1 // SPDX-License-Identifier: GPL-2.0-or-later
5 * Node local data allocation
7 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
11 #include <linux/types.h>
12 #include <linux/slab.h>
13 #include <linux/highmem.h>
14 #include <linux/bitops.h>
16 #include <cluster/masklog.h>
21 #include "blockcheck.h"
25 #include "localalloc.h"
29 #include "ocfs2_trace.h"
31 #include "buffer_head_io.h"
33 #define OCFS2_LOCAL_ALLOC(dinode) (&((dinode)->id2.i_lab))
35 static u32 ocfs2_local_alloc_count_bits(struct ocfs2_dinode *alloc);
37 static int ocfs2_local_alloc_find_clear_bits(struct ocfs2_super *osb,
38 struct ocfs2_dinode *alloc,
40 struct ocfs2_alloc_reservation *resv);
42 static void ocfs2_clear_local_alloc(struct ocfs2_dinode *alloc);
44 static int ocfs2_sync_local_to_main(struct ocfs2_super *osb,
46 struct ocfs2_dinode *alloc,
47 struct inode *main_bm_inode,
48 struct buffer_head *main_bm_bh);
50 static 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);
55 static int ocfs2_local_alloc_new_window(struct ocfs2_super *osb,
57 struct ocfs2_alloc_context *ac);
59 static int ocfs2_local_alloc_slide_window(struct ocfs2_super *osb,
60 struct inode *local_alloc_inode);
63 * ocfs2_la_default_mb() - determine a default size, in megabytes of
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
72 * Some things work against us when trying to choose a large local alloc:
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.
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.
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.
86 * The following reference table shows group descriptor and local
87 * alloc maximums at various cluster sizes (4k blocksize)
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
99 #define OCFS2_LA_MAX_DEFAULT_MB 256
100 #define OCFS2_LA_OLD_DEFAULT 8
101 unsigned int ocfs2_la_default_mb(struct ocfs2_super *osb)
105 unsigned int la_max_mb;
106 unsigned int megs_per_slot;
107 struct super_block *sb = osb->sb;
109 gd_mb = ocfs2_clusters_to_megabytes(osb->sb,
110 8 * ocfs2_group_bitmap_size(sb, 0, osb->s_feature_incompat));
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.
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;
122 * Leave enough room for some block groups and make the final
123 * value we work from a multiple of 4.
131 * Keep window sizes down to a reasonable default
133 if (la_mb > OCFS2_LA_MAX_DEFAULT_MB) {
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.
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.
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.
152 if (gd_mb > (2 * OCFS2_LA_MAX_DEFAULT_MB))
155 unsigned int gd_mult = gd_mb;
157 while (gd_mult > 256)
158 gd_mult = gd_mult >> 1;
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;
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)
179 void ocfs2_la_set_sizes(struct ocfs2_super *osb, int requested_mb)
181 struct super_block *sb = osb->sb;
182 unsigned int la_default_mb = ocfs2_la_default_mb(osb);
183 unsigned int la_max_mb;
185 la_max_mb = ocfs2_clusters_to_megabytes(sb,
186 ocfs2_local_alloc_size(sb) * 8);
188 trace_ocfs2_la_set_sizes(requested_mb, la_max_mb, la_default_mb);
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);
199 osb->local_alloc_default_bits =
200 ocfs2_megabytes_to_clusters(sb, requested_mb);
203 osb->local_alloc_bits = osb->local_alloc_default_bits;
206 static inline int ocfs2_la_state_enabled(struct ocfs2_super *osb)
208 return (osb->local_alloc_state == OCFS2_LA_THROTTLED ||
209 osb->local_alloc_state == OCFS2_LA_ENABLED);
212 void ocfs2_local_alloc_seen_free_bits(struct ocfs2_super *osb,
213 unsigned int num_clusters)
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;
222 spin_unlock(&osb->osb_lock);
226 void ocfs2_la_enable_worker(struct work_struct *work)
228 struct ocfs2_super *osb =
229 container_of(work, struct ocfs2_super,
231 spin_lock(&osb->osb_lock);
232 osb->local_alloc_state = OCFS2_LA_ENABLED;
233 spin_unlock(&osb->osb_lock);
237 * Tell us whether a given allocation should use the local alloc
238 * file. Otherwise, it has to go to the main bitmap.
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.
243 int ocfs2_alloc_should_use_local(struct ocfs2_super *osb, u64 bits)
248 spin_lock(&osb->osb_lock);
249 la_bits = osb->local_alloc_bits;
251 if (!ocfs2_la_state_enabled(osb))
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))
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);
269 int ocfs2_load_local_alloc(struct ocfs2_super *osb)
272 struct ocfs2_dinode *alloc = NULL;
273 struct buffer_head *alloc_bh = NULL;
275 struct inode *inode = NULL;
276 struct ocfs2_local_alloc *la;
278 if (osb->local_alloc_bits == 0)
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));
290 /* read the alloc off disk */
291 inode = ocfs2_get_system_file_inode(osb, LOCAL_ALLOC_SYSTEM_INODE,
299 status = ocfs2_read_inode_block_full(inode, &alloc_bh,
300 OCFS2_BH_IGNORE_CACHE);
306 alloc = (struct ocfs2_dinode *) alloc_bh->b_data;
307 la = OCFS2_LOCAL_ALLOC(alloc);
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);
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));
325 /* do a little verification. */
326 num_used = ocfs2_local_alloc_count_bits(alloc);
328 /* hopefully the local alloc has always been recovered before
331 || alloc->id1.bitmap1.i_used
332 || alloc->id1.bitmap1.i_total
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));
345 osb->local_alloc_bh = alloc_bh;
346 osb->local_alloc_state = OCFS2_LA_ENABLED;
353 trace_ocfs2_load_local_alloc(osb->local_alloc_bits);
361 * return any unused bits to the bitmap and write out a clean
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.*/
367 void ocfs2_shutdown_local_alloc(struct ocfs2_super *osb)
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;
378 cancel_delayed_work(&osb->la_enable_wq);
380 flush_workqueue(osb->ocfs2_wq);
382 if (osb->local_alloc_state == OCFS2_LA_UNUSED)
386 ocfs2_get_system_file_inode(osb,
387 LOCAL_ALLOC_SYSTEM_INODE,
389 if (!local_alloc_inode) {
395 osb->local_alloc_state = OCFS2_LA_DISABLED;
397 ocfs2_resmap_uninit(&osb->osb_la_resmap);
399 main_bm_inode = ocfs2_get_system_file_inode(osb,
400 GLOBAL_BITMAP_SYSTEM_INODE,
402 if (!main_bm_inode) {
408 inode_lock(main_bm_inode);
410 status = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 1);
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));
424 bh = osb->local_alloc_bh;
425 alloc = (struct ocfs2_dinode *) bh->b_data;
427 alloc_copy = kmemdup(alloc, bh->b_size, GFP_NOFS);
433 status = ocfs2_journal_access_di(handle, INODE_CACHE(local_alloc_inode),
434 bh, OCFS2_JOURNAL_ACCESS_WRITE);
440 ocfs2_clear_local_alloc(alloc);
441 ocfs2_journal_dirty(handle, bh);
444 osb->local_alloc_bh = NULL;
445 osb->local_alloc_state = OCFS2_LA_UNUSED;
447 status = ocfs2_sync_local_to_main(osb, handle, alloc_copy,
448 main_bm_inode, main_bm_bh);
453 ocfs2_commit_trans(osb, handle);
458 ocfs2_inode_unlock(main_bm_inode, 1);
461 inode_unlock(main_bm_inode);
465 iput(local_alloc_inode);
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
477 int ocfs2_begin_local_alloc_recovery(struct ocfs2_super *osb,
479 struct ocfs2_dinode **alloc_copy)
482 struct buffer_head *alloc_bh = NULL;
483 struct inode *inode = NULL;
484 struct ocfs2_dinode *alloc;
486 trace_ocfs2_begin_local_alloc_recovery(slot_num);
490 inode = ocfs2_get_system_file_inode(osb,
491 LOCAL_ALLOC_SYSTEM_INODE,
501 status = ocfs2_read_inode_block_full(inode, &alloc_bh,
502 OCFS2_BH_IGNORE_CACHE);
508 *alloc_copy = kmalloc(alloc_bh->b_size, GFP_KERNEL);
509 if (!(*alloc_copy)) {
513 memcpy((*alloc_copy), alloc_bh->b_data, alloc_bh->b_size);
515 alloc = (struct ocfs2_dinode *) alloc_bh->b_data;
516 ocfs2_clear_local_alloc(alloc);
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));
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
547 int ocfs2_complete_local_alloc_recovery(struct ocfs2_super *osb,
548 struct ocfs2_dinode *alloc)
552 struct buffer_head *main_bm_bh = NULL;
553 struct inode *main_bm_inode;
555 main_bm_inode = ocfs2_get_system_file_inode(osb,
556 GLOBAL_BITMAP_SYSTEM_INODE,
558 if (!main_bm_inode) {
564 inode_lock(main_bm_inode);
566 status = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 1);
572 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
573 if (IS_ERR(handle)) {
574 status = PTR_ERR(handle);
580 /* we want the bitmap change to be recorded on disk asap */
583 status = ocfs2_sync_local_to_main(osb, handle, alloc,
584 main_bm_inode, main_bm_bh);
588 ocfs2_commit_trans(osb, handle);
591 ocfs2_inode_unlock(main_bm_inode, 1);
594 inode_unlock(main_bm_inode);
602 ocfs2_init_steal_slots(osb);
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.
612 * We will add ourselves to the transaction passed in, but may start
613 * our own in order to shift windows.
615 int ocfs2_reserve_local_alloc_bits(struct ocfs2_super *osb,
617 struct ocfs2_alloc_context *ac)
620 struct ocfs2_dinode *alloc;
621 struct inode *local_alloc_inode;
622 unsigned int free_bits;
627 ocfs2_get_system_file_inode(osb,
628 LOCAL_ALLOC_SYSTEM_INODE,
630 if (!local_alloc_inode) {
636 inode_lock(local_alloc_inode);
639 * We must double check state and allocator bits because
640 * another process may have changed them while holding i_rwsem.
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);
649 spin_unlock(&osb->osb_lock);
651 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
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));
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. */
669 ocfs2_local_alloc_slide_window(osb, local_alloc_inode);
671 if (status != -ENOSPC)
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.
683 if (!ocfs2_la_state_enabled(osb))
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)
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;
700 if (status < 0 && local_alloc_inode) {
701 inode_unlock(local_alloc_inode);
702 iput(local_alloc_inode);
705 trace_ocfs2_reserve_local_alloc_bits(
706 (unsigned long long)ac->ac_max_block,
707 bits_wanted, osb->slot_num, status);
714 int ocfs2_claim_local_alloc_bits(struct ocfs2_super *osb,
716 struct ocfs2_alloc_context *ac,
722 struct inode *local_alloc_inode;
724 struct ocfs2_dinode *alloc;
725 struct ocfs2_local_alloc *la;
727 BUG_ON(ac->ac_which != OCFS2_AC_USE_LOCAL);
729 local_alloc_inode = ac->ac_inode;
730 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
731 la = OCFS2_LOCAL_ALLOC(alloc);
733 start = ocfs2_local_alloc_find_clear_bits(osb, alloc, &bits_wanted,
736 /* TODO: Shouldn't we just BUG here? */
742 bitmap = la->la_bitmap;
743 *bit_off = le32_to_cpu(la->la_bm_off) + start;
744 *num_bits = bits_wanted;
746 status = ocfs2_journal_access_di(handle,
747 INODE_CACHE(local_alloc_inode),
749 OCFS2_JOURNAL_ACCESS_WRITE);
755 ocfs2_resmap_claimed_bits(&osb->osb_la_resmap, ac->ac_resv, start,
759 ocfs2_set_bit(start++, bitmap);
761 le32_add_cpu(&alloc->id1.bitmap1.i_used, *num_bits);
762 ocfs2_journal_dirty(handle, osb->local_alloc_bh);
770 int ocfs2_free_local_alloc_bits(struct ocfs2_super *osb,
772 struct ocfs2_alloc_context *ac,
778 struct inode *local_alloc_inode;
780 struct ocfs2_dinode *alloc;
781 struct ocfs2_local_alloc *la;
783 BUG_ON(ac->ac_which != OCFS2_AC_USE_LOCAL);
785 local_alloc_inode = ac->ac_inode;
786 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
787 la = OCFS2_LOCAL_ALLOC(alloc);
789 bitmap = la->la_bitmap;
790 start = bit_off - le32_to_cpu(la->la_bm_off);
791 clear_bits = num_bits;
793 status = ocfs2_journal_access_di(handle,
794 INODE_CACHE(local_alloc_inode),
796 OCFS2_JOURNAL_ACCESS_WRITE);
803 ocfs2_clear_bit(start++, bitmap);
805 le32_add_cpu(&alloc->id1.bitmap1.i_used, -num_bits);
806 ocfs2_journal_dirty(handle, osb->local_alloc_bh);
812 static u32 ocfs2_local_alloc_count_bits(struct ocfs2_dinode *alloc)
815 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
817 count = memweight(la->la_bitmap, le16_to_cpu(la->la_size));
819 trace_ocfs2_local_alloc_count_bits(count);
823 static int ocfs2_local_alloc_find_clear_bits(struct ocfs2_super *osb,
824 struct ocfs2_dinode *alloc,
826 struct ocfs2_alloc_reservation *resv)
828 int numfound = 0, bitoff, left, startoff;
830 struct ocfs2_alloc_reservation r;
832 struct ocfs2_reservation_map *resmap = &osb->osb_la_resmap;
834 if (!alloc->id1.bitmap1.i_total) {
841 ocfs2_resv_init_once(&r);
842 ocfs2_resv_set_type(&r, OCFS2_RESV_FLAG_TMP);
847 if (ocfs2_resmap_resv_bits(resmap, resv, &bitoff, &numfound) == 0) {
848 if (numfound < *numbits)
854 * Code error. While reservations are enabled, local
855 * allocation should _always_ go through them.
857 BUG_ON(osb->osb_resv_level != 0);
860 * Reservations are disabled. Handle this the old way.
863 bitmap = OCFS2_LOCAL_ALLOC(alloc)->la_bitmap;
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)) <
869 /* Ok, we found a zero bit... is it contig. or do we
871 if (bitoff == startoff) {
872 /* we found a zero */
876 /* got a zero after some ones */
880 /* we got everything we needed */
881 if (numfound == *numbits) {
882 /* mlog(0, "Found it all!\n"); */
887 trace_ocfs2_local_alloc_find_clear_bits_search_bitmap(bitoff, numfound);
889 if (numfound == *numbits)
890 bitoff = startoff - numfound;
896 ocfs2_resv_discard(resmap, resv);
898 trace_ocfs2_local_alloc_find_clear_bits(*numbits,
899 le32_to_cpu(alloc->id1.bitmap1.i_total),
905 static void ocfs2_clear_local_alloc(struct ocfs2_dinode *alloc)
907 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
910 alloc->id1.bitmap1.i_total = 0;
911 alloc->id1.bitmap1.i_used = 0;
913 for(i = 0; i < le16_to_cpu(la->la_size); i++)
914 la->la_bitmap[i] = 0;
918 /* turn this on and uncomment below to aid debugging window shifts. */
919 static void ocfs2_verify_zero_bits(unsigned long *bitmap,
923 unsigned int tmp = count;
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!",
937 * sync the local alloc to main bitmap.
939 * assumes you've already locked the main bitmap -- the bitmap inode
940 * passed is used for caching.
942 static int ocfs2_sync_local_to_main(struct ocfs2_super *osb,
944 struct ocfs2_dinode *alloc,
945 struct inode *main_bm_inode,
946 struct buffer_head *main_bm_bh)
949 int bit_off, left, count, start;
953 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
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));
959 if (!alloc->id1.bitmap1.i_total) {
963 if (le32_to_cpu(alloc->id1.bitmap1.i_used) ==
964 le32_to_cpu(alloc->id1.bitmap1.i_total)) {
968 la_start_blk = ocfs2_clusters_to_blocks(osb->sb,
969 le32_to_cpu(la->la_bm_off));
970 bitmap = la->la_bitmap;
972 left = le32_to_cpu(alloc->id1.bitmap1.i_total);
975 bit_off = ocfs2_find_next_zero_bit(bitmap, left, start);
976 if ((bit_off < left) && (bit_off == start)) {
982 blkno = la_start_blk +
983 ocfs2_clusters_to_blocks(osb->sb,
986 trace_ocfs2_sync_local_to_main_free(
987 count, start - count,
988 (unsigned long long)la_start_blk,
989 (unsigned long long)blkno);
991 status = ocfs2_release_clusters(handle,
1001 if (bit_off >= left)
1004 start = bit_off + 1;
1013 enum 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
1020 OCFS2_LA_EVENT_ENOSPC, /* Global bitmap doesn't have
1021 * enough bits free to satisfy
1024 #define OCFS2_LA_ENABLE_INTERVAL (30 * HZ)
1026 * Given an event, calculate the size of our next local alloc window.
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.
1032 * Returns the state which the local alloc was left in. This value can
1033 * be ignored by some paths.
1035 static int ocfs2_recalc_la_window(struct ocfs2_super *osb,
1036 enum ocfs2_la_event event)
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);
1048 * ENOSPC and fragmentation are treated similarly for now.
1050 if (event == OCFS2_LA_EVENT_ENOSPC ||
1051 event == OCFS2_LA_EVENT_FRAGMENTED) {
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.
1057 bits = osb->local_alloc_bits >> 1;
1058 if (bits > ocfs2_megabytes_to_clusters(osb->sb, 1)) {
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
1066 osb->local_alloc_state = OCFS2_LA_THROTTLED;
1067 osb->local_alloc_bits = bits;
1069 osb->local_alloc_state = OCFS2_LA_DISABLED;
1071 queue_delayed_work(osb->ocfs2_wq, &osb->la_enable_wq,
1072 OCFS2_LA_ENABLE_INTERVAL);
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
1082 if (osb->local_alloc_state != OCFS2_LA_THROTTLED)
1083 osb->local_alloc_bits = osb->local_alloc_default_bits;
1086 state = osb->local_alloc_state;
1087 spin_unlock(&osb->osb_lock);
1092 static 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)
1099 *ac = kzalloc(sizeof(struct ocfs2_alloc_context), GFP_KERNEL);
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) ==
1114 ocfs2_free_ac_resource(*ac);
1115 memset(*ac, 0, sizeof(struct ocfs2_alloc_context));
1123 *bitmap_inode = (*ac)->ac_inode;
1124 igrab(*bitmap_inode);
1125 *bitmap_bh = (*ac)->ac_bh;
1129 if ((status < 0) && *ac) {
1130 ocfs2_free_alloc_context(*ac);
1140 * pass it the bitmap lock in lock_bh if you have it.
1142 static int ocfs2_local_alloc_new_window(struct ocfs2_super *osb,
1144 struct ocfs2_alloc_context *ac)
1147 u32 cluster_off, cluster_count;
1148 struct ocfs2_dinode *alloc = NULL;
1149 struct ocfs2_local_alloc *la;
1151 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
1152 la = OCFS2_LOCAL_ALLOC(alloc);
1154 trace_ocfs2_local_alloc_new_window(
1155 le32_to_cpu(alloc->id1.bitmap1.i_total),
1156 osb->local_alloc_bits);
1158 /* Instruct the allocation code to try the most recently used
1159 * cluster group. We'll re-record the group used this pass
1161 ac->ac_last_group = osb->la_last_gd;
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) {
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
1175 if (ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_FRAGMENTED) ==
1179 ac->ac_bits_wanted = osb->local_alloc_bits;
1180 status = ocfs2_claim_clusters(handle, ac,
1181 osb->local_alloc_bits,
1184 if (status == -ENOSPC)
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.
1192 spin_lock(&osb->osb_lock);
1193 osb->local_alloc_bits = cluster_count;
1194 spin_unlock(&osb->osb_lock);
1198 if (status != -ENOSPC)
1203 osb->la_last_gd = ac->ac_last_group;
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));
1215 ocfs2_resmap_restart(&osb->osb_la_resmap, cluster_count,
1216 OCFS2_LOCAL_ALLOC(alloc)->la_bitmap);
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));
1228 /* Note that we do *NOT* lock the local alloc inode here as
1229 * it's been locked already for us. */
1230 static int ocfs2_local_alloc_slide_window(struct ocfs2_super *osb,
1231 struct inode *local_alloc_inode)
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;
1241 ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_SLIDE);
1243 /* This will lock the main bitmap for us. */
1244 status = ocfs2_local_alloc_reserve_for_window(osb,
1249 if (status != -ENOSPC)
1254 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
1255 if (IS_ERR(handle)) {
1256 status = PTR_ERR(handle);
1262 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
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
1269 alloc_copy = kmemdup(alloc, osb->local_alloc_bh->b_size, GFP_NOFS);
1276 status = ocfs2_journal_access_di(handle,
1277 INODE_CACHE(local_alloc_inode),
1278 osb->local_alloc_bh,
1279 OCFS2_JOURNAL_ACCESS_WRITE);
1285 ocfs2_clear_local_alloc(alloc);
1286 ocfs2_journal_dirty(handle, osb->local_alloc_bh);
1288 status = ocfs2_sync_local_to_main(osb, handle, alloc_copy,
1289 main_bm_inode, main_bm_bh);
1295 status = ocfs2_local_alloc_new_window(osb, handle, ac);
1297 if (status != -ENOSPC)
1302 atomic_inc(&osb->alloc_stats.moves);
1306 ocfs2_commit_trans(osb, handle);
1310 iput(main_bm_inode);
1314 ocfs2_free_alloc_context(ac);