Loading...
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * localalloc.c
5 *
6 * Node local data allocation
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/bitops.h>
31
32#include <cluster/masklog.h>
33
34#include "ocfs2.h"
35
36#include "alloc.h"
37#include "blockcheck.h"
38#include "dlmglue.h"
39#include "inode.h"
40#include "journal.h"
41#include "localalloc.h"
42#include "suballoc.h"
43#include "super.h"
44#include "sysfile.h"
45#include "ocfs2_trace.h"
46
47#include "buffer_head_io.h"
48
49#define OCFS2_LOCAL_ALLOC(dinode) (&((dinode)->id2.i_lab))
50
51static u32 ocfs2_local_alloc_count_bits(struct ocfs2_dinode *alloc);
52
53static int ocfs2_local_alloc_find_clear_bits(struct ocfs2_super *osb,
54 struct ocfs2_dinode *alloc,
55 u32 *numbits,
56 struct ocfs2_alloc_reservation *resv);
57
58static void ocfs2_clear_local_alloc(struct ocfs2_dinode *alloc);
59
60static int ocfs2_sync_local_to_main(struct ocfs2_super *osb,
61 handle_t *handle,
62 struct ocfs2_dinode *alloc,
63 struct inode *main_bm_inode,
64 struct buffer_head *main_bm_bh);
65
66static int ocfs2_local_alloc_reserve_for_window(struct ocfs2_super *osb,
67 struct ocfs2_alloc_context **ac,
68 struct inode **bitmap_inode,
69 struct buffer_head **bitmap_bh);
70
71static int ocfs2_local_alloc_new_window(struct ocfs2_super *osb,
72 handle_t *handle,
73 struct ocfs2_alloc_context *ac);
74
75static int ocfs2_local_alloc_slide_window(struct ocfs2_super *osb,
76 struct inode *local_alloc_inode);
77
78/*
79 * ocfs2_la_default_mb() - determine a default size, in megabytes of
80 * the local alloc.
81 *
82 * Generally, we'd like to pick as large a local alloc as
83 * possible. Performance on large workloads tends to scale
84 * proportionally to la size. In addition to that, the reservations
85 * code functions more efficiently as it can reserve more windows for
86 * write.
87 *
88 * Some things work against us when trying to choose a large local alloc:
89 *
90 * - We need to ensure our sizing is picked to leave enough space in
91 * group descriptors for other allocations (such as block groups,
92 * etc). Picking default sizes which are a multiple of 4 could help
93 * - block groups are allocated in 2mb and 4mb chunks.
94 *
95 * - Likewise, we don't want to starve other nodes of bits on small
96 * file systems. This can easily be taken care of by limiting our
97 * default to a reasonable size (256M) on larger cluster sizes.
98 *
99 * - Some file systems can't support very large sizes - 4k and 8k in
100 * particular are limited to less than 128 and 256 megabytes respectively.
101 *
102 * The following reference table shows group descriptor and local
103 * alloc maximums at various cluster sizes (4k blocksize)
104 *
105 * csize: 4K group: 126M la: 121M
106 * csize: 8K group: 252M la: 243M
107 * csize: 16K group: 504M la: 486M
108 * csize: 32K group: 1008M la: 972M
109 * csize: 64K group: 2016M la: 1944M
110 * csize: 128K group: 4032M la: 3888M
111 * csize: 256K group: 8064M la: 7776M
112 * csize: 512K group: 16128M la: 15552M
113 * csize: 1024K group: 32256M la: 31104M
114 */
115#define OCFS2_LA_MAX_DEFAULT_MB 256
116#define OCFS2_LA_OLD_DEFAULT 8
117unsigned int ocfs2_la_default_mb(struct ocfs2_super *osb)
118{
119 unsigned int la_mb;
120 unsigned int gd_mb;
121 unsigned int la_max_mb;
122 unsigned int megs_per_slot;
123 struct super_block *sb = osb->sb;
124
125 gd_mb = ocfs2_clusters_to_megabytes(osb->sb,
126 8 * ocfs2_group_bitmap_size(sb, 0, osb->s_feature_incompat));
127
128 /*
129 * This takes care of files systems with very small group
130 * descriptors - 512 byte blocksize at cluster sizes lower
131 * than 16K and also 1k blocksize with 4k cluster size.
132 */
133 if ((sb->s_blocksize == 512 && osb->s_clustersize <= 8192)
134 || (sb->s_blocksize == 1024 && osb->s_clustersize == 4096))
135 return OCFS2_LA_OLD_DEFAULT;
136
137 /*
138 * Leave enough room for some block groups and make the final
139 * value we work from a multiple of 4.
140 */
141 gd_mb -= 16;
142 gd_mb &= 0xFFFFFFFB;
143
144 la_mb = gd_mb;
145
146 /*
147 * Keep window sizes down to a reasonable default
148 */
149 if (la_mb > OCFS2_LA_MAX_DEFAULT_MB) {
150 /*
151 * Some clustersize / blocksize combinations will have
152 * given us a larger than OCFS2_LA_MAX_DEFAULT_MB
153 * default size, but get poor distribution when
154 * limited to exactly 256 megabytes.
155 *
156 * As an example, 16K clustersize at 4K blocksize
157 * gives us a cluster group size of 504M. Paring the
158 * local alloc size down to 256 however, would give us
159 * only one window and around 200MB left in the
160 * cluster group. Instead, find the first size below
161 * 256 which would give us an even distribution.
162 *
163 * Larger cluster group sizes actually work out pretty
164 * well when pared to 256, so we don't have to do this
165 * for any group that fits more than two
166 * OCFS2_LA_MAX_DEFAULT_MB windows.
167 */
168 if (gd_mb > (2 * OCFS2_LA_MAX_DEFAULT_MB))
169 la_mb = 256;
170 else {
171 unsigned int gd_mult = gd_mb;
172
173 while (gd_mult > 256)
174 gd_mult = gd_mult >> 1;
175
176 la_mb = gd_mult;
177 }
178 }
179
180 megs_per_slot = osb->osb_clusters_at_boot / osb->max_slots;
181 megs_per_slot = ocfs2_clusters_to_megabytes(osb->sb, megs_per_slot);
182 /* Too many nodes, too few disk clusters. */
183 if (megs_per_slot < la_mb)
184 la_mb = megs_per_slot;
185
186 /* We can't store more bits than we can in a block. */
187 la_max_mb = ocfs2_clusters_to_megabytes(osb->sb,
188 ocfs2_local_alloc_size(sb) * 8);
189 if (la_mb > la_max_mb)
190 la_mb = la_max_mb;
191
192 return la_mb;
193}
194
195void ocfs2_la_set_sizes(struct ocfs2_super *osb, int requested_mb)
196{
197 struct super_block *sb = osb->sb;
198 unsigned int la_default_mb = ocfs2_la_default_mb(osb);
199 unsigned int la_max_mb;
200
201 la_max_mb = ocfs2_clusters_to_megabytes(sb,
202 ocfs2_local_alloc_size(sb) * 8);
203
204 trace_ocfs2_la_set_sizes(requested_mb, la_max_mb, la_default_mb);
205
206 if (requested_mb == -1) {
207 /* No user request - use defaults */
208 osb->local_alloc_default_bits =
209 ocfs2_megabytes_to_clusters(sb, la_default_mb);
210 } else if (requested_mb > la_max_mb) {
211 /* Request is too big, we give the maximum available */
212 osb->local_alloc_default_bits =
213 ocfs2_megabytes_to_clusters(sb, la_max_mb);
214 } else {
215 osb->local_alloc_default_bits =
216 ocfs2_megabytes_to_clusters(sb, requested_mb);
217 }
218
219 osb->local_alloc_bits = osb->local_alloc_default_bits;
220}
221
222static inline int ocfs2_la_state_enabled(struct ocfs2_super *osb)
223{
224 return (osb->local_alloc_state == OCFS2_LA_THROTTLED ||
225 osb->local_alloc_state == OCFS2_LA_ENABLED);
226}
227
228void ocfs2_local_alloc_seen_free_bits(struct ocfs2_super *osb,
229 unsigned int num_clusters)
230{
231 spin_lock(&osb->osb_lock);
232 if (osb->local_alloc_state == OCFS2_LA_DISABLED ||
233 osb->local_alloc_state == OCFS2_LA_THROTTLED)
234 if (num_clusters >= osb->local_alloc_default_bits) {
235 cancel_delayed_work(&osb->la_enable_wq);
236 osb->local_alloc_state = OCFS2_LA_ENABLED;
237 }
238 spin_unlock(&osb->osb_lock);
239}
240
241void ocfs2_la_enable_worker(struct work_struct *work)
242{
243 struct ocfs2_super *osb =
244 container_of(work, struct ocfs2_super,
245 la_enable_wq.work);
246 spin_lock(&osb->osb_lock);
247 osb->local_alloc_state = OCFS2_LA_ENABLED;
248 spin_unlock(&osb->osb_lock);
249}
250
251/*
252 * Tell us whether a given allocation should use the local alloc
253 * file. Otherwise, it has to go to the main bitmap.
254 *
255 * This function does semi-dirty reads of local alloc size and state!
256 * This is ok however, as the values are re-checked once under mutex.
257 */
258int ocfs2_alloc_should_use_local(struct ocfs2_super *osb, u64 bits)
259{
260 int ret = 0;
261 int la_bits;
262
263 spin_lock(&osb->osb_lock);
264 la_bits = osb->local_alloc_bits;
265
266 if (!ocfs2_la_state_enabled(osb))
267 goto bail;
268
269 /* la_bits should be at least twice the size (in clusters) of
270 * a new block group. We want to be sure block group
271 * allocations go through the local alloc, so allow an
272 * allocation to take up to half the bitmap. */
273 if (bits > (la_bits / 2))
274 goto bail;
275
276 ret = 1;
277bail:
278 trace_ocfs2_alloc_should_use_local(
279 (unsigned long long)bits, osb->local_alloc_state, la_bits, ret);
280 spin_unlock(&osb->osb_lock);
281 return ret;
282}
283
284int ocfs2_load_local_alloc(struct ocfs2_super *osb)
285{
286 int status = 0;
287 struct ocfs2_dinode *alloc = NULL;
288 struct buffer_head *alloc_bh = NULL;
289 u32 num_used;
290 struct inode *inode = NULL;
291 struct ocfs2_local_alloc *la;
292
293 if (osb->local_alloc_bits == 0)
294 goto bail;
295
296 if (osb->local_alloc_bits >= osb->bitmap_cpg) {
297 mlog(ML_NOTICE, "Requested local alloc window %d is larger "
298 "than max possible %u. Using defaults.\n",
299 osb->local_alloc_bits, (osb->bitmap_cpg - 1));
300 osb->local_alloc_bits =
301 ocfs2_megabytes_to_clusters(osb->sb,
302 ocfs2_la_default_mb(osb));
303 }
304
305 /* read the alloc off disk */
306 inode = ocfs2_get_system_file_inode(osb, LOCAL_ALLOC_SYSTEM_INODE,
307 osb->slot_num);
308 if (!inode) {
309 status = -EINVAL;
310 mlog_errno(status);
311 goto bail;
312 }
313
314 status = ocfs2_read_inode_block_full(inode, &alloc_bh,
315 OCFS2_BH_IGNORE_CACHE);
316 if (status < 0) {
317 mlog_errno(status);
318 goto bail;
319 }
320
321 alloc = (struct ocfs2_dinode *) alloc_bh->b_data;
322 la = OCFS2_LOCAL_ALLOC(alloc);
323
324 if (!(le32_to_cpu(alloc->i_flags) &
325 (OCFS2_LOCAL_ALLOC_FL|OCFS2_BITMAP_FL))) {
326 mlog(ML_ERROR, "Invalid local alloc inode, %llu\n",
327 (unsigned long long)OCFS2_I(inode)->ip_blkno);
328 status = -EINVAL;
329 goto bail;
330 }
331
332 if ((la->la_size == 0) ||
333 (le16_to_cpu(la->la_size) > ocfs2_local_alloc_size(inode->i_sb))) {
334 mlog(ML_ERROR, "Local alloc size is invalid (la_size = %u)\n",
335 le16_to_cpu(la->la_size));
336 status = -EINVAL;
337 goto bail;
338 }
339
340 /* do a little verification. */
341 num_used = ocfs2_local_alloc_count_bits(alloc);
342
343 /* hopefully the local alloc has always been recovered before
344 * we load it. */
345 if (num_used
346 || alloc->id1.bitmap1.i_used
347 || alloc->id1.bitmap1.i_total
348 || la->la_bm_off)
349 mlog(ML_ERROR, "Local alloc hasn't been recovered!\n"
350 "found = %u, set = %u, taken = %u, off = %u\n",
351 num_used, le32_to_cpu(alloc->id1.bitmap1.i_used),
352 le32_to_cpu(alloc->id1.bitmap1.i_total),
353 OCFS2_LOCAL_ALLOC(alloc)->la_bm_off);
354
355 osb->local_alloc_bh = alloc_bh;
356 osb->local_alloc_state = OCFS2_LA_ENABLED;
357
358bail:
359 if (status < 0)
360 brelse(alloc_bh);
361 iput(inode);
362
363 trace_ocfs2_load_local_alloc(osb->local_alloc_bits);
364
365 if (status)
366 mlog_errno(status);
367 return status;
368}
369
370/*
371 * return any unused bits to the bitmap and write out a clean
372 * local_alloc.
373 *
374 * local_alloc_bh is optional. If not passed, we will simply use the
375 * one off osb. If you do pass it however, be warned that it *will* be
376 * returned brelse'd and NULL'd out.*/
377void ocfs2_shutdown_local_alloc(struct ocfs2_super *osb)
378{
379 int status;
380 handle_t *handle;
381 struct inode *local_alloc_inode = NULL;
382 struct buffer_head *bh = NULL;
383 struct buffer_head *main_bm_bh = NULL;
384 struct inode *main_bm_inode = NULL;
385 struct ocfs2_dinode *alloc_copy = NULL;
386 struct ocfs2_dinode *alloc = NULL;
387
388 cancel_delayed_work(&osb->la_enable_wq);
389 flush_workqueue(osb->ocfs2_wq);
390
391 if (osb->local_alloc_state == OCFS2_LA_UNUSED)
392 goto out;
393
394 local_alloc_inode =
395 ocfs2_get_system_file_inode(osb,
396 LOCAL_ALLOC_SYSTEM_INODE,
397 osb->slot_num);
398 if (!local_alloc_inode) {
399 status = -ENOENT;
400 mlog_errno(status);
401 goto out;
402 }
403
404 osb->local_alloc_state = OCFS2_LA_DISABLED;
405
406 ocfs2_resmap_uninit(&osb->osb_la_resmap);
407
408 main_bm_inode = ocfs2_get_system_file_inode(osb,
409 GLOBAL_BITMAP_SYSTEM_INODE,
410 OCFS2_INVALID_SLOT);
411 if (!main_bm_inode) {
412 status = -EINVAL;
413 mlog_errno(status);
414 goto out;
415 }
416
417 inode_lock(main_bm_inode);
418
419 status = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 1);
420 if (status < 0) {
421 mlog_errno(status);
422 goto out_mutex;
423 }
424
425 /* WINDOW_MOVE_CREDITS is a bit heavy... */
426 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
427 if (IS_ERR(handle)) {
428 mlog_errno(PTR_ERR(handle));
429 handle = NULL;
430 goto out_unlock;
431 }
432
433 bh = osb->local_alloc_bh;
434 alloc = (struct ocfs2_dinode *) bh->b_data;
435
436 alloc_copy = kmalloc(bh->b_size, GFP_NOFS);
437 if (!alloc_copy) {
438 status = -ENOMEM;
439 goto out_commit;
440 }
441 memcpy(alloc_copy, alloc, bh->b_size);
442
443 status = ocfs2_journal_access_di(handle, INODE_CACHE(local_alloc_inode),
444 bh, OCFS2_JOURNAL_ACCESS_WRITE);
445 if (status < 0) {
446 mlog_errno(status);
447 goto out_commit;
448 }
449
450 ocfs2_clear_local_alloc(alloc);
451 ocfs2_journal_dirty(handle, bh);
452
453 brelse(bh);
454 osb->local_alloc_bh = NULL;
455 osb->local_alloc_state = OCFS2_LA_UNUSED;
456
457 status = ocfs2_sync_local_to_main(osb, handle, alloc_copy,
458 main_bm_inode, main_bm_bh);
459 if (status < 0)
460 mlog_errno(status);
461
462out_commit:
463 ocfs2_commit_trans(osb, handle);
464
465out_unlock:
466 brelse(main_bm_bh);
467
468 ocfs2_inode_unlock(main_bm_inode, 1);
469
470out_mutex:
471 inode_unlock(main_bm_inode);
472 iput(main_bm_inode);
473
474out:
475 iput(local_alloc_inode);
476
477 kfree(alloc_copy);
478}
479
480/*
481 * We want to free the bitmap bits outside of any recovery context as
482 * we'll need a cluster lock to do so, but we must clear the local
483 * alloc before giving up the recovered nodes journal. To solve this,
484 * we kmalloc a copy of the local alloc before it's change for the
485 * caller to process with ocfs2_complete_local_alloc_recovery
486 */
487int ocfs2_begin_local_alloc_recovery(struct ocfs2_super *osb,
488 int slot_num,
489 struct ocfs2_dinode **alloc_copy)
490{
491 int status = 0;
492 struct buffer_head *alloc_bh = NULL;
493 struct inode *inode = NULL;
494 struct ocfs2_dinode *alloc;
495
496 trace_ocfs2_begin_local_alloc_recovery(slot_num);
497
498 *alloc_copy = NULL;
499
500 inode = ocfs2_get_system_file_inode(osb,
501 LOCAL_ALLOC_SYSTEM_INODE,
502 slot_num);
503 if (!inode) {
504 status = -EINVAL;
505 mlog_errno(status);
506 goto bail;
507 }
508
509 inode_lock(inode);
510
511 status = ocfs2_read_inode_block_full(inode, &alloc_bh,
512 OCFS2_BH_IGNORE_CACHE);
513 if (status < 0) {
514 mlog_errno(status);
515 goto bail;
516 }
517
518 *alloc_copy = kmalloc(alloc_bh->b_size, GFP_KERNEL);
519 if (!(*alloc_copy)) {
520 status = -ENOMEM;
521 goto bail;
522 }
523 memcpy((*alloc_copy), alloc_bh->b_data, alloc_bh->b_size);
524
525 alloc = (struct ocfs2_dinode *) alloc_bh->b_data;
526 ocfs2_clear_local_alloc(alloc);
527
528 ocfs2_compute_meta_ecc(osb->sb, alloc_bh->b_data, &alloc->i_check);
529 status = ocfs2_write_block(osb, alloc_bh, INODE_CACHE(inode));
530 if (status < 0)
531 mlog_errno(status);
532
533bail:
534 if (status < 0) {
535 kfree(*alloc_copy);
536 *alloc_copy = NULL;
537 }
538
539 brelse(alloc_bh);
540
541 if (inode) {
542 inode_unlock(inode);
543 iput(inode);
544 }
545
546 if (status)
547 mlog_errno(status);
548 return status;
549}
550
551/*
552 * Step 2: By now, we've completed the journal recovery, we've stamped
553 * a clean local alloc on disk and dropped the node out of the
554 * recovery map. Dlm locks will no longer stall, so lets clear out the
555 * main bitmap.
556 */
557int ocfs2_complete_local_alloc_recovery(struct ocfs2_super *osb,
558 struct ocfs2_dinode *alloc)
559{
560 int status;
561 handle_t *handle;
562 struct buffer_head *main_bm_bh = NULL;
563 struct inode *main_bm_inode;
564
565 main_bm_inode = ocfs2_get_system_file_inode(osb,
566 GLOBAL_BITMAP_SYSTEM_INODE,
567 OCFS2_INVALID_SLOT);
568 if (!main_bm_inode) {
569 status = -EINVAL;
570 mlog_errno(status);
571 goto out;
572 }
573
574 inode_lock(main_bm_inode);
575
576 status = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 1);
577 if (status < 0) {
578 mlog_errno(status);
579 goto out_mutex;
580 }
581
582 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
583 if (IS_ERR(handle)) {
584 status = PTR_ERR(handle);
585 handle = NULL;
586 mlog_errno(status);
587 goto out_unlock;
588 }
589
590 /* we want the bitmap change to be recorded on disk asap */
591 handle->h_sync = 1;
592
593 status = ocfs2_sync_local_to_main(osb, handle, alloc,
594 main_bm_inode, main_bm_bh);
595 if (status < 0)
596 mlog_errno(status);
597
598 ocfs2_commit_trans(osb, handle);
599
600out_unlock:
601 ocfs2_inode_unlock(main_bm_inode, 1);
602
603out_mutex:
604 inode_unlock(main_bm_inode);
605
606 brelse(main_bm_bh);
607
608 iput(main_bm_inode);
609
610out:
611 if (!status)
612 ocfs2_init_steal_slots(osb);
613 if (status)
614 mlog_errno(status);
615 return status;
616}
617
618/*
619 * make sure we've got at least bits_wanted contiguous bits in the
620 * local alloc. You lose them when you drop i_mutex.
621 *
622 * We will add ourselves to the transaction passed in, but may start
623 * our own in order to shift windows.
624 */
625int ocfs2_reserve_local_alloc_bits(struct ocfs2_super *osb,
626 u32 bits_wanted,
627 struct ocfs2_alloc_context *ac)
628{
629 int status;
630 struct ocfs2_dinode *alloc;
631 struct inode *local_alloc_inode;
632 unsigned int free_bits;
633
634 BUG_ON(!ac);
635
636 local_alloc_inode =
637 ocfs2_get_system_file_inode(osb,
638 LOCAL_ALLOC_SYSTEM_INODE,
639 osb->slot_num);
640 if (!local_alloc_inode) {
641 status = -ENOENT;
642 mlog_errno(status);
643 goto bail;
644 }
645
646 inode_lock(local_alloc_inode);
647
648 /*
649 * We must double check state and allocator bits because
650 * another process may have changed them while holding i_mutex.
651 */
652 spin_lock(&osb->osb_lock);
653 if (!ocfs2_la_state_enabled(osb) ||
654 (bits_wanted > osb->local_alloc_bits)) {
655 spin_unlock(&osb->osb_lock);
656 status = -ENOSPC;
657 goto bail;
658 }
659 spin_unlock(&osb->osb_lock);
660
661 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
662
663#ifdef CONFIG_OCFS2_DEBUG_FS
664 if (le32_to_cpu(alloc->id1.bitmap1.i_used) !=
665 ocfs2_local_alloc_count_bits(alloc)) {
666 ocfs2_error(osb->sb, "local alloc inode %llu says it has %u used bits, but a count shows %u\n",
667 (unsigned long long)le64_to_cpu(alloc->i_blkno),
668 le32_to_cpu(alloc->id1.bitmap1.i_used),
669 ocfs2_local_alloc_count_bits(alloc));
670 status = -EIO;
671 goto bail;
672 }
673#endif
674
675 free_bits = le32_to_cpu(alloc->id1.bitmap1.i_total) -
676 le32_to_cpu(alloc->id1.bitmap1.i_used);
677 if (bits_wanted > free_bits) {
678 /* uhoh, window change time. */
679 status =
680 ocfs2_local_alloc_slide_window(osb, local_alloc_inode);
681 if (status < 0) {
682 if (status != -ENOSPC)
683 mlog_errno(status);
684 goto bail;
685 }
686
687 /*
688 * Under certain conditions, the window slide code
689 * might have reduced the number of bits available or
690 * disabled the the local alloc entirely. Re-check
691 * here and return -ENOSPC if necessary.
692 */
693 status = -ENOSPC;
694 if (!ocfs2_la_state_enabled(osb))
695 goto bail;
696
697 free_bits = le32_to_cpu(alloc->id1.bitmap1.i_total) -
698 le32_to_cpu(alloc->id1.bitmap1.i_used);
699 if (bits_wanted > free_bits)
700 goto bail;
701 }
702
703 ac->ac_inode = local_alloc_inode;
704 /* We should never use localalloc from another slot */
705 ac->ac_alloc_slot = osb->slot_num;
706 ac->ac_which = OCFS2_AC_USE_LOCAL;
707 get_bh(osb->local_alloc_bh);
708 ac->ac_bh = osb->local_alloc_bh;
709 status = 0;
710bail:
711 if (status < 0 && local_alloc_inode) {
712 inode_unlock(local_alloc_inode);
713 iput(local_alloc_inode);
714 }
715
716 trace_ocfs2_reserve_local_alloc_bits(
717 (unsigned long long)ac->ac_max_block,
718 bits_wanted, osb->slot_num, status);
719
720 if (status)
721 mlog_errno(status);
722 return status;
723}
724
725int ocfs2_claim_local_alloc_bits(struct ocfs2_super *osb,
726 handle_t *handle,
727 struct ocfs2_alloc_context *ac,
728 u32 bits_wanted,
729 u32 *bit_off,
730 u32 *num_bits)
731{
732 int status, start;
733 struct inode *local_alloc_inode;
734 void *bitmap;
735 struct ocfs2_dinode *alloc;
736 struct ocfs2_local_alloc *la;
737
738 BUG_ON(ac->ac_which != OCFS2_AC_USE_LOCAL);
739
740 local_alloc_inode = ac->ac_inode;
741 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
742 la = OCFS2_LOCAL_ALLOC(alloc);
743
744 start = ocfs2_local_alloc_find_clear_bits(osb, alloc, &bits_wanted,
745 ac->ac_resv);
746 if (start == -1) {
747 /* TODO: Shouldn't we just BUG here? */
748 status = -ENOSPC;
749 mlog_errno(status);
750 goto bail;
751 }
752
753 bitmap = la->la_bitmap;
754 *bit_off = le32_to_cpu(la->la_bm_off) + start;
755 *num_bits = bits_wanted;
756
757 status = ocfs2_journal_access_di(handle,
758 INODE_CACHE(local_alloc_inode),
759 osb->local_alloc_bh,
760 OCFS2_JOURNAL_ACCESS_WRITE);
761 if (status < 0) {
762 mlog_errno(status);
763 goto bail;
764 }
765
766 ocfs2_resmap_claimed_bits(&osb->osb_la_resmap, ac->ac_resv, start,
767 bits_wanted);
768
769 while(bits_wanted--)
770 ocfs2_set_bit(start++, bitmap);
771
772 le32_add_cpu(&alloc->id1.bitmap1.i_used, *num_bits);
773 ocfs2_journal_dirty(handle, osb->local_alloc_bh);
774
775bail:
776 if (status)
777 mlog_errno(status);
778 return status;
779}
780
781int ocfs2_free_local_alloc_bits(struct ocfs2_super *osb,
782 handle_t *handle,
783 struct ocfs2_alloc_context *ac,
784 u32 bit_off,
785 u32 num_bits)
786{
787 int status, start;
788 u32 clear_bits;
789 struct inode *local_alloc_inode;
790 void *bitmap;
791 struct ocfs2_dinode *alloc;
792 struct ocfs2_local_alloc *la;
793
794 BUG_ON(ac->ac_which != OCFS2_AC_USE_LOCAL);
795
796 local_alloc_inode = ac->ac_inode;
797 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
798 la = OCFS2_LOCAL_ALLOC(alloc);
799
800 bitmap = la->la_bitmap;
801 start = bit_off - le32_to_cpu(la->la_bm_off);
802 clear_bits = num_bits;
803
804 status = ocfs2_journal_access_di(handle,
805 INODE_CACHE(local_alloc_inode),
806 osb->local_alloc_bh,
807 OCFS2_JOURNAL_ACCESS_WRITE);
808 if (status < 0) {
809 mlog_errno(status);
810 goto bail;
811 }
812
813 while (clear_bits--)
814 ocfs2_clear_bit(start++, bitmap);
815
816 le32_add_cpu(&alloc->id1.bitmap1.i_used, -num_bits);
817 ocfs2_journal_dirty(handle, osb->local_alloc_bh);
818
819bail:
820 return status;
821}
822
823static u32 ocfs2_local_alloc_count_bits(struct ocfs2_dinode *alloc)
824{
825 u32 count;
826 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
827
828 count = memweight(la->la_bitmap, le16_to_cpu(la->la_size));
829
830 trace_ocfs2_local_alloc_count_bits(count);
831 return count;
832}
833
834static int ocfs2_local_alloc_find_clear_bits(struct ocfs2_super *osb,
835 struct ocfs2_dinode *alloc,
836 u32 *numbits,
837 struct ocfs2_alloc_reservation *resv)
838{
839 int numfound = 0, bitoff, left, startoff, lastzero;
840 int local_resv = 0;
841 struct ocfs2_alloc_reservation r;
842 void *bitmap = NULL;
843 struct ocfs2_reservation_map *resmap = &osb->osb_la_resmap;
844
845 if (!alloc->id1.bitmap1.i_total) {
846 bitoff = -1;
847 goto bail;
848 }
849
850 if (!resv) {
851 local_resv = 1;
852 ocfs2_resv_init_once(&r);
853 ocfs2_resv_set_type(&r, OCFS2_RESV_FLAG_TMP);
854 resv = &r;
855 }
856
857 numfound = *numbits;
858 if (ocfs2_resmap_resv_bits(resmap, resv, &bitoff, &numfound) == 0) {
859 if (numfound < *numbits)
860 *numbits = numfound;
861 goto bail;
862 }
863
864 /*
865 * Code error. While reservations are enabled, local
866 * allocation should _always_ go through them.
867 */
868 BUG_ON(osb->osb_resv_level != 0);
869
870 /*
871 * Reservations are disabled. Handle this the old way.
872 */
873
874 bitmap = OCFS2_LOCAL_ALLOC(alloc)->la_bitmap;
875
876 numfound = bitoff = startoff = 0;
877 lastzero = -1;
878 left = le32_to_cpu(alloc->id1.bitmap1.i_total);
879 while ((bitoff = ocfs2_find_next_zero_bit(bitmap, left, startoff)) != -1) {
880 if (bitoff == left) {
881 /* mlog(0, "bitoff (%d) == left", bitoff); */
882 break;
883 }
884 /* mlog(0, "Found a zero: bitoff = %d, startoff = %d, "
885 "numfound = %d\n", bitoff, startoff, numfound);*/
886
887 /* Ok, we found a zero bit... is it contig. or do we
888 * start over?*/
889 if (bitoff == startoff) {
890 /* we found a zero */
891 numfound++;
892 startoff++;
893 } else {
894 /* got a zero after some ones */
895 numfound = 1;
896 startoff = bitoff+1;
897 }
898 /* we got everything we needed */
899 if (numfound == *numbits) {
900 /* mlog(0, "Found it all!\n"); */
901 break;
902 }
903 }
904
905 trace_ocfs2_local_alloc_find_clear_bits_search_bitmap(bitoff, numfound);
906
907 if (numfound == *numbits)
908 bitoff = startoff - numfound;
909 else
910 bitoff = -1;
911
912bail:
913 if (local_resv)
914 ocfs2_resv_discard(resmap, resv);
915
916 trace_ocfs2_local_alloc_find_clear_bits(*numbits,
917 le32_to_cpu(alloc->id1.bitmap1.i_total),
918 bitoff, numfound);
919
920 return bitoff;
921}
922
923static void ocfs2_clear_local_alloc(struct ocfs2_dinode *alloc)
924{
925 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
926 int i;
927
928 alloc->id1.bitmap1.i_total = 0;
929 alloc->id1.bitmap1.i_used = 0;
930 la->la_bm_off = 0;
931 for(i = 0; i < le16_to_cpu(la->la_size); i++)
932 la->la_bitmap[i] = 0;
933}
934
935#if 0
936/* turn this on and uncomment below to aid debugging window shifts. */
937static void ocfs2_verify_zero_bits(unsigned long *bitmap,
938 unsigned int start,
939 unsigned int count)
940{
941 unsigned int tmp = count;
942 while(tmp--) {
943 if (ocfs2_test_bit(start + tmp, bitmap)) {
944 printk("ocfs2_verify_zero_bits: start = %u, count = "
945 "%u\n", start, count);
946 printk("ocfs2_verify_zero_bits: bit %u is set!",
947 start + tmp);
948 BUG();
949 }
950 }
951}
952#endif
953
954/*
955 * sync the local alloc to main bitmap.
956 *
957 * assumes you've already locked the main bitmap -- the bitmap inode
958 * passed is used for caching.
959 */
960static int ocfs2_sync_local_to_main(struct ocfs2_super *osb,
961 handle_t *handle,
962 struct ocfs2_dinode *alloc,
963 struct inode *main_bm_inode,
964 struct buffer_head *main_bm_bh)
965{
966 int status = 0;
967 int bit_off, left, count, start;
968 u64 la_start_blk;
969 u64 blkno;
970 void *bitmap;
971 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
972
973 trace_ocfs2_sync_local_to_main(
974 le32_to_cpu(alloc->id1.bitmap1.i_total),
975 le32_to_cpu(alloc->id1.bitmap1.i_used));
976
977 if (!alloc->id1.bitmap1.i_total) {
978 goto bail;
979 }
980
981 if (le32_to_cpu(alloc->id1.bitmap1.i_used) ==
982 le32_to_cpu(alloc->id1.bitmap1.i_total)) {
983 goto bail;
984 }
985
986 la_start_blk = ocfs2_clusters_to_blocks(osb->sb,
987 le32_to_cpu(la->la_bm_off));
988 bitmap = la->la_bitmap;
989 start = count = bit_off = 0;
990 left = le32_to_cpu(alloc->id1.bitmap1.i_total);
991
992 while ((bit_off = ocfs2_find_next_zero_bit(bitmap, left, start))
993 != -1) {
994 if ((bit_off < left) && (bit_off == start)) {
995 count++;
996 start++;
997 continue;
998 }
999 if (count) {
1000 blkno = la_start_blk +
1001 ocfs2_clusters_to_blocks(osb->sb,
1002 start - count);
1003
1004 trace_ocfs2_sync_local_to_main_free(
1005 count, start - count,
1006 (unsigned long long)la_start_blk,
1007 (unsigned long long)blkno);
1008
1009 status = ocfs2_release_clusters(handle,
1010 main_bm_inode,
1011 main_bm_bh, blkno,
1012 count);
1013 if (status < 0) {
1014 mlog_errno(status);
1015 goto bail;
1016 }
1017 }
1018 if (bit_off >= left)
1019 break;
1020 count = 1;
1021 start = bit_off + 1;
1022 }
1023
1024bail:
1025 if (status)
1026 mlog_errno(status);
1027 return status;
1028}
1029
1030enum ocfs2_la_event {
1031 OCFS2_LA_EVENT_SLIDE, /* Normal window slide. */
1032 OCFS2_LA_EVENT_FRAGMENTED, /* The global bitmap has
1033 * enough bits theoretically
1034 * free, but a contiguous
1035 * allocation could not be
1036 * found. */
1037 OCFS2_LA_EVENT_ENOSPC, /* Global bitmap doesn't have
1038 * enough bits free to satisfy
1039 * our request. */
1040};
1041#define OCFS2_LA_ENABLE_INTERVAL (30 * HZ)
1042/*
1043 * Given an event, calculate the size of our next local alloc window.
1044 *
1045 * This should always be called under i_mutex of the local alloc inode
1046 * so that local alloc disabling doesn't race with processes trying to
1047 * use the allocator.
1048 *
1049 * Returns the state which the local alloc was left in. This value can
1050 * be ignored by some paths.
1051 */
1052static int ocfs2_recalc_la_window(struct ocfs2_super *osb,
1053 enum ocfs2_la_event event)
1054{
1055 unsigned int bits;
1056 int state;
1057
1058 spin_lock(&osb->osb_lock);
1059 if (osb->local_alloc_state == OCFS2_LA_DISABLED) {
1060 WARN_ON_ONCE(osb->local_alloc_state == OCFS2_LA_DISABLED);
1061 goto out_unlock;
1062 }
1063
1064 /*
1065 * ENOSPC and fragmentation are treated similarly for now.
1066 */
1067 if (event == OCFS2_LA_EVENT_ENOSPC ||
1068 event == OCFS2_LA_EVENT_FRAGMENTED) {
1069 /*
1070 * We ran out of contiguous space in the primary
1071 * bitmap. Drastically reduce the number of bits used
1072 * by local alloc until we have to disable it.
1073 */
1074 bits = osb->local_alloc_bits >> 1;
1075 if (bits > ocfs2_megabytes_to_clusters(osb->sb, 1)) {
1076 /*
1077 * By setting state to THROTTLED, we'll keep
1078 * the number of local alloc bits used down
1079 * until an event occurs which would give us
1080 * reason to assume the bitmap situation might
1081 * have changed.
1082 */
1083 osb->local_alloc_state = OCFS2_LA_THROTTLED;
1084 osb->local_alloc_bits = bits;
1085 } else {
1086 osb->local_alloc_state = OCFS2_LA_DISABLED;
1087 }
1088 queue_delayed_work(osb->ocfs2_wq, &osb->la_enable_wq,
1089 OCFS2_LA_ENABLE_INTERVAL);
1090 goto out_unlock;
1091 }
1092
1093 /*
1094 * Don't increase the size of the local alloc window until we
1095 * know we might be able to fulfill the request. Otherwise, we
1096 * risk bouncing around the global bitmap during periods of
1097 * low space.
1098 */
1099 if (osb->local_alloc_state != OCFS2_LA_THROTTLED)
1100 osb->local_alloc_bits = osb->local_alloc_default_bits;
1101
1102out_unlock:
1103 state = osb->local_alloc_state;
1104 spin_unlock(&osb->osb_lock);
1105
1106 return state;
1107}
1108
1109static int ocfs2_local_alloc_reserve_for_window(struct ocfs2_super *osb,
1110 struct ocfs2_alloc_context **ac,
1111 struct inode **bitmap_inode,
1112 struct buffer_head **bitmap_bh)
1113{
1114 int status;
1115
1116 *ac = kzalloc(sizeof(struct ocfs2_alloc_context), GFP_KERNEL);
1117 if (!(*ac)) {
1118 status = -ENOMEM;
1119 mlog_errno(status);
1120 goto bail;
1121 }
1122
1123retry_enospc:
1124 (*ac)->ac_bits_wanted = osb->local_alloc_bits;
1125 status = ocfs2_reserve_cluster_bitmap_bits(osb, *ac);
1126 if (status == -ENOSPC) {
1127 if (ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_ENOSPC) ==
1128 OCFS2_LA_DISABLED)
1129 goto bail;
1130
1131 ocfs2_free_ac_resource(*ac);
1132 memset(*ac, 0, sizeof(struct ocfs2_alloc_context));
1133 goto retry_enospc;
1134 }
1135 if (status < 0) {
1136 mlog_errno(status);
1137 goto bail;
1138 }
1139
1140 *bitmap_inode = (*ac)->ac_inode;
1141 igrab(*bitmap_inode);
1142 *bitmap_bh = (*ac)->ac_bh;
1143 get_bh(*bitmap_bh);
1144 status = 0;
1145bail:
1146 if ((status < 0) && *ac) {
1147 ocfs2_free_alloc_context(*ac);
1148 *ac = NULL;
1149 }
1150
1151 if (status)
1152 mlog_errno(status);
1153 return status;
1154}
1155
1156/*
1157 * pass it the bitmap lock in lock_bh if you have it.
1158 */
1159static int ocfs2_local_alloc_new_window(struct ocfs2_super *osb,
1160 handle_t *handle,
1161 struct ocfs2_alloc_context *ac)
1162{
1163 int status = 0;
1164 u32 cluster_off, cluster_count;
1165 struct ocfs2_dinode *alloc = NULL;
1166 struct ocfs2_local_alloc *la;
1167
1168 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
1169 la = OCFS2_LOCAL_ALLOC(alloc);
1170
1171 trace_ocfs2_local_alloc_new_window(
1172 le32_to_cpu(alloc->id1.bitmap1.i_total),
1173 osb->local_alloc_bits);
1174
1175 /* Instruct the allocation code to try the most recently used
1176 * cluster group. We'll re-record the group used this pass
1177 * below. */
1178 ac->ac_last_group = osb->la_last_gd;
1179
1180 /* we used the generic suballoc reserve function, but we set
1181 * everything up nicely, so there's no reason why we can't use
1182 * the more specific cluster api to claim bits. */
1183 status = ocfs2_claim_clusters(handle, ac, osb->local_alloc_bits,
1184 &cluster_off, &cluster_count);
1185 if (status == -ENOSPC) {
1186retry_enospc:
1187 /*
1188 * Note: We could also try syncing the journal here to
1189 * allow use of any free bits which the current
1190 * transaction can't give us access to. --Mark
1191 */
1192 if (ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_FRAGMENTED) ==
1193 OCFS2_LA_DISABLED)
1194 goto bail;
1195
1196 ac->ac_bits_wanted = osb->local_alloc_bits;
1197 status = ocfs2_claim_clusters(handle, ac,
1198 osb->local_alloc_bits,
1199 &cluster_off,
1200 &cluster_count);
1201 if (status == -ENOSPC)
1202 goto retry_enospc;
1203 /*
1204 * We only shrunk the *minimum* number of in our
1205 * request - it's entirely possible that the allocator
1206 * might give us more than we asked for.
1207 */
1208 if (status == 0) {
1209 spin_lock(&osb->osb_lock);
1210 osb->local_alloc_bits = cluster_count;
1211 spin_unlock(&osb->osb_lock);
1212 }
1213 }
1214 if (status < 0) {
1215 if (status != -ENOSPC)
1216 mlog_errno(status);
1217 goto bail;
1218 }
1219
1220 osb->la_last_gd = ac->ac_last_group;
1221
1222 la->la_bm_off = cpu_to_le32(cluster_off);
1223 alloc->id1.bitmap1.i_total = cpu_to_le32(cluster_count);
1224 /* just in case... In the future when we find space ourselves,
1225 * we don't have to get all contiguous -- but we'll have to
1226 * set all previously used bits in bitmap and update
1227 * la_bits_set before setting the bits in the main bitmap. */
1228 alloc->id1.bitmap1.i_used = 0;
1229 memset(OCFS2_LOCAL_ALLOC(alloc)->la_bitmap, 0,
1230 le16_to_cpu(la->la_size));
1231
1232 ocfs2_resmap_restart(&osb->osb_la_resmap, cluster_count,
1233 OCFS2_LOCAL_ALLOC(alloc)->la_bitmap);
1234
1235 trace_ocfs2_local_alloc_new_window_result(
1236 OCFS2_LOCAL_ALLOC(alloc)->la_bm_off,
1237 le32_to_cpu(alloc->id1.bitmap1.i_total));
1238
1239bail:
1240 if (status)
1241 mlog_errno(status);
1242 return status;
1243}
1244
1245/* Note that we do *NOT* lock the local alloc inode here as
1246 * it's been locked already for us. */
1247static int ocfs2_local_alloc_slide_window(struct ocfs2_super *osb,
1248 struct inode *local_alloc_inode)
1249{
1250 int status = 0;
1251 struct buffer_head *main_bm_bh = NULL;
1252 struct inode *main_bm_inode = NULL;
1253 handle_t *handle = NULL;
1254 struct ocfs2_dinode *alloc;
1255 struct ocfs2_dinode *alloc_copy = NULL;
1256 struct ocfs2_alloc_context *ac = NULL;
1257
1258 ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_SLIDE);
1259
1260 /* This will lock the main bitmap for us. */
1261 status = ocfs2_local_alloc_reserve_for_window(osb,
1262 &ac,
1263 &main_bm_inode,
1264 &main_bm_bh);
1265 if (status < 0) {
1266 if (status != -ENOSPC)
1267 mlog_errno(status);
1268 goto bail;
1269 }
1270
1271 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
1272 if (IS_ERR(handle)) {
1273 status = PTR_ERR(handle);
1274 handle = NULL;
1275 mlog_errno(status);
1276 goto bail;
1277 }
1278
1279 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
1280
1281 /* We want to clear the local alloc before doing anything
1282 * else, so that if we error later during this operation,
1283 * local alloc shutdown won't try to double free main bitmap
1284 * bits. Make a copy so the sync function knows which bits to
1285 * free. */
1286 alloc_copy = kmalloc(osb->local_alloc_bh->b_size, GFP_NOFS);
1287 if (!alloc_copy) {
1288 status = -ENOMEM;
1289 mlog_errno(status);
1290 goto bail;
1291 }
1292 memcpy(alloc_copy, alloc, osb->local_alloc_bh->b_size);
1293
1294 status = ocfs2_journal_access_di(handle,
1295 INODE_CACHE(local_alloc_inode),
1296 osb->local_alloc_bh,
1297 OCFS2_JOURNAL_ACCESS_WRITE);
1298 if (status < 0) {
1299 mlog_errno(status);
1300 goto bail;
1301 }
1302
1303 ocfs2_clear_local_alloc(alloc);
1304 ocfs2_journal_dirty(handle, osb->local_alloc_bh);
1305
1306 status = ocfs2_sync_local_to_main(osb, handle, alloc_copy,
1307 main_bm_inode, main_bm_bh);
1308 if (status < 0) {
1309 mlog_errno(status);
1310 goto bail;
1311 }
1312
1313 status = ocfs2_local_alloc_new_window(osb, handle, ac);
1314 if (status < 0) {
1315 if (status != -ENOSPC)
1316 mlog_errno(status);
1317 goto bail;
1318 }
1319
1320 atomic_inc(&osb->alloc_stats.moves);
1321
1322bail:
1323 if (handle)
1324 ocfs2_commit_trans(osb, handle);
1325
1326 brelse(main_bm_bh);
1327
1328 iput(main_bm_inode);
1329 kfree(alloc_copy);
1330
1331 if (ac)
1332 ocfs2_free_alloc_context(ac);
1333
1334 if (status)
1335 mlog_errno(status);
1336 return status;
1337}
1338
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * localalloc.c
4 *
5 * Node local data allocation
6 *
7 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
8 */
9
10#include <linux/fs.h>
11#include <linux/types.h>
12#include <linux/slab.h>
13#include <linux/highmem.h>
14#include <linux/bitops.h>
15
16#include <cluster/masklog.h>
17
18#include "ocfs2.h"
19
20#include "alloc.h"
21#include "blockcheck.h"
22#include "dlmglue.h"
23#include "inode.h"
24#include "journal.h"
25#include "localalloc.h"
26#include "suballoc.h"
27#include "super.h"
28#include "sysfile.h"
29#include "ocfs2_trace.h"
30
31#include "buffer_head_io.h"
32
33#define OCFS2_LOCAL_ALLOC(dinode) (&((dinode)->id2.i_lab))
34
35static u32 ocfs2_local_alloc_count_bits(struct ocfs2_dinode *alloc);
36
37static int ocfs2_local_alloc_find_clear_bits(struct ocfs2_super *osb,
38 struct ocfs2_dinode *alloc,
39 u32 *numbits,
40 struct ocfs2_alloc_reservation *resv);
41
42static void ocfs2_clear_local_alloc(struct ocfs2_dinode *alloc);
43
44static int ocfs2_sync_local_to_main(struct ocfs2_super *osb,
45 handle_t *handle,
46 struct ocfs2_dinode *alloc,
47 struct inode *main_bm_inode,
48 struct buffer_head *main_bm_bh);
49
50static int ocfs2_local_alloc_reserve_for_window(struct ocfs2_super *osb,
51 struct ocfs2_alloc_context **ac,
52 struct inode **bitmap_inode,
53 struct buffer_head **bitmap_bh);
54
55static int ocfs2_local_alloc_new_window(struct ocfs2_super *osb,
56 handle_t *handle,
57 struct ocfs2_alloc_context *ac);
58
59static int ocfs2_local_alloc_slide_window(struct ocfs2_super *osb,
60 struct inode *local_alloc_inode);
61
62/*
63 * ocfs2_la_default_mb() - determine a default size, in megabytes of
64 * the local alloc.
65 *
66 * Generally, we'd like to pick as large a local alloc as
67 * possible. Performance on large workloads tends to scale
68 * proportionally to la size. In addition to that, the reservations
69 * code functions more efficiently as it can reserve more windows for
70 * write.
71 *
72 * Some things work against us when trying to choose a large local alloc:
73 *
74 * - We need to ensure our sizing is picked to leave enough space in
75 * group descriptors for other allocations (such as block groups,
76 * etc). Picking default sizes which are a multiple of 4 could help
77 * - block groups are allocated in 2mb and 4mb chunks.
78 *
79 * - Likewise, we don't want to starve other nodes of bits on small
80 * file systems. This can easily be taken care of by limiting our
81 * default to a reasonable size (256M) on larger cluster sizes.
82 *
83 * - Some file systems can't support very large sizes - 4k and 8k in
84 * particular are limited to less than 128 and 256 megabytes respectively.
85 *
86 * The following reference table shows group descriptor and local
87 * alloc maximums at various cluster sizes (4k blocksize)
88 *
89 * csize: 4K group: 126M la: 121M
90 * csize: 8K group: 252M la: 243M
91 * csize: 16K group: 504M la: 486M
92 * csize: 32K group: 1008M la: 972M
93 * csize: 64K group: 2016M la: 1944M
94 * csize: 128K group: 4032M la: 3888M
95 * csize: 256K group: 8064M la: 7776M
96 * csize: 512K group: 16128M la: 15552M
97 * csize: 1024K group: 32256M la: 31104M
98 */
99#define OCFS2_LA_MAX_DEFAULT_MB 256
100#define OCFS2_LA_OLD_DEFAULT 8
101unsigned int ocfs2_la_default_mb(struct ocfs2_super *osb)
102{
103 unsigned int la_mb;
104 unsigned int gd_mb;
105 unsigned int la_max_mb;
106 unsigned int megs_per_slot;
107 struct super_block *sb = osb->sb;
108
109 gd_mb = ocfs2_clusters_to_megabytes(osb->sb,
110 8 * ocfs2_group_bitmap_size(sb, 0, osb->s_feature_incompat));
111
112 /*
113 * This takes care of files systems with very small group
114 * descriptors - 512 byte blocksize at cluster sizes lower
115 * than 16K and also 1k blocksize with 4k cluster size.
116 */
117 if ((sb->s_blocksize == 512 && osb->s_clustersize <= 8192)
118 || (sb->s_blocksize == 1024 && osb->s_clustersize == 4096))
119 return OCFS2_LA_OLD_DEFAULT;
120
121 /*
122 * Leave enough room for some block groups and make the final
123 * value we work from a multiple of 4.
124 */
125 gd_mb -= 16;
126 gd_mb &= 0xFFFFFFFB;
127
128 la_mb = gd_mb;
129
130 /*
131 * Keep window sizes down to a reasonable default
132 */
133 if (la_mb > OCFS2_LA_MAX_DEFAULT_MB) {
134 /*
135 * Some clustersize / blocksize combinations will have
136 * given us a larger than OCFS2_LA_MAX_DEFAULT_MB
137 * default size, but get poor distribution when
138 * limited to exactly 256 megabytes.
139 *
140 * As an example, 16K clustersize at 4K blocksize
141 * gives us a cluster group size of 504M. Paring the
142 * local alloc size down to 256 however, would give us
143 * only one window and around 200MB left in the
144 * cluster group. Instead, find the first size below
145 * 256 which would give us an even distribution.
146 *
147 * Larger cluster group sizes actually work out pretty
148 * well when pared to 256, so we don't have to do this
149 * for any group that fits more than two
150 * OCFS2_LA_MAX_DEFAULT_MB windows.
151 */
152 if (gd_mb > (2 * OCFS2_LA_MAX_DEFAULT_MB))
153 la_mb = 256;
154 else {
155 unsigned int gd_mult = gd_mb;
156
157 while (gd_mult > 256)
158 gd_mult = gd_mult >> 1;
159
160 la_mb = gd_mult;
161 }
162 }
163
164 megs_per_slot = osb->osb_clusters_at_boot / osb->max_slots;
165 megs_per_slot = ocfs2_clusters_to_megabytes(osb->sb, megs_per_slot);
166 /* Too many nodes, too few disk clusters. */
167 if (megs_per_slot < la_mb)
168 la_mb = megs_per_slot;
169
170 /* We can't store more bits than we can in a block. */
171 la_max_mb = ocfs2_clusters_to_megabytes(osb->sb,
172 ocfs2_local_alloc_size(sb) * 8);
173 if (la_mb > la_max_mb)
174 la_mb = la_max_mb;
175
176 return la_mb;
177}
178
179void ocfs2_la_set_sizes(struct ocfs2_super *osb, int requested_mb)
180{
181 struct super_block *sb = osb->sb;
182 unsigned int la_default_mb = ocfs2_la_default_mb(osb);
183 unsigned int la_max_mb;
184
185 la_max_mb = ocfs2_clusters_to_megabytes(sb,
186 ocfs2_local_alloc_size(sb) * 8);
187
188 trace_ocfs2_la_set_sizes(requested_mb, la_max_mb, la_default_mb);
189
190 if (requested_mb == -1) {
191 /* No user request - use defaults */
192 osb->local_alloc_default_bits =
193 ocfs2_megabytes_to_clusters(sb, la_default_mb);
194 } else if (requested_mb > la_max_mb) {
195 /* Request is too big, we give the maximum available */
196 osb->local_alloc_default_bits =
197 ocfs2_megabytes_to_clusters(sb, la_max_mb);
198 } else {
199 osb->local_alloc_default_bits =
200 ocfs2_megabytes_to_clusters(sb, requested_mb);
201 }
202
203 osb->local_alloc_bits = osb->local_alloc_default_bits;
204}
205
206static inline int ocfs2_la_state_enabled(struct ocfs2_super *osb)
207{
208 return (osb->local_alloc_state == OCFS2_LA_THROTTLED ||
209 osb->local_alloc_state == OCFS2_LA_ENABLED);
210}
211
212void ocfs2_local_alloc_seen_free_bits(struct ocfs2_super *osb,
213 unsigned int num_clusters)
214{
215 spin_lock(&osb->osb_lock);
216 if (osb->local_alloc_state == OCFS2_LA_DISABLED ||
217 osb->local_alloc_state == OCFS2_LA_THROTTLED)
218 if (num_clusters >= osb->local_alloc_default_bits) {
219 cancel_delayed_work(&osb->la_enable_wq);
220 osb->local_alloc_state = OCFS2_LA_ENABLED;
221 }
222 spin_unlock(&osb->osb_lock);
223}
224
225void ocfs2_la_enable_worker(struct work_struct *work)
226{
227 struct ocfs2_super *osb =
228 container_of(work, struct ocfs2_super,
229 la_enable_wq.work);
230 spin_lock(&osb->osb_lock);
231 osb->local_alloc_state = OCFS2_LA_ENABLED;
232 spin_unlock(&osb->osb_lock);
233}
234
235/*
236 * Tell us whether a given allocation should use the local alloc
237 * file. Otherwise, it has to go to the main bitmap.
238 *
239 * This function does semi-dirty reads of local alloc size and state!
240 * This is ok however, as the values are re-checked once under mutex.
241 */
242int ocfs2_alloc_should_use_local(struct ocfs2_super *osb, u64 bits)
243{
244 int ret = 0;
245 int la_bits;
246
247 spin_lock(&osb->osb_lock);
248 la_bits = osb->local_alloc_bits;
249
250 if (!ocfs2_la_state_enabled(osb))
251 goto bail;
252
253 /* la_bits should be at least twice the size (in clusters) of
254 * a new block group. We want to be sure block group
255 * allocations go through the local alloc, so allow an
256 * allocation to take up to half the bitmap. */
257 if (bits > (la_bits / 2))
258 goto bail;
259
260 ret = 1;
261bail:
262 trace_ocfs2_alloc_should_use_local(
263 (unsigned long long)bits, osb->local_alloc_state, la_bits, ret);
264 spin_unlock(&osb->osb_lock);
265 return ret;
266}
267
268int ocfs2_load_local_alloc(struct ocfs2_super *osb)
269{
270 int status = 0;
271 struct ocfs2_dinode *alloc = NULL;
272 struct buffer_head *alloc_bh = NULL;
273 u32 num_used;
274 struct inode *inode = NULL;
275 struct ocfs2_local_alloc *la;
276
277 if (osb->local_alloc_bits == 0)
278 goto bail;
279
280 if (osb->local_alloc_bits >= osb->bitmap_cpg) {
281 mlog(ML_NOTICE, "Requested local alloc window %d is larger "
282 "than max possible %u. Using defaults.\n",
283 osb->local_alloc_bits, (osb->bitmap_cpg - 1));
284 osb->local_alloc_bits =
285 ocfs2_megabytes_to_clusters(osb->sb,
286 ocfs2_la_default_mb(osb));
287 }
288
289 /* read the alloc off disk */
290 inode = ocfs2_get_system_file_inode(osb, LOCAL_ALLOC_SYSTEM_INODE,
291 osb->slot_num);
292 if (!inode) {
293 status = -EINVAL;
294 mlog_errno(status);
295 goto bail;
296 }
297
298 status = ocfs2_read_inode_block_full(inode, &alloc_bh,
299 OCFS2_BH_IGNORE_CACHE);
300 if (status < 0) {
301 mlog_errno(status);
302 goto bail;
303 }
304
305 alloc = (struct ocfs2_dinode *) alloc_bh->b_data;
306 la = OCFS2_LOCAL_ALLOC(alloc);
307
308 if (!(le32_to_cpu(alloc->i_flags) &
309 (OCFS2_LOCAL_ALLOC_FL|OCFS2_BITMAP_FL))) {
310 mlog(ML_ERROR, "Invalid local alloc inode, %llu\n",
311 (unsigned long long)OCFS2_I(inode)->ip_blkno);
312 status = -EINVAL;
313 goto bail;
314 }
315
316 if ((la->la_size == 0) ||
317 (le16_to_cpu(la->la_size) > ocfs2_local_alloc_size(inode->i_sb))) {
318 mlog(ML_ERROR, "Local alloc size is invalid (la_size = %u)\n",
319 le16_to_cpu(la->la_size));
320 status = -EINVAL;
321 goto bail;
322 }
323
324 /* do a little verification. */
325 num_used = ocfs2_local_alloc_count_bits(alloc);
326
327 /* hopefully the local alloc has always been recovered before
328 * we load it. */
329 if (num_used
330 || alloc->id1.bitmap1.i_used
331 || alloc->id1.bitmap1.i_total
332 || la->la_bm_off) {
333 mlog(ML_ERROR, "inconsistent detected, clean journal with"
334 " unrecovered local alloc, please run fsck.ocfs2!\n"
335 "found = %u, set = %u, taken = %u, off = %u\n",
336 num_used, le32_to_cpu(alloc->id1.bitmap1.i_used),
337 le32_to_cpu(alloc->id1.bitmap1.i_total),
338 OCFS2_LOCAL_ALLOC(alloc)->la_bm_off);
339
340 status = -EINVAL;
341 goto bail;
342 }
343
344 osb->local_alloc_bh = alloc_bh;
345 osb->local_alloc_state = OCFS2_LA_ENABLED;
346
347bail:
348 if (status < 0)
349 brelse(alloc_bh);
350 iput(inode);
351
352 trace_ocfs2_load_local_alloc(osb->local_alloc_bits);
353
354 if (status)
355 mlog_errno(status);
356 return status;
357}
358
359/*
360 * return any unused bits to the bitmap and write out a clean
361 * local_alloc.
362 *
363 * local_alloc_bh is optional. If not passed, we will simply use the
364 * one off osb. If you do pass it however, be warned that it *will* be
365 * returned brelse'd and NULL'd out.*/
366void ocfs2_shutdown_local_alloc(struct ocfs2_super *osb)
367{
368 int status;
369 handle_t *handle;
370 struct inode *local_alloc_inode = NULL;
371 struct buffer_head *bh = NULL;
372 struct buffer_head *main_bm_bh = NULL;
373 struct inode *main_bm_inode = NULL;
374 struct ocfs2_dinode *alloc_copy = NULL;
375 struct ocfs2_dinode *alloc = NULL;
376
377 cancel_delayed_work(&osb->la_enable_wq);
378 if (osb->ocfs2_wq)
379 flush_workqueue(osb->ocfs2_wq);
380
381 if (osb->local_alloc_state == OCFS2_LA_UNUSED)
382 goto out;
383
384 local_alloc_inode =
385 ocfs2_get_system_file_inode(osb,
386 LOCAL_ALLOC_SYSTEM_INODE,
387 osb->slot_num);
388 if (!local_alloc_inode) {
389 status = -ENOENT;
390 mlog_errno(status);
391 goto out;
392 }
393
394 osb->local_alloc_state = OCFS2_LA_DISABLED;
395
396 ocfs2_resmap_uninit(&osb->osb_la_resmap);
397
398 main_bm_inode = ocfs2_get_system_file_inode(osb,
399 GLOBAL_BITMAP_SYSTEM_INODE,
400 OCFS2_INVALID_SLOT);
401 if (!main_bm_inode) {
402 status = -EINVAL;
403 mlog_errno(status);
404 goto out;
405 }
406
407 inode_lock(main_bm_inode);
408
409 status = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 1);
410 if (status < 0) {
411 mlog_errno(status);
412 goto out_mutex;
413 }
414
415 /* WINDOW_MOVE_CREDITS is a bit heavy... */
416 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
417 if (IS_ERR(handle)) {
418 mlog_errno(PTR_ERR(handle));
419 handle = NULL;
420 goto out_unlock;
421 }
422
423 bh = osb->local_alloc_bh;
424 alloc = (struct ocfs2_dinode *) bh->b_data;
425
426 alloc_copy = kmemdup(alloc, bh->b_size, GFP_NOFS);
427 if (!alloc_copy) {
428 status = -ENOMEM;
429 goto out_commit;
430 }
431
432 status = ocfs2_journal_access_di(handle, INODE_CACHE(local_alloc_inode),
433 bh, OCFS2_JOURNAL_ACCESS_WRITE);
434 if (status < 0) {
435 mlog_errno(status);
436 goto out_commit;
437 }
438
439 ocfs2_clear_local_alloc(alloc);
440 ocfs2_journal_dirty(handle, bh);
441
442 brelse(bh);
443 osb->local_alloc_bh = NULL;
444 osb->local_alloc_state = OCFS2_LA_UNUSED;
445
446 status = ocfs2_sync_local_to_main(osb, handle, alloc_copy,
447 main_bm_inode, main_bm_bh);
448 if (status < 0)
449 mlog_errno(status);
450
451out_commit:
452 ocfs2_commit_trans(osb, handle);
453
454out_unlock:
455 brelse(main_bm_bh);
456
457 ocfs2_inode_unlock(main_bm_inode, 1);
458
459out_mutex:
460 inode_unlock(main_bm_inode);
461 iput(main_bm_inode);
462
463out:
464 iput(local_alloc_inode);
465
466 kfree(alloc_copy);
467}
468
469/*
470 * We want to free the bitmap bits outside of any recovery context as
471 * we'll need a cluster lock to do so, but we must clear the local
472 * alloc before giving up the recovered nodes journal. To solve this,
473 * we kmalloc a copy of the local alloc before it's change for the
474 * caller to process with ocfs2_complete_local_alloc_recovery
475 */
476int ocfs2_begin_local_alloc_recovery(struct ocfs2_super *osb,
477 int slot_num,
478 struct ocfs2_dinode **alloc_copy)
479{
480 int status = 0;
481 struct buffer_head *alloc_bh = NULL;
482 struct inode *inode = NULL;
483 struct ocfs2_dinode *alloc;
484
485 trace_ocfs2_begin_local_alloc_recovery(slot_num);
486
487 *alloc_copy = NULL;
488
489 inode = ocfs2_get_system_file_inode(osb,
490 LOCAL_ALLOC_SYSTEM_INODE,
491 slot_num);
492 if (!inode) {
493 status = -EINVAL;
494 mlog_errno(status);
495 goto bail;
496 }
497
498 inode_lock(inode);
499
500 status = ocfs2_read_inode_block_full(inode, &alloc_bh,
501 OCFS2_BH_IGNORE_CACHE);
502 if (status < 0) {
503 mlog_errno(status);
504 goto bail;
505 }
506
507 *alloc_copy = kmalloc(alloc_bh->b_size, GFP_KERNEL);
508 if (!(*alloc_copy)) {
509 status = -ENOMEM;
510 goto bail;
511 }
512 memcpy((*alloc_copy), alloc_bh->b_data, alloc_bh->b_size);
513
514 alloc = (struct ocfs2_dinode *) alloc_bh->b_data;
515 ocfs2_clear_local_alloc(alloc);
516
517 ocfs2_compute_meta_ecc(osb->sb, alloc_bh->b_data, &alloc->i_check);
518 status = ocfs2_write_block(osb, alloc_bh, INODE_CACHE(inode));
519 if (status < 0)
520 mlog_errno(status);
521
522bail:
523 if (status < 0) {
524 kfree(*alloc_copy);
525 *alloc_copy = NULL;
526 }
527
528 brelse(alloc_bh);
529
530 if (inode) {
531 inode_unlock(inode);
532 iput(inode);
533 }
534
535 if (status)
536 mlog_errno(status);
537 return status;
538}
539
540/*
541 * Step 2: By now, we've completed the journal recovery, we've stamped
542 * a clean local alloc on disk and dropped the node out of the
543 * recovery map. Dlm locks will no longer stall, so lets clear out the
544 * main bitmap.
545 */
546int ocfs2_complete_local_alloc_recovery(struct ocfs2_super *osb,
547 struct ocfs2_dinode *alloc)
548{
549 int status;
550 handle_t *handle;
551 struct buffer_head *main_bm_bh = NULL;
552 struct inode *main_bm_inode;
553
554 main_bm_inode = ocfs2_get_system_file_inode(osb,
555 GLOBAL_BITMAP_SYSTEM_INODE,
556 OCFS2_INVALID_SLOT);
557 if (!main_bm_inode) {
558 status = -EINVAL;
559 mlog_errno(status);
560 goto out;
561 }
562
563 inode_lock(main_bm_inode);
564
565 status = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 1);
566 if (status < 0) {
567 mlog_errno(status);
568 goto out_mutex;
569 }
570
571 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
572 if (IS_ERR(handle)) {
573 status = PTR_ERR(handle);
574 handle = NULL;
575 mlog_errno(status);
576 goto out_unlock;
577 }
578
579 /* we want the bitmap change to be recorded on disk asap */
580 handle->h_sync = 1;
581
582 status = ocfs2_sync_local_to_main(osb, handle, alloc,
583 main_bm_inode, main_bm_bh);
584 if (status < 0)
585 mlog_errno(status);
586
587 ocfs2_commit_trans(osb, handle);
588
589out_unlock:
590 ocfs2_inode_unlock(main_bm_inode, 1);
591
592out_mutex:
593 inode_unlock(main_bm_inode);
594
595 brelse(main_bm_bh);
596
597 iput(main_bm_inode);
598
599out:
600 if (!status)
601 ocfs2_init_steal_slots(osb);
602 if (status)
603 mlog_errno(status);
604 return status;
605}
606
607/*
608 * make sure we've got at least bits_wanted contiguous bits in the
609 * local alloc. You lose them when you drop i_rwsem.
610 *
611 * We will add ourselves to the transaction passed in, but may start
612 * our own in order to shift windows.
613 */
614int ocfs2_reserve_local_alloc_bits(struct ocfs2_super *osb,
615 u32 bits_wanted,
616 struct ocfs2_alloc_context *ac)
617{
618 int status;
619 struct ocfs2_dinode *alloc;
620 struct inode *local_alloc_inode;
621 unsigned int free_bits;
622
623 BUG_ON(!ac);
624
625 local_alloc_inode =
626 ocfs2_get_system_file_inode(osb,
627 LOCAL_ALLOC_SYSTEM_INODE,
628 osb->slot_num);
629 if (!local_alloc_inode) {
630 status = -ENOENT;
631 mlog_errno(status);
632 goto bail;
633 }
634
635 inode_lock(local_alloc_inode);
636
637 /*
638 * We must double check state and allocator bits because
639 * another process may have changed them while holding i_rwsem.
640 */
641 spin_lock(&osb->osb_lock);
642 if (!ocfs2_la_state_enabled(osb) ||
643 (bits_wanted > osb->local_alloc_bits)) {
644 spin_unlock(&osb->osb_lock);
645 status = -ENOSPC;
646 goto bail;
647 }
648 spin_unlock(&osb->osb_lock);
649
650 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
651
652#ifdef CONFIG_OCFS2_DEBUG_FS
653 if (le32_to_cpu(alloc->id1.bitmap1.i_used) !=
654 ocfs2_local_alloc_count_bits(alloc)) {
655 status = ocfs2_error(osb->sb, "local alloc inode %llu says it has %u used bits, but a count shows %u\n",
656 (unsigned long long)le64_to_cpu(alloc->i_blkno),
657 le32_to_cpu(alloc->id1.bitmap1.i_used),
658 ocfs2_local_alloc_count_bits(alloc));
659 goto bail;
660 }
661#endif
662
663 free_bits = le32_to_cpu(alloc->id1.bitmap1.i_total) -
664 le32_to_cpu(alloc->id1.bitmap1.i_used);
665 if (bits_wanted > free_bits) {
666 /* uhoh, window change time. */
667 status =
668 ocfs2_local_alloc_slide_window(osb, local_alloc_inode);
669 if (status < 0) {
670 if (status != -ENOSPC)
671 mlog_errno(status);
672 goto bail;
673 }
674
675 /*
676 * Under certain conditions, the window slide code
677 * might have reduced the number of bits available or
678 * disabled the local alloc entirely. Re-check
679 * here and return -ENOSPC if necessary.
680 */
681 status = -ENOSPC;
682 if (!ocfs2_la_state_enabled(osb))
683 goto bail;
684
685 free_bits = le32_to_cpu(alloc->id1.bitmap1.i_total) -
686 le32_to_cpu(alloc->id1.bitmap1.i_used);
687 if (bits_wanted > free_bits)
688 goto bail;
689 }
690
691 ac->ac_inode = local_alloc_inode;
692 /* We should never use localalloc from another slot */
693 ac->ac_alloc_slot = osb->slot_num;
694 ac->ac_which = OCFS2_AC_USE_LOCAL;
695 get_bh(osb->local_alloc_bh);
696 ac->ac_bh = osb->local_alloc_bh;
697 status = 0;
698bail:
699 if (status < 0 && local_alloc_inode) {
700 inode_unlock(local_alloc_inode);
701 iput(local_alloc_inode);
702 }
703
704 trace_ocfs2_reserve_local_alloc_bits(
705 (unsigned long long)ac->ac_max_block,
706 bits_wanted, osb->slot_num, status);
707
708 if (status)
709 mlog_errno(status);
710 return status;
711}
712
713int ocfs2_claim_local_alloc_bits(struct ocfs2_super *osb,
714 handle_t *handle,
715 struct ocfs2_alloc_context *ac,
716 u32 bits_wanted,
717 u32 *bit_off,
718 u32 *num_bits)
719{
720 int status, start;
721 struct inode *local_alloc_inode;
722 void *bitmap;
723 struct ocfs2_dinode *alloc;
724 struct ocfs2_local_alloc *la;
725
726 BUG_ON(ac->ac_which != OCFS2_AC_USE_LOCAL);
727
728 local_alloc_inode = ac->ac_inode;
729 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
730 la = OCFS2_LOCAL_ALLOC(alloc);
731
732 start = ocfs2_local_alloc_find_clear_bits(osb, alloc, &bits_wanted,
733 ac->ac_resv);
734 if (start == -1) {
735 /* TODO: Shouldn't we just BUG here? */
736 status = -ENOSPC;
737 mlog_errno(status);
738 goto bail;
739 }
740
741 bitmap = la->la_bitmap;
742 *bit_off = le32_to_cpu(la->la_bm_off) + start;
743 *num_bits = bits_wanted;
744
745 status = ocfs2_journal_access_di(handle,
746 INODE_CACHE(local_alloc_inode),
747 osb->local_alloc_bh,
748 OCFS2_JOURNAL_ACCESS_WRITE);
749 if (status < 0) {
750 mlog_errno(status);
751 goto bail;
752 }
753
754 ocfs2_resmap_claimed_bits(&osb->osb_la_resmap, ac->ac_resv, start,
755 bits_wanted);
756
757 while(bits_wanted--)
758 ocfs2_set_bit(start++, bitmap);
759
760 le32_add_cpu(&alloc->id1.bitmap1.i_used, *num_bits);
761 ocfs2_journal_dirty(handle, osb->local_alloc_bh);
762
763bail:
764 if (status)
765 mlog_errno(status);
766 return status;
767}
768
769int ocfs2_free_local_alloc_bits(struct ocfs2_super *osb,
770 handle_t *handle,
771 struct ocfs2_alloc_context *ac,
772 u32 bit_off,
773 u32 num_bits)
774{
775 int status, start;
776 u32 clear_bits;
777 struct inode *local_alloc_inode;
778 void *bitmap;
779 struct ocfs2_dinode *alloc;
780 struct ocfs2_local_alloc *la;
781
782 BUG_ON(ac->ac_which != OCFS2_AC_USE_LOCAL);
783
784 local_alloc_inode = ac->ac_inode;
785 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
786 la = OCFS2_LOCAL_ALLOC(alloc);
787
788 bitmap = la->la_bitmap;
789 start = bit_off - le32_to_cpu(la->la_bm_off);
790 clear_bits = num_bits;
791
792 status = ocfs2_journal_access_di(handle,
793 INODE_CACHE(local_alloc_inode),
794 osb->local_alloc_bh,
795 OCFS2_JOURNAL_ACCESS_WRITE);
796 if (status < 0) {
797 mlog_errno(status);
798 goto bail;
799 }
800
801 while (clear_bits--)
802 ocfs2_clear_bit(start++, bitmap);
803
804 le32_add_cpu(&alloc->id1.bitmap1.i_used, -num_bits);
805 ocfs2_journal_dirty(handle, osb->local_alloc_bh);
806
807bail:
808 return status;
809}
810
811static u32 ocfs2_local_alloc_count_bits(struct ocfs2_dinode *alloc)
812{
813 u32 count;
814 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
815
816 count = memweight(la->la_bitmap, le16_to_cpu(la->la_size));
817
818 trace_ocfs2_local_alloc_count_bits(count);
819 return count;
820}
821
822static int ocfs2_local_alloc_find_clear_bits(struct ocfs2_super *osb,
823 struct ocfs2_dinode *alloc,
824 u32 *numbits,
825 struct ocfs2_alloc_reservation *resv)
826{
827 int numfound = 0, bitoff, left, startoff;
828 int local_resv = 0;
829 struct ocfs2_alloc_reservation r;
830 void *bitmap = NULL;
831 struct ocfs2_reservation_map *resmap = &osb->osb_la_resmap;
832
833 if (!alloc->id1.bitmap1.i_total) {
834 bitoff = -1;
835 goto bail;
836 }
837
838 if (!resv) {
839 local_resv = 1;
840 ocfs2_resv_init_once(&r);
841 ocfs2_resv_set_type(&r, OCFS2_RESV_FLAG_TMP);
842 resv = &r;
843 }
844
845 numfound = *numbits;
846 if (ocfs2_resmap_resv_bits(resmap, resv, &bitoff, &numfound) == 0) {
847 if (numfound < *numbits)
848 *numbits = numfound;
849 goto bail;
850 }
851
852 /*
853 * Code error. While reservations are enabled, local
854 * allocation should _always_ go through them.
855 */
856 BUG_ON(osb->osb_resv_level != 0);
857
858 /*
859 * Reservations are disabled. Handle this the old way.
860 */
861
862 bitmap = OCFS2_LOCAL_ALLOC(alloc)->la_bitmap;
863
864 numfound = bitoff = startoff = 0;
865 left = le32_to_cpu(alloc->id1.bitmap1.i_total);
866 while ((bitoff = ocfs2_find_next_zero_bit(bitmap, left, startoff)) <
867 left) {
868 /* Ok, we found a zero bit... is it contig. or do we
869 * start over?*/
870 if (bitoff == startoff) {
871 /* we found a zero */
872 numfound++;
873 startoff++;
874 } else {
875 /* got a zero after some ones */
876 numfound = 1;
877 startoff = bitoff+1;
878 }
879 /* we got everything we needed */
880 if (numfound == *numbits) {
881 /* mlog(0, "Found it all!\n"); */
882 break;
883 }
884 }
885
886 trace_ocfs2_local_alloc_find_clear_bits_search_bitmap(bitoff, numfound);
887
888 if (numfound == *numbits)
889 bitoff = startoff - numfound;
890 else
891 bitoff = -1;
892
893bail:
894 if (local_resv)
895 ocfs2_resv_discard(resmap, resv);
896
897 trace_ocfs2_local_alloc_find_clear_bits(*numbits,
898 le32_to_cpu(alloc->id1.bitmap1.i_total),
899 bitoff, numfound);
900
901 return bitoff;
902}
903
904static void ocfs2_clear_local_alloc(struct ocfs2_dinode *alloc)
905{
906 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
907 int i;
908
909 alloc->id1.bitmap1.i_total = 0;
910 alloc->id1.bitmap1.i_used = 0;
911 la->la_bm_off = 0;
912 for(i = 0; i < le16_to_cpu(la->la_size); i++)
913 la->la_bitmap[i] = 0;
914}
915
916#if 0
917/* turn this on and uncomment below to aid debugging window shifts. */
918static void ocfs2_verify_zero_bits(unsigned long *bitmap,
919 unsigned int start,
920 unsigned int count)
921{
922 unsigned int tmp = count;
923 while(tmp--) {
924 if (ocfs2_test_bit(start + tmp, bitmap)) {
925 printk("ocfs2_verify_zero_bits: start = %u, count = "
926 "%u\n", start, count);
927 printk("ocfs2_verify_zero_bits: bit %u is set!",
928 start + tmp);
929 BUG();
930 }
931 }
932}
933#endif
934
935/*
936 * sync the local alloc to main bitmap.
937 *
938 * assumes you've already locked the main bitmap -- the bitmap inode
939 * passed is used for caching.
940 */
941static int ocfs2_sync_local_to_main(struct ocfs2_super *osb,
942 handle_t *handle,
943 struct ocfs2_dinode *alloc,
944 struct inode *main_bm_inode,
945 struct buffer_head *main_bm_bh)
946{
947 int status = 0;
948 int bit_off, left, count, start;
949 u64 la_start_blk;
950 u64 blkno;
951 void *bitmap;
952 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
953
954 trace_ocfs2_sync_local_to_main(
955 le32_to_cpu(alloc->id1.bitmap1.i_total),
956 le32_to_cpu(alloc->id1.bitmap1.i_used));
957
958 if (!alloc->id1.bitmap1.i_total) {
959 goto bail;
960 }
961
962 if (le32_to_cpu(alloc->id1.bitmap1.i_used) ==
963 le32_to_cpu(alloc->id1.bitmap1.i_total)) {
964 goto bail;
965 }
966
967 la_start_blk = ocfs2_clusters_to_blocks(osb->sb,
968 le32_to_cpu(la->la_bm_off));
969 bitmap = la->la_bitmap;
970 start = count = 0;
971 left = le32_to_cpu(alloc->id1.bitmap1.i_total);
972
973 while ((bit_off = ocfs2_find_next_zero_bit(bitmap, left, start)) <
974 left) {
975 if (bit_off == start) {
976 count++;
977 start++;
978 continue;
979 }
980 if (count) {
981 blkno = la_start_blk +
982 ocfs2_clusters_to_blocks(osb->sb,
983 start - count);
984
985 trace_ocfs2_sync_local_to_main_free(
986 count, start - count,
987 (unsigned long long)la_start_blk,
988 (unsigned long long)blkno);
989
990 status = ocfs2_release_clusters(handle,
991 main_bm_inode,
992 main_bm_bh, blkno,
993 count);
994 if (status < 0) {
995 mlog_errno(status);
996 goto bail;
997 }
998 }
999
1000 count = 1;
1001 start = bit_off + 1;
1002 }
1003
1004bail:
1005 if (status)
1006 mlog_errno(status);
1007 return status;
1008}
1009
1010enum ocfs2_la_event {
1011 OCFS2_LA_EVENT_SLIDE, /* Normal window slide. */
1012 OCFS2_LA_EVENT_FRAGMENTED, /* The global bitmap has
1013 * enough bits theoretically
1014 * free, but a contiguous
1015 * allocation could not be
1016 * found. */
1017 OCFS2_LA_EVENT_ENOSPC, /* Global bitmap doesn't have
1018 * enough bits free to satisfy
1019 * our request. */
1020};
1021#define OCFS2_LA_ENABLE_INTERVAL (30 * HZ)
1022/*
1023 * Given an event, calculate the size of our next local alloc window.
1024 *
1025 * This should always be called under i_rwsem of the local alloc inode
1026 * so that local alloc disabling doesn't race with processes trying to
1027 * use the allocator.
1028 *
1029 * Returns the state which the local alloc was left in. This value can
1030 * be ignored by some paths.
1031 */
1032static int ocfs2_recalc_la_window(struct ocfs2_super *osb,
1033 enum ocfs2_la_event event)
1034{
1035 unsigned int bits;
1036 int state;
1037
1038 spin_lock(&osb->osb_lock);
1039 if (osb->local_alloc_state == OCFS2_LA_DISABLED) {
1040 WARN_ON_ONCE(osb->local_alloc_state == OCFS2_LA_DISABLED);
1041 goto out_unlock;
1042 }
1043
1044 /*
1045 * ENOSPC and fragmentation are treated similarly for now.
1046 */
1047 if (event == OCFS2_LA_EVENT_ENOSPC ||
1048 event == OCFS2_LA_EVENT_FRAGMENTED) {
1049 /*
1050 * We ran out of contiguous space in the primary
1051 * bitmap. Drastically reduce the number of bits used
1052 * by local alloc until we have to disable it.
1053 */
1054 bits = osb->local_alloc_bits >> 1;
1055 if (bits > ocfs2_megabytes_to_clusters(osb->sb, 1)) {
1056 /*
1057 * By setting state to THROTTLED, we'll keep
1058 * the number of local alloc bits used down
1059 * until an event occurs which would give us
1060 * reason to assume the bitmap situation might
1061 * have changed.
1062 */
1063 osb->local_alloc_state = OCFS2_LA_THROTTLED;
1064 osb->local_alloc_bits = bits;
1065 } else {
1066 osb->local_alloc_state = OCFS2_LA_DISABLED;
1067 }
1068 queue_delayed_work(osb->ocfs2_wq, &osb->la_enable_wq,
1069 OCFS2_LA_ENABLE_INTERVAL);
1070 goto out_unlock;
1071 }
1072
1073 /*
1074 * Don't increase the size of the local alloc window until we
1075 * know we might be able to fulfill the request. Otherwise, we
1076 * risk bouncing around the global bitmap during periods of
1077 * low space.
1078 */
1079 if (osb->local_alloc_state != OCFS2_LA_THROTTLED)
1080 osb->local_alloc_bits = osb->local_alloc_default_bits;
1081
1082out_unlock:
1083 state = osb->local_alloc_state;
1084 spin_unlock(&osb->osb_lock);
1085
1086 return state;
1087}
1088
1089static int ocfs2_local_alloc_reserve_for_window(struct ocfs2_super *osb,
1090 struct ocfs2_alloc_context **ac,
1091 struct inode **bitmap_inode,
1092 struct buffer_head **bitmap_bh)
1093{
1094 int status;
1095
1096 *ac = kzalloc(sizeof(struct ocfs2_alloc_context), GFP_KERNEL);
1097 if (!(*ac)) {
1098 status = -ENOMEM;
1099 mlog_errno(status);
1100 goto bail;
1101 }
1102
1103retry_enospc:
1104 (*ac)->ac_bits_wanted = osb->local_alloc_bits;
1105 status = ocfs2_reserve_cluster_bitmap_bits(osb, *ac);
1106 if (status == -ENOSPC) {
1107 if (ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_ENOSPC) ==
1108 OCFS2_LA_DISABLED)
1109 goto bail;
1110
1111 ocfs2_free_ac_resource(*ac);
1112 memset(*ac, 0, sizeof(struct ocfs2_alloc_context));
1113 goto retry_enospc;
1114 }
1115 if (status < 0) {
1116 mlog_errno(status);
1117 goto bail;
1118 }
1119
1120 *bitmap_inode = (*ac)->ac_inode;
1121 igrab(*bitmap_inode);
1122 *bitmap_bh = (*ac)->ac_bh;
1123 get_bh(*bitmap_bh);
1124 status = 0;
1125bail:
1126 if ((status < 0) && *ac) {
1127 ocfs2_free_alloc_context(*ac);
1128 *ac = NULL;
1129 }
1130
1131 if (status)
1132 mlog_errno(status);
1133 return status;
1134}
1135
1136/*
1137 * pass it the bitmap lock in lock_bh if you have it.
1138 */
1139static int ocfs2_local_alloc_new_window(struct ocfs2_super *osb,
1140 handle_t *handle,
1141 struct ocfs2_alloc_context *ac)
1142{
1143 int status = 0;
1144 u32 cluster_off, cluster_count;
1145 struct ocfs2_dinode *alloc = NULL;
1146 struct ocfs2_local_alloc *la;
1147
1148 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
1149 la = OCFS2_LOCAL_ALLOC(alloc);
1150
1151 trace_ocfs2_local_alloc_new_window(
1152 le32_to_cpu(alloc->id1.bitmap1.i_total),
1153 osb->local_alloc_bits);
1154
1155 /* Instruct the allocation code to try the most recently used
1156 * cluster group. We'll re-record the group used this pass
1157 * below. */
1158 ac->ac_last_group = osb->la_last_gd;
1159
1160 /* we used the generic suballoc reserve function, but we set
1161 * everything up nicely, so there's no reason why we can't use
1162 * the more specific cluster api to claim bits. */
1163 status = ocfs2_claim_clusters(handle, ac, osb->local_alloc_bits,
1164 &cluster_off, &cluster_count);
1165 if (status == -ENOSPC) {
1166retry_enospc:
1167 /*
1168 * Note: We could also try syncing the journal here to
1169 * allow use of any free bits which the current
1170 * transaction can't give us access to. --Mark
1171 */
1172 if (ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_FRAGMENTED) ==
1173 OCFS2_LA_DISABLED)
1174 goto bail;
1175
1176 ac->ac_bits_wanted = osb->local_alloc_bits;
1177 status = ocfs2_claim_clusters(handle, ac,
1178 osb->local_alloc_bits,
1179 &cluster_off,
1180 &cluster_count);
1181 if (status == -ENOSPC)
1182 goto retry_enospc;
1183 /*
1184 * We only shrunk the *minimum* number of in our
1185 * request - it's entirely possible that the allocator
1186 * might give us more than we asked for.
1187 */
1188 if (status == 0) {
1189 spin_lock(&osb->osb_lock);
1190 osb->local_alloc_bits = cluster_count;
1191 spin_unlock(&osb->osb_lock);
1192 }
1193 }
1194 if (status < 0) {
1195 if (status != -ENOSPC)
1196 mlog_errno(status);
1197 goto bail;
1198 }
1199
1200 osb->la_last_gd = ac->ac_last_group;
1201
1202 la->la_bm_off = cpu_to_le32(cluster_off);
1203 alloc->id1.bitmap1.i_total = cpu_to_le32(cluster_count);
1204 /* just in case... In the future when we find space ourselves,
1205 * we don't have to get all contiguous -- but we'll have to
1206 * set all previously used bits in bitmap and update
1207 * la_bits_set before setting the bits in the main bitmap. */
1208 alloc->id1.bitmap1.i_used = 0;
1209 memset(OCFS2_LOCAL_ALLOC(alloc)->la_bitmap, 0,
1210 le16_to_cpu(la->la_size));
1211
1212 ocfs2_resmap_restart(&osb->osb_la_resmap, cluster_count,
1213 OCFS2_LOCAL_ALLOC(alloc)->la_bitmap);
1214
1215 trace_ocfs2_local_alloc_new_window_result(
1216 OCFS2_LOCAL_ALLOC(alloc)->la_bm_off,
1217 le32_to_cpu(alloc->id1.bitmap1.i_total));
1218
1219bail:
1220 if (status)
1221 mlog_errno(status);
1222 return status;
1223}
1224
1225/* Note that we do *NOT* lock the local alloc inode here as
1226 * it's been locked already for us. */
1227static int ocfs2_local_alloc_slide_window(struct ocfs2_super *osb,
1228 struct inode *local_alloc_inode)
1229{
1230 int status = 0;
1231 struct buffer_head *main_bm_bh = NULL;
1232 struct inode *main_bm_inode = NULL;
1233 handle_t *handle = NULL;
1234 struct ocfs2_dinode *alloc;
1235 struct ocfs2_dinode *alloc_copy = NULL;
1236 struct ocfs2_alloc_context *ac = NULL;
1237
1238 ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_SLIDE);
1239
1240 /* This will lock the main bitmap for us. */
1241 status = ocfs2_local_alloc_reserve_for_window(osb,
1242 &ac,
1243 &main_bm_inode,
1244 &main_bm_bh);
1245 if (status < 0) {
1246 if (status != -ENOSPC)
1247 mlog_errno(status);
1248 goto bail;
1249 }
1250
1251 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
1252 if (IS_ERR(handle)) {
1253 status = PTR_ERR(handle);
1254 handle = NULL;
1255 mlog_errno(status);
1256 goto bail;
1257 }
1258
1259 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
1260
1261 /* We want to clear the local alloc before doing anything
1262 * else, so that if we error later during this operation,
1263 * local alloc shutdown won't try to double free main bitmap
1264 * bits. Make a copy so the sync function knows which bits to
1265 * free. */
1266 alloc_copy = kmemdup(alloc, osb->local_alloc_bh->b_size, GFP_NOFS);
1267 if (!alloc_copy) {
1268 status = -ENOMEM;
1269 mlog_errno(status);
1270 goto bail;
1271 }
1272
1273 status = ocfs2_journal_access_di(handle,
1274 INODE_CACHE(local_alloc_inode),
1275 osb->local_alloc_bh,
1276 OCFS2_JOURNAL_ACCESS_WRITE);
1277 if (status < 0) {
1278 mlog_errno(status);
1279 goto bail;
1280 }
1281
1282 ocfs2_clear_local_alloc(alloc);
1283 ocfs2_journal_dirty(handle, osb->local_alloc_bh);
1284
1285 status = ocfs2_sync_local_to_main(osb, handle, alloc_copy,
1286 main_bm_inode, main_bm_bh);
1287 if (status < 0) {
1288 mlog_errno(status);
1289 goto bail;
1290 }
1291
1292 status = ocfs2_local_alloc_new_window(osb, handle, ac);
1293 if (status < 0) {
1294 if (status != -ENOSPC)
1295 mlog_errno(status);
1296 goto bail;
1297 }
1298
1299 atomic_inc(&osb->alloc_stats.moves);
1300
1301bail:
1302 if (handle)
1303 ocfs2_commit_trans(osb, handle);
1304
1305 brelse(main_bm_bh);
1306
1307 iput(main_bm_inode);
1308 kfree(alloc_copy);
1309
1310 if (ac)
1311 ocfs2_free_alloc_context(ac);
1312
1313 if (status)
1314 mlog_errno(status);
1315 return status;
1316}
1317