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(sb: osb->sb,
110 clusters: 8 * ocfs2_group_bitmap_size(sb, suballocator: 0, feature_incompat: 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(sb: osb->sb, clusters: 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(sb: osb->sb,
172 clusters: 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 clusters: ocfs2_local_alloc_size(sb) * 8);
187
188 trace_ocfs2_la_set_sizes(value1: requested_mb, value2: la_max_mb, value3: 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, megs: 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, megs: la_max_mb);
198 } else {
199 osb->local_alloc_default_bits =
200 ocfs2_megabytes_to_clusters(sb, megs: requested_mb);
201 }
202
203 osb->local_alloc_bits = osb->local_alloc_default_bits;
204}
205
206static inline int ocfs2_la_state_enabled(struct ocfs2_super *osb)
207{
208 return (osb->local_alloc_state == OCFS2_LA_THROTTLED ||
209 osb->local_alloc_state == OCFS2_LA_ENABLED);
210}
211
212void ocfs2_local_alloc_seen_free_bits(struct ocfs2_super *osb,
213 unsigned int num_clusters)
214{
215 spin_lock(lock: &osb->osb_lock);
216 if (osb->local_alloc_state == OCFS2_LA_DISABLED ||
217 osb->local_alloc_state == OCFS2_LA_THROTTLED)
218 if (num_clusters >= osb->local_alloc_default_bits) {
219 cancel_delayed_work(dwork: &osb->la_enable_wq);
220 osb->local_alloc_state = OCFS2_LA_ENABLED;
221 }
222 spin_unlock(lock: &osb->osb_lock);
223}
224
225void ocfs2_la_enable_worker(struct work_struct *work)
226{
227 struct ocfs2_super *osb =
228 container_of(work, struct ocfs2_super,
229 la_enable_wq.work);
230 spin_lock(lock: &osb->osb_lock);
231 osb->local_alloc_state = OCFS2_LA_ENABLED;
232 spin_unlock(lock: &osb->osb_lock);
233}
234
235/*
236 * Tell us whether a given allocation should use the local alloc
237 * file. Otherwise, it has to go to the main bitmap.
238 *
239 * This function does semi-dirty reads of local alloc size and state!
240 * This is ok however, as the values are re-checked once under mutex.
241 */
242int ocfs2_alloc_should_use_local(struct ocfs2_super *osb, u64 bits)
243{
244 int ret = 0;
245 int la_bits;
246
247 spin_lock(lock: &osb->osb_lock);
248 la_bits = osb->local_alloc_bits;
249
250 if (!ocfs2_la_state_enabled(osb))
251 goto bail;
252
253 /* la_bits should be at least twice the size (in clusters) of
254 * a new block group. We want to be sure block group
255 * allocations go through the local alloc, so allow an
256 * allocation to take up to half the bitmap. */
257 if (bits > (la_bits / 2))
258 goto bail;
259
260 ret = 1;
261bail:
262 trace_ocfs2_alloc_should_use_local(
263 ull: (unsigned long long)bits, value1: osb->local_alloc_state, value2: la_bits, value3: ret);
264 spin_unlock(lock: &osb->osb_lock);
265 return ret;
266}
267
268int ocfs2_load_local_alloc(struct ocfs2_super *osb)
269{
270 int status = 0;
271 struct ocfs2_dinode *alloc = NULL;
272 struct buffer_head *alloc_bh = NULL;
273 u32 num_used;
274 struct inode *inode = NULL;
275 struct ocfs2_local_alloc *la;
276
277 if (osb->local_alloc_bits == 0)
278 goto bail;
279
280 if (osb->local_alloc_bits >= osb->bitmap_cpg) {
281 mlog(ML_NOTICE, "Requested local alloc window %d is larger "
282 "than max possible %u. Using defaults.\n",
283 osb->local_alloc_bits, (osb->bitmap_cpg - 1));
284 osb->local_alloc_bits =
285 ocfs2_megabytes_to_clusters(sb: osb->sb,
286 megs: ocfs2_la_default_mb(osb));
287 }
288
289 /* read the alloc off disk */
290 inode = ocfs2_get_system_file_inode(osb, type: LOCAL_ALLOC_SYSTEM_INODE,
291 slot: osb->slot_num);
292 if (!inode) {
293 status = -EINVAL;
294 mlog_errno(status);
295 goto bail;
296 }
297
298 status = ocfs2_read_inode_block_full(inode, bh: &alloc_bh,
299 OCFS2_BH_IGNORE_CACHE);
300 if (status < 0) {
301 mlog_errno(status);
302 goto bail;
303 }
304
305 alloc = (struct ocfs2_dinode *) alloc_bh->b_data;
306 la = OCFS2_LOCAL_ALLOC(alloc);
307
308 if (!(le32_to_cpu(alloc->i_flags) &
309 (OCFS2_LOCAL_ALLOC_FL|OCFS2_BITMAP_FL))) {
310 mlog(ML_ERROR, "Invalid local alloc inode, %llu\n",
311 (unsigned long long)OCFS2_I(inode)->ip_blkno);
312 status = -EINVAL;
313 goto bail;
314 }
315
316 if ((la->la_size == 0) ||
317 (le16_to_cpu(la->la_size) > ocfs2_local_alloc_size(sb: inode->i_sb))) {
318 mlog(ML_ERROR, "Local alloc size is invalid (la_size = %u)\n",
319 le16_to_cpu(la->la_size));
320 status = -EINVAL;
321 goto bail;
322 }
323
324 /* do a little verification. */
325 num_used = ocfs2_local_alloc_count_bits(alloc);
326
327 /* hopefully the local alloc has always been recovered before
328 * we load it. */
329 if (num_used
330 || alloc->id1.bitmap1.i_used
331 || alloc->id1.bitmap1.i_total
332 || la->la_bm_off) {
333 mlog(ML_ERROR, "inconsistent detected, clean journal with"
334 " unrecovered local alloc, please run fsck.ocfs2!\n"
335 "found = %u, set = %u, taken = %u, off = %u\n",
336 num_used, le32_to_cpu(alloc->id1.bitmap1.i_used),
337 le32_to_cpu(alloc->id1.bitmap1.i_total),
338 OCFS2_LOCAL_ALLOC(alloc)->la_bm_off);
339
340 status = -EINVAL;
341 goto bail;
342 }
343
344 osb->local_alloc_bh = alloc_bh;
345 osb->local_alloc_state = OCFS2_LA_ENABLED;
346
347bail:
348 if (status < 0)
349 brelse(bh: alloc_bh);
350 iput(inode);
351
352 trace_ocfs2_load_local_alloc(num: osb->local_alloc_bits);
353
354 if (status)
355 mlog_errno(status);
356 return status;
357}
358
359/*
360 * return any unused bits to the bitmap and write out a clean
361 * local_alloc.
362 *
363 * local_alloc_bh is optional. If not passed, we will simply use the
364 * one off osb. If you do pass it however, be warned that it *will* be
365 * returned brelse'd and NULL'd out.*/
366void ocfs2_shutdown_local_alloc(struct ocfs2_super *osb)
367{
368 int status;
369 handle_t *handle;
370 struct inode *local_alloc_inode = NULL;
371 struct buffer_head *bh = NULL;
372 struct buffer_head *main_bm_bh = NULL;
373 struct inode *main_bm_inode = NULL;
374 struct ocfs2_dinode *alloc_copy = NULL;
375 struct ocfs2_dinode *alloc = NULL;
376
377 cancel_delayed_work(dwork: &osb->la_enable_wq);
378 if (osb->ocfs2_wq)
379 flush_workqueue(osb->ocfs2_wq);
380
381 if (osb->local_alloc_state == OCFS2_LA_UNUSED)
382 goto out;
383
384 local_alloc_inode =
385 ocfs2_get_system_file_inode(osb,
386 type: LOCAL_ALLOC_SYSTEM_INODE,
387 slot: osb->slot_num);
388 if (!local_alloc_inode) {
389 status = -ENOENT;
390 mlog_errno(status);
391 goto out;
392 }
393
394 osb->local_alloc_state = OCFS2_LA_DISABLED;
395
396 ocfs2_resmap_uninit(resmap: &osb->osb_la_resmap);
397
398 main_bm_inode = ocfs2_get_system_file_inode(osb,
399 type: GLOBAL_BITMAP_SYSTEM_INODE,
400 OCFS2_INVALID_SLOT);
401 if (!main_bm_inode) {
402 status = -EINVAL;
403 mlog_errno(status);
404 goto out;
405 }
406
407 inode_lock(inode: main_bm_inode);
408
409 status = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 1);
410 if (status < 0) {
411 mlog_errno(status);
412 goto out_mutex;
413 }
414
415 /* WINDOW_MOVE_CREDITS is a bit heavy... */
416 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
417 if (IS_ERR(ptr: handle)) {
418 mlog_errno(PTR_ERR(handle));
419 handle = NULL;
420 goto out_unlock;
421 }
422
423 bh = osb->local_alloc_bh;
424 alloc = (struct ocfs2_dinode *) bh->b_data;
425
426 alloc_copy = kmemdup(p: alloc, size: bh->b_size, GFP_NOFS);
427 if (!alloc_copy) {
428 status = -ENOMEM;
429 goto out_commit;
430 }
431
432 status = ocfs2_journal_access_di(handle, ci: INODE_CACHE(inode: local_alloc_inode),
433 bh, OCFS2_JOURNAL_ACCESS_WRITE);
434 if (status < 0) {
435 mlog_errno(status);
436 goto out_commit;
437 }
438
439 ocfs2_clear_local_alloc(alloc);
440 ocfs2_journal_dirty(handle, bh);
441
442 brelse(bh);
443 osb->local_alloc_bh = NULL;
444 osb->local_alloc_state = OCFS2_LA_UNUSED;
445
446 status = ocfs2_sync_local_to_main(osb, handle, alloc: alloc_copy,
447 main_bm_inode, main_bm_bh);
448 if (status < 0)
449 mlog_errno(status);
450
451out_commit:
452 ocfs2_commit_trans(osb, handle);
453
454out_unlock:
455 brelse(bh: main_bm_bh);
456
457 ocfs2_inode_unlock(inode: main_bm_inode, ex: 1);
458
459out_mutex:
460 inode_unlock(inode: main_bm_inode);
461 iput(main_bm_inode);
462
463out:
464 iput(local_alloc_inode);
465
466 kfree(objp: alloc_copy);
467}
468
469/*
470 * We want to free the bitmap bits outside of any recovery context as
471 * we'll need a cluster lock to do so, but we must clear the local
472 * alloc before giving up the recovered nodes journal. To solve this,
473 * we kmalloc a copy of the local alloc before it's change for the
474 * caller to process with ocfs2_complete_local_alloc_recovery
475 */
476int ocfs2_begin_local_alloc_recovery(struct ocfs2_super *osb,
477 int slot_num,
478 struct ocfs2_dinode **alloc_copy)
479{
480 int status = 0;
481 struct buffer_head *alloc_bh = NULL;
482 struct inode *inode = NULL;
483 struct ocfs2_dinode *alloc;
484
485 trace_ocfs2_begin_local_alloc_recovery(num: slot_num);
486
487 *alloc_copy = NULL;
488
489 inode = ocfs2_get_system_file_inode(osb,
490 type: LOCAL_ALLOC_SYSTEM_INODE,
491 slot: slot_num);
492 if (!inode) {
493 status = -EINVAL;
494 mlog_errno(status);
495 goto bail;
496 }
497
498 inode_lock(inode);
499
500 status = ocfs2_read_inode_block_full(inode, bh: &alloc_bh,
501 OCFS2_BH_IGNORE_CACHE);
502 if (status < 0) {
503 mlog_errno(status);
504 goto bail;
505 }
506
507 *alloc_copy = kmalloc(size: alloc_bh->b_size, GFP_KERNEL);
508 if (!(*alloc_copy)) {
509 status = -ENOMEM;
510 goto bail;
511 }
512 memcpy((*alloc_copy), alloc_bh->b_data, alloc_bh->b_size);
513
514 alloc = (struct ocfs2_dinode *) alloc_bh->b_data;
515 ocfs2_clear_local_alloc(alloc);
516
517 ocfs2_compute_meta_ecc(sb: osb->sb, data: alloc_bh->b_data, bc: &alloc->i_check);
518 status = ocfs2_write_block(osb, bh: alloc_bh, ci: INODE_CACHE(inode));
519 if (status < 0)
520 mlog_errno(status);
521
522bail:
523 if (status < 0) {
524 kfree(objp: *alloc_copy);
525 *alloc_copy = NULL;
526 }
527
528 brelse(bh: alloc_bh);
529
530 if (inode) {
531 inode_unlock(inode);
532 iput(inode);
533 }
534
535 if (status)
536 mlog_errno(status);
537 return status;
538}
539
540/*
541 * Step 2: By now, we've completed the journal recovery, we've stamped
542 * a clean local alloc on disk and dropped the node out of the
543 * recovery map. Dlm locks will no longer stall, so lets clear out the
544 * main bitmap.
545 */
546int ocfs2_complete_local_alloc_recovery(struct ocfs2_super *osb,
547 struct ocfs2_dinode *alloc)
548{
549 int status;
550 handle_t *handle;
551 struct buffer_head *main_bm_bh = NULL;
552 struct inode *main_bm_inode;
553
554 main_bm_inode = ocfs2_get_system_file_inode(osb,
555 type: GLOBAL_BITMAP_SYSTEM_INODE,
556 OCFS2_INVALID_SLOT);
557 if (!main_bm_inode) {
558 status = -EINVAL;
559 mlog_errno(status);
560 goto out;
561 }
562
563 inode_lock(inode: main_bm_inode);
564
565 status = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 1);
566 if (status < 0) {
567 mlog_errno(status);
568 goto out_mutex;
569 }
570
571 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
572 if (IS_ERR(ptr: handle)) {
573 status = PTR_ERR(ptr: handle);
574 handle = NULL;
575 mlog_errno(status);
576 goto out_unlock;
577 }
578
579 /* we want the bitmap change to be recorded on disk asap */
580 handle->h_sync = 1;
581
582 status = ocfs2_sync_local_to_main(osb, handle, alloc,
583 main_bm_inode, main_bm_bh);
584 if (status < 0)
585 mlog_errno(status);
586
587 ocfs2_commit_trans(osb, handle);
588
589out_unlock:
590 ocfs2_inode_unlock(inode: main_bm_inode, ex: 1);
591
592out_mutex:
593 inode_unlock(inode: main_bm_inode);
594
595 brelse(bh: main_bm_bh);
596
597 iput(main_bm_inode);
598
599out:
600 if (!status)
601 ocfs2_init_steal_slots(osb);
602 if (status)
603 mlog_errno(status);
604 return status;
605}
606
607/*
608 * make sure we've got at least bits_wanted contiguous bits in the
609 * local alloc. You lose them when you drop i_rwsem.
610 *
611 * We will add ourselves to the transaction passed in, but may start
612 * our own in order to shift windows.
613 */
614int ocfs2_reserve_local_alloc_bits(struct ocfs2_super *osb,
615 u32 bits_wanted,
616 struct ocfs2_alloc_context *ac)
617{
618 int status;
619 struct ocfs2_dinode *alloc;
620 struct inode *local_alloc_inode;
621 unsigned int free_bits;
622
623 BUG_ON(!ac);
624
625 local_alloc_inode =
626 ocfs2_get_system_file_inode(osb,
627 type: LOCAL_ALLOC_SYSTEM_INODE,
628 slot: osb->slot_num);
629 if (!local_alloc_inode) {
630 status = -ENOENT;
631 mlog_errno(status);
632 goto bail;
633 }
634
635 inode_lock(inode: local_alloc_inode);
636
637 /*
638 * We must double check state and allocator bits because
639 * another process may have changed them while holding i_rwsem.
640 */
641 spin_lock(lock: &osb->osb_lock);
642 if (!ocfs2_la_state_enabled(osb) ||
643 (bits_wanted > osb->local_alloc_bits)) {
644 spin_unlock(lock: &osb->osb_lock);
645 status = -ENOSPC;
646 goto bail;
647 }
648 spin_unlock(lock: &osb->osb_lock);
649
650 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
651
652#ifdef CONFIG_OCFS2_DEBUG_FS
653 if (le32_to_cpu(alloc->id1.bitmap1.i_used) !=
654 ocfs2_local_alloc_count_bits(alloc)) {
655 status = ocfs2_error(osb->sb, "local alloc inode %llu says it has %u used bits, but a count shows %u\n",
656 (unsigned long long)le64_to_cpu(alloc->i_blkno),
657 le32_to_cpu(alloc->id1.bitmap1.i_used),
658 ocfs2_local_alloc_count_bits(alloc));
659 goto bail;
660 }
661#endif
662
663 free_bits = le32_to_cpu(alloc->id1.bitmap1.i_total) -
664 le32_to_cpu(alloc->id1.bitmap1.i_used);
665 if (bits_wanted > free_bits) {
666 /* uhoh, window change time. */
667 status =
668 ocfs2_local_alloc_slide_window(osb, local_alloc_inode);
669 if (status < 0) {
670 if (status != -ENOSPC)
671 mlog_errno(status);
672 goto bail;
673 }
674
675 /*
676 * Under certain conditions, the window slide code
677 * might have reduced the number of bits available or
678 * disabled the local alloc entirely. Re-check
679 * here and return -ENOSPC if necessary.
680 */
681 status = -ENOSPC;
682 if (!ocfs2_la_state_enabled(osb))
683 goto bail;
684
685 free_bits = le32_to_cpu(alloc->id1.bitmap1.i_total) -
686 le32_to_cpu(alloc->id1.bitmap1.i_used);
687 if (bits_wanted > free_bits)
688 goto bail;
689 }
690
691 ac->ac_inode = local_alloc_inode;
692 /* We should never use localalloc from another slot */
693 ac->ac_alloc_slot = osb->slot_num;
694 ac->ac_which = OCFS2_AC_USE_LOCAL;
695 get_bh(bh: osb->local_alloc_bh);
696 ac->ac_bh = osb->local_alloc_bh;
697 status = 0;
698bail:
699 if (status < 0 && local_alloc_inode) {
700 inode_unlock(inode: local_alloc_inode);
701 iput(local_alloc_inode);
702 }
703
704 trace_ocfs2_reserve_local_alloc_bits(
705 ull: (unsigned long long)ac->ac_max_block,
706 value1: bits_wanted, value2: osb->slot_num, value3: status);
707
708 if (status)
709 mlog_errno(status);
710 return status;
711}
712
713int ocfs2_claim_local_alloc_bits(struct ocfs2_super *osb,
714 handle_t *handle,
715 struct ocfs2_alloc_context *ac,
716 u32 bits_wanted,
717 u32 *bit_off,
718 u32 *num_bits)
719{
720 int status, start;
721 struct inode *local_alloc_inode;
722 void *bitmap;
723 struct ocfs2_dinode *alloc;
724 struct ocfs2_local_alloc *la;
725
726 BUG_ON(ac->ac_which != OCFS2_AC_USE_LOCAL);
727
728 local_alloc_inode = ac->ac_inode;
729 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
730 la = OCFS2_LOCAL_ALLOC(alloc);
731
732 start = ocfs2_local_alloc_find_clear_bits(osb, alloc, numbits: &bits_wanted,
733 resv: ac->ac_resv);
734 if (start == -1) {
735 /* TODO: Shouldn't we just BUG here? */
736 status = -ENOSPC;
737 mlog_errno(status);
738 goto bail;
739 }
740
741 bitmap = la->la_bitmap;
742 *bit_off = le32_to_cpu(la->la_bm_off) + start;
743 *num_bits = bits_wanted;
744
745 status = ocfs2_journal_access_di(handle,
746 ci: INODE_CACHE(inode: local_alloc_inode),
747 bh: osb->local_alloc_bh,
748 OCFS2_JOURNAL_ACCESS_WRITE);
749 if (status < 0) {
750 mlog_errno(status);
751 goto bail;
752 }
753
754 ocfs2_resmap_claimed_bits(resmap: &osb->osb_la_resmap, resv: ac->ac_resv, cstart: start,
755 clen: bits_wanted);
756
757 while(bits_wanted--)
758 ocfs2_set_bit(start++, bitmap);
759
760 le32_add_cpu(var: &alloc->id1.bitmap1.i_used, val: *num_bits);
761 ocfs2_journal_dirty(handle, bh: osb->local_alloc_bh);
762
763bail:
764 if (status)
765 mlog_errno(status);
766 return status;
767}
768
769int ocfs2_free_local_alloc_bits(struct ocfs2_super *osb,
770 handle_t *handle,
771 struct ocfs2_alloc_context *ac,
772 u32 bit_off,
773 u32 num_bits)
774{
775 int status, start;
776 u32 clear_bits;
777 struct inode *local_alloc_inode;
778 void *bitmap;
779 struct ocfs2_dinode *alloc;
780 struct ocfs2_local_alloc *la;
781
782 BUG_ON(ac->ac_which != OCFS2_AC_USE_LOCAL);
783
784 local_alloc_inode = ac->ac_inode;
785 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
786 la = OCFS2_LOCAL_ALLOC(alloc);
787
788 bitmap = la->la_bitmap;
789 start = bit_off - le32_to_cpu(la->la_bm_off);
790 clear_bits = num_bits;
791
792 status = ocfs2_journal_access_di(handle,
793 ci: INODE_CACHE(inode: local_alloc_inode),
794 bh: osb->local_alloc_bh,
795 OCFS2_JOURNAL_ACCESS_WRITE);
796 if (status < 0) {
797 mlog_errno(status);
798 goto bail;
799 }
800
801 while (clear_bits--)
802 ocfs2_clear_bit(start++, bitmap);
803
804 le32_add_cpu(var: &alloc->id1.bitmap1.i_used, val: -num_bits);
805 ocfs2_journal_dirty(handle, bh: osb->local_alloc_bh);
806
807bail:
808 return status;
809}
810
811static u32 ocfs2_local_alloc_count_bits(struct ocfs2_dinode *alloc)
812{
813 u32 count;
814 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
815
816 count = memweight(ptr: la->la_bitmap, le16_to_cpu(la->la_size));
817
818 trace_ocfs2_local_alloc_count_bits(num: count);
819 return count;
820}
821
822static int ocfs2_local_alloc_find_clear_bits(struct ocfs2_super *osb,
823 struct ocfs2_dinode *alloc,
824 u32 *numbits,
825 struct ocfs2_alloc_reservation *resv)
826{
827 int numfound = 0, bitoff, left, startoff;
828 int local_resv = 0;
829 struct ocfs2_alloc_reservation r;
830 void *bitmap = NULL;
831 struct ocfs2_reservation_map *resmap = &osb->osb_la_resmap;
832
833 if (!alloc->id1.bitmap1.i_total) {
834 bitoff = -1;
835 goto bail;
836 }
837
838 if (!resv) {
839 local_resv = 1;
840 ocfs2_resv_init_once(resv: &r);
841 ocfs2_resv_set_type(resv: &r, OCFS2_RESV_FLAG_TMP);
842 resv = &r;
843 }
844
845 numfound = *numbits;
846 if (ocfs2_resmap_resv_bits(resmap, resv, cstart: &bitoff, clen: &numfound) == 0) {
847 if (numfound < *numbits)
848 *numbits = numfound;
849 goto bail;
850 }
851
852 /*
853 * Code error. While reservations are enabled, local
854 * allocation should _always_ go through them.
855 */
856 BUG_ON(osb->osb_resv_level != 0);
857
858 /*
859 * Reservations are disabled. Handle this the old way.
860 */
861
862 bitmap = OCFS2_LOCAL_ALLOC(alloc)->la_bitmap;
863
864 numfound = bitoff = startoff = 0;
865 left = le32_to_cpu(alloc->id1.bitmap1.i_total);
866 while ((bitoff = ocfs2_find_next_zero_bit(addr: bitmap, size: left, offset: startoff)) != -1) {
867 if (bitoff == left) {
868 /* mlog(0, "bitoff (%d) == left", bitoff); */
869 break;
870 }
871 /* mlog(0, "Found a zero: bitoff = %d, startoff = %d, "
872 "numfound = %d\n", bitoff, startoff, numfound);*/
873
874 /* Ok, we found a zero bit... is it contig. or do we
875 * start over?*/
876 if (bitoff == startoff) {
877 /* we found a zero */
878 numfound++;
879 startoff++;
880 } else {
881 /* got a zero after some ones */
882 numfound = 1;
883 startoff = bitoff+1;
884 }
885 /* we got everything we needed */
886 if (numfound == *numbits) {
887 /* mlog(0, "Found it all!\n"); */
888 break;
889 }
890 }
891
892 trace_ocfs2_local_alloc_find_clear_bits_search_bitmap(val1: bitoff, val2: numfound);
893
894 if (numfound == *numbits)
895 bitoff = startoff - numfound;
896 else
897 bitoff = -1;
898
899bail:
900 if (local_resv)
901 ocfs2_resv_discard(resmap, resv);
902
903 trace_ocfs2_local_alloc_find_clear_bits(ull: *numbits,
904 le32_to_cpu(alloc->id1.bitmap1.i_total),
905 value2: bitoff, value3: numfound);
906
907 return bitoff;
908}
909
910static void ocfs2_clear_local_alloc(struct ocfs2_dinode *alloc)
911{
912 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
913 int i;
914
915 alloc->id1.bitmap1.i_total = 0;
916 alloc->id1.bitmap1.i_used = 0;
917 la->la_bm_off = 0;
918 for(i = 0; i < le16_to_cpu(la->la_size); i++)
919 la->la_bitmap[i] = 0;
920}
921
922#if 0
923/* turn this on and uncomment below to aid debugging window shifts. */
924static void ocfs2_verify_zero_bits(unsigned long *bitmap,
925 unsigned int start,
926 unsigned int count)
927{
928 unsigned int tmp = count;
929 while(tmp--) {
930 if (ocfs2_test_bit(start + tmp, bitmap)) {
931 printk("ocfs2_verify_zero_bits: start = %u, count = "
932 "%u\n", start, count);
933 printk("ocfs2_verify_zero_bits: bit %u is set!",
934 start + tmp);
935 BUG();
936 }
937 }
938}
939#endif
940
941/*
942 * sync the local alloc to main bitmap.
943 *
944 * assumes you've already locked the main bitmap -- the bitmap inode
945 * passed is used for caching.
946 */
947static int ocfs2_sync_local_to_main(struct ocfs2_super *osb,
948 handle_t *handle,
949 struct ocfs2_dinode *alloc,
950 struct inode *main_bm_inode,
951 struct buffer_head *main_bm_bh)
952{
953 int status = 0;
954 int bit_off, left, count, start;
955 u64 la_start_blk;
956 u64 blkno;
957 void *bitmap;
958 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
959
960 trace_ocfs2_sync_local_to_main(
961 le32_to_cpu(alloc->id1.bitmap1.i_total),
962 le32_to_cpu(alloc->id1.bitmap1.i_used));
963
964 if (!alloc->id1.bitmap1.i_total) {
965 goto bail;
966 }
967
968 if (le32_to_cpu(alloc->id1.bitmap1.i_used) ==
969 le32_to_cpu(alloc->id1.bitmap1.i_total)) {
970 goto bail;
971 }
972
973 la_start_blk = ocfs2_clusters_to_blocks(sb: osb->sb,
974 le32_to_cpu(la->la_bm_off));
975 bitmap = la->la_bitmap;
976 start = count = 0;
977 left = le32_to_cpu(alloc->id1.bitmap1.i_total);
978
979 while ((bit_off = ocfs2_find_next_zero_bit(addr: bitmap, size: left, offset: start))
980 != -1) {
981 if ((bit_off < left) && (bit_off == start)) {
982 count++;
983 start++;
984 continue;
985 }
986 if (count) {
987 blkno = la_start_blk +
988 ocfs2_clusters_to_blocks(sb: osb->sb,
989 clusters: start - count);
990
991 trace_ocfs2_sync_local_to_main_free(
992 count, bit: start - count,
993 start_blk: (unsigned long long)la_start_blk,
994 blkno: (unsigned long long)blkno);
995
996 status = ocfs2_release_clusters(handle,
997 bitmap_inode: main_bm_inode,
998 bitmap_bh: main_bm_bh, start_blk: blkno,
999 num_clusters: count);
1000 if (status < 0) {
1001 mlog_errno(status);
1002 goto bail;
1003 }
1004 }
1005 if (bit_off >= left)
1006 break;
1007 count = 1;
1008 start = bit_off + 1;
1009 }
1010
1011bail:
1012 if (status)
1013 mlog_errno(status);
1014 return status;
1015}
1016
1017enum ocfs2_la_event {
1018 OCFS2_LA_EVENT_SLIDE, /* Normal window slide. */
1019 OCFS2_LA_EVENT_FRAGMENTED, /* The global bitmap has
1020 * enough bits theoretically
1021 * free, but a contiguous
1022 * allocation could not be
1023 * found. */
1024 OCFS2_LA_EVENT_ENOSPC, /* Global bitmap doesn't have
1025 * enough bits free to satisfy
1026 * our request. */
1027};
1028#define OCFS2_LA_ENABLE_INTERVAL (30 * HZ)
1029/*
1030 * Given an event, calculate the size of our next local alloc window.
1031 *
1032 * This should always be called under i_rwsem of the local alloc inode
1033 * so that local alloc disabling doesn't race with processes trying to
1034 * use the allocator.
1035 *
1036 * Returns the state which the local alloc was left in. This value can
1037 * be ignored by some paths.
1038 */
1039static int ocfs2_recalc_la_window(struct ocfs2_super *osb,
1040 enum ocfs2_la_event event)
1041{
1042 unsigned int bits;
1043 int state;
1044
1045 spin_lock(lock: &osb->osb_lock);
1046 if (osb->local_alloc_state == OCFS2_LA_DISABLED) {
1047 WARN_ON_ONCE(osb->local_alloc_state == OCFS2_LA_DISABLED);
1048 goto out_unlock;
1049 }
1050
1051 /*
1052 * ENOSPC and fragmentation are treated similarly for now.
1053 */
1054 if (event == OCFS2_LA_EVENT_ENOSPC ||
1055 event == OCFS2_LA_EVENT_FRAGMENTED) {
1056 /*
1057 * We ran out of contiguous space in the primary
1058 * bitmap. Drastically reduce the number of bits used
1059 * by local alloc until we have to disable it.
1060 */
1061 bits = osb->local_alloc_bits >> 1;
1062 if (bits > ocfs2_megabytes_to_clusters(sb: osb->sb, megs: 1)) {
1063 /*
1064 * By setting state to THROTTLED, we'll keep
1065 * the number of local alloc bits used down
1066 * until an event occurs which would give us
1067 * reason to assume the bitmap situation might
1068 * have changed.
1069 */
1070 osb->local_alloc_state = OCFS2_LA_THROTTLED;
1071 osb->local_alloc_bits = bits;
1072 } else {
1073 osb->local_alloc_state = OCFS2_LA_DISABLED;
1074 }
1075 queue_delayed_work(wq: osb->ocfs2_wq, dwork: &osb->la_enable_wq,
1076 OCFS2_LA_ENABLE_INTERVAL);
1077 goto out_unlock;
1078 }
1079
1080 /*
1081 * Don't increase the size of the local alloc window until we
1082 * know we might be able to fulfill the request. Otherwise, we
1083 * risk bouncing around the global bitmap during periods of
1084 * low space.
1085 */
1086 if (osb->local_alloc_state != OCFS2_LA_THROTTLED)
1087 osb->local_alloc_bits = osb->local_alloc_default_bits;
1088
1089out_unlock:
1090 state = osb->local_alloc_state;
1091 spin_unlock(lock: &osb->osb_lock);
1092
1093 return state;
1094}
1095
1096static int ocfs2_local_alloc_reserve_for_window(struct ocfs2_super *osb,
1097 struct ocfs2_alloc_context **ac,
1098 struct inode **bitmap_inode,
1099 struct buffer_head **bitmap_bh)
1100{
1101 int status;
1102
1103 *ac = kzalloc(size: sizeof(struct ocfs2_alloc_context), GFP_KERNEL);
1104 if (!(*ac)) {
1105 status = -ENOMEM;
1106 mlog_errno(status);
1107 goto bail;
1108 }
1109
1110retry_enospc:
1111 (*ac)->ac_bits_wanted = osb->local_alloc_bits;
1112 status = ocfs2_reserve_cluster_bitmap_bits(osb, ac: *ac);
1113 if (status == -ENOSPC) {
1114 if (ocfs2_recalc_la_window(osb, event: OCFS2_LA_EVENT_ENOSPC) ==
1115 OCFS2_LA_DISABLED)
1116 goto bail;
1117
1118 ocfs2_free_ac_resource(ac: *ac);
1119 memset(*ac, 0, sizeof(struct ocfs2_alloc_context));
1120 goto retry_enospc;
1121 }
1122 if (status < 0) {
1123 mlog_errno(status);
1124 goto bail;
1125 }
1126
1127 *bitmap_inode = (*ac)->ac_inode;
1128 igrab(*bitmap_inode);
1129 *bitmap_bh = (*ac)->ac_bh;
1130 get_bh(bh: *bitmap_bh);
1131 status = 0;
1132bail:
1133 if ((status < 0) && *ac) {
1134 ocfs2_free_alloc_context(ac: *ac);
1135 *ac = NULL;
1136 }
1137
1138 if (status)
1139 mlog_errno(status);
1140 return status;
1141}
1142
1143/*
1144 * pass it the bitmap lock in lock_bh if you have it.
1145 */
1146static int ocfs2_local_alloc_new_window(struct ocfs2_super *osb,
1147 handle_t *handle,
1148 struct ocfs2_alloc_context *ac)
1149{
1150 int status = 0;
1151 u32 cluster_off, cluster_count;
1152 struct ocfs2_dinode *alloc = NULL;
1153 struct ocfs2_local_alloc *la;
1154
1155 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
1156 la = OCFS2_LOCAL_ALLOC(alloc);
1157
1158 trace_ocfs2_local_alloc_new_window(
1159 le32_to_cpu(alloc->id1.bitmap1.i_total),
1160 val2: osb->local_alloc_bits);
1161
1162 /* Instruct the allocation code to try the most recently used
1163 * cluster group. We'll re-record the group used this pass
1164 * below. */
1165 ac->ac_last_group = osb->la_last_gd;
1166
1167 /* we used the generic suballoc reserve function, but we set
1168 * everything up nicely, so there's no reason why we can't use
1169 * the more specific cluster api to claim bits. */
1170 status = ocfs2_claim_clusters(handle, ac, min_clusters: osb->local_alloc_bits,
1171 cluster_start: &cluster_off, num_clusters: &cluster_count);
1172 if (status == -ENOSPC) {
1173retry_enospc:
1174 /*
1175 * Note: We could also try syncing the journal here to
1176 * allow use of any free bits which the current
1177 * transaction can't give us access to. --Mark
1178 */
1179 if (ocfs2_recalc_la_window(osb, event: OCFS2_LA_EVENT_FRAGMENTED) ==
1180 OCFS2_LA_DISABLED)
1181 goto bail;
1182
1183 ac->ac_bits_wanted = osb->local_alloc_bits;
1184 status = ocfs2_claim_clusters(handle, ac,
1185 min_clusters: osb->local_alloc_bits,
1186 cluster_start: &cluster_off,
1187 num_clusters: &cluster_count);
1188 if (status == -ENOSPC)
1189 goto retry_enospc;
1190 /*
1191 * We only shrunk the *minimum* number of in our
1192 * request - it's entirely possible that the allocator
1193 * might give us more than we asked for.
1194 */
1195 if (status == 0) {
1196 spin_lock(lock: &osb->osb_lock);
1197 osb->local_alloc_bits = cluster_count;
1198 spin_unlock(lock: &osb->osb_lock);
1199 }
1200 }
1201 if (status < 0) {
1202 if (status != -ENOSPC)
1203 mlog_errno(status);
1204 goto bail;
1205 }
1206
1207 osb->la_last_gd = ac->ac_last_group;
1208
1209 la->la_bm_off = cpu_to_le32(cluster_off);
1210 alloc->id1.bitmap1.i_total = cpu_to_le32(cluster_count);
1211 /* just in case... In the future when we find space ourselves,
1212 * we don't have to get all contiguous -- but we'll have to
1213 * set all previously used bits in bitmap and update
1214 * la_bits_set before setting the bits in the main bitmap. */
1215 alloc->id1.bitmap1.i_used = 0;
1216 memset(OCFS2_LOCAL_ALLOC(alloc)->la_bitmap, 0,
1217 le16_to_cpu(la->la_size));
1218
1219 ocfs2_resmap_restart(resmap: &osb->osb_la_resmap, clen: cluster_count,
1220 OCFS2_LOCAL_ALLOC(alloc)->la_bitmap);
1221
1222 trace_ocfs2_local_alloc_new_window_result(
1223 OCFS2_LOCAL_ALLOC(alloc)->la_bm_off,
1224 le32_to_cpu(alloc->id1.bitmap1.i_total));
1225
1226bail:
1227 if (status)
1228 mlog_errno(status);
1229 return status;
1230}
1231
1232/* Note that we do *NOT* lock the local alloc inode here as
1233 * it's been locked already for us. */
1234static int ocfs2_local_alloc_slide_window(struct ocfs2_super *osb,
1235 struct inode *local_alloc_inode)
1236{
1237 int status = 0;
1238 struct buffer_head *main_bm_bh = NULL;
1239 struct inode *main_bm_inode = NULL;
1240 handle_t *handle = NULL;
1241 struct ocfs2_dinode *alloc;
1242 struct ocfs2_dinode *alloc_copy = NULL;
1243 struct ocfs2_alloc_context *ac = NULL;
1244
1245 ocfs2_recalc_la_window(osb, event: OCFS2_LA_EVENT_SLIDE);
1246
1247 /* This will lock the main bitmap for us. */
1248 status = ocfs2_local_alloc_reserve_for_window(osb,
1249 ac: &ac,
1250 bitmap_inode: &main_bm_inode,
1251 bitmap_bh: &main_bm_bh);
1252 if (status < 0) {
1253 if (status != -ENOSPC)
1254 mlog_errno(status);
1255 goto bail;
1256 }
1257
1258 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
1259 if (IS_ERR(ptr: handle)) {
1260 status = PTR_ERR(ptr: handle);
1261 handle = NULL;
1262 mlog_errno(status);
1263 goto bail;
1264 }
1265
1266 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
1267
1268 /* We want to clear the local alloc before doing anything
1269 * else, so that if we error later during this operation,
1270 * local alloc shutdown won't try to double free main bitmap
1271 * bits. Make a copy so the sync function knows which bits to
1272 * free. */
1273 alloc_copy = kmemdup(p: alloc, size: osb->local_alloc_bh->b_size, GFP_NOFS);
1274 if (!alloc_copy) {
1275 status = -ENOMEM;
1276 mlog_errno(status);
1277 goto bail;
1278 }
1279
1280 status = ocfs2_journal_access_di(handle,
1281 ci: INODE_CACHE(inode: local_alloc_inode),
1282 bh: osb->local_alloc_bh,
1283 OCFS2_JOURNAL_ACCESS_WRITE);
1284 if (status < 0) {
1285 mlog_errno(status);
1286 goto bail;
1287 }
1288
1289 ocfs2_clear_local_alloc(alloc);
1290 ocfs2_journal_dirty(handle, bh: osb->local_alloc_bh);
1291
1292 status = ocfs2_sync_local_to_main(osb, handle, alloc: alloc_copy,
1293 main_bm_inode, main_bm_bh);
1294 if (status < 0) {
1295 mlog_errno(status);
1296 goto bail;
1297 }
1298
1299 status = ocfs2_local_alloc_new_window(osb, handle, ac);
1300 if (status < 0) {
1301 if (status != -ENOSPC)
1302 mlog_errno(status);
1303 goto bail;
1304 }
1305
1306 atomic_inc(v: &osb->alloc_stats.moves);
1307
1308bail:
1309 if (handle)
1310 ocfs2_commit_trans(osb, handle);
1311
1312 brelse(bh: main_bm_bh);
1313
1314 iput(main_bm_inode);
1315 kfree(objp: alloc_copy);
1316
1317 if (ac)
1318 ocfs2_free_alloc_context(ac);
1319
1320 if (status)
1321 mlog_errno(status);
1322 return status;
1323}
1324
1325

source code of linux/fs/ocfs2/localalloc.c