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