Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/* -*- mode: c; c-basic-offset: 8; -*-
3 * vim: noexpandtab sw=8 ts=8 sts=0:
4 *
5 * localalloc.c
6 *
7 * Node local data allocation
8 *
9 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
10 */
11
12#include <linux/fs.h>
13#include <linux/types.h>
14#include <linux/slab.h>
15#include <linux/highmem.h>
16#include <linux/bitops.h>
17
18#include <cluster/masklog.h>
19
20#include "ocfs2.h"
21
22#include "alloc.h"
23#include "blockcheck.h"
24#include "dlmglue.h"
25#include "inode.h"
26#include "journal.h"
27#include "localalloc.h"
28#include "suballoc.h"
29#include "super.h"
30#include "sysfile.h"
31#include "ocfs2_trace.h"
32
33#include "buffer_head_io.h"
34
35#define OCFS2_LOCAL_ALLOC(dinode) (&((dinode)->id2.i_lab))
36
37static u32 ocfs2_local_alloc_count_bits(struct ocfs2_dinode *alloc);
38
39static int ocfs2_local_alloc_find_clear_bits(struct ocfs2_super *osb,
40 struct ocfs2_dinode *alloc,
41 u32 *numbits,
42 struct ocfs2_alloc_reservation *resv);
43
44static void ocfs2_clear_local_alloc(struct ocfs2_dinode *alloc);
45
46static int ocfs2_sync_local_to_main(struct ocfs2_super *osb,
47 handle_t *handle,
48 struct ocfs2_dinode *alloc,
49 struct inode *main_bm_inode,
50 struct buffer_head *main_bm_bh);
51
52static int ocfs2_local_alloc_reserve_for_window(struct ocfs2_super *osb,
53 struct ocfs2_alloc_context **ac,
54 struct inode **bitmap_inode,
55 struct buffer_head **bitmap_bh);
56
57static int ocfs2_local_alloc_new_window(struct ocfs2_super *osb,
58 handle_t *handle,
59 struct ocfs2_alloc_context *ac);
60
61static int ocfs2_local_alloc_slide_window(struct ocfs2_super *osb,
62 struct inode *local_alloc_inode);
63
64/*
65 * ocfs2_la_default_mb() - determine a default size, in megabytes of
66 * the local alloc.
67 *
68 * Generally, we'd like to pick as large a local alloc as
69 * possible. Performance on large workloads tends to scale
70 * proportionally to la size. In addition to that, the reservations
71 * code functions more efficiently as it can reserve more windows for
72 * write.
73 *
74 * Some things work against us when trying to choose a large local alloc:
75 *
76 * - We need to ensure our sizing is picked to leave enough space in
77 * group descriptors for other allocations (such as block groups,
78 * etc). Picking default sizes which are a multiple of 4 could help
79 * - block groups are allocated in 2mb and 4mb chunks.
80 *
81 * - Likewise, we don't want to starve other nodes of bits on small
82 * file systems. This can easily be taken care of by limiting our
83 * default to a reasonable size (256M) on larger cluster sizes.
84 *
85 * - Some file systems can't support very large sizes - 4k and 8k in
86 * particular are limited to less than 128 and 256 megabytes respectively.
87 *
88 * The following reference table shows group descriptor and local
89 * alloc maximums at various cluster sizes (4k blocksize)
90 *
91 * csize: 4K group: 126M la: 121M
92 * csize: 8K group: 252M la: 243M
93 * csize: 16K group: 504M la: 486M
94 * csize: 32K group: 1008M la: 972M
95 * csize: 64K group: 2016M la: 1944M
96 * csize: 128K group: 4032M la: 3888M
97 * csize: 256K group: 8064M la: 7776M
98 * csize: 512K group: 16128M la: 15552M
99 * csize: 1024K group: 32256M la: 31104M
100 */
101#define OCFS2_LA_MAX_DEFAULT_MB 256
102#define OCFS2_LA_OLD_DEFAULT 8
103unsigned int ocfs2_la_default_mb(struct ocfs2_super *osb)
104{
105 unsigned int la_mb;
106 unsigned int gd_mb;
107 unsigned int la_max_mb;
108 unsigned int megs_per_slot;
109 struct super_block *sb = osb->sb;
110
111 gd_mb = ocfs2_clusters_to_megabytes(osb->sb,
112 8 * ocfs2_group_bitmap_size(sb, 0, osb->s_feature_incompat));
113
114 /*
115 * This takes care of files systems with very small group
116 * descriptors - 512 byte blocksize at cluster sizes lower
117 * than 16K and also 1k blocksize with 4k cluster size.
118 */
119 if ((sb->s_blocksize == 512 && osb->s_clustersize <= 8192)
120 || (sb->s_blocksize == 1024 && osb->s_clustersize == 4096))
121 return OCFS2_LA_OLD_DEFAULT;
122
123 /*
124 * Leave enough room for some block groups and make the final
125 * value we work from a multiple of 4.
126 */
127 gd_mb -= 16;
128 gd_mb &= 0xFFFFFFFB;
129
130 la_mb = gd_mb;
131
132 /*
133 * Keep window sizes down to a reasonable default
134 */
135 if (la_mb > OCFS2_LA_MAX_DEFAULT_MB) {
136 /*
137 * Some clustersize / blocksize combinations will have
138 * given us a larger than OCFS2_LA_MAX_DEFAULT_MB
139 * default size, but get poor distribution when
140 * limited to exactly 256 megabytes.
141 *
142 * As an example, 16K clustersize at 4K blocksize
143 * gives us a cluster group size of 504M. Paring the
144 * local alloc size down to 256 however, would give us
145 * only one window and around 200MB left in the
146 * cluster group. Instead, find the first size below
147 * 256 which would give us an even distribution.
148 *
149 * Larger cluster group sizes actually work out pretty
150 * well when pared to 256, so we don't have to do this
151 * for any group that fits more than two
152 * OCFS2_LA_MAX_DEFAULT_MB windows.
153 */
154 if (gd_mb > (2 * OCFS2_LA_MAX_DEFAULT_MB))
155 la_mb = 256;
156 else {
157 unsigned int gd_mult = gd_mb;
158
159 while (gd_mult > 256)
160 gd_mult = gd_mult >> 1;
161
162 la_mb = gd_mult;
163 }
164 }
165
166 megs_per_slot = osb->osb_clusters_at_boot / osb->max_slots;
167 megs_per_slot = ocfs2_clusters_to_megabytes(osb->sb, megs_per_slot);
168 /* Too many nodes, too few disk clusters. */
169 if (megs_per_slot < la_mb)
170 la_mb = megs_per_slot;
171
172 /* We can't store more bits than we can in a block. */
173 la_max_mb = ocfs2_clusters_to_megabytes(osb->sb,
174 ocfs2_local_alloc_size(sb) * 8);
175 if (la_mb > la_max_mb)
176 la_mb = la_max_mb;
177
178 return la_mb;
179}
180
181void ocfs2_la_set_sizes(struct ocfs2_super *osb, int requested_mb)
182{
183 struct super_block *sb = osb->sb;
184 unsigned int la_default_mb = ocfs2_la_default_mb(osb);
185 unsigned int la_max_mb;
186
187 la_max_mb = ocfs2_clusters_to_megabytes(sb,
188 ocfs2_local_alloc_size(sb) * 8);
189
190 trace_ocfs2_la_set_sizes(requested_mb, la_max_mb, la_default_mb);
191
192 if (requested_mb == -1) {
193 /* No user request - use defaults */
194 osb->local_alloc_default_bits =
195 ocfs2_megabytes_to_clusters(sb, la_default_mb);
196 } else if (requested_mb > la_max_mb) {
197 /* Request is too big, we give the maximum available */
198 osb->local_alloc_default_bits =
199 ocfs2_megabytes_to_clusters(sb, la_max_mb);
200 } else {
201 osb->local_alloc_default_bits =
202 ocfs2_megabytes_to_clusters(sb, requested_mb);
203 }
204
205 osb->local_alloc_bits = osb->local_alloc_default_bits;
206}
207
208static inline int ocfs2_la_state_enabled(struct ocfs2_super *osb)
209{
210 return (osb->local_alloc_state == OCFS2_LA_THROTTLED ||
211 osb->local_alloc_state == OCFS2_LA_ENABLED);
212}
213
214void ocfs2_local_alloc_seen_free_bits(struct ocfs2_super *osb,
215 unsigned int num_clusters)
216{
217 spin_lock(&osb->osb_lock);
218 if (osb->local_alloc_state == OCFS2_LA_DISABLED ||
219 osb->local_alloc_state == OCFS2_LA_THROTTLED)
220 if (num_clusters >= osb->local_alloc_default_bits) {
221 cancel_delayed_work(&osb->la_enable_wq);
222 osb->local_alloc_state = OCFS2_LA_ENABLED;
223 }
224 spin_unlock(&osb->osb_lock);
225}
226
227void ocfs2_la_enable_worker(struct work_struct *work)
228{
229 struct ocfs2_super *osb =
230 container_of(work, struct ocfs2_super,
231 la_enable_wq.work);
232 spin_lock(&osb->osb_lock);
233 osb->local_alloc_state = OCFS2_LA_ENABLED;
234 spin_unlock(&osb->osb_lock);
235}
236
237/*
238 * Tell us whether a given allocation should use the local alloc
239 * file. Otherwise, it has to go to the main bitmap.
240 *
241 * This function does semi-dirty reads of local alloc size and state!
242 * This is ok however, as the values are re-checked once under mutex.
243 */
244int ocfs2_alloc_should_use_local(struct ocfs2_super *osb, u64 bits)
245{
246 int ret = 0;
247 int la_bits;
248
249 spin_lock(&osb->osb_lock);
250 la_bits = osb->local_alloc_bits;
251
252 if (!ocfs2_la_state_enabled(osb))
253 goto bail;
254
255 /* la_bits should be at least twice the size (in clusters) of
256 * a new block group. We want to be sure block group
257 * allocations go through the local alloc, so allow an
258 * allocation to take up to half the bitmap. */
259 if (bits > (la_bits / 2))
260 goto bail;
261
262 ret = 1;
263bail:
264 trace_ocfs2_alloc_should_use_local(
265 (unsigned long long)bits, osb->local_alloc_state, la_bits, ret);
266 spin_unlock(&osb->osb_lock);
267 return ret;
268}
269
270int ocfs2_load_local_alloc(struct ocfs2_super *osb)
271{
272 int status = 0;
273 struct ocfs2_dinode *alloc = NULL;
274 struct buffer_head *alloc_bh = NULL;
275 u32 num_used;
276 struct inode *inode = NULL;
277 struct ocfs2_local_alloc *la;
278
279 if (osb->local_alloc_bits == 0)
280 goto bail;
281
282 if (osb->local_alloc_bits >= osb->bitmap_cpg) {
283 mlog(ML_NOTICE, "Requested local alloc window %d is larger "
284 "than max possible %u. Using defaults.\n",
285 osb->local_alloc_bits, (osb->bitmap_cpg - 1));
286 osb->local_alloc_bits =
287 ocfs2_megabytes_to_clusters(osb->sb,
288 ocfs2_la_default_mb(osb));
289 }
290
291 /* read the alloc off disk */
292 inode = ocfs2_get_system_file_inode(osb, LOCAL_ALLOC_SYSTEM_INODE,
293 osb->slot_num);
294 if (!inode) {
295 status = -EINVAL;
296 mlog_errno(status);
297 goto bail;
298 }
299
300 status = ocfs2_read_inode_block_full(inode, &alloc_bh,
301 OCFS2_BH_IGNORE_CACHE);
302 if (status < 0) {
303 mlog_errno(status);
304 goto bail;
305 }
306
307 alloc = (struct ocfs2_dinode *) alloc_bh->b_data;
308 la = OCFS2_LOCAL_ALLOC(alloc);
309
310 if (!(le32_to_cpu(alloc->i_flags) &
311 (OCFS2_LOCAL_ALLOC_FL|OCFS2_BITMAP_FL))) {
312 mlog(ML_ERROR, "Invalid local alloc inode, %llu\n",
313 (unsigned long long)OCFS2_I(inode)->ip_blkno);
314 status = -EINVAL;
315 goto bail;
316 }
317
318 if ((la->la_size == 0) ||
319 (le16_to_cpu(la->la_size) > ocfs2_local_alloc_size(inode->i_sb))) {
320 mlog(ML_ERROR, "Local alloc size is invalid (la_size = %u)\n",
321 le16_to_cpu(la->la_size));
322 status = -EINVAL;
323 goto bail;
324 }
325
326 /* do a little verification. */
327 num_used = ocfs2_local_alloc_count_bits(alloc);
328
329 /* hopefully the local alloc has always been recovered before
330 * we load it. */
331 if (num_used
332 || alloc->id1.bitmap1.i_used
333 || alloc->id1.bitmap1.i_total
334 || la->la_bm_off) {
335 mlog(ML_ERROR, "inconsistent detected, clean journal with"
336 " unrecovered local alloc, please run fsck.ocfs2!\n"
337 "found = %u, set = %u, taken = %u, off = %u\n",
338 num_used, le32_to_cpu(alloc->id1.bitmap1.i_used),
339 le32_to_cpu(alloc->id1.bitmap1.i_total),
340 OCFS2_LOCAL_ALLOC(alloc)->la_bm_off);
341
342 status = -EINVAL;
343 goto bail;
344 }
345
346 osb->local_alloc_bh = alloc_bh;
347 osb->local_alloc_state = OCFS2_LA_ENABLED;
348
349bail:
350 if (status < 0)
351 brelse(alloc_bh);
352 iput(inode);
353
354 trace_ocfs2_load_local_alloc(osb->local_alloc_bits);
355
356 if (status)
357 mlog_errno(status);
358 return status;
359}
360
361/*
362 * return any unused bits to the bitmap and write out a clean
363 * local_alloc.
364 *
365 * local_alloc_bh is optional. If not passed, we will simply use the
366 * one off osb. If you do pass it however, be warned that it *will* be
367 * returned brelse'd and NULL'd out.*/
368void ocfs2_shutdown_local_alloc(struct ocfs2_super *osb)
369{
370 int status;
371 handle_t *handle;
372 struct inode *local_alloc_inode = NULL;
373 struct buffer_head *bh = NULL;
374 struct buffer_head *main_bm_bh = NULL;
375 struct inode *main_bm_inode = NULL;
376 struct ocfs2_dinode *alloc_copy = NULL;
377 struct ocfs2_dinode *alloc = NULL;
378
379 cancel_delayed_work(&osb->la_enable_wq);
380 if (osb->ocfs2_wq)
381 flush_workqueue(osb->ocfs2_wq);
382
383 if (osb->local_alloc_state == OCFS2_LA_UNUSED)
384 goto out;
385
386 local_alloc_inode =
387 ocfs2_get_system_file_inode(osb,
388 LOCAL_ALLOC_SYSTEM_INODE,
389 osb->slot_num);
390 if (!local_alloc_inode) {
391 status = -ENOENT;
392 mlog_errno(status);
393 goto out;
394 }
395
396 osb->local_alloc_state = OCFS2_LA_DISABLED;
397
398 ocfs2_resmap_uninit(&osb->osb_la_resmap);
399
400 main_bm_inode = ocfs2_get_system_file_inode(osb,
401 GLOBAL_BITMAP_SYSTEM_INODE,
402 OCFS2_INVALID_SLOT);
403 if (!main_bm_inode) {
404 status = -EINVAL;
405 mlog_errno(status);
406 goto out;
407 }
408
409 inode_lock(main_bm_inode);
410
411 status = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 1);
412 if (status < 0) {
413 mlog_errno(status);
414 goto out_mutex;
415 }
416
417 /* WINDOW_MOVE_CREDITS is a bit heavy... */
418 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
419 if (IS_ERR(handle)) {
420 mlog_errno(PTR_ERR(handle));
421 handle = NULL;
422 goto out_unlock;
423 }
424
425 bh = osb->local_alloc_bh;
426 alloc = (struct ocfs2_dinode *) bh->b_data;
427
428 alloc_copy = kmemdup(alloc, bh->b_size, GFP_NOFS);
429 if (!alloc_copy) {
430 status = -ENOMEM;
431 goto out_commit;
432 }
433
434 status = ocfs2_journal_access_di(handle, INODE_CACHE(local_alloc_inode),
435 bh, OCFS2_JOURNAL_ACCESS_WRITE);
436 if (status < 0) {
437 mlog_errno(status);
438 goto out_commit;
439 }
440
441 ocfs2_clear_local_alloc(alloc);
442 ocfs2_journal_dirty(handle, bh);
443
444 brelse(bh);
445 osb->local_alloc_bh = NULL;
446 osb->local_alloc_state = OCFS2_LA_UNUSED;
447
448 status = ocfs2_sync_local_to_main(osb, handle, alloc_copy,
449 main_bm_inode, main_bm_bh);
450 if (status < 0)
451 mlog_errno(status);
452
453out_commit:
454 ocfs2_commit_trans(osb, handle);
455
456out_unlock:
457 brelse(main_bm_bh);
458
459 ocfs2_inode_unlock(main_bm_inode, 1);
460
461out_mutex:
462 inode_unlock(main_bm_inode);
463 iput(main_bm_inode);
464
465out:
466 iput(local_alloc_inode);
467
468 kfree(alloc_copy);
469}
470
471/*
472 * We want to free the bitmap bits outside of any recovery context as
473 * we'll need a cluster lock to do so, but we must clear the local
474 * alloc before giving up the recovered nodes journal. To solve this,
475 * we kmalloc a copy of the local alloc before it's change for the
476 * caller to process with ocfs2_complete_local_alloc_recovery
477 */
478int ocfs2_begin_local_alloc_recovery(struct ocfs2_super *osb,
479 int slot_num,
480 struct ocfs2_dinode **alloc_copy)
481{
482 int status = 0;
483 struct buffer_head *alloc_bh = NULL;
484 struct inode *inode = NULL;
485 struct ocfs2_dinode *alloc;
486
487 trace_ocfs2_begin_local_alloc_recovery(slot_num);
488
489 *alloc_copy = NULL;
490
491 inode = ocfs2_get_system_file_inode(osb,
492 LOCAL_ALLOC_SYSTEM_INODE,
493 slot_num);
494 if (!inode) {
495 status = -EINVAL;
496 mlog_errno(status);
497 goto bail;
498 }
499
500 inode_lock(inode);
501
502 status = ocfs2_read_inode_block_full(inode, &alloc_bh,
503 OCFS2_BH_IGNORE_CACHE);
504 if (status < 0) {
505 mlog_errno(status);
506 goto bail;
507 }
508
509 *alloc_copy = kmalloc(alloc_bh->b_size, GFP_KERNEL);
510 if (!(*alloc_copy)) {
511 status = -ENOMEM;
512 goto bail;
513 }
514 memcpy((*alloc_copy), alloc_bh->b_data, alloc_bh->b_size);
515
516 alloc = (struct ocfs2_dinode *) alloc_bh->b_data;
517 ocfs2_clear_local_alloc(alloc);
518
519 ocfs2_compute_meta_ecc(osb->sb, alloc_bh->b_data, &alloc->i_check);
520 status = ocfs2_write_block(osb, alloc_bh, INODE_CACHE(inode));
521 if (status < 0)
522 mlog_errno(status);
523
524bail:
525 if (status < 0) {
526 kfree(*alloc_copy);
527 *alloc_copy = NULL;
528 }
529
530 brelse(alloc_bh);
531
532 if (inode) {
533 inode_unlock(inode);
534 iput(inode);
535 }
536
537 if (status)
538 mlog_errno(status);
539 return status;
540}
541
542/*
543 * Step 2: By now, we've completed the journal recovery, we've stamped
544 * a clean local alloc on disk and dropped the node out of the
545 * recovery map. Dlm locks will no longer stall, so lets clear out the
546 * main bitmap.
547 */
548int ocfs2_complete_local_alloc_recovery(struct ocfs2_super *osb,
549 struct ocfs2_dinode *alloc)
550{
551 int status;
552 handle_t *handle;
553 struct buffer_head *main_bm_bh = NULL;
554 struct inode *main_bm_inode;
555
556 main_bm_inode = ocfs2_get_system_file_inode(osb,
557 GLOBAL_BITMAP_SYSTEM_INODE,
558 OCFS2_INVALID_SLOT);
559 if (!main_bm_inode) {
560 status = -EINVAL;
561 mlog_errno(status);
562 goto out;
563 }
564
565 inode_lock(main_bm_inode);
566
567 status = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 1);
568 if (status < 0) {
569 mlog_errno(status);
570 goto out_mutex;
571 }
572
573 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
574 if (IS_ERR(handle)) {
575 status = PTR_ERR(handle);
576 handle = NULL;
577 mlog_errno(status);
578 goto out_unlock;
579 }
580
581 /* we want the bitmap change to be recorded on disk asap */
582 handle->h_sync = 1;
583
584 status = ocfs2_sync_local_to_main(osb, handle, alloc,
585 main_bm_inode, main_bm_bh);
586 if (status < 0)
587 mlog_errno(status);
588
589 ocfs2_commit_trans(osb, handle);
590
591out_unlock:
592 ocfs2_inode_unlock(main_bm_inode, 1);
593
594out_mutex:
595 inode_unlock(main_bm_inode);
596
597 brelse(main_bm_bh);
598
599 iput(main_bm_inode);
600
601out:
602 if (!status)
603 ocfs2_init_steal_slots(osb);
604 if (status)
605 mlog_errno(status);
606 return status;
607}
608
609/*
610 * make sure we've got at least bits_wanted contiguous bits in the
611 * local alloc. You lose them when you drop i_mutex.
612 *
613 * We will add ourselves to the transaction passed in, but may start
614 * our own in order to shift windows.
615 */
616int ocfs2_reserve_local_alloc_bits(struct ocfs2_super *osb,
617 u32 bits_wanted,
618 struct ocfs2_alloc_context *ac)
619{
620 int status;
621 struct ocfs2_dinode *alloc;
622 struct inode *local_alloc_inode;
623 unsigned int free_bits;
624
625 BUG_ON(!ac);
626
627 local_alloc_inode =
628 ocfs2_get_system_file_inode(osb,
629 LOCAL_ALLOC_SYSTEM_INODE,
630 osb->slot_num);
631 if (!local_alloc_inode) {
632 status = -ENOENT;
633 mlog_errno(status);
634 goto bail;
635 }
636
637 inode_lock(local_alloc_inode);
638
639 /*
640 * We must double check state and allocator bits because
641 * another process may have changed them while holding i_mutex.
642 */
643 spin_lock(&osb->osb_lock);
644 if (!ocfs2_la_state_enabled(osb) ||
645 (bits_wanted > osb->local_alloc_bits)) {
646 spin_unlock(&osb->osb_lock);
647 status = -ENOSPC;
648 goto bail;
649 }
650 spin_unlock(&osb->osb_lock);
651
652 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
653
654#ifdef CONFIG_OCFS2_DEBUG_FS
655 if (le32_to_cpu(alloc->id1.bitmap1.i_used) !=
656 ocfs2_local_alloc_count_bits(alloc)) {
657 status = ocfs2_error(osb->sb, "local alloc inode %llu says it has %u used bits, but a count shows %u\n",
658 (unsigned long long)le64_to_cpu(alloc->i_blkno),
659 le32_to_cpu(alloc->id1.bitmap1.i_used),
660 ocfs2_local_alloc_count_bits(alloc));
661 goto bail;
662 }
663#endif
664
665 free_bits = le32_to_cpu(alloc->id1.bitmap1.i_total) -
666 le32_to_cpu(alloc->id1.bitmap1.i_used);
667 if (bits_wanted > free_bits) {
668 /* uhoh, window change time. */
669 status =
670 ocfs2_local_alloc_slide_window(osb, local_alloc_inode);
671 if (status < 0) {
672 if (status != -ENOSPC)
673 mlog_errno(status);
674 goto bail;
675 }
676
677 /*
678 * Under certain conditions, the window slide code
679 * might have reduced the number of bits available or
680 * disabled the the local alloc entirely. Re-check
681 * here and return -ENOSPC if necessary.
682 */
683 status = -ENOSPC;
684 if (!ocfs2_la_state_enabled(osb))
685 goto bail;
686
687 free_bits = le32_to_cpu(alloc->id1.bitmap1.i_total) -
688 le32_to_cpu(alloc->id1.bitmap1.i_used);
689 if (bits_wanted > free_bits)
690 goto bail;
691 }
692
693 ac->ac_inode = local_alloc_inode;
694 /* We should never use localalloc from another slot */
695 ac->ac_alloc_slot = osb->slot_num;
696 ac->ac_which = OCFS2_AC_USE_LOCAL;
697 get_bh(osb->local_alloc_bh);
698 ac->ac_bh = osb->local_alloc_bh;
699 status = 0;
700bail:
701 if (status < 0 && local_alloc_inode) {
702 inode_unlock(local_alloc_inode);
703 iput(local_alloc_inode);
704 }
705
706 trace_ocfs2_reserve_local_alloc_bits(
707 (unsigned long long)ac->ac_max_block,
708 bits_wanted, osb->slot_num, status);
709
710 if (status)
711 mlog_errno(status);
712 return status;
713}
714
715int ocfs2_claim_local_alloc_bits(struct ocfs2_super *osb,
716 handle_t *handle,
717 struct ocfs2_alloc_context *ac,
718 u32 bits_wanted,
719 u32 *bit_off,
720 u32 *num_bits)
721{
722 int status, start;
723 struct inode *local_alloc_inode;
724 void *bitmap;
725 struct ocfs2_dinode *alloc;
726 struct ocfs2_local_alloc *la;
727
728 BUG_ON(ac->ac_which != OCFS2_AC_USE_LOCAL);
729
730 local_alloc_inode = ac->ac_inode;
731 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
732 la = OCFS2_LOCAL_ALLOC(alloc);
733
734 start = ocfs2_local_alloc_find_clear_bits(osb, alloc, &bits_wanted,
735 ac->ac_resv);
736 if (start == -1) {
737 /* TODO: Shouldn't we just BUG here? */
738 status = -ENOSPC;
739 mlog_errno(status);
740 goto bail;
741 }
742
743 bitmap = la->la_bitmap;
744 *bit_off = le32_to_cpu(la->la_bm_off) + start;
745 *num_bits = bits_wanted;
746
747 status = ocfs2_journal_access_di(handle,
748 INODE_CACHE(local_alloc_inode),
749 osb->local_alloc_bh,
750 OCFS2_JOURNAL_ACCESS_WRITE);
751 if (status < 0) {
752 mlog_errno(status);
753 goto bail;
754 }
755
756 ocfs2_resmap_claimed_bits(&osb->osb_la_resmap, ac->ac_resv, start,
757 bits_wanted);
758
759 while(bits_wanted--)
760 ocfs2_set_bit(start++, bitmap);
761
762 le32_add_cpu(&alloc->id1.bitmap1.i_used, *num_bits);
763 ocfs2_journal_dirty(handle, osb->local_alloc_bh);
764
765bail:
766 if (status)
767 mlog_errno(status);
768 return status;
769}
770
771int ocfs2_free_local_alloc_bits(struct ocfs2_super *osb,
772 handle_t *handle,
773 struct ocfs2_alloc_context *ac,
774 u32 bit_off,
775 u32 num_bits)
776{
777 int status, start;
778 u32 clear_bits;
779 struct inode *local_alloc_inode;
780 void *bitmap;
781 struct ocfs2_dinode *alloc;
782 struct ocfs2_local_alloc *la;
783
784 BUG_ON(ac->ac_which != OCFS2_AC_USE_LOCAL);
785
786 local_alloc_inode = ac->ac_inode;
787 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
788 la = OCFS2_LOCAL_ALLOC(alloc);
789
790 bitmap = la->la_bitmap;
791 start = bit_off - le32_to_cpu(la->la_bm_off);
792 clear_bits = num_bits;
793
794 status = ocfs2_journal_access_di(handle,
795 INODE_CACHE(local_alloc_inode),
796 osb->local_alloc_bh,
797 OCFS2_JOURNAL_ACCESS_WRITE);
798 if (status < 0) {
799 mlog_errno(status);
800 goto bail;
801 }
802
803 while (clear_bits--)
804 ocfs2_clear_bit(start++, bitmap);
805
806 le32_add_cpu(&alloc->id1.bitmap1.i_used, -num_bits);
807 ocfs2_journal_dirty(handle, osb->local_alloc_bh);
808
809bail:
810 return status;
811}
812
813static u32 ocfs2_local_alloc_count_bits(struct ocfs2_dinode *alloc)
814{
815 u32 count;
816 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
817
818 count = memweight(la->la_bitmap, le16_to_cpu(la->la_size));
819
820 trace_ocfs2_local_alloc_count_bits(count);
821 return count;
822}
823
824static int ocfs2_local_alloc_find_clear_bits(struct ocfs2_super *osb,
825 struct ocfs2_dinode *alloc,
826 u32 *numbits,
827 struct ocfs2_alloc_reservation *resv)
828{
829 int numfound = 0, bitoff, left, startoff;
830 int local_resv = 0;
831 struct ocfs2_alloc_reservation r;
832 void *bitmap = NULL;
833 struct ocfs2_reservation_map *resmap = &osb->osb_la_resmap;
834
835 if (!alloc->id1.bitmap1.i_total) {
836 bitoff = -1;
837 goto bail;
838 }
839
840 if (!resv) {
841 local_resv = 1;
842 ocfs2_resv_init_once(&r);
843 ocfs2_resv_set_type(&r, OCFS2_RESV_FLAG_TMP);
844 resv = &r;
845 }
846
847 numfound = *numbits;
848 if (ocfs2_resmap_resv_bits(resmap, resv, &bitoff, &numfound) == 0) {
849 if (numfound < *numbits)
850 *numbits = numfound;
851 goto bail;
852 }
853
854 /*
855 * Code error. While reservations are enabled, local
856 * allocation should _always_ go through them.
857 */
858 BUG_ON(osb->osb_resv_level != 0);
859
860 /*
861 * Reservations are disabled. Handle this the old way.
862 */
863
864 bitmap = OCFS2_LOCAL_ALLOC(alloc)->la_bitmap;
865
866 numfound = bitoff = startoff = 0;
867 left = le32_to_cpu(alloc->id1.bitmap1.i_total);
868 while ((bitoff = ocfs2_find_next_zero_bit(bitmap, left, startoff)) != -1) {
869 if (bitoff == left) {
870 /* mlog(0, "bitoff (%d) == left", bitoff); */
871 break;
872 }
873 /* mlog(0, "Found a zero: bitoff = %d, startoff = %d, "
874 "numfound = %d\n", bitoff, startoff, numfound);*/
875
876 /* Ok, we found a zero bit... is it contig. or do we
877 * start over?*/
878 if (bitoff == startoff) {
879 /* we found a zero */
880 numfound++;
881 startoff++;
882 } else {
883 /* got a zero after some ones */
884 numfound = 1;
885 startoff = bitoff+1;
886 }
887 /* we got everything we needed */
888 if (numfound == *numbits) {
889 /* mlog(0, "Found it all!\n"); */
890 break;
891 }
892 }
893
894 trace_ocfs2_local_alloc_find_clear_bits_search_bitmap(bitoff, numfound);
895
896 if (numfound == *numbits)
897 bitoff = startoff - numfound;
898 else
899 bitoff = -1;
900
901bail:
902 if (local_resv)
903 ocfs2_resv_discard(resmap, resv);
904
905 trace_ocfs2_local_alloc_find_clear_bits(*numbits,
906 le32_to_cpu(alloc->id1.bitmap1.i_total),
907 bitoff, numfound);
908
909 return bitoff;
910}
911
912static void ocfs2_clear_local_alloc(struct ocfs2_dinode *alloc)
913{
914 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
915 int i;
916
917 alloc->id1.bitmap1.i_total = 0;
918 alloc->id1.bitmap1.i_used = 0;
919 la->la_bm_off = 0;
920 for(i = 0; i < le16_to_cpu(la->la_size); i++)
921 la->la_bitmap[i] = 0;
922}
923
924#if 0
925/* turn this on and uncomment below to aid debugging window shifts. */
926static void ocfs2_verify_zero_bits(unsigned long *bitmap,
927 unsigned int start,
928 unsigned int count)
929{
930 unsigned int tmp = count;
931 while(tmp--) {
932 if (ocfs2_test_bit(start + tmp, bitmap)) {
933 printk("ocfs2_verify_zero_bits: start = %u, count = "
934 "%u\n", start, count);
935 printk("ocfs2_verify_zero_bits: bit %u is set!",
936 start + tmp);
937 BUG();
938 }
939 }
940}
941#endif
942
943/*
944 * sync the local alloc to main bitmap.
945 *
946 * assumes you've already locked the main bitmap -- the bitmap inode
947 * passed is used for caching.
948 */
949static int ocfs2_sync_local_to_main(struct ocfs2_super *osb,
950 handle_t *handle,
951 struct ocfs2_dinode *alloc,
952 struct inode *main_bm_inode,
953 struct buffer_head *main_bm_bh)
954{
955 int status = 0;
956 int bit_off, left, count, start;
957 u64 la_start_blk;
958 u64 blkno;
959 void *bitmap;
960 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
961
962 trace_ocfs2_sync_local_to_main(
963 le32_to_cpu(alloc->id1.bitmap1.i_total),
964 le32_to_cpu(alloc->id1.bitmap1.i_used));
965
966 if (!alloc->id1.bitmap1.i_total) {
967 goto bail;
968 }
969
970 if (le32_to_cpu(alloc->id1.bitmap1.i_used) ==
971 le32_to_cpu(alloc->id1.bitmap1.i_total)) {
972 goto bail;
973 }
974
975 la_start_blk = ocfs2_clusters_to_blocks(osb->sb,
976 le32_to_cpu(la->la_bm_off));
977 bitmap = la->la_bitmap;
978 start = count = bit_off = 0;
979 left = le32_to_cpu(alloc->id1.bitmap1.i_total);
980
981 while ((bit_off = ocfs2_find_next_zero_bit(bitmap, left, start))
982 != -1) {
983 if ((bit_off < left) && (bit_off == start)) {
984 count++;
985 start++;
986 continue;
987 }
988 if (count) {
989 blkno = la_start_blk +
990 ocfs2_clusters_to_blocks(osb->sb,
991 start - count);
992
993 trace_ocfs2_sync_local_to_main_free(
994 count, start - count,
995 (unsigned long long)la_start_blk,
996 (unsigned long long)blkno);
997
998 status = ocfs2_release_clusters(handle,
999 main_bm_inode,
1000 main_bm_bh, blkno,
1001 count);
1002 if (status < 0) {
1003 mlog_errno(status);
1004 goto bail;
1005 }
1006 }
1007 if (bit_off >= left)
1008 break;
1009 count = 1;
1010 start = bit_off + 1;
1011 }
1012
1013bail:
1014 if (status)
1015 mlog_errno(status);
1016 return status;
1017}
1018
1019enum ocfs2_la_event {
1020 OCFS2_LA_EVENT_SLIDE, /* Normal window slide. */
1021 OCFS2_LA_EVENT_FRAGMENTED, /* The global bitmap has
1022 * enough bits theoretically
1023 * free, but a contiguous
1024 * allocation could not be
1025 * found. */
1026 OCFS2_LA_EVENT_ENOSPC, /* Global bitmap doesn't have
1027 * enough bits free to satisfy
1028 * our request. */
1029};
1030#define OCFS2_LA_ENABLE_INTERVAL (30 * HZ)
1031/*
1032 * Given an event, calculate the size of our next local alloc window.
1033 *
1034 * This should always be called under i_mutex of the local alloc inode
1035 * so that local alloc disabling doesn't race with processes trying to
1036 * use the allocator.
1037 *
1038 * Returns the state which the local alloc was left in. This value can
1039 * be ignored by some paths.
1040 */
1041static int ocfs2_recalc_la_window(struct ocfs2_super *osb,
1042 enum ocfs2_la_event event)
1043{
1044 unsigned int bits;
1045 int state;
1046
1047 spin_lock(&osb->osb_lock);
1048 if (osb->local_alloc_state == OCFS2_LA_DISABLED) {
1049 WARN_ON_ONCE(osb->local_alloc_state == OCFS2_LA_DISABLED);
1050 goto out_unlock;
1051 }
1052
1053 /*
1054 * ENOSPC and fragmentation are treated similarly for now.
1055 */
1056 if (event == OCFS2_LA_EVENT_ENOSPC ||
1057 event == OCFS2_LA_EVENT_FRAGMENTED) {
1058 /*
1059 * We ran out of contiguous space in the primary
1060 * bitmap. Drastically reduce the number of bits used
1061 * by local alloc until we have to disable it.
1062 */
1063 bits = osb->local_alloc_bits >> 1;
1064 if (bits > ocfs2_megabytes_to_clusters(osb->sb, 1)) {
1065 /*
1066 * By setting state to THROTTLED, we'll keep
1067 * the number of local alloc bits used down
1068 * until an event occurs which would give us
1069 * reason to assume the bitmap situation might
1070 * have changed.
1071 */
1072 osb->local_alloc_state = OCFS2_LA_THROTTLED;
1073 osb->local_alloc_bits = bits;
1074 } else {
1075 osb->local_alloc_state = OCFS2_LA_DISABLED;
1076 }
1077 queue_delayed_work(osb->ocfs2_wq, &osb->la_enable_wq,
1078 OCFS2_LA_ENABLE_INTERVAL);
1079 goto out_unlock;
1080 }
1081
1082 /*
1083 * Don't increase the size of the local alloc window until we
1084 * know we might be able to fulfill the request. Otherwise, we
1085 * risk bouncing around the global bitmap during periods of
1086 * low space.
1087 */
1088 if (osb->local_alloc_state != OCFS2_LA_THROTTLED)
1089 osb->local_alloc_bits = osb->local_alloc_default_bits;
1090
1091out_unlock:
1092 state = osb->local_alloc_state;
1093 spin_unlock(&osb->osb_lock);
1094
1095 return state;
1096}
1097
1098static int ocfs2_local_alloc_reserve_for_window(struct ocfs2_super *osb,
1099 struct ocfs2_alloc_context **ac,
1100 struct inode **bitmap_inode,
1101 struct buffer_head **bitmap_bh)
1102{
1103 int status;
1104
1105 *ac = kzalloc(sizeof(struct ocfs2_alloc_context), GFP_KERNEL);
1106 if (!(*ac)) {
1107 status = -ENOMEM;
1108 mlog_errno(status);
1109 goto bail;
1110 }
1111
1112retry_enospc:
1113 (*ac)->ac_bits_wanted = osb->local_alloc_bits;
1114 status = ocfs2_reserve_cluster_bitmap_bits(osb, *ac);
1115 if (status == -ENOSPC) {
1116 if (ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_ENOSPC) ==
1117 OCFS2_LA_DISABLED)
1118 goto bail;
1119
1120 ocfs2_free_ac_resource(*ac);
1121 memset(*ac, 0, sizeof(struct ocfs2_alloc_context));
1122 goto retry_enospc;
1123 }
1124 if (status < 0) {
1125 mlog_errno(status);
1126 goto bail;
1127 }
1128
1129 *bitmap_inode = (*ac)->ac_inode;
1130 igrab(*bitmap_inode);
1131 *bitmap_bh = (*ac)->ac_bh;
1132 get_bh(*bitmap_bh);
1133 status = 0;
1134bail:
1135 if ((status < 0) && *ac) {
1136 ocfs2_free_alloc_context(*ac);
1137 *ac = NULL;
1138 }
1139
1140 if (status)
1141 mlog_errno(status);
1142 return status;
1143}
1144
1145/*
1146 * pass it the bitmap lock in lock_bh if you have it.
1147 */
1148static int ocfs2_local_alloc_new_window(struct ocfs2_super *osb,
1149 handle_t *handle,
1150 struct ocfs2_alloc_context *ac)
1151{
1152 int status = 0;
1153 u32 cluster_off, cluster_count;
1154 struct ocfs2_dinode *alloc = NULL;
1155 struct ocfs2_local_alloc *la;
1156
1157 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
1158 la = OCFS2_LOCAL_ALLOC(alloc);
1159
1160 trace_ocfs2_local_alloc_new_window(
1161 le32_to_cpu(alloc->id1.bitmap1.i_total),
1162 osb->local_alloc_bits);
1163
1164 /* Instruct the allocation code to try the most recently used
1165 * cluster group. We'll re-record the group used this pass
1166 * below. */
1167 ac->ac_last_group = osb->la_last_gd;
1168
1169 /* we used the generic suballoc reserve function, but we set
1170 * everything up nicely, so there's no reason why we can't use
1171 * the more specific cluster api to claim bits. */
1172 status = ocfs2_claim_clusters(handle, ac, osb->local_alloc_bits,
1173 &cluster_off, &cluster_count);
1174 if (status == -ENOSPC) {
1175retry_enospc:
1176 /*
1177 * Note: We could also try syncing the journal here to
1178 * allow use of any free bits which the current
1179 * transaction can't give us access to. --Mark
1180 */
1181 if (ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_FRAGMENTED) ==
1182 OCFS2_LA_DISABLED)
1183 goto bail;
1184
1185 ac->ac_bits_wanted = osb->local_alloc_bits;
1186 status = ocfs2_claim_clusters(handle, ac,
1187 osb->local_alloc_bits,
1188 &cluster_off,
1189 &cluster_count);
1190 if (status == -ENOSPC)
1191 goto retry_enospc;
1192 /*
1193 * We only shrunk the *minimum* number of in our
1194 * request - it's entirely possible that the allocator
1195 * might give us more than we asked for.
1196 */
1197 if (status == 0) {
1198 spin_lock(&osb->osb_lock);
1199 osb->local_alloc_bits = cluster_count;
1200 spin_unlock(&osb->osb_lock);
1201 }
1202 }
1203 if (status < 0) {
1204 if (status != -ENOSPC)
1205 mlog_errno(status);
1206 goto bail;
1207 }
1208
1209 osb->la_last_gd = ac->ac_last_group;
1210
1211 la->la_bm_off = cpu_to_le32(cluster_off);
1212 alloc->id1.bitmap1.i_total = cpu_to_le32(cluster_count);
1213 /* just in case... In the future when we find space ourselves,
1214 * we don't have to get all contiguous -- but we'll have to
1215 * set all previously used bits in bitmap and update
1216 * la_bits_set before setting the bits in the main bitmap. */
1217 alloc->id1.bitmap1.i_used = 0;
1218 memset(OCFS2_LOCAL_ALLOC(alloc)->la_bitmap, 0,
1219 le16_to_cpu(la->la_size));
1220
1221 ocfs2_resmap_restart(&osb->osb_la_resmap, cluster_count,
1222 OCFS2_LOCAL_ALLOC(alloc)->la_bitmap);
1223
1224 trace_ocfs2_local_alloc_new_window_result(
1225 OCFS2_LOCAL_ALLOC(alloc)->la_bm_off,
1226 le32_to_cpu(alloc->id1.bitmap1.i_total));
1227
1228bail:
1229 if (status)
1230 mlog_errno(status);
1231 return status;
1232}
1233
1234/* Note that we do *NOT* lock the local alloc inode here as
1235 * it's been locked already for us. */
1236static int ocfs2_local_alloc_slide_window(struct ocfs2_super *osb,
1237 struct inode *local_alloc_inode)
1238{
1239 int status = 0;
1240 struct buffer_head *main_bm_bh = NULL;
1241 struct inode *main_bm_inode = NULL;
1242 handle_t *handle = NULL;
1243 struct ocfs2_dinode *alloc;
1244 struct ocfs2_dinode *alloc_copy = NULL;
1245 struct ocfs2_alloc_context *ac = NULL;
1246
1247 ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_SLIDE);
1248
1249 /* This will lock the main bitmap for us. */
1250 status = ocfs2_local_alloc_reserve_for_window(osb,
1251 &ac,
1252 &main_bm_inode,
1253 &main_bm_bh);
1254 if (status < 0) {
1255 if (status != -ENOSPC)
1256 mlog_errno(status);
1257 goto bail;
1258 }
1259
1260 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
1261 if (IS_ERR(handle)) {
1262 status = PTR_ERR(handle);
1263 handle = NULL;
1264 mlog_errno(status);
1265 goto bail;
1266 }
1267
1268 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
1269
1270 /* We want to clear the local alloc before doing anything
1271 * else, so that if we error later during this operation,
1272 * local alloc shutdown won't try to double free main bitmap
1273 * bits. Make a copy so the sync function knows which bits to
1274 * free. */
1275 alloc_copy = kmemdup(alloc, osb->local_alloc_bh->b_size, GFP_NOFS);
1276 if (!alloc_copy) {
1277 status = -ENOMEM;
1278 mlog_errno(status);
1279 goto bail;
1280 }
1281
1282 status = ocfs2_journal_access_di(handle,
1283 INODE_CACHE(local_alloc_inode),
1284 osb->local_alloc_bh,
1285 OCFS2_JOURNAL_ACCESS_WRITE);
1286 if (status < 0) {
1287 mlog_errno(status);
1288 goto bail;
1289 }
1290
1291 ocfs2_clear_local_alloc(alloc);
1292 ocfs2_journal_dirty(handle, osb->local_alloc_bh);
1293
1294 status = ocfs2_sync_local_to_main(osb, handle, alloc_copy,
1295 main_bm_inode, main_bm_bh);
1296 if (status < 0) {
1297 mlog_errno(status);
1298 goto bail;
1299 }
1300
1301 status = ocfs2_local_alloc_new_window(osb, handle, ac);
1302 if (status < 0) {
1303 if (status != -ENOSPC)
1304 mlog_errno(status);
1305 goto bail;
1306 }
1307
1308 atomic_inc(&osb->alloc_stats.moves);
1309
1310bail:
1311 if (handle)
1312 ocfs2_commit_trans(osb, handle);
1313
1314 brelse(main_bm_bh);
1315
1316 iput(main_bm_inode);
1317 kfree(alloc_copy);
1318
1319 if (ac)
1320 ocfs2_free_alloc_context(ac);
1321
1322 if (status)
1323 mlog_errno(status);
1324 return status;
1325}
1326
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 if (inode)
362 iput(inode);
363
364 trace_ocfs2_load_local_alloc(osb->local_alloc_bits);
365
366 if (status)
367 mlog_errno(status);
368 return status;
369}
370
371/*
372 * return any unused bits to the bitmap and write out a clean
373 * local_alloc.
374 *
375 * local_alloc_bh is optional. If not passed, we will simply use the
376 * one off osb. If you do pass it however, be warned that it *will* be
377 * returned brelse'd and NULL'd out.*/
378void ocfs2_shutdown_local_alloc(struct ocfs2_super *osb)
379{
380 int status;
381 handle_t *handle;
382 struct inode *local_alloc_inode = NULL;
383 struct buffer_head *bh = NULL;
384 struct buffer_head *main_bm_bh = NULL;
385 struct inode *main_bm_inode = NULL;
386 struct ocfs2_dinode *alloc_copy = NULL;
387 struct ocfs2_dinode *alloc = NULL;
388
389 cancel_delayed_work(&osb->la_enable_wq);
390 flush_workqueue(ocfs2_wq);
391
392 if (osb->local_alloc_state == OCFS2_LA_UNUSED)
393 goto out;
394
395 local_alloc_inode =
396 ocfs2_get_system_file_inode(osb,
397 LOCAL_ALLOC_SYSTEM_INODE,
398 osb->slot_num);
399 if (!local_alloc_inode) {
400 status = -ENOENT;
401 mlog_errno(status);
402 goto out;
403 }
404
405 osb->local_alloc_state = OCFS2_LA_DISABLED;
406
407 ocfs2_resmap_uninit(&osb->osb_la_resmap);
408
409 main_bm_inode = ocfs2_get_system_file_inode(osb,
410 GLOBAL_BITMAP_SYSTEM_INODE,
411 OCFS2_INVALID_SLOT);
412 if (!main_bm_inode) {
413 status = -EINVAL;
414 mlog_errno(status);
415 goto out;
416 }
417
418 mutex_lock(&main_bm_inode->i_mutex);
419
420 status = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 1);
421 if (status < 0) {
422 mlog_errno(status);
423 goto out_mutex;
424 }
425
426 /* WINDOW_MOVE_CREDITS is a bit heavy... */
427 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
428 if (IS_ERR(handle)) {
429 mlog_errno(PTR_ERR(handle));
430 handle = NULL;
431 goto out_unlock;
432 }
433
434 bh = osb->local_alloc_bh;
435 alloc = (struct ocfs2_dinode *) bh->b_data;
436
437 alloc_copy = kmalloc(bh->b_size, GFP_NOFS);
438 if (!alloc_copy) {
439 status = -ENOMEM;
440 goto out_commit;
441 }
442 memcpy(alloc_copy, alloc, bh->b_size);
443
444 status = ocfs2_journal_access_di(handle, INODE_CACHE(local_alloc_inode),
445 bh, OCFS2_JOURNAL_ACCESS_WRITE);
446 if (status < 0) {
447 mlog_errno(status);
448 goto out_commit;
449 }
450
451 ocfs2_clear_local_alloc(alloc);
452 ocfs2_journal_dirty(handle, bh);
453
454 brelse(bh);
455 osb->local_alloc_bh = NULL;
456 osb->local_alloc_state = OCFS2_LA_UNUSED;
457
458 status = ocfs2_sync_local_to_main(osb, handle, alloc_copy,
459 main_bm_inode, main_bm_bh);
460 if (status < 0)
461 mlog_errno(status);
462
463out_commit:
464 ocfs2_commit_trans(osb, handle);
465
466out_unlock:
467 brelse(main_bm_bh);
468
469 ocfs2_inode_unlock(main_bm_inode, 1);
470
471out_mutex:
472 mutex_unlock(&main_bm_inode->i_mutex);
473 iput(main_bm_inode);
474
475out:
476 if (local_alloc_inode)
477 iput(local_alloc_inode);
478
479 kfree(alloc_copy);
480}
481
482/*
483 * We want to free the bitmap bits outside of any recovery context as
484 * we'll need a cluster lock to do so, but we must clear the local
485 * alloc before giving up the recovered nodes journal. To solve this,
486 * we kmalloc a copy of the local alloc before it's change for the
487 * caller to process with ocfs2_complete_local_alloc_recovery
488 */
489int ocfs2_begin_local_alloc_recovery(struct ocfs2_super *osb,
490 int slot_num,
491 struct ocfs2_dinode **alloc_copy)
492{
493 int status = 0;
494 struct buffer_head *alloc_bh = NULL;
495 struct inode *inode = NULL;
496 struct ocfs2_dinode *alloc;
497
498 trace_ocfs2_begin_local_alloc_recovery(slot_num);
499
500 *alloc_copy = NULL;
501
502 inode = ocfs2_get_system_file_inode(osb,
503 LOCAL_ALLOC_SYSTEM_INODE,
504 slot_num);
505 if (!inode) {
506 status = -EINVAL;
507 mlog_errno(status);
508 goto bail;
509 }
510
511 mutex_lock(&inode->i_mutex);
512
513 status = ocfs2_read_inode_block_full(inode, &alloc_bh,
514 OCFS2_BH_IGNORE_CACHE);
515 if (status < 0) {
516 mlog_errno(status);
517 goto bail;
518 }
519
520 *alloc_copy = kmalloc(alloc_bh->b_size, GFP_KERNEL);
521 if (!(*alloc_copy)) {
522 status = -ENOMEM;
523 goto bail;
524 }
525 memcpy((*alloc_copy), alloc_bh->b_data, alloc_bh->b_size);
526
527 alloc = (struct ocfs2_dinode *) alloc_bh->b_data;
528 ocfs2_clear_local_alloc(alloc);
529
530 ocfs2_compute_meta_ecc(osb->sb, alloc_bh->b_data, &alloc->i_check);
531 status = ocfs2_write_block(osb, alloc_bh, INODE_CACHE(inode));
532 if (status < 0)
533 mlog_errno(status);
534
535bail:
536 if (status < 0) {
537 kfree(*alloc_copy);
538 *alloc_copy = NULL;
539 }
540
541 brelse(alloc_bh);
542
543 if (inode) {
544 mutex_unlock(&inode->i_mutex);
545 iput(inode);
546 }
547
548 if (status)
549 mlog_errno(status);
550 return status;
551}
552
553/*
554 * Step 2: By now, we've completed the journal recovery, we've stamped
555 * a clean local alloc on disk and dropped the node out of the
556 * recovery map. Dlm locks will no longer stall, so lets clear out the
557 * main bitmap.
558 */
559int ocfs2_complete_local_alloc_recovery(struct ocfs2_super *osb,
560 struct ocfs2_dinode *alloc)
561{
562 int status;
563 handle_t *handle;
564 struct buffer_head *main_bm_bh = NULL;
565 struct inode *main_bm_inode;
566
567 main_bm_inode = ocfs2_get_system_file_inode(osb,
568 GLOBAL_BITMAP_SYSTEM_INODE,
569 OCFS2_INVALID_SLOT);
570 if (!main_bm_inode) {
571 status = -EINVAL;
572 mlog_errno(status);
573 goto out;
574 }
575
576 mutex_lock(&main_bm_inode->i_mutex);
577
578 status = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 1);
579 if (status < 0) {
580 mlog_errno(status);
581 goto out_mutex;
582 }
583
584 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
585 if (IS_ERR(handle)) {
586 status = PTR_ERR(handle);
587 handle = NULL;
588 mlog_errno(status);
589 goto out_unlock;
590 }
591
592 /* we want the bitmap change to be recorded on disk asap */
593 handle->h_sync = 1;
594
595 status = ocfs2_sync_local_to_main(osb, handle, alloc,
596 main_bm_inode, main_bm_bh);
597 if (status < 0)
598 mlog_errno(status);
599
600 ocfs2_commit_trans(osb, handle);
601
602out_unlock:
603 ocfs2_inode_unlock(main_bm_inode, 1);
604
605out_mutex:
606 mutex_unlock(&main_bm_inode->i_mutex);
607
608 brelse(main_bm_bh);
609
610 iput(main_bm_inode);
611
612out:
613 if (!status)
614 ocfs2_init_steal_slots(osb);
615 if (status)
616 mlog_errno(status);
617 return status;
618}
619
620/*
621 * make sure we've got at least bits_wanted contiguous bits in the
622 * local alloc. You lose them when you drop i_mutex.
623 *
624 * We will add ourselves to the transaction passed in, but may start
625 * our own in order to shift windows.
626 */
627int ocfs2_reserve_local_alloc_bits(struct ocfs2_super *osb,
628 u32 bits_wanted,
629 struct ocfs2_alloc_context *ac)
630{
631 int status;
632 struct ocfs2_dinode *alloc;
633 struct inode *local_alloc_inode;
634 unsigned int free_bits;
635
636 BUG_ON(!ac);
637
638 local_alloc_inode =
639 ocfs2_get_system_file_inode(osb,
640 LOCAL_ALLOC_SYSTEM_INODE,
641 osb->slot_num);
642 if (!local_alloc_inode) {
643 status = -ENOENT;
644 mlog_errno(status);
645 goto bail;
646 }
647
648 mutex_lock(&local_alloc_inode->i_mutex);
649
650 /*
651 * We must double check state and allocator bits because
652 * another process may have changed them while holding i_mutex.
653 */
654 spin_lock(&osb->osb_lock);
655 if (!ocfs2_la_state_enabled(osb) ||
656 (bits_wanted > osb->local_alloc_bits)) {
657 spin_unlock(&osb->osb_lock);
658 status = -ENOSPC;
659 goto bail;
660 }
661 spin_unlock(&osb->osb_lock);
662
663 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
664
665#ifdef CONFIG_OCFS2_DEBUG_FS
666 if (le32_to_cpu(alloc->id1.bitmap1.i_used) !=
667 ocfs2_local_alloc_count_bits(alloc)) {
668 ocfs2_error(osb->sb, "local alloc inode %llu says it has "
669 "%u free bits, but a count shows %u",
670 (unsigned long long)le64_to_cpu(alloc->i_blkno),
671 le32_to_cpu(alloc->id1.bitmap1.i_used),
672 ocfs2_local_alloc_count_bits(alloc));
673 status = -EIO;
674 goto bail;
675 }
676#endif
677
678 free_bits = le32_to_cpu(alloc->id1.bitmap1.i_total) -
679 le32_to_cpu(alloc->id1.bitmap1.i_used);
680 if (bits_wanted > free_bits) {
681 /* uhoh, window change time. */
682 status =
683 ocfs2_local_alloc_slide_window(osb, local_alloc_inode);
684 if (status < 0) {
685 if (status != -ENOSPC)
686 mlog_errno(status);
687 goto bail;
688 }
689
690 /*
691 * Under certain conditions, the window slide code
692 * might have reduced the number of bits available or
693 * disabled the the local alloc entirely. Re-check
694 * here and return -ENOSPC if necessary.
695 */
696 status = -ENOSPC;
697 if (!ocfs2_la_state_enabled(osb))
698 goto bail;
699
700 free_bits = le32_to_cpu(alloc->id1.bitmap1.i_total) -
701 le32_to_cpu(alloc->id1.bitmap1.i_used);
702 if (bits_wanted > free_bits)
703 goto bail;
704 }
705
706 ac->ac_inode = local_alloc_inode;
707 /* We should never use localalloc from another slot */
708 ac->ac_alloc_slot = osb->slot_num;
709 ac->ac_which = OCFS2_AC_USE_LOCAL;
710 get_bh(osb->local_alloc_bh);
711 ac->ac_bh = osb->local_alloc_bh;
712 status = 0;
713bail:
714 if (status < 0 && local_alloc_inode) {
715 mutex_unlock(&local_alloc_inode->i_mutex);
716 iput(local_alloc_inode);
717 }
718
719 trace_ocfs2_reserve_local_alloc_bits(
720 (unsigned long long)ac->ac_max_block,
721 bits_wanted, osb->slot_num, status);
722
723 if (status)
724 mlog_errno(status);
725 return status;
726}
727
728int ocfs2_claim_local_alloc_bits(struct ocfs2_super *osb,
729 handle_t *handle,
730 struct ocfs2_alloc_context *ac,
731 u32 bits_wanted,
732 u32 *bit_off,
733 u32 *num_bits)
734{
735 int status, start;
736 struct inode *local_alloc_inode;
737 void *bitmap;
738 struct ocfs2_dinode *alloc;
739 struct ocfs2_local_alloc *la;
740
741 BUG_ON(ac->ac_which != OCFS2_AC_USE_LOCAL);
742
743 local_alloc_inode = ac->ac_inode;
744 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
745 la = OCFS2_LOCAL_ALLOC(alloc);
746
747 start = ocfs2_local_alloc_find_clear_bits(osb, alloc, &bits_wanted,
748 ac->ac_resv);
749 if (start == -1) {
750 /* TODO: Shouldn't we just BUG here? */
751 status = -ENOSPC;
752 mlog_errno(status);
753 goto bail;
754 }
755
756 bitmap = la->la_bitmap;
757 *bit_off = le32_to_cpu(la->la_bm_off) + start;
758 *num_bits = bits_wanted;
759
760 status = ocfs2_journal_access_di(handle,
761 INODE_CACHE(local_alloc_inode),
762 osb->local_alloc_bh,
763 OCFS2_JOURNAL_ACCESS_WRITE);
764 if (status < 0) {
765 mlog_errno(status);
766 goto bail;
767 }
768
769 ocfs2_resmap_claimed_bits(&osb->osb_la_resmap, ac->ac_resv, start,
770 bits_wanted);
771
772 while(bits_wanted--)
773 ocfs2_set_bit(start++, bitmap);
774
775 le32_add_cpu(&alloc->id1.bitmap1.i_used, *num_bits);
776 ocfs2_journal_dirty(handle, osb->local_alloc_bh);
777
778bail:
779 if (status)
780 mlog_errno(status);
781 return status;
782}
783
784int ocfs2_free_local_alloc_bits(struct ocfs2_super *osb,
785 handle_t *handle,
786 struct ocfs2_alloc_context *ac,
787 u32 bit_off,
788 u32 num_bits)
789{
790 int status, start;
791 u32 clear_bits;
792 struct inode *local_alloc_inode;
793 void *bitmap;
794 struct ocfs2_dinode *alloc;
795 struct ocfs2_local_alloc *la;
796
797 BUG_ON(ac->ac_which != OCFS2_AC_USE_LOCAL);
798
799 local_alloc_inode = ac->ac_inode;
800 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
801 la = OCFS2_LOCAL_ALLOC(alloc);
802
803 bitmap = la->la_bitmap;
804 start = bit_off - le32_to_cpu(la->la_bm_off);
805 clear_bits = num_bits;
806
807 status = ocfs2_journal_access_di(handle,
808 INODE_CACHE(local_alloc_inode),
809 osb->local_alloc_bh,
810 OCFS2_JOURNAL_ACCESS_WRITE);
811 if (status < 0) {
812 mlog_errno(status);
813 goto bail;
814 }
815
816 while (clear_bits--)
817 ocfs2_clear_bit(start++, bitmap);
818
819 le32_add_cpu(&alloc->id1.bitmap1.i_used, -num_bits);
820 ocfs2_journal_dirty(handle, osb->local_alloc_bh);
821
822bail:
823 return status;
824}
825
826static u32 ocfs2_local_alloc_count_bits(struct ocfs2_dinode *alloc)
827{
828 u32 count;
829 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
830
831 count = memweight(la->la_bitmap, le16_to_cpu(la->la_size));
832
833 trace_ocfs2_local_alloc_count_bits(count);
834 return count;
835}
836
837static int ocfs2_local_alloc_find_clear_bits(struct ocfs2_super *osb,
838 struct ocfs2_dinode *alloc,
839 u32 *numbits,
840 struct ocfs2_alloc_reservation *resv)
841{
842 int numfound, bitoff, left, startoff, lastzero;
843 int local_resv = 0;
844 struct ocfs2_alloc_reservation r;
845 void *bitmap = NULL;
846 struct ocfs2_reservation_map *resmap = &osb->osb_la_resmap;
847
848 if (!alloc->id1.bitmap1.i_total) {
849 bitoff = -1;
850 goto bail;
851 }
852
853 if (!resv) {
854 local_resv = 1;
855 ocfs2_resv_init_once(&r);
856 ocfs2_resv_set_type(&r, OCFS2_RESV_FLAG_TMP);
857 resv = &r;
858 }
859
860 numfound = *numbits;
861 if (ocfs2_resmap_resv_bits(resmap, resv, &bitoff, &numfound) == 0) {
862 if (numfound < *numbits)
863 *numbits = numfound;
864 goto bail;
865 }
866
867 /*
868 * Code error. While reservations are enabled, local
869 * allocation should _always_ go through them.
870 */
871 BUG_ON(osb->osb_resv_level != 0);
872
873 /*
874 * Reservations are disabled. Handle this the old way.
875 */
876
877 bitmap = OCFS2_LOCAL_ALLOC(alloc)->la_bitmap;
878
879 numfound = bitoff = startoff = 0;
880 lastzero = -1;
881 left = le32_to_cpu(alloc->id1.bitmap1.i_total);
882 while ((bitoff = ocfs2_find_next_zero_bit(bitmap, left, startoff)) != -1) {
883 if (bitoff == left) {
884 /* mlog(0, "bitoff (%d) == left", bitoff); */
885 break;
886 }
887 /* mlog(0, "Found a zero: bitoff = %d, startoff = %d, "
888 "numfound = %d\n", bitoff, startoff, numfound);*/
889
890 /* Ok, we found a zero bit... is it contig. or do we
891 * start over?*/
892 if (bitoff == startoff) {
893 /* we found a zero */
894 numfound++;
895 startoff++;
896 } else {
897 /* got a zero after some ones */
898 numfound = 1;
899 startoff = bitoff+1;
900 }
901 /* we got everything we needed */
902 if (numfound == *numbits) {
903 /* mlog(0, "Found it all!\n"); */
904 break;
905 }
906 }
907
908 trace_ocfs2_local_alloc_find_clear_bits_search_bitmap(bitoff, numfound);
909
910 if (numfound == *numbits)
911 bitoff = startoff - numfound;
912 else
913 bitoff = -1;
914
915bail:
916 if (local_resv)
917 ocfs2_resv_discard(resmap, resv);
918
919 trace_ocfs2_local_alloc_find_clear_bits(*numbits,
920 le32_to_cpu(alloc->id1.bitmap1.i_total),
921 bitoff, numfound);
922
923 return bitoff;
924}
925
926static void ocfs2_clear_local_alloc(struct ocfs2_dinode *alloc)
927{
928 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
929 int i;
930
931 alloc->id1.bitmap1.i_total = 0;
932 alloc->id1.bitmap1.i_used = 0;
933 la->la_bm_off = 0;
934 for(i = 0; i < le16_to_cpu(la->la_size); i++)
935 la->la_bitmap[i] = 0;
936}
937
938#if 0
939/* turn this on and uncomment below to aid debugging window shifts. */
940static void ocfs2_verify_zero_bits(unsigned long *bitmap,
941 unsigned int start,
942 unsigned int count)
943{
944 unsigned int tmp = count;
945 while(tmp--) {
946 if (ocfs2_test_bit(start + tmp, bitmap)) {
947 printk("ocfs2_verify_zero_bits: start = %u, count = "
948 "%u\n", start, count);
949 printk("ocfs2_verify_zero_bits: bit %u is set!",
950 start + tmp);
951 BUG();
952 }
953 }
954}
955#endif
956
957/*
958 * sync the local alloc to main bitmap.
959 *
960 * assumes you've already locked the main bitmap -- the bitmap inode
961 * passed is used for caching.
962 */
963static int ocfs2_sync_local_to_main(struct ocfs2_super *osb,
964 handle_t *handle,
965 struct ocfs2_dinode *alloc,
966 struct inode *main_bm_inode,
967 struct buffer_head *main_bm_bh)
968{
969 int status = 0;
970 int bit_off, left, count, start;
971 u64 la_start_blk;
972 u64 blkno;
973 void *bitmap;
974 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
975
976 trace_ocfs2_sync_local_to_main(
977 le32_to_cpu(alloc->id1.bitmap1.i_total),
978 le32_to_cpu(alloc->id1.bitmap1.i_used));
979
980 if (!alloc->id1.bitmap1.i_total) {
981 goto bail;
982 }
983
984 if (le32_to_cpu(alloc->id1.bitmap1.i_used) ==
985 le32_to_cpu(alloc->id1.bitmap1.i_total)) {
986 goto bail;
987 }
988
989 la_start_blk = ocfs2_clusters_to_blocks(osb->sb,
990 le32_to_cpu(la->la_bm_off));
991 bitmap = la->la_bitmap;
992 start = count = bit_off = 0;
993 left = le32_to_cpu(alloc->id1.bitmap1.i_total);
994
995 while ((bit_off = ocfs2_find_next_zero_bit(bitmap, left, start))
996 != -1) {
997 if ((bit_off < left) && (bit_off == start)) {
998 count++;
999 start++;
1000 continue;
1001 }
1002 if (count) {
1003 blkno = la_start_blk +
1004 ocfs2_clusters_to_blocks(osb->sb,
1005 start - count);
1006
1007 trace_ocfs2_sync_local_to_main_free(
1008 count, start - count,
1009 (unsigned long long)la_start_blk,
1010 (unsigned long long)blkno);
1011
1012 status = ocfs2_release_clusters(handle,
1013 main_bm_inode,
1014 main_bm_bh, blkno,
1015 count);
1016 if (status < 0) {
1017 mlog_errno(status);
1018 goto bail;
1019 }
1020 }
1021 if (bit_off >= left)
1022 break;
1023 count = 1;
1024 start = bit_off + 1;
1025 }
1026
1027bail:
1028 if (status)
1029 mlog_errno(status);
1030 return status;
1031}
1032
1033enum ocfs2_la_event {
1034 OCFS2_LA_EVENT_SLIDE, /* Normal window slide. */
1035 OCFS2_LA_EVENT_FRAGMENTED, /* The global bitmap has
1036 * enough bits theoretically
1037 * free, but a contiguous
1038 * allocation could not be
1039 * found. */
1040 OCFS2_LA_EVENT_ENOSPC, /* Global bitmap doesn't have
1041 * enough bits free to satisfy
1042 * our request. */
1043};
1044#define OCFS2_LA_ENABLE_INTERVAL (30 * HZ)
1045/*
1046 * Given an event, calculate the size of our next local alloc window.
1047 *
1048 * This should always be called under i_mutex of the local alloc inode
1049 * so that local alloc disabling doesn't race with processes trying to
1050 * use the allocator.
1051 *
1052 * Returns the state which the local alloc was left in. This value can
1053 * be ignored by some paths.
1054 */
1055static int ocfs2_recalc_la_window(struct ocfs2_super *osb,
1056 enum ocfs2_la_event event)
1057{
1058 unsigned int bits;
1059 int state;
1060
1061 spin_lock(&osb->osb_lock);
1062 if (osb->local_alloc_state == OCFS2_LA_DISABLED) {
1063 WARN_ON_ONCE(osb->local_alloc_state == OCFS2_LA_DISABLED);
1064 goto out_unlock;
1065 }
1066
1067 /*
1068 * ENOSPC and fragmentation are treated similarly for now.
1069 */
1070 if (event == OCFS2_LA_EVENT_ENOSPC ||
1071 event == OCFS2_LA_EVENT_FRAGMENTED) {
1072 /*
1073 * We ran out of contiguous space in the primary
1074 * bitmap. Drastically reduce the number of bits used
1075 * by local alloc until we have to disable it.
1076 */
1077 bits = osb->local_alloc_bits >> 1;
1078 if (bits > ocfs2_megabytes_to_clusters(osb->sb, 1)) {
1079 /*
1080 * By setting state to THROTTLED, we'll keep
1081 * the number of local alloc bits used down
1082 * until an event occurs which would give us
1083 * reason to assume the bitmap situation might
1084 * have changed.
1085 */
1086 osb->local_alloc_state = OCFS2_LA_THROTTLED;
1087 osb->local_alloc_bits = bits;
1088 } else {
1089 osb->local_alloc_state = OCFS2_LA_DISABLED;
1090 }
1091 queue_delayed_work(ocfs2_wq, &osb->la_enable_wq,
1092 OCFS2_LA_ENABLE_INTERVAL);
1093 goto out_unlock;
1094 }
1095
1096 /*
1097 * Don't increase the size of the local alloc window until we
1098 * know we might be able to fulfill the request. Otherwise, we
1099 * risk bouncing around the global bitmap during periods of
1100 * low space.
1101 */
1102 if (osb->local_alloc_state != OCFS2_LA_THROTTLED)
1103 osb->local_alloc_bits = osb->local_alloc_default_bits;
1104
1105out_unlock:
1106 state = osb->local_alloc_state;
1107 spin_unlock(&osb->osb_lock);
1108
1109 return state;
1110}
1111
1112static int ocfs2_local_alloc_reserve_for_window(struct ocfs2_super *osb,
1113 struct ocfs2_alloc_context **ac,
1114 struct inode **bitmap_inode,
1115 struct buffer_head **bitmap_bh)
1116{
1117 int status;
1118
1119 *ac = kzalloc(sizeof(struct ocfs2_alloc_context), GFP_KERNEL);
1120 if (!(*ac)) {
1121 status = -ENOMEM;
1122 mlog_errno(status);
1123 goto bail;
1124 }
1125
1126retry_enospc:
1127 (*ac)->ac_bits_wanted = osb->local_alloc_bits;
1128 status = ocfs2_reserve_cluster_bitmap_bits(osb, *ac);
1129 if (status == -ENOSPC) {
1130 if (ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_ENOSPC) ==
1131 OCFS2_LA_DISABLED)
1132 goto bail;
1133
1134 ocfs2_free_ac_resource(*ac);
1135 memset(*ac, 0, sizeof(struct ocfs2_alloc_context));
1136 goto retry_enospc;
1137 }
1138 if (status < 0) {
1139 mlog_errno(status);
1140 goto bail;
1141 }
1142
1143 *bitmap_inode = (*ac)->ac_inode;
1144 igrab(*bitmap_inode);
1145 *bitmap_bh = (*ac)->ac_bh;
1146 get_bh(*bitmap_bh);
1147 status = 0;
1148bail:
1149 if ((status < 0) && *ac) {
1150 ocfs2_free_alloc_context(*ac);
1151 *ac = NULL;
1152 }
1153
1154 if (status)
1155 mlog_errno(status);
1156 return status;
1157}
1158
1159/*
1160 * pass it the bitmap lock in lock_bh if you have it.
1161 */
1162static int ocfs2_local_alloc_new_window(struct ocfs2_super *osb,
1163 handle_t *handle,
1164 struct ocfs2_alloc_context *ac)
1165{
1166 int status = 0;
1167 u32 cluster_off, cluster_count;
1168 struct ocfs2_dinode *alloc = NULL;
1169 struct ocfs2_local_alloc *la;
1170
1171 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
1172 la = OCFS2_LOCAL_ALLOC(alloc);
1173
1174 trace_ocfs2_local_alloc_new_window(
1175 le32_to_cpu(alloc->id1.bitmap1.i_total),
1176 osb->local_alloc_bits);
1177
1178 /* Instruct the allocation code to try the most recently used
1179 * cluster group. We'll re-record the group used this pass
1180 * below. */
1181 ac->ac_last_group = osb->la_last_gd;
1182
1183 /* we used the generic suballoc reserve function, but we set
1184 * everything up nicely, so there's no reason why we can't use
1185 * the more specific cluster api to claim bits. */
1186 status = ocfs2_claim_clusters(handle, ac, osb->local_alloc_bits,
1187 &cluster_off, &cluster_count);
1188 if (status == -ENOSPC) {
1189retry_enospc:
1190 /*
1191 * Note: We could also try syncing the journal here to
1192 * allow use of any free bits which the current
1193 * transaction can't give us access to. --Mark
1194 */
1195 if (ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_FRAGMENTED) ==
1196 OCFS2_LA_DISABLED)
1197 goto bail;
1198
1199 ac->ac_bits_wanted = osb->local_alloc_bits;
1200 status = ocfs2_claim_clusters(handle, ac,
1201 osb->local_alloc_bits,
1202 &cluster_off,
1203 &cluster_count);
1204 if (status == -ENOSPC)
1205 goto retry_enospc;
1206 /*
1207 * We only shrunk the *minimum* number of in our
1208 * request - it's entirely possible that the allocator
1209 * might give us more than we asked for.
1210 */
1211 if (status == 0) {
1212 spin_lock(&osb->osb_lock);
1213 osb->local_alloc_bits = cluster_count;
1214 spin_unlock(&osb->osb_lock);
1215 }
1216 }
1217 if (status < 0) {
1218 if (status != -ENOSPC)
1219 mlog_errno(status);
1220 goto bail;
1221 }
1222
1223 osb->la_last_gd = ac->ac_last_group;
1224
1225 la->la_bm_off = cpu_to_le32(cluster_off);
1226 alloc->id1.bitmap1.i_total = cpu_to_le32(cluster_count);
1227 /* just in case... In the future when we find space ourselves,
1228 * we don't have to get all contiguous -- but we'll have to
1229 * set all previously used bits in bitmap and update
1230 * la_bits_set before setting the bits in the main bitmap. */
1231 alloc->id1.bitmap1.i_used = 0;
1232 memset(OCFS2_LOCAL_ALLOC(alloc)->la_bitmap, 0,
1233 le16_to_cpu(la->la_size));
1234
1235 ocfs2_resmap_restart(&osb->osb_la_resmap, cluster_count,
1236 OCFS2_LOCAL_ALLOC(alloc)->la_bitmap);
1237
1238 trace_ocfs2_local_alloc_new_window_result(
1239 OCFS2_LOCAL_ALLOC(alloc)->la_bm_off,
1240 le32_to_cpu(alloc->id1.bitmap1.i_total));
1241
1242bail:
1243 if (status)
1244 mlog_errno(status);
1245 return status;
1246}
1247
1248/* Note that we do *NOT* lock the local alloc inode here as
1249 * it's been locked already for us. */
1250static int ocfs2_local_alloc_slide_window(struct ocfs2_super *osb,
1251 struct inode *local_alloc_inode)
1252{
1253 int status = 0;
1254 struct buffer_head *main_bm_bh = NULL;
1255 struct inode *main_bm_inode = NULL;
1256 handle_t *handle = NULL;
1257 struct ocfs2_dinode *alloc;
1258 struct ocfs2_dinode *alloc_copy = NULL;
1259 struct ocfs2_alloc_context *ac = NULL;
1260
1261 ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_SLIDE);
1262
1263 /* This will lock the main bitmap for us. */
1264 status = ocfs2_local_alloc_reserve_for_window(osb,
1265 &ac,
1266 &main_bm_inode,
1267 &main_bm_bh);
1268 if (status < 0) {
1269 if (status != -ENOSPC)
1270 mlog_errno(status);
1271 goto bail;
1272 }
1273
1274 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
1275 if (IS_ERR(handle)) {
1276 status = PTR_ERR(handle);
1277 handle = NULL;
1278 mlog_errno(status);
1279 goto bail;
1280 }
1281
1282 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
1283
1284 /* We want to clear the local alloc before doing anything
1285 * else, so that if we error later during this operation,
1286 * local alloc shutdown won't try to double free main bitmap
1287 * bits. Make a copy so the sync function knows which bits to
1288 * free. */
1289 alloc_copy = kmalloc(osb->local_alloc_bh->b_size, GFP_NOFS);
1290 if (!alloc_copy) {
1291 status = -ENOMEM;
1292 mlog_errno(status);
1293 goto bail;
1294 }
1295 memcpy(alloc_copy, alloc, osb->local_alloc_bh->b_size);
1296
1297 status = ocfs2_journal_access_di(handle,
1298 INODE_CACHE(local_alloc_inode),
1299 osb->local_alloc_bh,
1300 OCFS2_JOURNAL_ACCESS_WRITE);
1301 if (status < 0) {
1302 mlog_errno(status);
1303 goto bail;
1304 }
1305
1306 ocfs2_clear_local_alloc(alloc);
1307 ocfs2_journal_dirty(handle, osb->local_alloc_bh);
1308
1309 status = ocfs2_sync_local_to_main(osb, handle, alloc_copy,
1310 main_bm_inode, main_bm_bh);
1311 if (status < 0) {
1312 mlog_errno(status);
1313 goto bail;
1314 }
1315
1316 status = ocfs2_local_alloc_new_window(osb, handle, ac);
1317 if (status < 0) {
1318 if (status != -ENOSPC)
1319 mlog_errno(status);
1320 goto bail;
1321 }
1322
1323 atomic_inc(&osb->alloc_stats.moves);
1324
1325bail:
1326 if (handle)
1327 ocfs2_commit_trans(osb, handle);
1328
1329 brelse(main_bm_bh);
1330
1331 if (main_bm_inode)
1332 iput(main_bm_inode);
1333
1334 kfree(alloc_copy);
1335
1336 if (ac)
1337 ocfs2_free_alloc_context(ac);
1338
1339 if (status)
1340 mlog_errno(status);
1341 return status;
1342}
1343