          */
2494                         *ret_left_path = left_path;
2495                         goto out_ret_path;
2496                 }
2497
2498                 /*
2499                  * There is no need to re-read the next right path
2500                  * as we know that it'll be our current left
2501                  * path. Optimize by copying values instead.
2502                  */
2503                 ocfs2_mv_path(right_path, left_path);
2504
2505                 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2506                 if (ret) {
2507                         mlog_errno(ret);
2508                         goto out;
2509                 }
2510         }
2511
2512 out:
2513         ocfs2_free_path(left_path);
2514
2515 out_ret_path:
2516         return ret;
2517 }
2518
2519 static int ocfs2_update_edge_lengths(handle_t *handle,
2520                                      struct ocfs2_extent_tree *et,
2521                                      int subtree_index, struct ocfs2_path *path)
2522 {
2523         int i, idx, ret;
2524         struct ocfs2_extent_rec *rec;
2525         struct ocfs2_extent_list *el;
2526         struct ocfs2_extent_block *eb;
2527         u32 range;
2528
2529         /*
2530          * In normal tree rotation process, we will never touch the
2531          * tree branch above subtree_index and ocfs2_extend_rotate_transaction
2532          * doesn't reserve the credits for them either.
2533          *
2534          * But we do have a special case here which will update the rightmost
2535          * records for all the bh in the path.
2536          * So we have to allocate extra credits and access them.
2537          */
2538         ret = ocfs2_extend_trans(handle, subtree_index);
2539         if (ret) {
2540                 mlog_errno(ret);
2541                 goto out;
2542         }
2543
2544         ret = ocfs2_journal_access_path(et->et_ci, handle, path);
2545         if (ret) {
2546                 mlog_errno(ret);
2547                 goto out;
2548         }
2549
2550         /* Path should always be rightmost. */
2551         eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2552         BUG_ON(eb->h_next_leaf_blk != 0ULL);
2553
2554         el = &eb->h_list;
2555         BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
2556         idx = le16_to_cpu(el->l_next_free_rec) - 1;
2557         rec = &el->l_recs[idx];
2558         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2559
2560         for (i = 0; i < path->p_tree_depth; i++) {
2561                 el = path->p_node[i].el;
2562                 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2563                 rec = &el->l_recs[idx];
2564
2565                 rec->e_int_clusters = cpu_to_le32(range);
2566                 le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));
2567
2568                 ocfs2_journal_dirty(handle, path->p_node[i].bh);
2569         }
2570 out:
2571         return ret;
2572 }
2573
2574 static void ocfs2_unlink_path(handle_t *handle,
2575                               struct ocfs2_extent_tree *et,
2576                               struct ocfs2_cached_dealloc_ctxt *dealloc,
2577                               struct ocfs2_path *path, int unlink_start)
2578 {
2579         int ret, i;
2580         struct ocfs2_extent_block *eb;
2581         struct ocfs2_extent_list *el;
2582         struct buffer_head *bh;
2583
2584         for(i = unlink_start; i < path_num_items(path); i++) {
2585                 bh = path->p_node[i].bh;
2586
2587                 eb = (struct ocfs2_extent_block *)bh->b_data;
2588                 /*
2589                  * Not all nodes might have had their final count
2590                  * decremented by the caller - handle this here.
2591                  */
2592                 el = &eb->h_list;
2593                 if (le16_to_cpu(el->l_next_free_rec) > 1) {
2594                         mlog(ML_ERROR,
2595                              "Inode %llu, attempted to remove extent block "
2596                              "%llu with %u records\n",
2597                              (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2598                              (unsigned long long)le64_to_cpu(eb->h_blkno),
2599                              le16_to_cpu(el->l_next_free_rec));
2600
2601                         ocfs2_journal_dirty(handle, bh);
2602                         ocfs2_remove_from_cache(et->et_ci, bh);
2603                         continue;
2604                 }
2605
2606                 el->l_next_free_rec = 0;
2607                 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2608
2609                 ocfs2_journal_dirty(handle, bh);
2610
2611                 ret = ocfs2_cache_extent_block_free(dealloc, eb);
2612                 if (ret)
2613                         mlog_errno(ret);
2614
2615                 ocfs2_remove_from_cache(et->et_ci, bh);
2616         }
2617 }
2618
2619 static void ocfs2_unlink_subtree(handle_t *handle,
2620                                  struct ocfs2_extent_tree *et,
2621                                  struct ocfs2_path *left_path,
2622                                  struct ocfs2_path *right_path,
2623                                  int subtree_index,
2624                                  struct ocfs2_cached_dealloc_ctxt *dealloc)
2625 {
2626         int i;
2627         struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2628         struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2629         struct ocfs2_extent_list *el;
2630         struct ocfs2_extent_block *eb;
2631
2632         el = path_leaf_el(left_path);
2633
2634         eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2635
2636         for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2637                 if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2638                         break;
2639
2640         BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2641
2642         memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2643         le16_add_cpu(&root_el->l_next_free_rec, -1);
2644
2645         eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2646         eb->h_next_leaf_blk = 0;
2647
2648         ocfs2_journal_dirty(handle, root_bh);
2649         ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2650
2651         ocfs2_unlink_path(handle, et, dealloc, right_path,
2652                           subtree_index + 1);
2653 }
2654
2655 static int ocfs2_rotate_subtree_left(handle_t *handle,
2656                                      struct ocfs2_extent_tree *et,
2657                                      struct ocfs2_path *left_path,
2658                                      struct ocfs2_path *right_path,
2659                                      int subtree_index,
2660                                      struct ocfs2_cached_dealloc_ctxt *dealloc,
2661                                      int *deleted)
2662 {
2663         int ret, i, del_right_subtree = 0, right_has_empty = 0;
2664         struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path);
2665         struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2666         struct ocfs2_extent_block *eb;
2667
2668         *deleted = 0;
2669
2670         right_leaf_el = path_leaf_el(right_path);
2671         left_leaf_el = path_leaf_el(left_path);
2672         root_bh = left_path->p_node[subtree_index].bh;
2673         BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2674
2675         if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2676                 return 0;
2677
2678         eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2679         if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2680                 /*
2681                  * It's legal for us to proceed if the right leaf is
2682                  * the rightmost one and it has an empty extent. There
2683                  * are two cases to handle - whether the leaf will be
2684                  * empty after removal or not. If the leaf isn't empty
2685                  * then just remove the empty extent up front. The
2686                  * next block will handle empty leaves by flagging
2687                  * them for unlink.
2688                  *
2689                  * Non rightmost leaves will throw -EAGAIN and the
2690                  * caller can manually move the subtree and retry.
2691                  */
2692
2693                 if (eb->h_next_leaf_blk != 0ULL)
2694                         return -EAGAIN;
2695
2696                 if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2697                         ret = ocfs2_journal_access_eb(handle, et->et_ci,
2698                                                       path_leaf_bh(right_path),
2699                                                       OCFS2_JOURNAL_ACCESS_WRITE);
2700                         if (ret) {
2701                                 mlog_errno(ret);
2702                                 goto out;
2703                         }
2704
2705                         ocfs2_remove_empty_extent(right_leaf_el);
2706                 } else
2707                         right_has_empty = 1;
2708         }
2709
2710         if (eb->h_next_leaf_blk == 0ULL &&
2711             le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2712                 /*
2713                  * We have to update i_last_eb_blk during the meta
2714                  * data delete.
2715                  */
2716                 ret = ocfs2_et_root_journal_access(handle, et,
2717                                                    OCFS2_JOURNAL_ACCESS_WRITE);
2718                 if (ret) {
2719                         mlog_errno(ret);
2720                         goto out;
2721                 }
2722
2723                 del_right_subtree = 1;
2724         }
2725
2726         /*
2727          * Getting here with an empty extent in the right path implies
2728          * that it's the rightmost path and will be deleted.
2729          */
2730         BUG_ON(right_has_empty && !del_right_subtree);
2731
2732         ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2733                                            subtree_index);
2734         if (ret) {
2735                 mlog_errno(ret);
2736                 goto out;
2737         }
2738
2739         for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2740                 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2741                                                    right_path, i);
2742                 if (ret) {
2743                         mlog_errno(ret);
2744                         goto out;
2745                 }
2746
2747                 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2748                                                    left_path, i);
2749                 if (ret) {
2750                         mlog_errno(ret);
2751                         goto out;
2752                 }
2753         }
2754
2755         if (!right_has_empty) {
2756                 /*
2757                  * Only do this if we're moving a real
2758                  * record. Otherwise, the action is delayed until
2759                  * after removal of the right path in which case we
2760                  * can do a simple shift to remove the empty extent.
2761                  */
2762                 ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2763                 memset(&right_leaf_el->l_recs[0], 0,
2764                        sizeof(struct ocfs2_extent_rec));
2765         }
2766         if (eb->h_next_leaf_blk == 0ULL) {
2767                 /*
2768                  * Move recs over to get rid of empty extent, decrease
2769                  * next_free. This is allowed to remove the last
2770                  * extent in our leaf (setting l_next_free_rec to
2771                  * zero) - the delete code below won't care.
2772                  */
2773                 ocfs2_remove_empty_extent(right_leaf_el);
2774         }
2775
2776         ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2777         ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2778
2779         if (del_right_subtree) {
2780                 ocfs2_unlink_subtree(handle, et, left_path, right_path,
2781                                      subtree_index, dealloc);
2782                 ret = ocfs2_update_edge_lengths(handle, et, subtree_index,
2783                                                 left_path);
2784                 if (ret) {
2785                         mlog_errno(ret);
2786                         goto out;
2787                 }
2788
2789                 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2790                 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2791
2792                 /*
2793                  * Removal of the extent in the left leaf was skipped
2794                  * above so we could delete the right path
2795                  * 1st.
2796                  */
2797                 if (right_has_empty)
2798                         ocfs2_remove_empty_extent(left_leaf_el);
2799
2800                 ocfs2_journal_dirty(handle, et_root_bh);
2801
2802                 *deleted = 1;
2803         } else
2804                 ocfs2_complete_edge_insert(handle, left_path, right_path,
2805                                            subtree_index);
2806
2807 out:
2808         return ret;
2809 }
2810
2811 /*
2812  * Given a full path, determine what cpos value would return us a path
2813  * containing the leaf immediately to the right of the current one.
2814  *
2815  * Will return zero if the path passed in is already the rightmost path.
2816  *
2817  * This looks similar, but is subtly different to
2818  * ocfs2_find_cpos_for_left_leaf().
2819  */
2820 int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
2821                                    struct ocfs2_path *path, u32 *cpos)
2822 {
2823         int i, j, ret = 0;
2824         u64 blkno;
2825         struct ocfs2_extent_list *el;
2826
2827         *cpos = 0;
2828
2829         if (path->p_tree_depth == 0)
2830                 return 0;
2831
2832         blkno = path_leaf_bh(path)->b_blocknr;
2833
2834         /* Start at the tree node just above the leaf and work our way up. */
2835         i = path->p_tree_depth - 1;
2836         while (i >= 0) {
2837                 int next_free;
2838
2839                 el = path->p_node[i].el;
2840
2841                 /*
2842                  * Find the extent record just after the one in our
2843                  * path.
2844                  */
2845                 next_free = le16_to_cpu(el->l_next_free_rec);
2846                 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2847                         if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2848                                 if (j == (next_free - 1)) {
2849                                         if (i == 0) {
2850                                                 /*
2851                                                  * We've determined that the
2852                                                  * path specified is already
2853                                                  * the rightmost one - return a
2854                                                  * cpos of zero.
2855                                                  */
2856                                                 goto out;
2857                                         }
2858                                         /*
2859                                          * The rightmost record points to our
2860                                          * leaf - we need to travel up the
2861                                          * tree one level.
2862                                          */
2863                                         goto next_node;
2864                                 }
2865
2866                                 *cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
2867                                 goto out;
2868                         }
2869                 }
2870
2871                 /*
2872                  * If we got here, we never found a valid node where
2873                  * the tree indicated one should be.
2874                  */
2875                 ocfs2_error(sb,
2876                             "Invalid extent tree at extent block %llu\n",
2877                             (unsigned long long)blkno);
2878                 ret = -EROFS;
2879                 goto out;
2880
2881 next_node:
2882                 blkno = path->p_node[i].bh->b_blocknr;
2883                 i--;
2884         }
2885
2886 out:
2887         return ret;
2888 }
2889
2890 static int ocfs2_rotate_rightmost_leaf_left(handle_t *handle,
2891                                             struct ocfs2_extent_tree *et,
2892                                             struct ocfs2_path *path)
2893 {
2894         int ret;
2895         struct buffer_head *bh = path_leaf_bh(path);
2896         struct ocfs2_extent_list *el = path_leaf_el(path);
2897
2898         if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2899                 return 0;
2900
2901         ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
2902                                            path_num_items(path) - 1);
2903         if (ret) {
2904                 mlog_errno(ret);
2905                 goto out;
2906         }
2907
2908         ocfs2_remove_empty_extent(el);
2909         ocfs2_journal_dirty(handle, bh);
2910
2911 out:
2912         return ret;
2913 }
2914
2915 static int __ocfs2_rotate_tree_left(handle_t *handle,
2916                                     struct ocfs2_extent_tree *et,
2917                                     int orig_credits,
2918                                     struct ocfs2_path *path,
2919                                     struct ocfs2_cached_dealloc_ctxt *dealloc,
2920                                     struct ocfs2_path **empty_extent_path)
2921 {
2922         int ret, subtree_root, deleted;
2923         u32 right_cpos;
2924         struct ocfs2_path *left_path = NULL;
2925         struct ocfs2_path *right_path = NULL;
2926         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2927
2928         BUG_ON(!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])));
2929
2930         *empty_extent_path = NULL;
2931
2932         ret = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
2933         if (ret) {
2934                 mlog_errno(ret);
2935                 goto out;
2936         }
2937
2938         left_path = ocfs2_new_path_from_path(path);
2939         if (!left_path) {
2940                 ret = -ENOMEM;
2941                 mlog_errno(ret);
2942                 goto out;
2943         }
2944
2945         ocfs2_cp_path(left_path, path);
2946
2947         right_path = ocfs2_new_path_from_path(path);
2948         if (!right_path) {
2949                 ret = -ENOMEM;
2950                 mlog_errno(ret);
2951                 goto out;
2952         }
2953
2954         while (right_cpos) {
2955                 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
2956                 if (ret) {
2957                         mlog_errno(ret);
2958                         goto out;
2959                 }
2960
2961                 subtree_root = ocfs2_find_subtree_root(et, left_path,
2962                                                        right_path);
2963
2964                 trace_ocfs2_rotate_subtree(subtree_root,
2965                      (unsigned long long)
2966                      right_path->p_node[subtree_root].bh->b_blocknr,
2967                      right_path->p_tree_depth);
2968
2969                 ret = ocfs2_extend_rotate_transaction(handle, subtree_root,
2970                                                       orig_credits, left_path);
2971                 if (ret) {
2972                         mlog_errno(ret);
2973                         goto out;
2974                 }
2975
2976                 /*
2977                  * Caller might still want to make changes to the
2978                  * tree root, so re-add it to the journal here.
2979                  */
2980                 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2981                                                    left_path, 0);
2982                 if (ret) {
2983                         mlog_errno(ret);
2984                         goto out;
2985                 }
2986
2987                 ret = ocfs2_rotate_subtree_left(handle, et, left_path,
2988                                                 right_path, subtree_root,
2989                                                 dealloc, &deleted);
2990                 if (ret == -EAGAIN) {
2991                         /*
2992                          * The rotation has to temporarily stop due to
2993                          * the right subtree having an empty
2994                          * extent. Pass it back to the caller for a
2995                          * fixup.
2996                          */
2997                         *empty_extent_path = right_path;
2998                         right_path = NULL;
2999                         goto out;
3000                 }
3001                 if (ret) {
3002                         mlog_errno(ret);
3003                         goto out;
3004                 }
3005
3006                 /*
3007                  * The subtree rotate might have removed records on
3008                  * the rightmost edge. If so, then rotation is
3009                  * complete.
3010                  */
3011                 if (deleted)
3012                         break;
3013
3014                 ocfs2_mv_path(left_path, right_path);
3015
3016                 ret = ocfs2_find_cpos_for_right_leaf(sb, left_path,
3017                                                      &right_cpos);
3018                 if (ret) {
3019                         mlog_errno(ret);
3020                         goto out;
3021                 }
3022         }
3023
3024 out:
3025         ocfs2_free_path(right_path);
3026         ocfs2_free_path(left_path);
3027
3028         return ret;
3029 }
3030
3031 static int ocfs2_remove_rightmost_path(handle_t *handle,
3032                                 struct ocfs2_extent_tree *et,
3033                                 struct ocfs2_path *path,
3034                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
3035 {
3036         int ret, subtree_index;
3037         u32 cpos;
3038         struct ocfs2_path *left_path = NULL;
3039         struct ocfs2_extent_block *eb;
3040         struct ocfs2_extent_list *el;
3041
3042
3043         ret = ocfs2_et_sanity_check(et);
3044         if (ret)
3045                 goto out;
3046         /*
3047          * There's two ways we handle this depending on
3048          * whether path is the only existing one.
3049          */
3050         ret = ocfs2_extend_rotate_transaction(handle, 0,
3051                                               handle->h_buffer_credits,
3052                                               path);
3053         if (ret) {
3054                 mlog_errno(ret);
3055                 goto out;
3056         }
3057
3058         ret = ocfs2_journal_access_path(et->et_ci, handle, path);
3059         if (ret) {
3060                 mlog_errno(ret);
3061                 goto out;
3062         }
3063
3064         ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3065                                             path, &cpos);
3066         if (ret) {
3067                 mlog_errno(ret);
3068                 goto out;
3069         }
3070
3071         if (cpos) {
3072                 /*
3073                  * We have a path to the left of this one - it needs
3074                  * an update too.
3075                  */
3076                 left_path = ocfs2_new_path_from_path(path);
3077                 if (!left_path) {
3078                         ret = -ENOMEM;
3079                         mlog_errno(ret);
3080                         goto out;
3081                 }
3082
3083                 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
3084                 if (ret) {
3085                         mlog_errno(ret);
3086                         goto out;
3087                 }
3088
3089                 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
3090                 if (ret) {
3091                         mlog_errno(ret);
3092                         goto out;
3093                 }
3094
3095                 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
3096
3097                 ocfs2_unlink_subtree(handle, et, left_path, path,
3098                                      subtree_index, dealloc);
3099                 ret = ocfs2_update_edge_lengths(handle, et, subtree_index,
3100                                                 left_path);
3101                 if (ret) {
3102                         mlog_errno(ret);
3103                         goto out;
3104                 }
3105
3106                 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
3107                 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
3108         } else {
3109                 /*
3110                  * 'path' is also the leftmost path which
3111                  * means it must be the only one. This gets
3112                  * handled differently because we want to
3113                  * revert the root back to having extents
3114                  * in-line.
3115                  */
3116                 ocfs2_unlink_path(handle, et, dealloc, path, 1);
3117
3118                 el = et->et_root_el;
3119                 el->l_tree_depth = 0;
3120                 el->l_next_free_rec = 0;
3121                 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3122
3123                 ocfs2_et_set_last_eb_blk(et, 0);
3124         }
3125
3126         ocfs2_journal_dirty(handle, path_root_bh(path));
3127
3128 out:
3129         ocfs2_free_path(left_path);
3130         return ret;
3131 }
3132
3133 /*
3134  * Left rotation of btree records.
3135  *
3136  * In many ways, this is (unsurprisingly) the opposite of right
3137  * rotation. We start at some non-rightmost path containing an empty
3138  * extent in the leaf block. The code works its way to the rightmost
3139  * path by rotating records to the left in every subtree.
3140  *
3141  * This is used by any code which reduces the number of extent records
3142  * in a leaf. After removal, an empty record should be placed in the
3143  * leftmost list position.
3144  *
3145  * This won't handle a length update of the rightmost path records if
3146  * the rightmost tree leaf record is removed so the caller is
3147  * responsible for detecting and correcting that.
3148  */
3149 static int ocfs2_rotate_tree_left(handle_t *handle,
3150                                   struct ocfs2_extent_tree *et,
3151                                   struct ocfs2_path *path,
3152                                   struct ocfs2_cached_dealloc_ctxt *dealloc)
3153 {
3154         int ret, orig_credits = handle->h_buffer_credits;
3155         struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
3156         struct ocfs2_extent_block *eb;
3157         struct ocfs2_extent_list *el;
3158
3159         el = path_leaf_el(path);
3160         if (!ocfs2_is_empty_extent(&el->l_recs[0]))
3161                 return 0;
3162
3163         if (path->p_tree_depth == 0) {
3164 rightmost_no_delete:
3165                 /*
3166                  * Inline extents. This is trivially handled, so do
3167                  * it up front.
3168                  */
3169                 ret = ocfs2_rotate_rightmost_leaf_left(handle, et, path);
3170                 if (ret)
3171                         mlog_errno(ret);
3172                 goto out;
3173         }
3174
3175         /*
3176          * Handle rightmost branch now. There's several cases:
3177          *  1) simple rotation leaving records in there. That's trivial.
3178          *  2) rotation requiring a branch delete - there's no more
3179          *     records left. Two cases of this:
3180          *     a) There are branches to the left.
3181          *     b) This is also the leftmost (the only) branch.
3182          *
3183          *  1) is handled via ocfs2_rotate_rightmost_leaf_left()
3184          *  2a) we need the left branch so that we can update it with the unlink
3185          *  2b) we need to bring the root back to inline extents.
3186          */
3187
3188         eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
3189         el = &eb->h_list;
3190         if (eb->h_next_leaf_blk == 0) {
3191                 /*
3192                  * This gets a bit tricky if we're going to delete the
3193                  * rightmost path. Get the other cases out of the way
3194                  * 1st.
3195                  */
3196                 if (le16_to_cpu(el->l_next_free_rec) > 1)
3197                         goto rightmost_no_delete;
3198
3199                 if (le16_to_cpu(el->l_next_free_rec) == 0) {
3200                         ret = -EIO;
3201                         ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3202                                     "Owner %llu has empty extent block at %llu",
3203                                     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
3204                                     (unsigned long long)le64_to_cpu(eb->h_blkno));
3205                         goto out;
3206                 }
3207
3208                 /*
3209                  * XXX: The caller can not trust "path" any more after
3210                  * this as it will have been deleted. What do we do?
3211                  *
3212                  * In theory the rotate-for-merge code will never get
3213                  * here because it'll always ask for a rotate in a
3214                  * nonempty list.
3215                  */
3216
3217                 ret = ocfs2_remove_rightmost_path(handle, et, path,
3218                                                   dealloc);
3219                 if (ret)
3220                         mlog_errno(ret);
3221                 goto out;
3222         }
3223
3224         /*
3225          * Now we can loop, remembering the path we get from -EAGAIN
3226          * and restarting from there.
3227          */
3228 try_rotate:
3229         ret = __ocfs2_rotate_tree_left(handle, et, orig_credits, path,
3230                                        dealloc, &restart_path);
3231         if (ret && ret != -EAGAIN) {
3232                 mlog_errno(ret);
3233                 goto out;
3234         }
3235
3236         while (ret == -EAGAIN) {
3237                 tmp_path = restart_path;
3238                 restart_path = NULL;
3239
3240                 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits,
3241                                                tmp_path, dealloc,
3242                                                &restart_path);
3243                 if (ret && ret != -EAGAIN) {
3244                         mlog_errno(ret);
3245                         goto out;
3246                 }
3247
3248                 ocfs2_free_path(tmp_path);
3249                 tmp_path = NULL;
3250
3251                 if (ret == 0)
3252                         goto try_rotate;
3253         }
3254
3255 out:
3256         ocfs2_free_path(tmp_path);
3257         ocfs2_free_path(restart_path);
3258         return ret;
3259 }
3260
3261 static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
3262                                 int index)
3263 {
3264         struct ocfs2_extent_rec *rec = &el->l_recs[index];
3265         unsigned int size;
3266
3267         if (rec->e_leaf_clusters == 0) {
3268                 /*
3269                  * We consumed all of the merged-from record. An empty
3270                  * extent cannot exist anywhere but the 1st array
3271                  * position, so move things over if the merged-from
3272                  * record doesn't occupy that position.
3273                  *
3274                  * This creates a new empty extent so the caller
3275                  * should be smart enough to have removed any existing
3276                  * ones.
3277                  */
3278                 if (index > 0) {
3279                         BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3280                         size = index * sizeof(struct ocfs2_extent_rec);
3281                         memmove(&el->l_recs[1], &el->l_recs[0], size);
3282                 }
3283
3284                 /*
3285                  * Always memset - the caller doesn't check whether it
3286                  * created an empty extent, so there could be junk in
3287                  * the other fields.
3288                  */
3289                 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3290         }
3291 }
3292
3293 static int ocfs2_get_right_path(struct ocfs2_extent_tree *et,
3294                                 struct ocfs2_path *left_path,
3295                                 struct ocfs2_path **ret_right_path)
3296 {
3297         int ret;
3298         u32 right_cpos;
3299         struct ocfs2_path *right_path = NULL;
3300         struct ocfs2_extent_list *left_el;
3301
3302         *ret_right_path = NULL;
3303
3304         /* This function shouldn't be called for non-trees. */
3305         BUG_ON(left_path->p_tree_depth == 0);
3306
3307         left_el = path_leaf_el(left_path);
3308         BUG_ON(left_el->l_next_free_rec != left_el->l_count);
3309
3310         ret = ocfs2_find_cpos_for_right_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3311                                              left_path, &right_cpos);
3312         if (ret) {
3313                 mlog_errno(ret);
3314                 goto out;
3315         }
3316
3317         /* This function shouldn't be called for the rightmost leaf. */
3318         BUG_ON(right_cpos == 0);
3319
3320         right_path = ocfs2_new_path_from_path(left_path);
3321         if (!right_path) {
3322                 ret = -ENOMEM;
3323                 mlog_errno(ret);
3324                 goto out;
3325         }
3326
3327         ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
3328         if (ret) {
3329                 mlog_errno(ret);
3330                 goto out;
3331         }
3332
3333         *ret_right_path = right_path;
3334 out:
3335         if (ret)
3336                 ocfs2_free_path(right_path);
3337         return ret;
3338 }
3339
3340 /*
3341  * Remove split_rec clusters from the record at index and merge them
3342  * onto the beginning of the record "next" to it.
3343  * For index < l_count - 1, the next means the extent rec at index + 1.
3344  * For index == l_count - 1, the "next" means the 1st extent rec of the
3345  * next extent block.
3346  */
3347 static int ocfs2_merge_rec_right(struct ocfs2_path *left_path,
3348                                  handle_t *handle,
3349                                  struct ocfs2_extent_tree *et,
3350                                  struct ocfs2_extent_rec *split_rec,
3351                                  int index)
3352 {
3353         int ret, next_free, i;
3354         unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3355         struct ocfs2_extent_rec *left_rec;
3356         struct ocfs2_extent_rec *right_rec;
3357         struct ocfs2_extent_list *right_el;
3358         struct ocfs2_path *right_path = NULL;
3359         int subtree_index = 0;
3360         struct ocfs2_extent_list *el = path_leaf_el(left_path);
3361         struct buffer_head *bh = path_leaf_bh(left_path);
3362         struct buffer_head *root_bh = NULL;
3363
3364         BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
3365         left_rec = &el->l_recs[index];
3366
3367         if (index == le16_to_cpu(el->l_next_free_rec) - 1 &&
3368             le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count)) {
3369                 /* we meet with a cross extent block merge. */
3370                 ret = ocfs2_get_right_path(et, left_path, &right_path);
3371                 if (ret) {
3372                         mlog_errno(ret);
3373                         goto out;
3374                 }
3375
3376                 right_el = path_leaf_el(right_path);
3377                 next_free = le16_to_cpu(right_el->l_next_free_rec);
3378                 BUG_ON(next_free <= 0);
3379                 right_rec = &right_el->l_recs[0];
3380                 if (ocfs2_is_empty_extent(right_rec)) {
3381                         BUG_ON(next_free <= 1);
3382                         right_rec = &right_el->l_recs[1];
3383                 }
3384
3385                 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3386                        le16_to_cpu(left_rec->e_leaf_clusters) !=
3387                        le32_to_cpu(right_rec->e_cpos));
3388
3389                 subtree_index = ocfs2_find_subtree_root(et, left_path,
3390                                                         right_path);
3391
3392                 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3393                                                       handle->h_buffer_credits,
3394                                                       right_path);
3395                 if (ret) {
3396                         mlog_errno(ret);
3397                         goto out;
3398                 }
3399
3400                 root_bh = left_path->p_node[subtree_index].bh;
3401                 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3402
3403                 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3404                                                    subtree_index);
3405                 if (ret) {
3406                         mlog_errno(ret);
3407                         goto out;
3408                 }
3409
3410                 for (i = subtree_index + 1;
3411                      i < path_num_items(right_path); i++) {
3412                         ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3413                                                            right_path, i);
3414                         if (ret) {
3415                                 mlog_errno(ret);
3416                                 goto out;
3417                         }
3418
3419                         ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3420                                                            left_path, i);
3421                         if (ret) {
3422                                 mlog_errno(ret);
3423                                 goto out;
3424                         }
3425                 }
3426
3427         } else {
3428                 BUG_ON(index == le16_to_cpu(el->l_next_free_rec) - 1);
3429                 right_rec = &el->l_recs[index + 1];
3430         }
3431
3432         ret = ocfs2_path_bh_journal_access(handle, et->et_ci, left_path,
3433                                            path_num_items(left_path) - 1);
3434         if (ret) {
3435                 mlog_errno(ret);
3436                 goto out;
3437         }
3438
3439         le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
3440
3441         le32_add_cpu(&right_rec->e_cpos, -split_clusters);
3442         le64_add_cpu(&right_rec->e_blkno,
3443                      -ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3444                                                split_clusters));
3445         le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
3446
3447         ocfs2_cleanup_merge(el, index);
3448
3449         ocfs2_journal_dirty(handle, bh);
3450         if (right_path) {
3451                 ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
3452                 ocfs2_complete_edge_insert(handle, left_path, right_path,
3453                                            subtree_index);
3454         }
3455 out:
3456         if (right_path)
3457                 ocfs2_free_path(right_path);
3458         return ret;
3459 }
3460
3461 static int ocfs2_get_left_path(struct ocfs2_extent_tree *et,
3462                                struct ocfs2_path *right_path,
3463                                struct ocfs2_path **ret_left_path)
3464 {
3465         int ret;
3466         u32 left_cpos;
3467         struct ocfs2_path *left_path = NULL;
3468
3469         *ret_left_path = NULL;
3470
3471         /* This function shouldn't be called for non-trees. */
3472         BUG_ON(right_path->p_tree_depth == 0);
3473
3474         ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3475                                             right_path, &left_cpos);
3476         if (ret) {
3477                 mlog_errno(ret);
3478                 goto out;
3479         }
3480
3481         /* This function shouldn't be called for the leftmost leaf. */
3482         BUG_ON(left_cpos == 0);
3483
3484         left_path = ocfs2_new_path_from_path(right_path);
3485         if (!left_path) {
3486                 ret = -ENOMEM;
3487                 mlog_errno(ret);
3488                 goto out;
3489         }
3490
3491         ret = ocfs2_find_path(et->et_ci, left_path, left_cpos);
3492         if (ret) {
3493                 mlog_errno(ret);
3494                 goto out;
3495         }
3496
3497         *ret_left_path = left_path;
3498 out:
3499         if (ret)
3500                 ocfs2_free_path(left_path);
3501         return ret;
3502 }
3503
3504 /*
3505  * Remove split_rec clusters from the record at index and merge them
3506  * onto the tail of the record "before" it.
3507  * For index > 0, the "before" means the extent rec at index - 1.
3508  *
3509  * For index == 0, the "before" means the last record of the previous
3510  * extent block. And there is also a situation that we may need to
3511  * remove the rightmost leaf extent block in the right_path and change
3512  * the right path to indicate the new rightmost path.
3513  */
3514 static int ocfs2_merge_rec_left(struct ocfs2_path *right_path,
3515                                 handle_t *handle,
3516                                 struct ocfs2_extent_tree *et,
3517                                 struct ocfs2_extent_rec *split_rec,
3518                                 struct ocfs2_cached_dealloc_ctxt *dealloc,
3519                                 int index)
3520 {
3521         int ret, i, subtree_index = 0, has_empty_extent = 0;
3522         unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3523         struct ocfs2_extent_rec *left_rec;
3524         struct ocfs2_extent_rec *right_rec;
3525         struct ocfs2_extent_list *el = path_leaf_el(right_path);
3526         struct buffer_head *bh = path_leaf_bh(right_path);
3527         struct buffer_head *root_bh = NULL;
3528         struct ocfs2_path *left_path = NULL;
3529         struct ocfs2_extent_list *left_el;
3530
3531         BUG_ON(index < 0);
3532
3533         right_rec = &el->l_recs[index];
3534         if (index == 0) {
3535                 /* we meet with a cross extent block merge. */
3536                 ret = ocfs2_get_left_path(et, right_path, &left_path);
3537                 if (ret) {
3538                         mlog_errno(ret);
3539                         goto out;
3540                 }
3541
3542                 left_el = path_leaf_el(left_path);
3543                 BUG_ON(le16_to_cpu(left_el->l_next_free_rec) !=
3544                        le16_to_cpu(left_el->l_count));
3545
3546                 left_rec = &left_el->l_recs[
3547                                 le16_to_cpu(left_el->l_next_free_rec) - 1];
3548                 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3549                        le16_to_cpu(left_rec->e_leaf_clusters) !=
3550                        le32_to_cpu(split_rec->e_cpos));
3551
3552                 subtree_index = ocfs2_find_subtree_root(et, left_path,
3553                                                         right_path);
3554
3555                 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3556                                                       handle->h_buffer_credits,
3557                                                       left_path);
3558                 if (ret) {
3559                         mlog_errno(ret);
3560                         goto out;
3561                 }
3562
3563                 root_bh = left_path->p_node[subtree_index].bh;
3564                 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3565
3566                 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3567                                                    subtree_index);
3568                 if (ret) {
3569                         mlog_errno(ret);
3570                         goto out;
3571                 }
3572
3573                 for (i = subtree_index + 1;
3574                      i < path_num_items(right_path); i++) {
3575                         ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3576                                                            right_path, i);
3577                         if (ret) {
3578                                 mlog_errno(ret);
3579                                 goto out;
3580                         }
3581
3582                         ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3583                                                            left_path, i);
3584                         if (ret) {
3585                                 mlog_errno(ret);
3586                                 goto out;
3587                         }
3588                 }
3589         } else {
3590                 left_rec = &el->l_recs[index - 1];
3591                 if (ocfs2_is_empty_extent(&el->l_recs[0]))
3592                         has_empty_extent = 1;
3593         }
3594
3595         ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3596                                            path_num_items(right_path) - 1);
3597         if (ret) {
3598                 mlog_errno(ret);
3599                 goto out;
3600         }
3601
3602         if (has_empty_extent && index == 1) {
3603                 /*
3604                  * The easy case - we can just plop the record right in.
3605                  */
3606                 *left_rec = *split_rec;
3607
3608                 has_empty_extent = 0;
3609         } else
3610                 le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
3611
3612         le32_add_cpu(&right_rec->e_cpos, split_clusters);
3613         le64_add_cpu(&right_rec->e_blkno,
3614                      ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3615                                               split_clusters));
3616         le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
3617
3618         ocfs2_cleanup_merge(el, index);
3619
3620         ocfs2_journal_dirty(handle, bh);
3621         if (left_path) {
3622                 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
3623
3624                 /*
3625                  * In the situation that the right_rec is empty and the extent
3626                  * block is empty also,  ocfs2_complete_edge_insert can't handle
3627                  * it and we need to delete the right extent block.
3628                  */
3629                 if (le16_to_cpu(right_rec->e_leaf_clusters) == 0 &&
3630                     le16_to_cpu(el->l_next_free_rec) == 1) {
3631
3632                         ret = ocfs2_remove_rightmost_path(handle, et,
3633                                                           right_path,
3634                                                           dealloc);
3635                         if (ret) {
3636                                 mlog_errno(ret);
3637                                 goto out;
3638                         }
3639
3640                         /* Now the rightmost extent block has been deleted.
3641                          * So we use the new rightmost path.
3642                          */
3643                         ocfs2_mv_path(right_path, left_path);
3644                         left_path = NULL;
3645                 } else
3646                         ocfs2_complete_edge_insert(handle, left_path,
3647                                                    right_path, subtree_index);
3648         }
3649 out:
3650         if (left_path)
3651                 ocfs2_free_path(left_path);
3652         return ret;
3653 }
3654
3655 static int ocfs2_try_to_merge_extent(handle_t *handle,
3656                                      struct ocfs2_extent_tree *et,
3657                                      struct ocfs2_path *path,
3658                                      int split_index,
3659                                      struct ocfs2_extent_rec *split_rec,
3660                                      struct ocfs2_cached_dealloc_ctxt *dealloc,
3661                                      struct ocfs2_merge_ctxt *ctxt)
3662 {
3663         int ret = 0;
3664         struct ocfs2_extent_list *el = path_leaf_el(path);
3665         struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3666
3667         BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
3668
3669         if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) {
3670                 /*
3671                  * The merge code will need to create an empty
3672                  * extent to take the place of the newly
3673                  * emptied slot. Remove any pre-existing empty
3674                  * extents - having more than one in a leaf is
3675                  * illegal.
3676                  */
3677                 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3678                 if (ret) {
3679                         mlog_errno(ret);
3680                         goto out;
3681                 }
3682                 split_index--;
3683                 rec = &el->l_recs[split_index];
3684         }
3685
3686         if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
3687                 /*
3688                  * Left-right contig implies this.
3689                  */
3690                 BUG_ON(!ctxt->c_split_covers_rec);
3691
3692                 /*
3693                  * Since the leftright insert always covers the entire
3694                  * extent, this call will delete the insert record
3695                  * entirely, resulting in an empty extent record added to
3696                  * the extent block.
3697                  *
3698                  * Since the adding of an empty extent shifts
3699                  * everything back to the right, there's no need to
3700                  * update split_index here.
3701                  *
3702                  * When the split_index is zero, we need to merge it to the
3703                  * prevoius extent block. It is more efficient and easier
3704                  * if we do merge_right first and merge_left later.
3705                  */
3706                 ret = ocfs2_merge_rec_right(path, handle, et, split_rec,
3707                                             split_index);
3708                 if (ret) {
3709                         mlog_errno(ret);
3710                         goto out;
3711                 }
3712
3713                 /*
3714                  * We can only get this from logic error above.
3715                  */
3716                 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
3717
3718                 /* The merge left us with an empty extent, remove it. */
3719                 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3720                 if (ret) {
3721                         mlog_errno(ret);
3722                         goto out;
3723                 }
3724
3725                 rec = &el->l_recs[split_index];
3726
3727                 /*
3728                  * Note that we don't pass split_rec here on purpose -
3729                  * we've merged it into the rec already.
3730                  */
3731                 ret = ocfs2_merge_rec_left(path, handle, et, rec,
3732                                            dealloc, split_index);
3733
3734                 if (ret) {
3735                         mlog_errno(ret);
3736                         goto out;
3737                 }
3738
3739                 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3740                 /*
3741                  * Error from this last rotate is not critical, so
3742                  * print but don't bubble it up.
3743                  */
3744                 if (ret)
3745                         mlog_errno(ret);
3746                 ret = 0;
3747         } else {
3748                 /*
3749                  * Merge a record to the left or right.
3750                  *
3751                  * 'contig_type' is relative to the existing record,
3752                  * so for example, if we're "right contig", it's to
3753                  * the record on the left (hence the left merge).
3754                  */
3755                 if (ctxt->c_contig_type == CONTIG_RIGHT) {
3756                         ret = ocfs2_merge_rec_left(path, handle, et,
3757                                                    split_rec, dealloc,
3758                                                    split_index);
3759                         if (ret) {
3760                                 mlog_errno(ret);
3761                                 goto out;
3762                         }
3763                 } else {
3764                         ret = ocfs2_merge_rec_right(path, handle,
3765                                                     et, split_rec,
3766                                                     split_index);
3767                         if (ret) {
3768                                 mlog_errno(ret);
3769                                 goto out;
3770                         }
3771                 }
3772
3773                 if (ctxt->c_split_covers_rec) {
3774                         /*
3775                          * The merge may have left an empty extent in
3776                          * our leaf. Try to rotate it away.
3777                          */
3778                         ret = ocfs2_rotate_tree_left(handle, et, path,
3779                                                      dealloc);
3780                         if (ret)
3781                                 mlog_errno(ret);
3782                         ret = 0;
3783                 }
3784         }
3785
3786 out:
3787         return ret;
3788 }
3789
3790 static void ocfs2_subtract_from_rec(struct super_block *sb,
3791                                     enum ocfs2_split_type split,
3792                                     struct ocfs2_extent_rec *rec,
3793                                     struct ocfs2_extent_rec *split_rec)
3794 {
3795         u64 len_blocks;
3796
3797         len_blocks = ocfs2_clusters_to_blocks(sb,
3798                                 le16_to_cpu(split_rec->e_leaf_clusters));
3799
3800         if (split == SPLIT_LEFT) {
3801                 /*
3802                  * Region is on the left edge of the existing
3803                  * record.
3804                  */
3805                 le32_add_cpu(&rec->e_cpos,
3806                              le16_to_cpu(split_rec->e_leaf_clusters));
3807                 le64_add_cpu(&rec->e_blkno, len_blocks);
3808                 le16_add_cpu(&rec->e_leaf_clusters,
3809                              -le16_to_cpu(split_rec->e_leaf_clusters));
3810         } else {
3811                 /*
3812                  * Region is on the right edge of the existing
3813                  * record.
3814                  */
3815                 le16_add_cpu(&rec->e_leaf_clusters,
3816                              -le16_to_cpu(split_rec->e_leaf_clusters));
3817         }
3818 }
3819
3820 /*
3821  * Do the final bits of extent record insertion at the target leaf
3822  * list. If this leaf is part of an allocation tree, it is assumed
3823  * that the tree above has been prepared.
3824  */
3825 static void ocfs2_insert_at_leaf(struct ocfs2_extent_tree *et,
3826                                  struct ocfs2_extent_rec *insert_rec,
3827                                  struct ocfs2_extent_list *el,
3828                                  struct ocfs2_insert_type *insert)
3829 {
3830         int i = insert->ins_contig_index;
3831         unsigned int range;
3832         struct ocfs2_extent_rec *rec;
3833
3834         BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3835
3836         if (insert->ins_split != SPLIT_NONE) {
3837                 i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
3838                 BUG_ON(i == -1);
3839                 rec = &el->l_recs[i];
3840                 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
3841                                         insert->ins_split, rec,
3842                                         insert_rec);
3843                 goto rotate;
3844         }
3845
3846         /*
3847          * Contiguous insert - either left or right.
3848          */
3849         if (insert->ins_contig != CONTIG_NONE) {
3850                 rec = &el->l_recs[i];
3851                 if (insert->ins_contig == CONTIG_LEFT) {
3852                         rec->e_blkno = insert_rec->e_blkno;
3853                         rec->e_cpos = insert_rec->e_cpos;
3854                 }
3855                 le16_add_cpu(&rec->e_leaf_clusters,
3856                              le16_to_cpu(insert_rec->e_leaf_clusters));
3857                 return;
3858         }
3859
3860         /*
3861          * Handle insert into an empty leaf.
3862          */
3863         if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3864             ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3865              ocfs2_is_empty_extent(&el->l_recs[0]))) {
3866                 el->l_recs[0] = *insert_rec;
3867                 el->l_next_free_rec = cpu_to_le16(1);
3868                 return;
3869         }
3870
3871         /*
3872          * Appending insert.
3873          */
3874         if (insert->ins_appending == APPEND_TAIL) {
3875                 i = le16_to_cpu(el->l_next_free_rec) - 1;
3876                 rec = &el->l_recs[i];
3877                 range = le32_to_cpu(rec->e_cpos)
3878                         + le16_to_cpu(rec->e_leaf_clusters);
3879                 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3880
3881                 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3882                                 le16_to_cpu(el->l_count),
3883                                 "owner %llu, depth %u, count %u, next free %u, "
3884                                 "rec.cpos %u, rec.clusters %u, "
3885                                 "insert.cpos %u, insert.clusters %u\n",
3886                                 ocfs2_metadata_cache_owner(et->et_ci),
3887                                 le16_to_cpu(el->l_tree_depth),
3888                                 le16_to_cpu(el->l_count),
3889                                 le16_to_cpu(el->l_next_free_rec),
3890                                 le32_to_cpu(el->l_recs[i].e_cpos),
3891                                 le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3892                                 le32_to_cpu(insert_rec->e_cpos),
3893                                 le16_to_cpu(insert_rec->e_leaf_clusters));
3894                 i++;
3895                 el->l_recs[i] = *insert_rec;
3896                 le16_add_cpu(&el->l_next_free_rec, 1);
3897                 return;
3898         }
3899
3900 rotate:
3901         /*
3902          * Ok, we have to rotate.
3903          *
3904          * At this point, it is safe to assume that inserting into an
3905          * empty leaf and appending to a leaf have both been handled
3906          * above.
3907          *
3908          * This leaf needs to have space, either by the empty 1st
3909          * extent record, or by virtue of an l_next_rec < l_count.
3910          */
3911         ocfs2_rotate_leaf(el, insert_rec);
3912 }
3913
3914 static void ocfs2_adjust_rightmost_records(handle_t *handle,
3915                                            struct ocfs2_extent_tree *et,
3916                                            struct ocfs2_path *path,
3917                                            struct ocfs2_extent_rec *insert_rec)
3918 {
3919         int ret, i, next_free;
3920         struct buffer_head *bh;
3921         struct ocfs2_extent_list *el;
3922         struct ocfs2_extent_rec *rec;
3923
3924         /*
3925          * Update everything except the leaf block.
3926          */
3927         for (i = 0; i < path->p_tree_depth; i++) {
3928                 bh = path->p_node[i].bh;
3929                 el = path->p_node[i].el;
3930
3931                 next_free = le16_to_cpu(el->l_next_free_rec);
3932                 if (next_free == 0) {
3933                         ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3934                                     "Owner %llu has a bad extent list",
3935                                     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
3936                         ret = -EIO;
3937                         return;
3938                 }
3939
3940                 rec = &el->l_recs[next_free - 1];
3941
3942                 rec->e_int_clusters = insert_rec->e_cpos;
3943                 le32_add_cpu(&rec->e_int_clusters,
3944                              le16_to_cpu(insert_rec->e_leaf_clusters));
3945                 le32_add_cpu(&rec->e_int_clusters,
3946                              -le32_to_cpu(rec->e_cpos));
3947
3948                 ocfs2_journal_dirty(handle, bh);
3949         }
3950 }
3951
3952 static int ocfs2_append_rec_to_path(handle_t *handle,
3953                                     struct ocfs2_extent_tree *et,
3954                                     struct ocfs2_extent_rec *insert_rec,
3955                                     struct ocfs2_path *right_path,
3956                                     struct ocfs2_path **ret_left_path)
3957 {
3958         int ret, next_free;
3959         struct ocfs2_extent_list *el;
3960         struct ocfs2_path *left_path = NULL;
3961
3962         *ret_left_path = NULL;
3963
3964         /*
3965          * This shouldn't happen for non-trees. The extent rec cluster
3966          * count manipulation below only works for interior nodes.
3967          */
3968         BUG_ON(right_path->p_tree_depth == 0);
3969
3970         /*
3971          * If our appending insert is at the leftmost edge of a leaf,
3972          * then we might need to update the rightmost records of the
3973          * neighboring path.
3974          */
3975         el = path_leaf_el(right_path);
3976         next_free = le16_to_cpu(el->l_next_free_rec);
3977         if (next_free == 0 ||
3978             (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
3979                 u32 left_cpos;
3980
3981                 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3982                                                     right_path, &left_cpos);
3983                 if (ret) {
3984                         mlog_errno(ret);
3985                         goto out;
3986                 }
3987
3988                 trace_ocfs2_append_rec_to_path(
3989                         (unsigned long long)
3990                         ocfs2_metadata_cache_owner(et->et_ci),
3991                         le32_to_cpu(insert_rec->e_cpos),
3992                         left_cpos);
3993
3994                 /*
3995                  * No need to worry if the append is already in the
3996                  * leftmost leaf.
3997                  */
3998                 if (left_cpos) {
3999                         left_path = ocfs2_new_path_from_path(right_path);
4000                         if (!left_path) {
4001                                 ret = -ENOMEM;
4002                                 mlog_errno(ret);
4003                                 goto out;
4004                         }
4005
4006                         ret = ocfs2_find_path(et->et_ci, left_path,
4007                                               left_cpos);
4008                         if (ret) {
4009                                 mlog_errno(ret);
4010                                 goto out;
4011                         }
4012
4013                         /*
4014                          * ocfs2_insert_path() will pass the left_path to the
4015                          * journal for us.
4016                          */
4017                 }
4018         }
4019
4020         ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4021         if (ret) {
4022                 mlog_errno(ret);
4023                 goto out;
4024         }
4025
4026         ocfs2_adjust_rightmost_records(handle, et, right_path, insert_rec);
4027
4028         *ret_left_path = left_path;
4029         ret = 0;
4030 out:
4031         if (ret != 0)
4032                 ocfs2_free_path(left_path);
4033
4034         return ret;
4035 }
4036
4037 static void ocfs2_split_record(struct ocfs2_extent_tree *et,
4038                                struct ocfs2_path *left_path,
4039                                struct ocfs2_path *right_path,
4040                                struct ocfs2_extent_rec *split_rec,
4041                                enum ocfs2_split_type split)
4042 {
4043         int index;
4044         u32 cpos = le32_to_cpu(split_rec->e_cpos);
4045         struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
4046         struct ocfs2_extent_rec *rec, *tmprec;
4047
4048         right_el = path_leaf_el(right_path);
4049         if (left_path)
4050                 left_el = path_leaf_el(left_path);
4051
4052         el = right_el;
4053         insert_el = right_el;
4054         index = ocfs2_search_extent_list(el, cpos);
4055         if (index != -1) {
4056                 if (index == 0 && left_path) {
4057                         BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
4058
4059                         /*
4060                          * This typically means that the record
4061                          * started in the left path but moved to the
4062                          * right as a result of rotation. We either
4063                          * move the existing record to the left, or we
4064                          * do the later insert there.
4065                          *
4066                          * In this case, the left path should always
4067                          * exist as the rotate code will have passed
4068                          * it back for a post-insert update.
4069                          */
4070
4071                         if (split == SPLIT_LEFT) {
4072                                 /*
4073                                  * It's a left split. Since we know
4074                                  * that the rotate code gave us an
4075                                  * empty extent in the left path, we
4076                                  * can just do the insert there.
4077                                  */
4078                                 insert_el = left_el;
4079                         } else {
4080                                 /*
4081                                  * Right split - we have to move the
4082                                  * existing record over to the left
4083                                  * leaf. The insert will be into the
4084                                  * newly created empty extent in the
4085                                  * right leaf.
4086                                  */
4087                                 tmprec = &right_el->l_recs[index];
4088                                 ocfs2_rotate_leaf(left_el, tmprec);
4089                                 el = left_el;
4090
4091                                 memset(tmprec, 0, sizeof(*tmprec));
4092                                 index = ocfs2_search_extent_list(left_el, cpos);
4093                                 BUG_ON(index == -1);
4094                         }
4095                 }
4096         } else {
4097                 BUG_ON(!left_path);
4098                 BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
4099                 /*
4100                  * Left path is easy - we can just allow the insert to
4101                  * happen.
4102                  */
4103                 el = left_el;
4104                 insert_el = left_el;
4105                 index = ocfs2_search_extent_list(el, cpos);
4106                 BUG_ON(index == -1);
4107         }
4108
4109         rec = &el->l_recs[index];
4110         ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4111                                 split, rec, split_rec);
4112         ocfs2_rotate_leaf(insert_el, split_rec);
4113 }
4114
4115 /*
4116  * This function only does inserts on an allocation b-tree. For tree
4117  * depth = 0, ocfs2_insert_at_leaf() is called directly.
4118  *
4119  * right_path is the path we want to do the actual insert
4120  * in. left_path should only be passed in if we need to update that
4121  * portion of the tree after an edge insert.
4122  */
4123 static int ocfs2_insert_path(handle_t *handle,
4124                              struct ocfs2_extent_tree *et,
4125                              struct ocfs2_path *left_path,
4126                              struct ocfs2_path *right_path,
4127                              struct ocfs2_extent_rec *insert_rec,
4128                              struct ocfs2_insert_type *insert)
4129 {
4130         int ret, subtree_index;
4131         struct buffer_head *leaf_bh = path_leaf_bh(right_path);
4132
4133         if (left_path) {
4134                 /*
4135                  * There's a chance that left_path got passed back to
4136                  * us without being accounted for in the
4137                  * journal. Extend our transaction here to be sure we
4138                  * can change those blocks.
4139                  */
4140                 ret = ocfs2_extend_trans(handle, left_path->p_tree_depth);
4141                 if (ret < 0) {
4142                         mlog_errno(ret);
4143                         goto out;
4144                 }
4145
4146                 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
4147                 if (ret < 0) {
4148                         mlog_errno(ret);
4149                         goto out;
4150                 }
4151         }
4152
4153         /*
4154          * Pass both paths to the journal. The majority of inserts
4155          * will be touching all components anyway.
4156          */
4157         ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4158         if (ret < 0) {
4159                 mlog_errno(ret);
4160                 goto out;
4161         }
4162
4163         if (insert->ins_split != SPLIT_NONE) {
4164                 /*
4165                  * We could call ocfs2_insert_at_leaf() for some types
4166                  * of splits, but it's easier to just let one separate
4167                  * function sort it all out.
4168                  */
4169                 ocfs2_split_record(et, left_path, right_path,
4170                                    insert_rec, insert->ins_split);
4171
4172                 /*
4173                  * Split might have modified either leaf and we don't
4174                  * have a guarantee that the later edge insert will
4175                  * dirty this for us.
4176                  */
4177                 if (left_path)
4178                         ocfs2_journal_dirty(handle,
4179                                             path_leaf_bh(left_path));
4180         } else
4181                 ocfs2_insert_at_leaf(et, insert_rec, path_leaf_el(right_path),
4182                                      insert);
4183
4184         ocfs2_journal_dirty(handle, leaf_bh);
4185
4186         if (left_path) {
4187                 /*
4188                  * The rotate code has indicated that we need to fix
4189                  * up portions of the tree after the insert.
4190                  *
4191                  * XXX: Should we extend the transaction here?
4192                  */
4193                 subtree_index = ocfs2_find_subtree_root(et, left_path,
4194                                                         right_path);
4195                 ocfs2_complete_edge_insert(handle, left_path, right_path,
4196                                            subtree_index);
4197         }
4198
4199         ret = 0;
4200 out:
4201         return ret;
4202 }
4203
4204 static int ocfs2_do_insert_extent(handle_t *handle,
4205                                   struct ocfs2_extent_tree *et,
4206                                   struct ocfs2_extent_rec *insert_rec,
4207                                   struct ocfs2_insert_type *type)
4208 {
4209         int ret, rotate = 0;
4210         u32 cpos;
4211         struct ocfs2_path *right_path = NULL;
4212         struct ocfs2_path *left_path = NULL;
4213         struct ocfs2_extent_list *el;
4214
4215         el = et->et_root_el;
4216
4217         ret = ocfs2_et_root_journal_access(handle, et,
4218                                            OCFS2_JOURNAL_ACCESS_WRITE);
4219         if (ret) {
4220                 mlog_errno(ret);
4221                 goto out;
4222         }
4223
4224         if (le16_to_cpu(el->l_tree_depth) == 0) {
4225                 ocfs2_insert_at_leaf(et, insert_rec, el, type);
4226                 goto out_update_clusters;
4227         }
4228
4229         right_path = ocfs2_new_path_from_et(et);
4230         if (!right_path) {
4231                 ret = -ENOMEM;
4232                 mlog_errno(ret);
4233                 goto out;
4234         }
4235
4236         /*
4237          * Determine the path to start with. Rotations need the
4238          * rightmost path, everything else can go directly to the
4239          * target leaf.
4240          */
4241         cpos = le32_to_cpu(insert_rec->e_cpos);
4242         if (type->ins_appending == APPEND_NONE &&
4243             type->ins_contig == CONTIG_NONE) {
4244                 rotate = 1;
4245                 cpos = UINT_MAX;
4246         }
4247
4248         ret = ocfs2_find_path(et->et_ci, right_path, cpos);
4249         if (ret) {
4250                 mlog_errno(ret);
4251                 goto out;
4252         }
4253
4254         /*
4255          * Rotations and appends need special treatment - they modify
4256          * parts of the tree's above them.
4257          *
4258          * Both might pass back a path immediate to the left of the
4259          * one being inserted to. This will be cause
4260          * ocfs2_insert_path() to modify the rightmost records of
4261          * left_path to account for an edge insert.
4262          *
4263          * XXX: When modifying this code, keep in mind that an insert
4264          * can wind up skipping both of these two special cases...
4265          */
4266         if (rotate) {
4267                 ret = ocfs2_rotate_tree_right(handle, et, type->ins_split,
4268                                               le32_to_cpu(insert_rec->e_cpos),
4269                                               right_path, &left_path);
4270                 if (ret) {
4271                         mlog_errno(ret);
4272                         goto out;
4273                 }
4274
4275                 /*
4276                  * ocfs2_rotate_tree_right() might have extended the
4277                  * transaction without re-journaling our tree root.
4278                  */
4279                 ret = ocfs2_et_root_journal_access(handle, et,
4280                                                    OCFS2_JOURNAL_ACCESS_WRITE);
4281                 if (ret) {
4282                         mlog_errno(ret);
4283                         goto out;
4284                 }
4285         } else if (type->ins_appending == APPEND_TAIL
4286                    && type->ins_contig != CONTIG_LEFT) {
4287                 ret = ocfs2_append_rec_to_path(handle, et, insert_rec,
4288                                                right_path, &left_path);
4289                 if (ret) {
4290                         mlog_errno(ret);
4291                         goto out;
4292                 }
4293         }
4294
4295         ret = ocfs2_insert_path(handle, et, left_path, right_path,
4296                                 insert_rec, type);
4297         if (ret) {
4298                 mlog_errno(ret);
4299                 goto out;
4300         }
4301
4302 out_update_clusters:
4303         if (type->ins_split == SPLIT_NONE)
4304                 ocfs2_et_update_clusters(et,
4305                                          le16_to_cpu(insert_rec->e_leaf_clusters));
4306
4307         ocfs2_journal_dirty(handle, et->et_root_bh);
4308
4309 out:
4310         ocfs2_free_path(left_path);
4311         ocfs2_free_path(right_path);
4312
4313         return ret;
4314 }
4315
4316 static enum ocfs2_contig_type
4317 ocfs2_figure_merge_contig_type(struct ocfs2_extent_tree *et,
4318                                struct ocfs2_path *path,
4319                                struct ocfs2_extent_list *el, int index,
4320                                struct ocfs2_extent_rec *split_rec)
4321 {
4322         int status;
4323         enum ocfs2_contig_type ret = CONTIG_NONE;
4324         u32 left_cpos, right_cpos;
4325         struct ocfs2_extent_rec *rec = NULL;
4326         struct ocfs2_extent_list *new_el;
4327         struct ocfs2_path *left_path = NULL, *right_path = NULL;
4328         struct buffer_head *bh;
4329         struct ocfs2_extent_block *eb;
4330         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
4331
4332         if (index > 0) {
4333                 rec = &el->l_recs[index - 1];
4334         } else if (path->p_tree_depth > 0) {
4335                 status = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
4336                 if (status)
4337                         goto out;
4338
4339                 if (left_cpos != 0) {
4340                         left_path = ocfs2_new_path_from_path(path);
4341                         if (!left_path)
4342                                 goto out;
4343
4344                         status = ocfs2_find_path(et->et_ci, left_path,
4345                                                  left_cpos);
4346                         if (status)
4347                                 goto out;
4348
4349                         new_el = path_leaf_el(left_path);
4350
4351                         if (le16_to_cpu(new_el->l_next_free_rec) !=
4352                             le16_to_cpu(new_el->l_count)) {
4353                                 bh = path_leaf_bh(left_path);
4354                                 eb = (struct ocfs2_extent_block *)bh->b_data;
4355                                 ocfs2_error(sb,
4356                                             "Extent block #%llu has an "
4357                                             "invalid l_next_free_rec of "
4358                                             "%d.  It should have "
4359                                             "matched the l_count of %d",
4360                                             (unsigned long long)le64_to_cpu(eb->h_blkno),
4361                                             le16_to_cpu(new_el->l_next_free_rec),
4362                                             le16_to_cpu(new_el->l_count));
4363                                 status = -EINVAL;
4364                                 goto out;
4365                         }
4366                         rec = &new_el->l_recs[
4367                                 le16_to_cpu(new_el->l_next_free_rec) - 1];
4368                 }
4369         }
4370
4371         /*
4372          * We're careful to check for an empty extent record here -
4373          * the merge code will know what to do if it sees one.
4374          */
4375         if (rec) {
4376                 if (index == 1 && ocfs2_is_empty_extent(rec)) {
4377                         if (split_rec->e_cpos == el->l_recs[index].e_cpos)
4378                                 ret = CONTIG_RIGHT;
4379                 } else {
4380                         ret = ocfs2_et_extent_contig(et, rec, split_rec);
4381                 }
4382         }
4383
4384         rec = NULL;
4385         if (index < (le16_to_cpu(el->l_next_free_rec) - 1))
4386                 rec = &el->l_recs[index + 1];
4387         else if (le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count) &&
4388                  path->p_tree_depth > 0) {
4389                 status = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
4390                 if (status)
4391                         goto out;
4392
4393                 if (right_cpos == 0)
4394                         goto out;
4395
4396                 right_path = ocfs2_new_path_from_path(path);
4397                 if (!right_path)
4398                         goto out;
4399
4400                 status = ocfs2_find_path(et->et_ci, right_path, right_cpos);
4401                 if (status)
4402                         goto out;
4403
4404                 new_el = path_leaf_el(right_path);
4405                 rec = &new_el->l_recs[0];
4406                 if (ocfs2_is_empty_extent(rec)) {
4407                         if (le16_to_cpu(new_el->l_next_free_rec) <= 1) {
4408                                 bh = path_leaf_bh(right_path);
4409                                 eb = (struct ocfs2_extent_block *)bh->b_data;
4410                                 ocfs2_error(sb,
4411                                             "Extent block #%llu has an "
4412                                             "invalid l_next_free_rec of %d",
4413                                             (unsigned long long)le64_to_cpu(eb->h_blkno),
4414                                             le16_to_cpu(new_el->l_next_free_rec));
4415                                 status = -EINVAL;
4416                                 goto out;
4417                         }
4418                         rec = &new_el->l_recs[1];
4419                 }
4420         }
4421
4422         if (rec) {
4423                 enum ocfs2_contig_type contig_type;
4424
4425                 contig_type = ocfs2_et_extent_contig(et, rec, split_rec);
4426
4427                 if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
4428                         ret = CONTIG_LEFTRIGHT;
4429                 else if (ret == CONTIG_NONE)
4430                         ret = contig_type;
4431         }
4432
4433 out:
4434         if (left_path)
4435                 ocfs2_free_path(left_path);
4436         if (right_path)
4437                 ocfs2_free_path(right_path);
4438
4439         return ret;
4440 }
4441
4442 static void ocfs2_figure_contig_type(struct ocfs2_extent_tree *et,
4443                                      struct ocfs2_insert_type *insert,
4444                                      struct ocfs2_extent_list *el,
4445                                      struct ocfs2_extent_rec *insert_rec)
4446 {
4447         int i;
4448         enum ocfs2_contig_type contig_type = CONTIG_NONE;
4449
4450         BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4451
4452         for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
4453                 contig_type = ocfs2_et_extent_contig(et, &el->l_recs[i],
4454                                                      insert_rec);
4455                 if (contig_type != CONTIG_NONE) {
4456                         insert->ins_contig_index = i;
4457                         break;
4458                 }
4459         }
4460         insert->ins_contig = contig_type;
4461
4462         if (insert->ins_contig != CONTIG_NONE) {
4463                 struct ocfs2_extent_rec *rec =
4464                                 &el->l_recs[insert->ins_contig_index];
4465                 unsigned int len = le16_to_cpu(rec->e_leaf_clusters) +
4466                                    le16_to_cpu(insert_rec->e_leaf_clusters);
4467
4468                 /*
4469                  * Caller might want us to limit the size of extents, don't
4470                  * calculate contiguousness if we might exceed that limit.
4471                  */
4472                 if (et->et_max_leaf_clusters &&
4473                     (len > et->et_max_leaf_clusters))
4474                         insert->ins_contig = CONTIG_NONE;
4475         }
4476 }
4477
4478 /*
4479  * This should only be called against the righmost leaf extent list.
4480  *
4481  * ocfs2_figure_appending_type() will figure out whether we'll have to
4482  * insert at the tail of the rightmost leaf.
4483  *
4484  * This should also work against the root extent list for tree's with 0
4485  * depth. If we consider the root extent list to be the rightmost leaf node
4486  * then the logic here makes sense.
4487  */
4488 static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
4489                                         struct ocfs2_extent_list *el,
4490                                         struct ocfs2_extent_rec *insert_rec)
4491 {
4492         int i;
4493         u32 cpos = le32_to_cpu(insert_rec->e_cpos);
4494         struct ocfs2_extent_rec *rec;
4495
4496         insert->ins_appending = APPEND_NONE;
4497
4498         BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4499
4500         if (!el->l_next_free_rec)
4501                 goto set_tail_append;
4502
4503         if (ocfs2_is_empty_extent(&el->l_recs[0])) {
4504                 /* Were all records empty? */
4505                 if (le16_to_cpu(el->l_next_free_rec) == 1)
4506                         goto set_tail_append;
4507         }
4508
4509         i = le16_to_cpu(el->l_next_free_rec) - 1;
4510         rec = &el->l_recs[i];
4511
4512         if (cpos >=
4513             (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
4514                 goto set_tail_append;
4515
4516         return;
4517
4518 set_tail_append:
4519         insert->ins_appending = APPEND_TAIL;
4520 }
4521
4522 /*
4523  * Helper function called at the beginning of an insert.
4524  *
4525  * This computes a few things that are commonly used in the process of
4526  * inserting into the btree:
4527  *   - Whether the new extent is contiguous with an existing one.
4528  *   - The current tree depth.
4529  *   - Whether the insert is an appending one.
4530  *   - The total # of free records in the tree.
4531  *
4532  * All of the information is stored on the ocfs2_insert_type
4533  * structure.
4534  */
4535 static int ocfs2_figure_insert_type(struct ocfs2_extent_tree *et,
4536                                     struct buffer_head **last_eb_bh,
4537                                     struct ocfs2_extent_rec *insert_rec,
4538                                     int *free_records,
4539                                     struct ocfs2_insert_type *insert)
4540 {
4541         int ret;
4542         struct ocfs2_extent_block *eb;
4543         struct ocfs2_extent_list *el;
4544         struct ocfs2_path *path = NULL;
4545         struct buffer_head *bh = NULL;
4546
4547         insert->ins_split = SPLIT_NONE;
4548
4549         el = et->et_root_el;
4550         insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
4551
4552         if (el->l_tree_depth) {
4553                 /*
4554                  * If we have tree depth, we read in the
4555                  * rightmost extent block ahead of time as
4556                  * ocfs2_figure_insert_type() and ocfs2_add_branch()
4557                  * may want it later.
4558                  */
4559                 ret = ocfs2_read_extent_block(et->et_ci,
4560                                               ocfs2_et_get_last_eb_blk(et),
4561                                               &bh);
4562                 if (ret) {
4563                         mlog_errno(ret);
4564                         goto out;
4565                 }
4566                 eb = (struct ocfs2_extent_block *) bh->b_data;
4567                 el = &eb->h_list;
4568         }
4569
4570         /*
4571          * Unless we have a contiguous insert, we'll need to know if
4572          * there is room left in our allocation tree for another
4573          * extent record.
4574          *
4575          * XXX: This test is simplistic, we can search for empty
4576          * extent records too.
4577          */
4578         *free_records = le16_to_cpu(el->l_count) -
4579                 le16_to_cpu(el->l_next_free_rec);
4580
4581         if (!insert->ins_tree_depth) {
4582                 ocfs2_figure_contig_type(et, insert, el, insert_rec);
4583                 ocfs2_figure_appending_type(insert, el, insert_rec);
4584                 return 0;
4585         }
4586
4587         path = ocfs2_new_path_from_et(et);
4588         if (!path) {
4589                 ret = -ENOMEM;
4590                 mlog_errno(ret);
4591                 goto out;
4592         }
4593
4594         /*
4595          * In the case that we're inserting past what the tree
4596          * currently accounts for, ocfs2_find_path() will return for
4597          * us the rightmost tree path. This is accounted for below in
4598          * the appending code.
4599          */
4600         ret = ocfs2_find_path(et->et_ci, path, le32_to_cpu(insert_rec->e_cpos));
4601         if (ret) {
4602                 mlog_errno(ret);
4603                 goto out;
4604         }
4605
4606         el = path_leaf_el(path);
4607
4608         /*
4609          * Now that we have the path, there's two things we want to determine:
4610          * 1) Contiguousness (also set contig_index if this is so)
4611          *
4612          * 2) Are we doing an append? We can trivially break this up
4613          *     into two types of appends: simple record append, or a
4614          *     rotate inside the tail leaf.
4615          */
4616         ocfs2_figure_contig_type(et, insert, el, insert_rec);
4617
4618         /*
4619          * The insert code isn't quite ready to deal with all cases of
4620          * left contiguousness. Specifically, if it's an insert into
4621          * the 1st record in a leaf, it will require the adjustment of
4622          * cluster count on the last record of the path directly to it's
4623          * left. For now, just catch that case and fool the layers
4624          * above us. This works just fine for tree_depth == 0, which
4625          * is why we allow that above.
4626          */
4627         if (insert->ins_contig == CONTIG_LEFT &&
4628             insert->ins_contig_index == 0)
4629                 insert->ins_contig = CONTIG_NONE;
4630
4631         /*
4632          * Ok, so we can simply compare against last_eb to figure out
4633          * whether the path doesn't exist. This will only happen in
4634          * the case that we're doing a tail append, so maybe we can
4635          * take advantage of that information somehow.
4636          */
4637         if (ocfs2_et_get_last_eb_blk(et) ==
4638             path_leaf_bh(path)->b_blocknr) {
4639                 /*
4640                  * Ok, ocfs2_find_path() returned us the rightmost
4641                  * tree path. This might be an appending insert. There are
4642                  * two cases:
4643                  *    1) We're doing a true append at the tail:
4644                  *      -This might even be off the end of the leaf
4645                  *    2) We're "appending" by rotating in the tail
4646                  */
4647                 ocfs2_figure_appending_type(insert, el, insert_rec);
4648         }
4649
4650 out:
4651         ocfs2_free_path(path);
4652
4653         if (ret == 0)
4654                 *last_eb_bh = bh;
4655         else
4656                 brelse(bh);
4657         return ret;
4658 }
4659
4660 /*
4661  * Insert an extent into a btree.
4662  *
4663  * The caller needs to update the owning btree's cluster count.
4664  */
4665 int ocfs2_insert_extent(handle_t *handle,
4666                         struct ocfs2_extent_tree *et,
4667                         u32 cpos,
4668                         u64 start_blk,
4669                         u32 new_clusters,
4670                         u8 flags,
4671                         struct ocfs2_alloc_context *meta_ac)
4672 {
4673         int status;
4674         int uninitialized_var(free_records);
4675         struct buffer_head *last_eb_bh = NULL;
4676         struct ocfs2_insert_type insert = {0, };
4677         struct ocfs2_extent_rec rec;
4678
4679         trace_ocfs2_insert_extent_start(
4680                 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4681                 cpos, new_clusters);
4682
4683         memset(&rec, 0, sizeof(rec));
4684         rec.e_cpos = cpu_to_le32(cpos);
4685         rec.e_blkno = cpu_to_le64(start_blk);
4686         rec.e_leaf_clusters = cpu_to_le16(new_clusters);
4687         rec.e_flags = flags;
4688         status = ocfs2_et_insert_check(et, &rec);
4689         if (status) {
4690                 mlog_errno(status);
4691                 goto bail;
4692         }
4693
4694         status = ocfs2_figure_insert_type(et, &last_eb_bh, &rec,
4695                                           &free_records, &insert);
4696         if (status < 0) {
4697                 mlog_errno(status);
4698                 goto bail;
4699         }
4700
4701         trace_ocfs2_insert_extent(insert.ins_appending, insert.ins_contig,
4702                                   insert.ins_contig_index, free_records,
4703                                   insert.ins_tree_depth);
4704
4705         if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
4706                 status = ocfs2_grow_tree(handle, et,
4707                                          &insert.ins_tree_depth, &last_eb_bh,
4708                                          meta_ac);
4709                 if (status) {
4710                         mlog_errno(status);
4711                         goto bail;
4712                 }
4713         }
4714
4715         /* Finally, we can add clusters. This might rotate the tree for us. */
4716         status = ocfs2_do_insert_extent(handle, et, &rec, &insert);
4717         if (status < 0)
4718                 mlog_errno(status);
4719         else
4720                 ocfs2_et_extent_map_insert(et, &rec);
4721
4722 bail:
4723         brelse(last_eb_bh);
4724
4725         return status;
4726 }
4727
4728 /*
4729  * Allcate and add clusters into the extent b-tree.
4730  * The new clusters(clusters_to_add) will be inserted at logical_offset.
4731  * The extent b-tree's root is specified by et, and
4732  * it is not limited to the file storage. Any extent tree can use this
4733  * function if it implements the proper ocfs2_extent_tree.
4734  */
4735 int ocfs2_add_clusters_in_btree(handle_t *handle,
4736                                 struct ocfs2_extent_tree *et,
4737                                 u32 *logical_offset,
4738                                 u32 clusters_to_add,
4739                                 int mark_unwritten,
4740                                 struct ocfs2_alloc_context *data_ac,
4741                                 struct ocfs2_alloc_context *meta_ac,
4742                                 enum ocfs2_alloc_restarted *reason_ret)
4743 {
4744         int status = 0, err = 0;
4745         int free_extents;
4746         enum ocfs2_alloc_restarted reason = RESTART_NONE;
4747         u32 bit_off, num_bits;
4748         u64 block;
4749         u8 flags = 0;
4750         struct ocfs2_super *osb =
4751                 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
4752
4753         BUG_ON(!clusters_to_add);
4754
4755         if (mark_unwritten)
4756                 flags = OCFS2_EXT_UNWRITTEN;
4757
4758         free_extents = ocfs2_num_free_extents(osb, et);
4759         if (free_extents < 0) {
4760                 status = free_extents;
4761                 mlog_errno(status);
4762                 goto leave;
4763         }
4764
4765         /* there are two cases which could cause us to EAGAIN in the
4766          * we-need-more-metadata case:
4767          * 1) we haven't reserved *any*
4768          * 2) we are so fragmented, we've needed to add metadata too
4769          *    many times. */
4770         if (!free_extents && !meta_ac) {
4771                 err = -1;
4772                 status = -EAGAIN;
4773                 reason = RESTART_META;
4774                 goto leave;
4775         } else if ((!free_extents)
4776                    && (ocfs2_alloc_context_bits_left(meta_ac)
4777                        < ocfs2_extend_meta_needed(et->et_root_el))) {
4778                 err = -2;
4779                 status = -EAGAIN;
4780                 reason = RESTART_META;
4781                 goto leave;
4782         }
4783
4784         status = __ocfs2_claim_clusters(handle, data_ac, 1,
4785                                         clusters_to_add, &bit_off, &num_bits);
4786         if (status < 0) {
4787                 if (status != -ENOSPC)
4788                         mlog_errno(status);
4789                 goto leave;
4790         }
4791
4792         BUG_ON(num_bits > clusters_to_add);
4793
4794         /* reserve our write early -- insert_extent may update the tree root */
4795         status = ocfs2_et_root_journal_access(handle, et,
4796                                               OCFS2_JOURNAL_ACCESS_WRITE);
4797         if (status < 0) {
4798                 mlog_errno(status);
4799                 goto leave;
4800         }
4801
4802         block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
4803         trace_ocfs2_add_clusters_in_btree(
4804              (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4805              bit_off, num_bits);
4806         status = ocfs2_insert_extent(handle, et, *logical_offset, block,
4807                                      num_bits, flags, meta_ac);
4808         if (status < 0) {
4809                 mlog_errno(status);
4810                 goto leave;
4811         }
4812
4813         ocfs2_journal_dirty(handle, et->et_root_bh);
4814
4815         clusters_to_add -= num_bits;
4816         *logical_offset += num_bits;
4817
4818         if (clusters_to_add) {
4819                 err = clusters_to_add;
4820                 status = -EAGAIN;
4821                 reason = RESTART_TRANS;
4822         }
4823
4824 leave:
4825         if (reason_ret)
4826                 *reason_ret = reason;
4827         trace_ocfs2_add_clusters_in_btree_ret(status, reason, err);
4828         return status;
4829 }
4830
4831 static void ocfs2_make_right_split_rec(struct super_block *sb,
4832                                        struct ocfs2_extent_rec *split_rec,
4833                                        u32 cpos,
4834                                        struct ocfs2_extent_rec *rec)
4835 {
4836         u32 rec_cpos = le32_to_cpu(rec->e_cpos);
4837         u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
4838
4839         memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
4840
4841         split_rec->e_cpos = cpu_to_le32(cpos);
4842         split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
4843
4844         split_rec->e_blkno = rec->e_blkno;
4845         le64_add_cpu(&split_rec->e_blkno,
4846                      ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
4847
4848         split_rec->e_flags = rec->e_flags;
4849 }
4850
4851 static int ocfs2_split_and_insert(handle_t *handle,
4852                                   struct ocfs2_extent_tree *et,
4853                                   struct ocfs2_path *path,
4854                                   struct buffer_head **last_eb_bh,
4855                                   int split_index,
4856                                   struct ocfs2_extent_rec *orig_split_rec,
4857                                   struct ocfs2_alloc_context *meta_ac)
4858 {
4859         int ret = 0, depth;
4860         unsigned int insert_range, rec_range, do_leftright = 0;
4861         struct ocfs2_extent_rec tmprec;
4862         struct ocfs2_extent_list *rightmost_el;
4863         struct ocfs2_extent_rec rec;
4864         struct ocfs2_extent_rec split_rec = *orig_split_rec;
4865         struct ocfs2_insert_type insert;
4866         struct ocfs2_extent_block *eb;
4867
4868 leftright:
4869         /*
4870          * Store a copy of the record on the stack - it might move
4871          * around as the tree is manipulated below.
4872          */
4873         rec = path_leaf_el(path)->l_recs[split_index];
4874
4875         rightmost_el = et->et_root_el;
4876
4877         depth = le16_to_cpu(rightmost_el->l_tree_depth);
4878         if (depth) {
4879                 BUG_ON(!(*last_eb_bh));
4880                 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
4881                 rightmost_el = &eb->h_list;
4882         }
4883
4884         if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4885             le16_to_cpu(rightmost_el->l_count)) {
4886                 ret = ocfs2_grow_tree(handle, et,
4887                                       &depth, last_eb_bh, meta_ac);
4888                 if (ret) {
4889                         mlog_errno(ret);
4890                         goto out;
4891                 }
4892         }
4893
4894         memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4895         insert.ins_appending = APPEND_NONE;
4896         insert.ins_contig = CONTIG_NONE;
4897         insert.ins_tree_depth = depth;
4898
4899         insert_range = le32_to_cpu(split_rec.e_cpos) +
4900                 le16_to_cpu(split_rec.e_leaf_clusters);
4901         rec_range = le32_to_cpu(rec.e_cpos) +
4902                 le16_to_cpu(rec.e_leaf_clusters);
4903
4904         if (split_rec.e_cpos == rec.e_cpos) {
4905                 insert.ins_split = SPLIT_LEFT;
4906         } else if (insert_range == rec_range) {
4907                 insert.ins_split = SPLIT_RIGHT;
4908         } else {
4909                 /*
4910                  * Left/right split. We fake this as a right split
4911                  * first and then make a second pass as a left split.
4912                  */
4913                 insert.ins_split = SPLIT_RIGHT;
4914
4915                 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4916                                            &tmprec, insert_range, &rec);
4917
4918                 split_rec = tmprec;
4919
4920                 BUG_ON(do_leftright);
4921                 do_leftright = 1;
4922         }
4923
4924         ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
4925         if (ret) {
4926                 mlog_errno(ret);
4927                 goto out;
4928         }
4929
4930         if (do_leftright == 1) {
4931                 u32 cpos;
4932                 struct ocfs2_extent_list *el;
4933
4934                 do_leftright++;
4935                 split_rec = *orig_split_rec;
4936
4937                 ocfs2_reinit_path(path, 1);
4938
4939                 cpos = le32_to_cpu(split_rec.e_cpos);
4940                 ret = ocfs2_find_path(et->et_ci, path, cpos);
4941                 if (ret) {
4942                         mlog_errno(ret);
4943                         goto out;
4944                 }
4945
4946                 el = path_leaf_el(path);
4947                 split_index = ocfs2_search_extent_list(el, cpos);
4948                 goto leftright;
4949         }
4950 out:
4951
4952         return ret;
4953 }
4954
4955 static int ocfs2_replace_extent_rec(handle_t *handle,
4956                                     struct ocfs2_extent_tree *et,
4957                                     struct ocfs2_path *path,
4958                                     struct ocfs2_extent_list *el,
4959                                     int split_index,
4960                                     struct ocfs2_extent_rec *split_rec)
4961 {
4962         int ret;
4963
4964         ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
4965                                            path_num_items(path) - 1);
4966         if (ret) {
4967                 mlog_errno(ret);
4968                 goto out;
4969         }
4970
4971         el->l_recs[split_index] = *split_rec;
4972
4973         ocfs2_journal_dirty(handle, path_leaf_bh(path));
4974 out:
4975         return ret;
4976 }
4977
4978 /*
4979  * Split part or all of the extent record at split_index in the leaf
4980  * pointed to by path. Merge with the contiguous extent record if needed.
4981  *
4982  * Care is taken to handle contiguousness so as to not grow the tree.
4983  *
4984  * meta_ac is not strictly necessary - we only truly need it if growth
4985  * of the tree is required. All other cases will degrade into a less
4986  * optimal tree layout.
4987  *
4988  * last_eb_bh should be the rightmost leaf block for any extent
4989  * btree. Since a split may grow the tree or a merge might shrink it,
4990  * the caller cannot trust the contents of that buffer after this call.
4991  *
4992  * This code is optimized for readability - several passes might be
4993  * made over certain portions of the tree. All of those blocks will
4994  * have been brought into cache (and pinned via the journal), so the
4995  * extra overhead is not expressed in terms of disk reads.
4996  */
4997 int ocfs2_split_extent(handle_t *handle,
4998                        struct ocfs2_extent_tree *et,
4999                        struct ocfs2_path *path,
5000                        int split_index,
5001                        struct ocfs2_extent_rec *split_rec,
5002                        struct ocfs2_alloc_context *meta_ac,
5003                        struct ocfs2_cached_dealloc_ctxt *dealloc)
5004 {
5005         int ret = 0;
5006         struct ocfs2_extent_list *el = path_leaf_el(path);
5007         struct buffer_head *last_eb_bh = NULL;
5008         struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
5009         struct ocfs2_merge_ctxt ctxt;
5010         struct ocfs2_extent_list *rightmost_el;
5011
5012         if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) ||
5013             ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) <
5014              (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) {
5015                 ret = -EIO;
5016                 mlog_errno(ret);
5017                 goto out;
5018         }
5019
5020         ctxt.c_contig_type = ocfs2_figure_merge_contig_type(et, path, el,
5021                                                             split_index,
5022                                                             split_rec);
5023
5024         /*
5025          * The core merge / split code wants to know how much room is
5026          * left in this allocation tree, so we pass the
5027          * rightmost extent list.
5028          */
5029         if (path->p_tree_depth) {
5030                 struct ocfs2_extent_block *eb;
5031
5032                 ret = ocfs2_read_extent_block(et->et_ci,
5033                                               ocfs2_et_get_last_eb_blk(et),
5034                                               &last_eb_bh);
5035                 if (ret) {
5036                         mlog_errno(ret);
5037                         goto out;
5038                 }
5039
5040                 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5041                 rightmost_el = &eb->h_list;
5042         } else
5043                 rightmost_el = path_root_el(path);
5044
5045         if (rec->e_cpos == split_rec->e_cpos &&
5046             rec->e_leaf_clusters == split_rec->e_leaf_clusters)
5047                 ctxt.c_split_covers_rec = 1;
5048         else
5049                 ctxt.c_split_covers_rec = 0;
5050
5051         ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
5052
5053         trace_ocfs2_split_extent(split_index, ctxt.c_contig_type,
5054                                  ctxt.c_has_empty_extent,
5055                                  ctxt.c_split_covers_rec);
5056
5057         if (ctxt.c_contig_type == CONTIG_NONE) {
5058                 if (ctxt.c_split_covers_rec)
5059                         ret = ocfs2_replace_extent_rec(handle, et, path, el,
5060                                                        split_index, split_rec);
5061                 else
5062                         ret = ocfs2_split_and_insert(handle, et, path,
5063                                                      &last_eb_bh, split_index,
5064                                                      split_rec, meta_ac);
5065                 if (ret)
5066                         mlog_errno(ret);
5067         } else {
5068                 ret = ocfs2_try_to_merge_extent(handle, et, path,
5069                                                 split_index, split_rec,
5070                                                 dealloc, &ctxt);
5071                 if (ret)
5072                         mlog_errno(ret);
5073         }
5074
5075 out:
5076         brelse(last_eb_bh);
5077         return ret;
5078 }
5079
5080 /*
5081  * Change the flags of the already-existing extent at cpos for len clusters.
5082  *
5083  * new_flags: the flags we want to set.
5084  * clear_flags: the flags we want to clear.
5085  * phys: the new physical offset we want this new extent starts from.
5086  *
5087  * If the existing extent is larger than the request, initiate a
5088  * split. An attempt will be made at merging with adjacent extents.
5089  *
5090  * The caller is responsible for passing down meta_ac if we'll need it.
5091  */
5092 int ocfs2_change_extent_flag(handle_t *handle,
5093                              struct ocfs2_extent_tree *et,
5094                              u32 cpos, u32 len, u32 phys,
5095                              struct ocfs2_alloc_context *meta_ac,
5096                              struct ocfs2_cached_dealloc_ctxt *dealloc,
5097                              int new_flags, int clear_flags)
5098 {
5099         int ret, index;
5100         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5101         u64 start_blkno = ocfs2_clusters_to_blocks(sb, phys);
5102         struct ocfs2_extent_rec split_rec;
5103         struct ocfs2_path *left_path = NULL;
5104         struct ocfs2_extent_list *el;
5105         struct ocfs2_extent_rec *rec;
5106
5107         left_path = ocfs2_new_path_from_et(et);
5108         if (!left_path) {
5109                 ret = -ENOMEM;
5110                 mlog_errno(ret);
5111                 goto out;
5112         }
5113
5114         ret = ocfs2_find_path(et->et_ci, left_path, cpos);
5115         if (ret) {
5116                 mlog_errno(ret);
5117                 goto out;
5118         }
5119         el = path_leaf_el(left_path);
5120
5121         index = ocfs2_search_extent_list(el, cpos);
5122         if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
5123                 ocfs2_error(sb,
5124                             "Owner %llu has an extent at cpos %u which can no "
5125                             "longer be found.\n",
5126                              (unsigned long long)
5127                              ocfs2_metadata_cache_owner(et->et_ci), cpos);
5128                 ret = -EROFS;
5129                 goto out;
5130         }
5131
5132         ret = -EIO;
5133         rec = &el->l_recs[index];
5134         if (new_flags && (rec->e_flags & new_flags)) {
5135                 mlog(ML_ERROR, "Owner %llu tried to set %d flags on an "
5136                      "extent that already had them",
5137                      (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5138                      new_flags);
5139                 goto out;
5140         }
5141
5142         if (clear_flags && !(rec->e_flags & clear_flags)) {
5143                 mlog(ML_ERROR, "Owner %llu tried to clear %d flags on an "
5144                      "extent that didn't have them",
5145                      (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5146                      clear_flags);
5147                 goto out;
5148         }
5149
5150         memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
5151         split_rec.e_cpos = cpu_to_le32(cpos);
5152         split_rec.e_leaf_clusters = cpu_to_le16(len);
5153         split_rec.e_blkno = cpu_to_le64(start_blkno);
5154         split_rec.e_flags = rec->e_flags;
5155         if (new_flags)
5156                 split_rec.e_flags |= new_flags;
5157         if (clear_flags)
5158                 split_rec.e_flags &= ~clear_flags;
5159
5160         ret = ocfs2_split_extent(handle, et, left_path,
5161                                  index, &split_rec, meta_ac,
5162                                  dealloc);
5163         if (ret)
5164                 mlog_errno(ret);
5165
5166 out:
5167         ocfs2_free_path(left_path);
5168         return ret;
5169
5170 }
5171
5172 /*
5173  * Mark the already-existing extent at cpos as written for len clusters.
5174  * This removes the unwritten extent flag.
5175  *
5176  * If the existing extent is larger than the request, initiate a
5177  * split. An attempt will be made at merging with adjacent extents.
5178  *
5179  * The caller is responsible for passing down meta_ac if we'll need it.
5180  */
5181 int ocfs2_mark_extent_written(struct inode *inode,
5182                               struct ocfs2_extent_tree *et,
5183                               handle_t *handle, u32 cpos, u32 len, u32 phys,
5184                               struct ocfs2_alloc_context *meta_ac,
5185                               struct ocfs2_cached_dealloc_ctxt *dealloc)
5186 {
5187         int ret;
5188
5189         trace_ocfs2_mark_extent_written(
5190                 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5191                 cpos, len, phys);
5192
5193         if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
5194                 ocfs2_error(inode->i_sb, "Inode %llu has unwritten extents "
5195                             "that are being written to, but the feature bit "
5196                             "is not set in the super block.",
5197                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
5198                 ret = -EROFS;
5199                 goto out;
5200         }
5201
5202         /*
5203          * XXX: This should be fixed up so that we just re-insert the
5204          * next extent records.
5205          */
5206         ocfs2_et_extent_map_truncate(et, 0);
5207
5208         ret = ocfs2_change_extent_flag(handle, et, cpos,
5209                                        len, phys, meta_ac, dealloc,
5210                                        0, OCFS2_EXT_UNWRITTEN);
5211         if (ret)
5212                 mlog_errno(ret);
5213
5214 out:
5215         return ret;
5216 }
5217
5218 static int ocfs2_split_tree(handle_t *handle, struct ocfs2_extent_tree *et,
5219                             struct ocfs2_path *path,
5220                             int index, u32 new_range,
5221                             struct ocfs2_alloc_context *meta_ac)
5222 {
5223         int ret, depth, credits;
5224         struct buffer_head *last_eb_bh = NULL;
5225         struct ocfs2_extent_block *eb;
5226         struct ocfs2_extent_list *rightmost_el, *el;
5227         struct ocfs2_extent_rec split_rec;
5228         struct ocfs2_extent_rec *rec;
5229         struct ocfs2_insert_type insert;
5230
5231         /*
5232          * Setup the record to split before we grow the tree.
5233          */
5234         el = path_leaf_el(path);
5235         rec = &el->l_recs[index];
5236         ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
5237                                    &split_rec, new_range, rec);
5238
5239         depth = path->p_tree_depth;
5240         if (depth > 0) {
5241                 ret = ocfs2_read_extent_block(et->et_ci,
5242                                               ocfs2_et_get_last_eb_blk(et),
5243                                               &last_eb_bh);
5244                 if (ret < 0) {
5245                         mlog_errno(ret);
5246                         goto out;
5247                 }
5248
5249                 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5250                 rightmost_el = &eb->h_list;
5251         } else
5252                 rightmost_el = path_leaf_el(path);
5253
5254         credits = path->p_tree_depth +
5255                   ocfs2_extend_meta_needed(et->et_root_el);
5256         ret = ocfs2_extend_trans(handle, credits);
5257         if (ret) {
5258                 mlog_errno(ret);
5259                 goto out;
5260         }
5261
5262         if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
5263             le16_to_cpu(rightmost_el->l_count)) {
5264                 ret = ocfs2_grow_tree(handle, et, &depth, &last_eb_bh,
5265                                       meta_ac);
5266                 if (ret) {
5267                         mlog_errno(ret);
5268                         goto out;
5269                 }
5270         }
5271
5272         memset(&insert, 0, sizeof(struct ocfs2_insert_type));
5273         insert.ins_appending = APPEND_NONE;
5274         insert.ins_contig = CONTIG_NONE;
5275         insert.ins_split = SPLIT_RIGHT;
5276         insert.ins_tree_depth = depth;
5277
5278         ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
5279         if (ret)
5280                 mlog_errno(ret);
5281
5282 out:
5283         brelse(last_eb_bh);
5284         return ret;
5285 }
5286
5287 static int ocfs2_truncate_rec(handle_t *handle,
5288                               struct ocfs2_extent_tree *et,
5289                               struct ocfs2_path *path, int index,
5290                               struct ocfs2_cached_dealloc_ctxt *dealloc,
5291                               u32 cpos, u32 len)
5292 {
5293         int ret;
5294         u32 left_cpos, rec_range, trunc_range;
5295         int wants_rotate = 0, is_rightmost_tree_rec = 0;
5296         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5297         struct ocfs2_path *left_path = NULL;
5298         struct ocfs2_extent_list *el = path_leaf_el(path);
5299         struct ocfs2_extent_rec *rec;
5300         struct ocfs2_extent_block *eb;
5301
5302         if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
5303                 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5304                 if (ret) {
5305                         mlog_errno(ret);
5306                         goto out;
5307                 }
5308
5309                 index--;
5310         }
5311
5312         if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
5313             path->p_tree_depth) {
5314                 /*
5315                  * Check whether this is the rightmost tree record. If
5316                  * we remove all of this record or part of its right
5317                  * edge then an update of the record lengths above it
5318                  * will be required.
5319                  */
5320                 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
5321                 if (eb->h_next_leaf_blk == 0)
5322                         is_rightmost_tree_rec = 1;
5323         }
5324
5325         rec = &el->l_recs[index];
5326         if (index == 0 && path->p_tree_depth &&
5327             le32_to_cpu(rec->e_cpos) == cpos) {
5328                 /*
5329                  * Changing the leftmost offset (via partial or whole
5330                  * record truncate) of an interior (or rightmost) path
5331                  * means we have to update the subtree that is formed
5332                  * by this leaf and the one to it's left.
5333                  *
5334                  * There are two cases we can skip:
5335                  *   1) Path is the leftmost one in our btree.
5336                  *   2) The leaf is rightmost and will be empty after
5337                  *      we remove the extent record - the rotate code
5338                  *      knows how to update the newly formed edge.
5339                  */
5340
5341                 ret = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
5342                 if (ret) {
5343                         mlog_errno(ret);
5344                         goto out;
5345                 }
5346
5347                 if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
5348                         left_path = ocfs2_new_path_from_path(path);
5349                         if (!left_path) {
5350                                 ret = -ENOMEM;
5351                                 mlog_errno(ret);
5352                                 goto out;
5353                         }
5354
5355                         ret = ocfs2_find_path(et->et_ci, left_path,
5356                                               left_cpos);
5357                         if (ret) {
5358                                 mlog_errno(ret);
5359                                 goto out;
5360                         }
5361                 }
5362         }
5363
5364         ret = ocfs2_extend_rotate_transaction(handle, 0,
5365                                               handle->h_buffer_credits,
5366                                               path);
5367         if (ret) {
5368                 mlog_errno(ret);
5369                 goto out;
5370         }
5371
5372         ret = ocfs2_journal_access_path(et->et_ci, handle, path);
5373         if (ret) {
5374                 mlog_errno(ret);
5375                 goto out;
5376         }
5377
5378         ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
5379         if (ret) {
5380                 mlog_errno(ret);
5381                 goto out;
5382         }
5383
5384         rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5385         trunc_range = cpos + len;
5386
5387         if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
5388                 int next_free;
5389
5390                 memset(rec, 0, sizeof(*rec));
5391                 ocfs2_cleanup_merge(el, index);
5392                 wants_rotate = 1;
5393
5394                 next_free = le16_to_cpu(el->l_next_free_rec);
5395                 if (is_rightmost_tree_rec && next_free > 1) {
5396                         /*
5397                          * We skip the edge update if this path will
5398                          * be deleted by the rotate code.
5399                          */
5400                         rec = &el->l_recs[next_free - 1];
5401                         ocfs2_adjust_rightmost_records(handle, et, path,
5402                                                        rec);
5403                 }
5404         } else if (le32_to_cpu(rec->e_cpos) == cpos) {
5405                 /* Remove leftmost portion of the record. */
5406                 le32_add_cpu(&rec->e_cpos, len);
5407                 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
5408                 le16_add_cpu(&rec->e_leaf_clusters, -len);
5409         } else if (rec_range == trunc_range) {
5410                 /* Remove rightmost portion of the record */
5411                 le16_add_cpu(&rec->e_leaf_clusters, -len);
5412                 if (is_rightmost_tree_rec)
5413                         ocfs2_adjust_rightmost_records(handle, et, path, rec);
5414         } else {
5415                 /* Caller should have trapped this. */
5416                 mlog(ML_ERROR, "Owner %llu: Invalid record truncate: (%u, %u) "
5417                      "(%u, %u)\n",
5418                      (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5419                      le32_to_cpu(rec->e_cpos),
5420                      le16_to_cpu(rec->e_leaf_clusters), cpos, len);
5421                 BUG();
5422         }
5423
5424         if (left_path) {
5425                 int subtree_index;
5426
5427                 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
5428                 ocfs2_complete_edge_insert(handle, left_path, path,
5429                                            subtree_index);
5430         }
5431
5432         ocfs2_journal_dirty(handle, path_leaf_bh(path));
5433
5434         ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5435         if (ret) {
5436                 mlog_errno(ret);
5437                 goto out;
5438         }
5439
5440 out:
5441         ocfs2_free_path(left_path);
5442         return ret;
5443 }
5444
5445 int ocfs2_remove_extent(handle_t *handle,
5446                         struct ocfs2_extent_tree *et,
5447                         u32 cpos, u32 len,
5448                         struct ocfs2_alloc_context *meta_ac,
5449                         struct ocfs2_cached_dealloc_ctxt *dealloc)
5450 {
5451         int ret, index;
5452         u32 rec_range, trunc_range;
5453         struct ocfs2_extent_rec *rec;
5454         struct ocfs2_extent_list *el;
5455         struct ocfs2_path *path = NULL;
5456
5457         /*
5458          * XXX: Why are we truncating to 0 instead of wherever this
5459          * affects us?
5460          */
5461         ocfs2_et_extent_map_truncate(et, 0);
5462
5463         path = ocfs2_new_path_from_et(et);
5464         if (!path) {
5465                 ret = -ENOMEM;
5466                 mlog_errno(ret);
5467                 goto out;
5468         }
5469
5470         ret = ocfs2_find_path(et->et_ci, path, cpos);
5471         if (ret) {
5472                 mlog_errno(ret);
5473                 goto out;
5474         }
5475
5476         el = path_leaf_el(path);
5477         index = ocfs2_search_extent_list(el, cpos);
5478         if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
5479                 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5480                             "Owner %llu has an extent at cpos %u which can no "
5481                             "longer be found.\n",
5482                             (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5483                             cpos);
5484                 ret = -EROFS;
5485                 goto out;
5486         }
5487
5488         /*
5489          * We have 3 cases of extent removal:
5490          *   1) Range covers the entire extent rec
5491          *   2) Range begins or ends on one edge of the extent rec
5492          *   3) Range is in the middle of the extent rec (no shared edges)
5493          *
5494          * For case 1 we remove the extent rec and left rotate to
5495          * fill the hole.
5496          *
5497          * For case 2 we just shrink the existing extent rec, with a
5498          * tree update if the shrinking edge is also the edge of an
5499          * extent block.
5500          *
5501          * For case 3 we do a right split to turn the extent rec into
5502          * something case 2 can handle.
5503          */
5504         rec = &el->l_recs[index];
5505         rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5506         trunc_range = cpos + len;
5507
5508         BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
5509
5510         trace_ocfs2_remove_extent(
5511                 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5512                 cpos, len, index, le32_to_cpu(rec->e_cpos),
5513                 ocfs2_rec_clusters(el, rec));
5514
5515         if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
5516                 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5517                                          cpos, len);
5518                 if (ret) {
5519                         mlog_errno(ret);
5520                         goto out;
5521                 }
5522         } else {
5523                 ret = ocfs2_split_tree(handle, et, path, index,
5524                                        trunc_range, meta_ac);
5525                 if (ret) {
5526                         mlog_errno(ret);
5527                         goto out;
5528                 }
5529
5530                 /*
5531                  * The split could have manipulated the tree enough to
5532                  * move the record location, so we have to look for it again.
5533                  */
5534                 ocfs2_reinit_path(path, 1);
5535
5536                 ret = ocfs2_find_path(et->et_ci, path, cpos);
5537                 if (ret) {
5538                         mlog_errno(ret);
5539                         goto out;
5540                 }
5541
5542                 el = path_leaf_el(path);
5543                 index = ocfs2_search_extent_list(el, cpos);
5544                 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
5545                         ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5546                                     "Owner %llu: split at cpos %u lost record.",
5547                                     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5548                                     cpos);
5549                         ret = -EROFS;
5550                         goto out;
5551                 }
5552
5553                 /*
5554                  * Double check our values here. If anything is fishy,
5555                  * it's easier to catch it at the top level.
5556                  */
5557                 rec = &el->l_recs[index];
5558                 rec_range = le32_to_cpu(rec->e_cpos) +
5559                         ocfs2_rec_clusters(el, rec);
5560                 if (rec_range != trunc_range) {
5561                         ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5562                                     "Owner %llu: error after split at cpos %u"
5563                                     "trunc len %u, existing record is (%u,%u)",
5564                                     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5565                                     cpos, len, le32_to_cpu(rec->e_cpos),
5566                                     ocfs2_rec_clusters(el, rec));
5567                         ret = -EROFS;
5568                         goto out;
5569                 }
5570
5571                 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5572                                          cpos, len);
5573                 if (ret) {
5574                         mlog_errno(ret);
5575                         goto out;
5576                 }
5577         }
5578
5579 out:
5580         ocfs2_free_path(path);
5581         return ret;
5582 }
5583
5584 /*
5585  * ocfs2_reserve_blocks_for_rec_trunc() would look basically the
5586  * same as ocfs2_lock_alloctors(), except for it accepts a blocks
5587  * number to reserve some extra blocks, and it only handles meta
5588  * data allocations.
5589  *
5590  * Currently, only ocfs2_remove_btree_range() uses it for truncating
5591  * and punching holes.
5592  */
5593 static int ocfs2_reserve_blocks_for_rec_trunc(struct inode *inode,
5594                                               struct ocfs2_extent_tree *et,
5595                                               u32 extents_to_split,
5596                                               struct ocfs2_alloc_context **ac,
5597                                               int extra_blocks)
5598 {
5599         int ret = 0, num_free_extents;
5600         unsigned int max_recs_needed = 2 * extents_to_split;
5601         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5602
5603         *ac = NULL;
5604
5605         num_free_extents = ocfs2_num_free_extents(osb, et);
5606         if (num_free_extents < 0) {
5607                 ret = num_free_extents;
5608                 mlog_errno(ret);
5609                 goto out;
5610         }
5611
5612         if (!num_free_extents ||
5613             (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
5614                 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
5615
5616         if (extra_blocks) {
5617                 ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, ac);
5618                 if (ret < 0) {
5619                         if (ret != -ENOSPC)
5620                                 mlog_errno(ret);
5621                         goto out;
5622                 }
5623         }
5624
5625 out:
5626         if (ret) {
5627                 if (*ac) {
5628                         ocfs2_free_alloc_context(*ac);
5629                         *ac = NULL;
5630                 }
5631         }
5632
5633         return ret;
5634 }
5635
5636 int ocfs2_remove_btree_range(struct inode *inode,
5637                              struct ocfs2_extent_tree *et,
5638                              u32 cpos, u32 phys_cpos, u32 len, int flags,
5639                              struct ocfs2_cached_dealloc_ctxt *dealloc,
5640                              u64 refcount_loc)
5641 {
5642         int ret, credits = 0, extra_blocks = 0;
5643         u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
5644         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5645         struct inode *tl_inode = osb->osb_tl_inode;
5646         handle_t *handle;
5647         struct ocfs2_alloc_context *meta_ac = NULL;
5648         struct ocfs2_refcount_tree *ref_tree = NULL;
5649
5650         if ((flags & OCFS2_EXT_REFCOUNTED) && len) {
5651                 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
5652                          OCFS2_HAS_REFCOUNT_FL));
5653
5654                 ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
5655                                                &ref_tree, NULL);
5656                 if (ret) {
5657                         mlog_errno(ret);
5658                         goto out;
5659                 }
5660
5661                 ret = ocfs2_prepare_refcount_change_for_del(inode,
5662                                                             refcount_loc,
5663                                                             phys_blkno,
5664                                                             len,
5665                                                             &credits,
5666                                                             &extra_blocks);
5667                 if (ret < 0) {
5668                         mlog_errno(ret);
5669                         goto out;
5670                 }
5671 &nbs