GNU Linux-libre 4.9.308-gnu1
[releases.git] / fs / ocfs2 / move_extents.c
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * move_extents.c
5  *
6  * Copyright (C) 2011 Oracle.  All rights reserved.
7  *
8  * This program is free software; you can redistribute it and/or
9  * modify it under the terms of the GNU General Public
10  * License version 2 as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License for more details.
16  */
17 #include <linux/fs.h>
18 #include <linux/types.h>
19 #include <linux/mount.h>
20 #include <linux/swap.h>
21
22 #include <cluster/masklog.h>
23
24 #include "ocfs2.h"
25 #include "ocfs2_ioctl.h"
26
27 #include "alloc.h"
28 #include "localalloc.h"
29 #include "aops.h"
30 #include "dlmglue.h"
31 #include "extent_map.h"
32 #include "inode.h"
33 #include "journal.h"
34 #include "suballoc.h"
35 #include "uptodate.h"
36 #include "super.h"
37 #include "dir.h"
38 #include "buffer_head_io.h"
39 #include "sysfile.h"
40 #include "refcounttree.h"
41 #include "move_extents.h"
42
43 struct ocfs2_move_extents_context {
44         struct inode *inode;
45         struct file *file;
46         int auto_defrag;
47         int partial;
48         int credits;
49         u32 new_phys_cpos;
50         u32 clusters_moved;
51         u64 refcount_loc;
52         struct ocfs2_move_extents *range;
53         struct ocfs2_extent_tree et;
54         struct ocfs2_alloc_context *meta_ac;
55         struct ocfs2_alloc_context *data_ac;
56         struct ocfs2_cached_dealloc_ctxt dealloc;
57 };
58
59 static int __ocfs2_move_extent(handle_t *handle,
60                                struct ocfs2_move_extents_context *context,
61                                u32 cpos, u32 len, u32 p_cpos, u32 new_p_cpos,
62                                int ext_flags)
63 {
64         int ret = 0, index;
65         struct inode *inode = context->inode;
66         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
67         struct ocfs2_extent_rec *rec, replace_rec;
68         struct ocfs2_path *path = NULL;
69         struct ocfs2_extent_list *el;
70         u64 ino = ocfs2_metadata_cache_owner(context->et.et_ci);
71         u64 old_blkno = ocfs2_clusters_to_blocks(inode->i_sb, p_cpos);
72
73         ret = ocfs2_duplicate_clusters_by_page(handle, inode, cpos,
74                                                p_cpos, new_p_cpos, len);
75         if (ret) {
76                 mlog_errno(ret);
77                 goto out;
78         }
79
80         memset(&replace_rec, 0, sizeof(replace_rec));
81         replace_rec.e_cpos = cpu_to_le32(cpos);
82         replace_rec.e_leaf_clusters = cpu_to_le16(len);
83         replace_rec.e_blkno = cpu_to_le64(ocfs2_clusters_to_blocks(inode->i_sb,
84                                                                    new_p_cpos));
85
86         path = ocfs2_new_path_from_et(&context->et);
87         if (!path) {
88                 ret = -ENOMEM;
89                 mlog_errno(ret);
90                 goto out;
91         }
92
93         ret = ocfs2_find_path(INODE_CACHE(inode), path, cpos);
94         if (ret) {
95                 mlog_errno(ret);
96                 goto out;
97         }
98
99         el = path_leaf_el(path);
100
101         index = ocfs2_search_extent_list(el, cpos);
102         if (index == -1) {
103                 ret = ocfs2_error(inode->i_sb,
104                                   "Inode %llu has an extent at cpos %u which can no longer be found\n",
105                                   (unsigned long long)ino, cpos);
106                 goto out;
107         }
108
109         rec = &el->l_recs[index];
110
111         BUG_ON(ext_flags != rec->e_flags);
112         /*
113          * after moving/defraging to new location, the extent is not going
114          * to be refcounted anymore.
115          */
116         replace_rec.e_flags = ext_flags & ~OCFS2_EXT_REFCOUNTED;
117
118         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode),
119                                       context->et.et_root_bh,
120                                       OCFS2_JOURNAL_ACCESS_WRITE);
121         if (ret) {
122                 mlog_errno(ret);
123                 goto out;
124         }
125
126         ret = ocfs2_split_extent(handle, &context->et, path, index,
127                                  &replace_rec, context->meta_ac,
128                                  &context->dealloc);
129         if (ret) {
130                 mlog_errno(ret);
131                 goto out;
132         }
133
134         ocfs2_journal_dirty(handle, context->et.et_root_bh);
135
136         context->new_phys_cpos = new_p_cpos;
137
138         /*
139          * need I to append truncate log for old clusters?
140          */
141         if (old_blkno) {
142                 if (ext_flags & OCFS2_EXT_REFCOUNTED)
143                         ret = ocfs2_decrease_refcount(inode, handle,
144                                         ocfs2_blocks_to_clusters(osb->sb,
145                                                                  old_blkno),
146                                         len, context->meta_ac,
147                                         &context->dealloc, 1);
148                 else
149                         ret = ocfs2_truncate_log_append(osb, handle,
150                                                         old_blkno, len);
151         }
152
153         ocfs2_update_inode_fsync_trans(handle, inode, 0);
154 out:
155         ocfs2_free_path(path);
156         return ret;
157 }
158
159 /*
160  * lock allocator, and reserve appropriate number of bits for
161  * meta blocks.
162  */
163 static int ocfs2_lock_meta_allocator_move_extents(struct inode *inode,
164                                         struct ocfs2_extent_tree *et,
165                                         u32 clusters_to_move,
166                                         u32 extents_to_split,
167                                         struct ocfs2_alloc_context **meta_ac,
168                                         int extra_blocks,
169                                         int *credits)
170 {
171         int ret, num_free_extents;
172         unsigned int max_recs_needed = 2 * extents_to_split + clusters_to_move;
173         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
174
175         num_free_extents = ocfs2_num_free_extents(osb, et);
176         if (num_free_extents < 0) {
177                 ret = num_free_extents;
178                 mlog_errno(ret);
179                 goto out;
180         }
181
182         if (!num_free_extents ||
183             (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
184                 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
185
186         ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, meta_ac);
187         if (ret) {
188                 mlog_errno(ret);
189                 goto out;
190         }
191
192
193         *credits += ocfs2_calc_extend_credits(osb->sb, et->et_root_el);
194
195         mlog(0, "reserve metadata_blocks: %d, data_clusters: %u, credits: %d\n",
196              extra_blocks, clusters_to_move, *credits);
197 out:
198         if (ret) {
199                 if (*meta_ac) {
200                         ocfs2_free_alloc_context(*meta_ac);
201                         *meta_ac = NULL;
202                 }
203         }
204
205         return ret;
206 }
207
208 /*
209  * Using one journal handle to guarantee the data consistency in case
210  * crash happens anywhere.
211  *
212  *  XXX: defrag can end up with finishing partial extent as requested,
213  * due to not enough contiguous clusters can be found in allocator.
214  */
215 static int ocfs2_defrag_extent(struct ocfs2_move_extents_context *context,
216                                u32 cpos, u32 phys_cpos, u32 *len, int ext_flags)
217 {
218         int ret, credits = 0, extra_blocks = 0, partial = context->partial;
219         handle_t *handle;
220         struct inode *inode = context->inode;
221         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
222         struct inode *tl_inode = osb->osb_tl_inode;
223         struct ocfs2_refcount_tree *ref_tree = NULL;
224         u32 new_phys_cpos, new_len;
225         u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
226         int need_free = 0;
227
228         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && *len) {
229
230                 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
231                          OCFS2_HAS_REFCOUNT_FL));
232
233                 BUG_ON(!context->refcount_loc);
234
235                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
236                                                &ref_tree, NULL);
237                 if (ret) {
238                         mlog_errno(ret);
239                         return ret;
240                 }
241
242                 ret = ocfs2_prepare_refcount_change_for_del(inode,
243                                                         context->refcount_loc,
244                                                         phys_blkno,
245                                                         *len,
246                                                         &credits,
247                                                         &extra_blocks);
248                 if (ret) {
249                         mlog_errno(ret);
250                         goto out;
251                 }
252         }
253
254         ret = ocfs2_lock_meta_allocator_move_extents(inode, &context->et,
255                                                 *len, 1,
256                                                 &context->meta_ac,
257                                                 extra_blocks, &credits);
258         if (ret) {
259                 mlog_errno(ret);
260                 goto out;
261         }
262
263         /*
264          * should be using allocation reservation strategy there?
265          *
266          * if (context->data_ac)
267          *      context->data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
268          */
269
270         inode_lock(tl_inode);
271
272         if (ocfs2_truncate_log_needs_flush(osb)) {
273                 ret = __ocfs2_flush_truncate_log(osb);
274                 if (ret < 0) {
275                         mlog_errno(ret);
276                         goto out_unlock_mutex;
277                 }
278         }
279
280         /*
281          * Make sure ocfs2_reserve_cluster is called after
282          * __ocfs2_flush_truncate_log, otherwise, dead lock may happen.
283          *
284          * If ocfs2_reserve_cluster is called
285          * before __ocfs2_flush_truncate_log, dead lock on global bitmap
286          * may happen.
287          *
288          */
289         ret = ocfs2_reserve_clusters(osb, *len, &context->data_ac);
290         if (ret) {
291                 mlog_errno(ret);
292                 goto out_unlock_mutex;
293         }
294
295         handle = ocfs2_start_trans(osb, credits);
296         if (IS_ERR(handle)) {
297                 ret = PTR_ERR(handle);
298                 mlog_errno(ret);
299                 goto out_unlock_mutex;
300         }
301
302         ret = __ocfs2_claim_clusters(handle, context->data_ac, 1, *len,
303                                      &new_phys_cpos, &new_len);
304         if (ret) {
305                 mlog_errno(ret);
306                 goto out_commit;
307         }
308
309         /*
310          * allowing partial extent moving is kind of 'pros and cons', it makes
311          * whole defragmentation less likely to fail, on the contrary, the bad
312          * thing is it may make the fs even more fragmented after moving, let
313          * userspace make a good decision here.
314          */
315         if (new_len != *len) {
316                 mlog(0, "len_claimed: %u, len: %u\n", new_len, *len);
317                 if (!partial) {
318                         context->range->me_flags &= ~OCFS2_MOVE_EXT_FL_COMPLETE;
319                         ret = -ENOSPC;
320                         need_free = 1;
321                         goto out_commit;
322                 }
323         }
324
325         mlog(0, "cpos: %u, phys_cpos: %u, new_phys_cpos: %u\n", cpos,
326              phys_cpos, new_phys_cpos);
327
328         ret = __ocfs2_move_extent(handle, context, cpos, new_len, phys_cpos,
329                                   new_phys_cpos, ext_flags);
330         if (ret)
331                 mlog_errno(ret);
332
333         if (partial && (new_len != *len))
334                 *len = new_len;
335
336         /*
337          * Here we should write the new page out first if we are
338          * in write-back mode.
339          */
340         ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, *len);
341         if (ret)
342                 mlog_errno(ret);
343
344 out_commit:
345         if (need_free && context->data_ac) {
346                 struct ocfs2_alloc_context *data_ac = context->data_ac;
347
348                 if (context->data_ac->ac_which == OCFS2_AC_USE_LOCAL)
349                         ocfs2_free_local_alloc_bits(osb, handle, data_ac,
350                                         new_phys_cpos, new_len);
351                 else
352                         ocfs2_free_clusters(handle,
353                                         data_ac->ac_inode,
354                                         data_ac->ac_bh,
355                                         ocfs2_clusters_to_blocks(osb->sb, new_phys_cpos),
356                                         new_len);
357         }
358
359         ocfs2_commit_trans(osb, handle);
360
361 out_unlock_mutex:
362         inode_unlock(tl_inode);
363
364         if (context->data_ac) {
365                 ocfs2_free_alloc_context(context->data_ac);
366                 context->data_ac = NULL;
367         }
368
369         if (context->meta_ac) {
370                 ocfs2_free_alloc_context(context->meta_ac);
371                 context->meta_ac = NULL;
372         }
373
374 out:
375         if (ref_tree)
376                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
377
378         return ret;
379 }
380
381 /*
382  * find the victim alloc group, where #blkno fits.
383  */
384 static int ocfs2_find_victim_alloc_group(struct inode *inode,
385                                          u64 vict_blkno,
386                                          int type, int slot,
387                                          int *vict_bit,
388                                          struct buffer_head **ret_bh)
389 {
390         int ret, i, bits_per_unit = 0;
391         u64 blkno;
392         char namebuf[40];
393
394         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
395         struct buffer_head *ac_bh = NULL, *gd_bh = NULL;
396         struct ocfs2_chain_list *cl;
397         struct ocfs2_chain_rec *rec;
398         struct ocfs2_dinode *ac_dinode;
399         struct ocfs2_group_desc *bg;
400
401         ocfs2_sprintf_system_inode_name(namebuf, sizeof(namebuf), type, slot);
402         ret = ocfs2_lookup_ino_from_name(osb->sys_root_inode, namebuf,
403                                          strlen(namebuf), &blkno);
404         if (ret) {
405                 ret = -ENOENT;
406                 goto out;
407         }
408
409         ret = ocfs2_read_blocks_sync(osb, blkno, 1, &ac_bh);
410         if (ret) {
411                 mlog_errno(ret);
412                 goto out;
413         }
414
415         ac_dinode = (struct ocfs2_dinode *)ac_bh->b_data;
416         cl = &(ac_dinode->id2.i_chain);
417         rec = &(cl->cl_recs[0]);
418
419         if (type == GLOBAL_BITMAP_SYSTEM_INODE)
420                 bits_per_unit = osb->s_clustersize_bits -
421                                         inode->i_sb->s_blocksize_bits;
422         /*
423          * 'vict_blkno' was out of the valid range.
424          */
425         if ((vict_blkno < le64_to_cpu(rec->c_blkno)) ||
426             (vict_blkno >= ((u64)le32_to_cpu(ac_dinode->id1.bitmap1.i_total) <<
427                                 bits_per_unit))) {
428                 ret = -EINVAL;
429                 goto out;
430         }
431
432         for (i = 0; i < le16_to_cpu(cl->cl_next_free_rec); i++) {
433
434                 rec = &(cl->cl_recs[i]);
435                 if (!rec)
436                         continue;
437
438                 bg = NULL;
439
440                 do {
441                         if (!bg)
442                                 blkno = le64_to_cpu(rec->c_blkno);
443                         else
444                                 blkno = le64_to_cpu(bg->bg_next_group);
445
446                         if (gd_bh) {
447                                 brelse(gd_bh);
448                                 gd_bh = NULL;
449                         }
450
451                         ret = ocfs2_read_blocks_sync(osb, blkno, 1, &gd_bh);
452                         if (ret) {
453                                 mlog_errno(ret);
454                                 goto out;
455                         }
456
457                         bg = (struct ocfs2_group_desc *)gd_bh->b_data;
458
459                         if (vict_blkno < (le64_to_cpu(bg->bg_blkno) +
460                                                 le16_to_cpu(bg->bg_bits))) {
461
462                                 *ret_bh = gd_bh;
463                                 *vict_bit = (vict_blkno - blkno) >>
464                                                         bits_per_unit;
465                                 mlog(0, "find the victim group: #%llu, "
466                                      "total_bits: %u, vict_bit: %u\n",
467                                      blkno, le16_to_cpu(bg->bg_bits),
468                                      *vict_bit);
469                                 goto out;
470                         }
471
472                 } while (le64_to_cpu(bg->bg_next_group));
473         }
474
475         ret = -EINVAL;
476 out:
477         brelse(ac_bh);
478
479         /*
480          * caller has to release the gd_bh properly.
481          */
482         return ret;
483 }
484
485 /*
486  * XXX: helper to validate and adjust moving goal.
487  */
488 static int ocfs2_validate_and_adjust_move_goal(struct inode *inode,
489                                                struct ocfs2_move_extents *range)
490 {
491         int ret, goal_bit = 0;
492
493         struct buffer_head *gd_bh = NULL;
494         struct ocfs2_group_desc *bg;
495         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
496         int c_to_b = 1 << (osb->s_clustersize_bits -
497                                         inode->i_sb->s_blocksize_bits);
498
499         /*
500          * make goal become cluster aligned.
501          */
502         range->me_goal = ocfs2_block_to_cluster_start(inode->i_sb,
503                                                       range->me_goal);
504         /*
505          * validate goal sits within global_bitmap, and return the victim
506          * group desc
507          */
508         ret = ocfs2_find_victim_alloc_group(inode, range->me_goal,
509                                             GLOBAL_BITMAP_SYSTEM_INODE,
510                                             OCFS2_INVALID_SLOT,
511                                             &goal_bit, &gd_bh);
512         if (ret)
513                 goto out;
514
515         bg = (struct ocfs2_group_desc *)gd_bh->b_data;
516
517         /*
518          * moving goal is not allowd to start with a group desc blok(#0 blk)
519          * let's compromise to the latter cluster.
520          */
521         if (range->me_goal == le64_to_cpu(bg->bg_blkno))
522                 range->me_goal += c_to_b;
523
524         /*
525          * movement is not gonna cross two groups.
526          */
527         if ((le16_to_cpu(bg->bg_bits) - goal_bit) * osb->s_clustersize <
528                                                                 range->me_len) {
529                 ret = -EINVAL;
530                 goto out;
531         }
532         /*
533          * more exact validations/adjustments will be performed later during
534          * moving operation for each extent range.
535          */
536         mlog(0, "extents get ready to be moved to #%llu block\n",
537              range->me_goal);
538
539 out:
540         brelse(gd_bh);
541
542         return ret;
543 }
544
545 static void ocfs2_probe_alloc_group(struct inode *inode, struct buffer_head *bh,
546                                     int *goal_bit, u32 move_len, u32 max_hop,
547                                     u32 *phys_cpos)
548 {
549         int i, used, last_free_bits = 0, base_bit = *goal_bit;
550         struct ocfs2_group_desc *gd = (struct ocfs2_group_desc *)bh->b_data;
551         u32 base_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
552                                                  le64_to_cpu(gd->bg_blkno));
553
554         for (i = base_bit; i < le16_to_cpu(gd->bg_bits); i++) {
555
556                 used = ocfs2_test_bit(i, (unsigned long *)gd->bg_bitmap);
557                 if (used) {
558                         /*
559                          * we even tried searching the free chunk by jumping
560                          * a 'max_hop' distance, but still failed.
561                          */
562                         if ((i - base_bit) > max_hop) {
563                                 *phys_cpos = 0;
564                                 break;
565                         }
566
567                         if (last_free_bits)
568                                 last_free_bits = 0;
569
570                         continue;
571                 } else
572                         last_free_bits++;
573
574                 if (last_free_bits == move_len) {
575                         *goal_bit = i;
576                         *phys_cpos = base_cpos + i;
577                         break;
578                 }
579         }
580
581         mlog(0, "found phys_cpos: %u to fit the wanted moving.\n", *phys_cpos);
582 }
583
584 static int ocfs2_move_extent(struct ocfs2_move_extents_context *context,
585                              u32 cpos, u32 phys_cpos, u32 *new_phys_cpos,
586                              u32 len, int ext_flags)
587 {
588         int ret, credits = 0, extra_blocks = 0, goal_bit = 0;
589         handle_t *handle;
590         struct inode *inode = context->inode;
591         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
592         struct inode *tl_inode = osb->osb_tl_inode;
593         struct inode *gb_inode = NULL;
594         struct buffer_head *gb_bh = NULL;
595         struct buffer_head *gd_bh = NULL;
596         struct ocfs2_group_desc *gd;
597         struct ocfs2_refcount_tree *ref_tree = NULL;
598         u32 move_max_hop = ocfs2_blocks_to_clusters(inode->i_sb,
599                                                     context->range->me_threshold);
600         u64 phys_blkno, new_phys_blkno;
601
602         phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
603
604         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && len) {
605
606                 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
607                          OCFS2_HAS_REFCOUNT_FL));
608
609                 BUG_ON(!context->refcount_loc);
610
611                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
612                                                &ref_tree, NULL);
613                 if (ret) {
614                         mlog_errno(ret);
615                         return ret;
616                 }
617
618                 ret = ocfs2_prepare_refcount_change_for_del(inode,
619                                                         context->refcount_loc,
620                                                         phys_blkno,
621                                                         len,
622                                                         &credits,
623                                                         &extra_blocks);
624                 if (ret) {
625                         mlog_errno(ret);
626                         goto out;
627                 }
628         }
629
630         ret = ocfs2_lock_meta_allocator_move_extents(inode, &context->et,
631                                                 len, 1,
632                                                 &context->meta_ac,
633                                                 extra_blocks, &credits);
634         if (ret) {
635                 mlog_errno(ret);
636                 goto out;
637         }
638
639         /*
640          * need to count 2 extra credits for global_bitmap inode and
641          * group descriptor.
642          */
643         credits += OCFS2_INODE_UPDATE_CREDITS + 1;
644
645         /*
646          * ocfs2_move_extent() didn't reserve any clusters in lock_allocators()
647          * logic, while we still need to lock the global_bitmap.
648          */
649         gb_inode = ocfs2_get_system_file_inode(osb, GLOBAL_BITMAP_SYSTEM_INODE,
650                                                OCFS2_INVALID_SLOT);
651         if (!gb_inode) {
652                 mlog(ML_ERROR, "unable to get global_bitmap inode\n");
653                 ret = -EIO;
654                 goto out;
655         }
656
657         inode_lock(gb_inode);
658
659         ret = ocfs2_inode_lock(gb_inode, &gb_bh, 1);
660         if (ret) {
661                 mlog_errno(ret);
662                 goto out_unlock_gb_mutex;
663         }
664
665         inode_lock(tl_inode);
666
667         handle = ocfs2_start_trans(osb, credits);
668         if (IS_ERR(handle)) {
669                 ret = PTR_ERR(handle);
670                 mlog_errno(ret);
671                 goto out_unlock_tl_inode;
672         }
673
674         new_phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *new_phys_cpos);
675         ret = ocfs2_find_victim_alloc_group(inode, new_phys_blkno,
676                                             GLOBAL_BITMAP_SYSTEM_INODE,
677                                             OCFS2_INVALID_SLOT,
678                                             &goal_bit, &gd_bh);
679         if (ret) {
680                 mlog_errno(ret);
681                 goto out_commit;
682         }
683
684         /*
685          * probe the victim cluster group to find a proper
686          * region to fit wanted movement, it even will perfrom
687          * a best-effort attempt by compromising to a threshold
688          * around the goal.
689          */
690         ocfs2_probe_alloc_group(inode, gd_bh, &goal_bit, len, move_max_hop,
691                                 new_phys_cpos);
692         if (!*new_phys_cpos) {
693                 ret = -ENOSPC;
694                 goto out_commit;
695         }
696
697         ret = __ocfs2_move_extent(handle, context, cpos, len, phys_cpos,
698                                   *new_phys_cpos, ext_flags);
699         if (ret) {
700                 mlog_errno(ret);
701                 goto out_commit;
702         }
703
704         gd = (struct ocfs2_group_desc *)gd_bh->b_data;
705         ret = ocfs2_alloc_dinode_update_counts(gb_inode, handle, gb_bh, len,
706                                                le16_to_cpu(gd->bg_chain));
707         if (ret) {
708                 mlog_errno(ret);
709                 goto out_commit;
710         }
711
712         ret = ocfs2_block_group_set_bits(handle, gb_inode, gd, gd_bh,
713                                          goal_bit, len);
714         if (ret) {
715                 ocfs2_rollback_alloc_dinode_counts(gb_inode, gb_bh, len,
716                                                le16_to_cpu(gd->bg_chain));
717                 mlog_errno(ret);
718         }
719
720         /*
721          * Here we should write the new page out first if we are
722          * in write-back mode.
723          */
724         ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, len);
725         if (ret)
726                 mlog_errno(ret);
727
728 out_commit:
729         ocfs2_commit_trans(osb, handle);
730         brelse(gd_bh);
731
732 out_unlock_tl_inode:
733         inode_unlock(tl_inode);
734
735         ocfs2_inode_unlock(gb_inode, 1);
736 out_unlock_gb_mutex:
737         inode_unlock(gb_inode);
738         brelse(gb_bh);
739         iput(gb_inode);
740
741 out:
742         if (context->meta_ac) {
743                 ocfs2_free_alloc_context(context->meta_ac);
744                 context->meta_ac = NULL;
745         }
746
747         if (ref_tree)
748                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
749
750         return ret;
751 }
752
753 /*
754  * Helper to calculate the defraging length in one run according to threshold.
755  */
756 static void ocfs2_calc_extent_defrag_len(u32 *alloc_size, u32 *len_defraged,
757                                          u32 threshold, int *skip)
758 {
759         if ((*alloc_size + *len_defraged) < threshold) {
760                 /*
761                  * proceed defragmentation until we meet the thresh
762                  */
763                 *len_defraged += *alloc_size;
764         } else if (*len_defraged == 0) {
765                 /*
766                  * XXX: skip a large extent.
767                  */
768                 *skip = 1;
769         } else {
770                 /*
771                  * split this extent to coalesce with former pieces as
772                  * to reach the threshold.
773                  *
774                  * we're done here with one cycle of defragmentation
775                  * in a size of 'thresh', resetting 'len_defraged'
776                  * forces a new defragmentation.
777                  */
778                 *alloc_size = threshold - *len_defraged;
779                 *len_defraged = 0;
780         }
781 }
782
783 static int __ocfs2_move_extents_range(struct buffer_head *di_bh,
784                                 struct ocfs2_move_extents_context *context)
785 {
786         int ret = 0, flags, do_defrag, skip = 0;
787         u32 cpos, phys_cpos, move_start, len_to_move, alloc_size;
788         u32 len_defraged = 0, defrag_thresh = 0, new_phys_cpos = 0;
789
790         struct inode *inode = context->inode;
791         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
792         struct ocfs2_move_extents *range = context->range;
793         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
794
795         if ((i_size_read(inode) == 0) || (range->me_len == 0))
796                 return 0;
797
798         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
799                 return 0;
800
801         context->refcount_loc = le64_to_cpu(di->i_refcount_loc);
802
803         ocfs2_init_dinode_extent_tree(&context->et, INODE_CACHE(inode), di_bh);
804         ocfs2_init_dealloc_ctxt(&context->dealloc);
805
806         /*
807          * TO-DO XXX:
808          *
809          * - xattr extents.
810          */
811
812         do_defrag = context->auto_defrag;
813
814         /*
815          * extents moving happens in unit of clusters, for the sake
816          * of simplicity, we may ignore two clusters where 'byte_start'
817          * and 'byte_start + len' were within.
818          */
819         move_start = ocfs2_clusters_for_bytes(osb->sb, range->me_start);
820         len_to_move = (range->me_start + range->me_len) >>
821                                                 osb->s_clustersize_bits;
822         if (len_to_move >= move_start)
823                 len_to_move -= move_start;
824         else
825                 len_to_move = 0;
826
827         if (do_defrag) {
828                 defrag_thresh = range->me_threshold >> osb->s_clustersize_bits;
829                 if (defrag_thresh <= 1)
830                         goto done;
831         } else
832                 new_phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
833                                                          range->me_goal);
834
835         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u, "
836              "thresh: %u\n",
837              (unsigned long long)OCFS2_I(inode)->ip_blkno,
838              (unsigned long long)range->me_start,
839              (unsigned long long)range->me_len,
840              move_start, len_to_move, defrag_thresh);
841
842         cpos = move_start;
843         while (len_to_move) {
844                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &alloc_size,
845                                          &flags);
846                 if (ret) {
847                         mlog_errno(ret);
848                         goto out;
849                 }
850
851                 if (alloc_size > len_to_move)
852                         alloc_size = len_to_move;
853
854                 /*
855                  * XXX: how to deal with a hole:
856                  *
857                  * - skip the hole of course
858                  * - force a new defragmentation
859                  */
860                 if (!phys_cpos) {
861                         if (do_defrag)
862                                 len_defraged = 0;
863
864                         goto next;
865                 }
866
867                 if (do_defrag) {
868                         ocfs2_calc_extent_defrag_len(&alloc_size, &len_defraged,
869                                                      defrag_thresh, &skip);
870                         /*
871                          * skip large extents
872                          */
873                         if (skip) {
874                                 skip = 0;
875                                 goto next;
876                         }
877
878                         mlog(0, "#Defrag: cpos: %u, phys_cpos: %u, "
879                              "alloc_size: %u, len_defraged: %u\n",
880                              cpos, phys_cpos, alloc_size, len_defraged);
881
882                         ret = ocfs2_defrag_extent(context, cpos, phys_cpos,
883                                                   &alloc_size, flags);
884                 } else {
885                         ret = ocfs2_move_extent(context, cpos, phys_cpos,
886                                                 &new_phys_cpos, alloc_size,
887                                                 flags);
888
889                         new_phys_cpos += alloc_size;
890                 }
891
892                 if (ret < 0) {
893                         mlog_errno(ret);
894                         goto out;
895                 }
896
897                 context->clusters_moved += alloc_size;
898 next:
899                 cpos += alloc_size;
900                 len_to_move -= alloc_size;
901         }
902
903 done:
904         range->me_flags |= OCFS2_MOVE_EXT_FL_COMPLETE;
905
906 out:
907         range->me_moved_len = ocfs2_clusters_to_bytes(osb->sb,
908                                                       context->clusters_moved);
909         range->me_new_offset = ocfs2_clusters_to_bytes(osb->sb,
910                                                        context->new_phys_cpos);
911
912         ocfs2_schedule_truncate_log_flush(osb, 1);
913         ocfs2_run_deallocs(osb, &context->dealloc);
914
915         return ret;
916 }
917
918 static int ocfs2_move_extents(struct ocfs2_move_extents_context *context)
919 {
920         int status;
921         handle_t *handle;
922         struct inode *inode = context->inode;
923         struct ocfs2_dinode *di;
924         struct buffer_head *di_bh = NULL;
925         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
926
927         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
928                 return -EROFS;
929
930         inode_lock(inode);
931
932         /*
933          * This prevents concurrent writes from other nodes
934          */
935         status = ocfs2_rw_lock(inode, 1);
936         if (status) {
937                 mlog_errno(status);
938                 goto out;
939         }
940
941         status = ocfs2_inode_lock(inode, &di_bh, 1);
942         if (status) {
943                 mlog_errno(status);
944                 goto out_rw_unlock;
945         }
946
947         /*
948          * rememer ip_xattr_sem also needs to be held if necessary
949          */
950         down_write(&OCFS2_I(inode)->ip_alloc_sem);
951
952         status = __ocfs2_move_extents_range(di_bh, context);
953
954         up_write(&OCFS2_I(inode)->ip_alloc_sem);
955         if (status) {
956                 mlog_errno(status);
957                 goto out_inode_unlock;
958         }
959
960         /*
961          * We update ctime for these changes
962          */
963         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
964         if (IS_ERR(handle)) {
965                 status = PTR_ERR(handle);
966                 mlog_errno(status);
967                 goto out_inode_unlock;
968         }
969
970         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
971                                          OCFS2_JOURNAL_ACCESS_WRITE);
972         if (status) {
973                 mlog_errno(status);
974                 goto out_commit;
975         }
976
977         di = (struct ocfs2_dinode *)di_bh->b_data;
978         inode->i_ctime = current_time(inode);
979         di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec);
980         di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
981         ocfs2_update_inode_fsync_trans(handle, inode, 0);
982
983         ocfs2_journal_dirty(handle, di_bh);
984
985 out_commit:
986         ocfs2_commit_trans(osb, handle);
987
988 out_inode_unlock:
989         brelse(di_bh);
990         ocfs2_inode_unlock(inode, 1);
991 out_rw_unlock:
992         ocfs2_rw_unlock(inode, 1);
993 out:
994         inode_unlock(inode);
995
996         return status;
997 }
998
999 int ocfs2_ioctl_move_extents(struct file *filp, void __user *argp)
1000 {
1001         int status;
1002
1003         struct inode *inode = file_inode(filp);
1004         struct ocfs2_move_extents range;
1005         struct ocfs2_move_extents_context *context;
1006
1007         if (!argp)
1008                 return -EINVAL;
1009
1010         status = mnt_want_write_file(filp);
1011         if (status)
1012                 return status;
1013
1014         if ((!S_ISREG(inode->i_mode)) || !(filp->f_mode & FMODE_WRITE)) {
1015                 status = -EPERM;
1016                 goto out_drop;
1017         }
1018
1019         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1020                 status = -EPERM;
1021                 goto out_drop;
1022         }
1023
1024         context = kzalloc(sizeof(struct ocfs2_move_extents_context), GFP_NOFS);
1025         if (!context) {
1026                 status = -ENOMEM;
1027                 mlog_errno(status);
1028                 goto out_drop;
1029         }
1030
1031         context->inode = inode;
1032         context->file = filp;
1033
1034         if (copy_from_user(&range, argp, sizeof(range))) {
1035                 status = -EFAULT;
1036                 goto out_free;
1037         }
1038
1039         if (range.me_start > i_size_read(inode)) {
1040                 status = -EINVAL;
1041                 goto out_free;
1042         }
1043
1044         if (range.me_start + range.me_len > i_size_read(inode))
1045                         range.me_len = i_size_read(inode) - range.me_start;
1046
1047         context->range = &range;
1048
1049         if (range.me_flags & OCFS2_MOVE_EXT_FL_AUTO_DEFRAG) {
1050                 context->auto_defrag = 1;
1051                 /*
1052                  * ok, the default theshold for the defragmentation
1053                  * is 1M, since our maximum clustersize was 1M also.
1054                  * any thought?
1055                  */
1056                 if (!range.me_threshold)
1057                         range.me_threshold = 1024 * 1024;
1058
1059                 if (range.me_threshold > i_size_read(inode))
1060                         range.me_threshold = i_size_read(inode);
1061
1062                 if (range.me_flags & OCFS2_MOVE_EXT_FL_PART_DEFRAG)
1063                         context->partial = 1;
1064         } else {
1065                 /*
1066                  * first best-effort attempt to validate and adjust the goal
1067                  * (physical address in block), while it can't guarantee later
1068                  * operation can succeed all the time since global_bitmap may
1069                  * change a bit over time.
1070                  */
1071
1072                 status = ocfs2_validate_and_adjust_move_goal(inode, &range);
1073                 if (status)
1074                         goto out_copy;
1075         }
1076
1077         status = ocfs2_move_extents(context);
1078         if (status)
1079                 mlog_errno(status);
1080 out_copy:
1081         /*
1082          * movement/defragmentation may end up being partially completed,
1083          * that's the reason why we need to return userspace the finished
1084          * length and new_offset even if failure happens somewhere.
1085          */
1086         if (copy_to_user(argp, &range, sizeof(range)))
1087                 status = -EFAULT;
1088
1089 out_free:
1090         kfree(context);
1091 out_drop:
1092         mnt_drop_write_file(filp);
1093
1094         return status;
1095 }