~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/fs/ocfs2/alloc.c

Version: ~ [ linux-5.13-rc1 ] ~ [ linux-5.12.2 ] ~ [ linux-5.11.19 ] ~ [ linux-5.10.35 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.117 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.190 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.232 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.268 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.268 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

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