1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * Node local data allocation
8 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
27 #include <linux/types.h>
28 #include <linux/slab.h>
29 #include <linux/highmem.h>
30 #include <linux/bitops.h>
32 #define MLOG_MASK_PREFIX ML_DISK_ALLOC
33 #include <cluster/masklog.h>
38 #include "blockcheck.h"
42 #include "localalloc.h"
46 #include "ocfs2_trace.h"
48 #include "buffer_head_io.h"
50 #define OCFS2_LOCAL_ALLOC(dinode) (&((dinode)->id2.i_lab))
52 static u32 ocfs2_local_alloc_count_bits(struct ocfs2_dinode *alloc);
54 static int ocfs2_local_alloc_find_clear_bits(struct ocfs2_super *osb,
55 struct ocfs2_dinode *alloc,
57 struct ocfs2_alloc_reservation *resv);
59 static void ocfs2_clear_local_alloc(struct ocfs2_dinode *alloc);
61 static int ocfs2_sync_local_to_main(struct ocfs2_super *osb,
63 struct ocfs2_dinode *alloc,
64 struct inode *main_bm_inode,
65 struct buffer_head *main_bm_bh);
67 static int ocfs2_local_alloc_reserve_for_window(struct ocfs2_super *osb,
68 struct ocfs2_alloc_context **ac,
69 struct inode **bitmap_inode,
70 struct buffer_head **bitmap_bh);
72 static int ocfs2_local_alloc_new_window(struct ocfs2_super *osb,
74 struct ocfs2_alloc_context *ac);
76 static int ocfs2_local_alloc_slide_window(struct ocfs2_super *osb,
77 struct inode *local_alloc_inode);
80 * ocfs2_la_default_mb() - determine a default size, in megabytes of
83 * Generally, we'd like to pick as large a local alloc as
84 * possible. Performance on large workloads tends to scale
85 * proportionally to la size. In addition to that, the reservations
86 * code functions more efficiently as it can reserve more windows for
89 * Some things work against us when trying to choose a large local alloc:
91 * - We need to ensure our sizing is picked to leave enough space in
92 * group descriptors for other allocations (such as block groups,
93 * etc). Picking default sizes which are a multiple of 4 could help
94 * - block groups are allocated in 2mb and 4mb chunks.
96 * - Likewise, we don't want to starve other nodes of bits on small
97 * file systems. This can easily be taken care of by limiting our
98 * default to a reasonable size (256M) on larger cluster sizes.
100 * - Some file systems can't support very large sizes - 4k and 8k in
101 * particular are limited to less than 128 and 256 megabytes respectively.
103 * The following reference table shows group descriptor and local
104 * alloc maximums at various cluster sizes (4k blocksize)
106 * csize: 4K group: 126M la: 121M
107 * csize: 8K group: 252M la: 243M
108 * csize: 16K group: 504M la: 486M
109 * csize: 32K group: 1008M la: 972M
110 * csize: 64K group: 2016M la: 1944M
111 * csize: 128K group: 4032M la: 3888M
112 * csize: 256K group: 8064M la: 7776M
113 * csize: 512K group: 16128M la: 15552M
114 * csize: 1024K group: 32256M la: 31104M
116 #define OCFS2_LA_MAX_DEFAULT_MB 256
117 #define OCFS2_LA_OLD_DEFAULT 8
118 unsigned int ocfs2_la_default_mb(struct ocfs2_super *osb)
122 unsigned int la_max_mb;
123 unsigned int megs_per_slot;
124 struct super_block *sb = osb->sb;
126 gd_mb = ocfs2_clusters_to_megabytes(osb->sb,
127 8 * ocfs2_group_bitmap_size(sb, 0, osb->s_feature_incompat));
130 * This takes care of files systems with very small group
131 * descriptors - 512 byte blocksize at cluster sizes lower
132 * than 16K and also 1k blocksize with 4k cluster size.
134 if ((sb->s_blocksize == 512 && osb->s_clustersize <= 8192)
135 || (sb->s_blocksize == 1024 && osb->s_clustersize == 4096))
136 return OCFS2_LA_OLD_DEFAULT;
139 * Leave enough room for some block groups and make the final
140 * value we work from a multiple of 4.
148 * Keep window sizes down to a reasonable default
150 if (la_mb > OCFS2_LA_MAX_DEFAULT_MB) {
152 * Some clustersize / blocksize combinations will have
153 * given us a larger than OCFS2_LA_MAX_DEFAULT_MB
154 * default size, but get poor distribution when
155 * limited to exactly 256 megabytes.
157 * As an example, 16K clustersize at 4K blocksize
158 * gives us a cluster group size of 504M. Paring the
159 * local alloc size down to 256 however, would give us
160 * only one window and around 200MB left in the
161 * cluster group. Instead, find the first size below
162 * 256 which would give us an even distribution.
164 * Larger cluster group sizes actually work out pretty
165 * well when pared to 256, so we don't have to do this
166 * for any group that fits more than two
167 * OCFS2_LA_MAX_DEFAULT_MB windows.
169 if (gd_mb > (2 * OCFS2_LA_MAX_DEFAULT_MB))
172 unsigned int gd_mult = gd_mb;
174 while (gd_mult > 256)
175 gd_mult = gd_mult >> 1;
181 megs_per_slot = osb->osb_clusters_at_boot / osb->max_slots;
182 megs_per_slot = ocfs2_clusters_to_megabytes(osb->sb, megs_per_slot);
183 /* Too many nodes, too few disk clusters. */
184 if (megs_per_slot < la_mb)
185 la_mb = megs_per_slot;
187 /* We can't store more bits than we can in a block. */
188 la_max_mb = ocfs2_clusters_to_megabytes(osb->sb,
189 ocfs2_local_alloc_size(sb) * 8);
190 if (la_mb > la_max_mb)
196 void ocfs2_la_set_sizes(struct ocfs2_super *osb, int requested_mb)
198 struct super_block *sb = osb->sb;
199 unsigned int la_default_mb = ocfs2_la_default_mb(osb);
200 unsigned int la_max_mb;
202 la_max_mb = ocfs2_clusters_to_megabytes(sb,
203 ocfs2_local_alloc_size(sb) * 8);
205 trace_ocfs2_la_set_sizes(requested_mb, la_max_mb, la_default_mb);
207 if (requested_mb == -1) {
208 /* No user request - use defaults */
209 osb->local_alloc_default_bits =
210 ocfs2_megabytes_to_clusters(sb, la_default_mb);
211 } else if (requested_mb > la_max_mb) {
212 /* Request is too big, we give the maximum available */
213 osb->local_alloc_default_bits =
214 ocfs2_megabytes_to_clusters(sb, la_max_mb);
216 osb->local_alloc_default_bits =
217 ocfs2_megabytes_to_clusters(sb, requested_mb);
220 osb->local_alloc_bits = osb->local_alloc_default_bits;
223 static inline int ocfs2_la_state_enabled(struct ocfs2_super *osb)
225 return (osb->local_alloc_state == OCFS2_LA_THROTTLED ||
226 osb->local_alloc_state == OCFS2_LA_ENABLED);
229 void ocfs2_local_alloc_seen_free_bits(struct ocfs2_super *osb,
230 unsigned int num_clusters)
232 spin_lock(&osb->osb_lock);
233 if (osb->local_alloc_state == OCFS2_LA_DISABLED ||
234 osb->local_alloc_state == OCFS2_LA_THROTTLED)
235 if (num_clusters >= osb->local_alloc_default_bits) {
236 cancel_delayed_work(&osb->la_enable_wq);
237 osb->local_alloc_state = OCFS2_LA_ENABLED;
239 spin_unlock(&osb->osb_lock);
242 void ocfs2_la_enable_worker(struct work_struct *work)
244 struct ocfs2_super *osb =
245 container_of(work, struct ocfs2_super,
247 spin_lock(&osb->osb_lock);
248 osb->local_alloc_state = OCFS2_LA_ENABLED;
249 spin_unlock(&osb->osb_lock);
253 * Tell us whether a given allocation should use the local alloc
254 * file. Otherwise, it has to go to the main bitmap.
256 * This function does semi-dirty reads of local alloc size and state!
257 * This is ok however, as the values are re-checked once under mutex.
259 int ocfs2_alloc_should_use_local(struct ocfs2_super *osb, u64 bits)
264 spin_lock(&osb->osb_lock);
265 la_bits = osb->local_alloc_bits;
267 if (!ocfs2_la_state_enabled(osb))
270 /* la_bits should be at least twice the size (in clusters) of
271 * a new block group. We want to be sure block group
272 * allocations go through the local alloc, so allow an
273 * allocation to take up to half the bitmap. */
274 if (bits > (la_bits / 2))
279 trace_ocfs2_alloc_should_use_local(
280 (unsigned long long)bits, osb->local_alloc_state, la_bits, ret);
281 spin_unlock(&osb->osb_lock);
285 int ocfs2_load_local_alloc(struct ocfs2_super *osb)
288 struct ocfs2_dinode *alloc = NULL;
289 struct buffer_head *alloc_bh = NULL;
291 struct inode *inode = NULL;
292 struct ocfs2_local_alloc *la;
294 if (osb->local_alloc_bits == 0)
297 if (osb->local_alloc_bits >= osb->bitmap_cpg) {
298 mlog(ML_NOTICE, "Requested local alloc window %d is larger "
299 "than max possible %u. Using defaults.\n",
300 osb->local_alloc_bits, (osb->bitmap_cpg - 1));
301 osb->local_alloc_bits =
302 ocfs2_megabytes_to_clusters(osb->sb,
303 ocfs2_la_default_mb(osb));
306 /* read the alloc off disk */
307 inode = ocfs2_get_system_file_inode(osb, LOCAL_ALLOC_SYSTEM_INODE,
315 status = ocfs2_read_inode_block_full(inode, &alloc_bh,
316 OCFS2_BH_IGNORE_CACHE);
322 alloc = (struct ocfs2_dinode *) alloc_bh->b_data;
323 la = OCFS2_LOCAL_ALLOC(alloc);
325 if (!(le32_to_cpu(alloc->i_flags) &
326 (OCFS2_LOCAL_ALLOC_FL|OCFS2_BITMAP_FL))) {
327 mlog(ML_ERROR, "Invalid local alloc inode, %llu\n",
328 (unsigned long long)OCFS2_I(inode)->ip_blkno);
333 if ((la->la_size == 0) ||
334 (le16_to_cpu(la->la_size) > ocfs2_local_alloc_size(inode->i_sb))) {
335 mlog(ML_ERROR, "Local alloc size is invalid (la_size = %u)\n",
336 le16_to_cpu(la->la_size));
341 /* do a little verification. */
342 num_used = ocfs2_local_alloc_count_bits(alloc);
344 /* hopefully the local alloc has always been recovered before
347 || alloc->id1.bitmap1.i_used
348 || alloc->id1.bitmap1.i_total
350 mlog(ML_ERROR, "Local alloc hasn't been recovered!\n"
351 "found = %u, set = %u, taken = %u, off = %u\n",
352 num_used, le32_to_cpu(alloc->id1.bitmap1.i_used),
353 le32_to_cpu(alloc->id1.bitmap1.i_total),
354 OCFS2_LOCAL_ALLOC(alloc)->la_bm_off);
356 osb->local_alloc_bh = alloc_bh;
357 osb->local_alloc_state = OCFS2_LA_ENABLED;
365 trace_ocfs2_load_local_alloc(osb->local_alloc_bits);
373 * return any unused bits to the bitmap and write out a clean
376 * local_alloc_bh is optional. If not passed, we will simply use the
377 * one off osb. If you do pass it however, be warned that it *will* be
378 * returned brelse'd and NULL'd out.*/
379 void ocfs2_shutdown_local_alloc(struct ocfs2_super *osb)
383 struct inode *local_alloc_inode = NULL;
384 struct buffer_head *bh = NULL;
385 struct buffer_head *main_bm_bh = NULL;
386 struct inode *main_bm_inode = NULL;
387 struct ocfs2_dinode *alloc_copy = NULL;
388 struct ocfs2_dinode *alloc = NULL;
390 cancel_delayed_work(&osb->la_enable_wq);
391 flush_workqueue(ocfs2_wq);
393 if (osb->local_alloc_state == OCFS2_LA_UNUSED)
397 ocfs2_get_system_file_inode(osb,
398 LOCAL_ALLOC_SYSTEM_INODE,
400 if (!local_alloc_inode) {
406 osb->local_alloc_state = OCFS2_LA_DISABLED;
408 ocfs2_resmap_uninit(&osb->osb_la_resmap);
410 main_bm_inode = ocfs2_get_system_file_inode(osb,
411 GLOBAL_BITMAP_SYSTEM_INODE,
413 if (!main_bm_inode) {
419 mutex_lock(&main_bm_inode->i_mutex);
421 status = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 1);
427 /* WINDOW_MOVE_CREDITS is a bit heavy... */
428 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
429 if (IS_ERR(handle)) {
430 mlog_errno(PTR_ERR(handle));
435 bh = osb->local_alloc_bh;
436 alloc = (struct ocfs2_dinode *) bh->b_data;
438 alloc_copy = kmalloc(bh->b_size, GFP_NOFS);
443 memcpy(alloc_copy, alloc, bh->b_size);
445 status = ocfs2_journal_access_di(handle, INODE_CACHE(local_alloc_inode),
446 bh, OCFS2_JOURNAL_ACCESS_WRITE);
452 ocfs2_clear_local_alloc(alloc);
453 ocfs2_journal_dirty(handle, bh);
456 osb->local_alloc_bh = NULL;
457 osb->local_alloc_state = OCFS2_LA_UNUSED;
459 status = ocfs2_sync_local_to_main(osb, handle, alloc_copy,
460 main_bm_inode, main_bm_bh);
465 ocfs2_commit_trans(osb, handle);
470 ocfs2_inode_unlock(main_bm_inode, 1);
473 mutex_unlock(&main_bm_inode->i_mutex);
477 if (local_alloc_inode)
478 iput(local_alloc_inode);
485 * We want to free the bitmap bits outside of any recovery context as
486 * we'll need a cluster lock to do so, but we must clear the local
487 * alloc before giving up the recovered nodes journal. To solve this,
488 * we kmalloc a copy of the local alloc before it's change for the
489 * caller to process with ocfs2_complete_local_alloc_recovery
491 int ocfs2_begin_local_alloc_recovery(struct ocfs2_super *osb,
493 struct ocfs2_dinode **alloc_copy)
496 struct buffer_head *alloc_bh = NULL;
497 struct inode *inode = NULL;
498 struct ocfs2_dinode *alloc;
500 trace_ocfs2_begin_local_alloc_recovery(slot_num);
504 inode = ocfs2_get_system_file_inode(osb,
505 LOCAL_ALLOC_SYSTEM_INODE,
513 mutex_lock(&inode->i_mutex);
515 status = ocfs2_read_inode_block_full(inode, &alloc_bh,
516 OCFS2_BH_IGNORE_CACHE);
522 *alloc_copy = kmalloc(alloc_bh->b_size, GFP_KERNEL);
523 if (!(*alloc_copy)) {
527 memcpy((*alloc_copy), alloc_bh->b_data, alloc_bh->b_size);
529 alloc = (struct ocfs2_dinode *) alloc_bh->b_data;
530 ocfs2_clear_local_alloc(alloc);
532 ocfs2_compute_meta_ecc(osb->sb, alloc_bh->b_data, &alloc->i_check);
533 status = ocfs2_write_block(osb, alloc_bh, INODE_CACHE(inode));
538 if ((status < 0) && (*alloc_copy)) {
546 mutex_unlock(&inode->i_mutex);
556 * Step 2: By now, we've completed the journal recovery, we've stamped
557 * a clean local alloc on disk and dropped the node out of the
558 * recovery map. Dlm locks will no longer stall, so lets clear out the
561 int ocfs2_complete_local_alloc_recovery(struct ocfs2_super *osb,
562 struct ocfs2_dinode *alloc)
566 struct buffer_head *main_bm_bh = NULL;
567 struct inode *main_bm_inode;
569 main_bm_inode = ocfs2_get_system_file_inode(osb,
570 GLOBAL_BITMAP_SYSTEM_INODE,
572 if (!main_bm_inode) {
578 mutex_lock(&main_bm_inode->i_mutex);
580 status = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 1);
586 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
587 if (IS_ERR(handle)) {
588 status = PTR_ERR(handle);
594 /* we want the bitmap change to be recorded on disk asap */
597 status = ocfs2_sync_local_to_main(osb, handle, alloc,
598 main_bm_inode, main_bm_bh);
602 ocfs2_commit_trans(osb, handle);
605 ocfs2_inode_unlock(main_bm_inode, 1);
608 mutex_unlock(&main_bm_inode->i_mutex);
616 ocfs2_init_steal_slots(osb);
623 * make sure we've got at least bits_wanted contiguous bits in the
624 * local alloc. You lose them when you drop i_mutex.
626 * We will add ourselves to the transaction passed in, but may start
627 * our own in order to shift windows.
629 int ocfs2_reserve_local_alloc_bits(struct ocfs2_super *osb,
631 struct ocfs2_alloc_context *ac)
634 struct ocfs2_dinode *alloc;
635 struct inode *local_alloc_inode;
636 unsigned int free_bits;
641 ocfs2_get_system_file_inode(osb,
642 LOCAL_ALLOC_SYSTEM_INODE,
644 if (!local_alloc_inode) {
650 mutex_lock(&local_alloc_inode->i_mutex);
653 * We must double check state and allocator bits because
654 * another process may have changed them while holding i_mutex.
656 spin_lock(&osb->osb_lock);
657 if (!ocfs2_la_state_enabled(osb) ||
658 (bits_wanted > osb->local_alloc_bits)) {
659 spin_unlock(&osb->osb_lock);
663 spin_unlock(&osb->osb_lock);
665 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
667 #ifdef CONFIG_OCFS2_DEBUG_FS
668 if (le32_to_cpu(alloc->id1.bitmap1.i_used) !=
669 ocfs2_local_alloc_count_bits(alloc)) {
670 ocfs2_error(osb->sb, "local alloc inode %llu says it has "
671 "%u free bits, but a count shows %u",
672 (unsigned long long)le64_to_cpu(alloc->i_blkno),
673 le32_to_cpu(alloc->id1.bitmap1.i_used),
674 ocfs2_local_alloc_count_bits(alloc));
680 free_bits = le32_to_cpu(alloc->id1.bitmap1.i_total) -
681 le32_to_cpu(alloc->id1.bitmap1.i_used);
682 if (bits_wanted > free_bits) {
683 /* uhoh, window change time. */
685 ocfs2_local_alloc_slide_window(osb, local_alloc_inode);
687 if (status != -ENOSPC)
693 * Under certain conditions, the window slide code
694 * might have reduced the number of bits available or
695 * disabled the the local alloc entirely. Re-check
696 * here and return -ENOSPC if necessary.
699 if (!ocfs2_la_state_enabled(osb))
702 free_bits = le32_to_cpu(alloc->id1.bitmap1.i_total) -
703 le32_to_cpu(alloc->id1.bitmap1.i_used);
704 if (bits_wanted > free_bits)
708 ac->ac_inode = local_alloc_inode;
709 /* We should never use localalloc from another slot */
710 ac->ac_alloc_slot = osb->slot_num;
711 ac->ac_which = OCFS2_AC_USE_LOCAL;
712 get_bh(osb->local_alloc_bh);
713 ac->ac_bh = osb->local_alloc_bh;
716 if (status < 0 && local_alloc_inode) {
717 mutex_unlock(&local_alloc_inode->i_mutex);
718 iput(local_alloc_inode);
721 trace_ocfs2_reserve_local_alloc_bits(
722 (unsigned long long)ac->ac_max_block,
723 bits_wanted, osb->slot_num, status);
730 int ocfs2_claim_local_alloc_bits(struct ocfs2_super *osb,
732 struct ocfs2_alloc_context *ac,
738 struct inode *local_alloc_inode;
740 struct ocfs2_dinode *alloc;
741 struct ocfs2_local_alloc *la;
743 BUG_ON(ac->ac_which != OCFS2_AC_USE_LOCAL);
745 local_alloc_inode = ac->ac_inode;
746 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
747 la = OCFS2_LOCAL_ALLOC(alloc);
749 start = ocfs2_local_alloc_find_clear_bits(osb, alloc, &bits_wanted,
752 /* TODO: Shouldn't we just BUG here? */
758 bitmap = la->la_bitmap;
759 *bit_off = le32_to_cpu(la->la_bm_off) + start;
760 *num_bits = bits_wanted;
762 status = ocfs2_journal_access_di(handle,
763 INODE_CACHE(local_alloc_inode),
765 OCFS2_JOURNAL_ACCESS_WRITE);
771 ocfs2_resmap_claimed_bits(&osb->osb_la_resmap, ac->ac_resv, start,
775 ocfs2_set_bit(start++, bitmap);
777 le32_add_cpu(&alloc->id1.bitmap1.i_used, *num_bits);
778 ocfs2_journal_dirty(handle, osb->local_alloc_bh);
786 static u32 ocfs2_local_alloc_count_bits(struct ocfs2_dinode *alloc)
791 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
793 buffer = la->la_bitmap;
794 for (i = 0; i < le16_to_cpu(la->la_size); i++)
795 count += hweight8(buffer[i]);
797 trace_ocfs2_local_alloc_count_bits(count);
801 static int ocfs2_local_alloc_find_clear_bits(struct ocfs2_super *osb,
802 struct ocfs2_dinode *alloc,
804 struct ocfs2_alloc_reservation *resv)
806 int numfound, bitoff, left, startoff, lastzero;
808 struct ocfs2_alloc_reservation r;
810 struct ocfs2_reservation_map *resmap = &osb->osb_la_resmap;
812 if (!alloc->id1.bitmap1.i_total) {
819 ocfs2_resv_init_once(&r);
820 ocfs2_resv_set_type(&r, OCFS2_RESV_FLAG_TMP);
825 if (ocfs2_resmap_resv_bits(resmap, resv, &bitoff, &numfound) == 0) {
826 if (numfound < *numbits)
832 * Code error. While reservations are enabled, local
833 * allocation should _always_ go through them.
835 BUG_ON(osb->osb_resv_level != 0);
838 * Reservations are disabled. Handle this the old way.
841 bitmap = OCFS2_LOCAL_ALLOC(alloc)->la_bitmap;
843 numfound = bitoff = startoff = 0;
845 left = le32_to_cpu(alloc->id1.bitmap1.i_total);
846 while ((bitoff = ocfs2_find_next_zero_bit(bitmap, left, startoff)) != -1) {
847 if (bitoff == left) {
848 /* mlog(0, "bitoff (%d) == left", bitoff); */
851 /* mlog(0, "Found a zero: bitoff = %d, startoff = %d, "
852 "numfound = %d\n", bitoff, startoff, numfound);*/
854 /* Ok, we found a zero bit... is it contig. or do we
856 if (bitoff == startoff) {
857 /* we found a zero */
861 /* got a zero after some ones */
865 /* we got everything we needed */
866 if (numfound == *numbits) {
867 /* mlog(0, "Found it all!\n"); */
872 trace_ocfs2_local_alloc_find_clear_bits_search_bitmap(bitoff, numfound);
874 if (numfound == *numbits)
875 bitoff = startoff - numfound;
881 ocfs2_resv_discard(resmap, resv);
883 trace_ocfs2_local_alloc_find_clear_bits(*numbits,
884 le32_to_cpu(alloc->id1.bitmap1.i_total),
890 static void ocfs2_clear_local_alloc(struct ocfs2_dinode *alloc)
892 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
895 alloc->id1.bitmap1.i_total = 0;
896 alloc->id1.bitmap1.i_used = 0;
898 for(i = 0; i < le16_to_cpu(la->la_size); i++)
899 la->la_bitmap[i] = 0;
903 /* turn this on and uncomment below to aid debugging window shifts. */
904 static void ocfs2_verify_zero_bits(unsigned long *bitmap,
908 unsigned int tmp = count;
910 if (ocfs2_test_bit(start + tmp, bitmap)) {
911 printk("ocfs2_verify_zero_bits: start = %u, count = "
912 "%u\n", start, count);
913 printk("ocfs2_verify_zero_bits: bit %u is set!",
922 * sync the local alloc to main bitmap.
924 * assumes you've already locked the main bitmap -- the bitmap inode
925 * passed is used for caching.
927 static int ocfs2_sync_local_to_main(struct ocfs2_super *osb,
929 struct ocfs2_dinode *alloc,
930 struct inode *main_bm_inode,
931 struct buffer_head *main_bm_bh)
934 int bit_off, left, count, start;
938 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
940 trace_ocfs2_sync_local_to_main(
941 le32_to_cpu(alloc->id1.bitmap1.i_total),
942 le32_to_cpu(alloc->id1.bitmap1.i_used));
944 if (!alloc->id1.bitmap1.i_total) {
948 if (le32_to_cpu(alloc->id1.bitmap1.i_used) ==
949 le32_to_cpu(alloc->id1.bitmap1.i_total)) {
953 la_start_blk = ocfs2_clusters_to_blocks(osb->sb,
954 le32_to_cpu(la->la_bm_off));
955 bitmap = la->la_bitmap;
956 start = count = bit_off = 0;
957 left = le32_to_cpu(alloc->id1.bitmap1.i_total);
959 while ((bit_off = ocfs2_find_next_zero_bit(bitmap, left, start))
961 if ((bit_off < left) && (bit_off == start)) {
967 blkno = la_start_blk +
968 ocfs2_clusters_to_blocks(osb->sb,
971 trace_ocfs2_sync_local_to_main_free(
972 count, start - count,
973 (unsigned long long)la_start_blk,
974 (unsigned long long)blkno);
976 status = ocfs2_release_clusters(handle,
997 enum ocfs2_la_event {
998 OCFS2_LA_EVENT_SLIDE, /* Normal window slide. */
999 OCFS2_LA_EVENT_FRAGMENTED, /* The global bitmap has
1000 * enough bits theoretically
1001 * free, but a contiguous
1002 * allocation could not be
1004 OCFS2_LA_EVENT_ENOSPC, /* Global bitmap doesn't have
1005 * enough bits free to satisfy
1008 #define OCFS2_LA_ENABLE_INTERVAL (30 * HZ)
1010 * Given an event, calculate the size of our next local alloc window.
1012 * This should always be called under i_mutex of the local alloc inode
1013 * so that local alloc disabling doesn't race with processes trying to
1014 * use the allocator.
1016 * Returns the state which the local alloc was left in. This value can
1017 * be ignored by some paths.
1019 static int ocfs2_recalc_la_window(struct ocfs2_super *osb,
1020 enum ocfs2_la_event event)
1025 spin_lock(&osb->osb_lock);
1026 if (osb->local_alloc_state == OCFS2_LA_DISABLED) {
1027 WARN_ON_ONCE(osb->local_alloc_state == OCFS2_LA_DISABLED);
1032 * ENOSPC and fragmentation are treated similarly for now.
1034 if (event == OCFS2_LA_EVENT_ENOSPC ||
1035 event == OCFS2_LA_EVENT_FRAGMENTED) {
1037 * We ran out of contiguous space in the primary
1038 * bitmap. Drastically reduce the number of bits used
1039 * by local alloc until we have to disable it.
1041 bits = osb->local_alloc_bits >> 1;
1042 if (bits > ocfs2_megabytes_to_clusters(osb->sb, 1)) {
1044 * By setting state to THROTTLED, we'll keep
1045 * the number of local alloc bits used down
1046 * until an event occurs which would give us
1047 * reason to assume the bitmap situation might
1050 osb->local_alloc_state = OCFS2_LA_THROTTLED;
1051 osb->local_alloc_bits = bits;
1053 osb->local_alloc_state = OCFS2_LA_DISABLED;
1055 queue_delayed_work(ocfs2_wq, &osb->la_enable_wq,
1056 OCFS2_LA_ENABLE_INTERVAL);
1061 * Don't increase the size of the local alloc window until we
1062 * know we might be able to fulfill the request. Otherwise, we
1063 * risk bouncing around the global bitmap during periods of
1066 if (osb->local_alloc_state != OCFS2_LA_THROTTLED)
1067 osb->local_alloc_bits = osb->local_alloc_default_bits;
1070 state = osb->local_alloc_state;
1071 spin_unlock(&osb->osb_lock);
1076 static int ocfs2_local_alloc_reserve_for_window(struct ocfs2_super *osb,
1077 struct ocfs2_alloc_context **ac,
1078 struct inode **bitmap_inode,
1079 struct buffer_head **bitmap_bh)
1083 *ac = kzalloc(sizeof(struct ocfs2_alloc_context), GFP_KERNEL);
1091 (*ac)->ac_bits_wanted = osb->local_alloc_default_bits;
1092 status = ocfs2_reserve_cluster_bitmap_bits(osb, *ac);
1093 if (status == -ENOSPC) {
1094 if (ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_ENOSPC) ==
1098 ocfs2_free_ac_resource(*ac);
1099 memset(*ac, 0, sizeof(struct ocfs2_alloc_context));
1107 *bitmap_inode = (*ac)->ac_inode;
1108 igrab(*bitmap_inode);
1109 *bitmap_bh = (*ac)->ac_bh;
1113 if ((status < 0) && *ac) {
1114 ocfs2_free_alloc_context(*ac);
1124 * pass it the bitmap lock in lock_bh if you have it.
1126 static int ocfs2_local_alloc_new_window(struct ocfs2_super *osb,
1128 struct ocfs2_alloc_context *ac)
1131 u32 cluster_off, cluster_count;
1132 struct ocfs2_dinode *alloc = NULL;
1133 struct ocfs2_local_alloc *la;
1135 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
1136 la = OCFS2_LOCAL_ALLOC(alloc);
1138 trace_ocfs2_local_alloc_new_window(
1139 le32_to_cpu(alloc->id1.bitmap1.i_total),
1140 osb->local_alloc_bits);
1142 /* Instruct the allocation code to try the most recently used
1143 * cluster group. We'll re-record the group used this pass
1145 ac->ac_last_group = osb->la_last_gd;
1147 /* we used the generic suballoc reserve function, but we set
1148 * everything up nicely, so there's no reason why we can't use
1149 * the more specific cluster api to claim bits. */
1150 status = ocfs2_claim_clusters(handle, ac, osb->local_alloc_bits,
1151 &cluster_off, &cluster_count);
1152 if (status == -ENOSPC) {
1155 * Note: We could also try syncing the journal here to
1156 * allow use of any free bits which the current
1157 * transaction can't give us access to. --Mark
1159 if (ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_FRAGMENTED) ==
1163 ac->ac_bits_wanted = osb->local_alloc_default_bits;
1164 status = ocfs2_claim_clusters(handle, ac,
1165 osb->local_alloc_bits,
1168 if (status == -ENOSPC)
1171 * We only shrunk the *minimum* number of in our
1172 * request - it's entirely possible that the allocator
1173 * might give us more than we asked for.
1176 spin_lock(&osb->osb_lock);
1177 osb->local_alloc_bits = cluster_count;
1178 spin_unlock(&osb->osb_lock);
1182 if (status != -ENOSPC)
1187 osb->la_last_gd = ac->ac_last_group;
1189 la->la_bm_off = cpu_to_le32(cluster_off);
1190 alloc->id1.bitmap1.i_total = cpu_to_le32(cluster_count);
1191 /* just in case... In the future when we find space ourselves,
1192 * we don't have to get all contiguous -- but we'll have to
1193 * set all previously used bits in bitmap and update
1194 * la_bits_set before setting the bits in the main bitmap. */
1195 alloc->id1.bitmap1.i_used = 0;
1196 memset(OCFS2_LOCAL_ALLOC(alloc)->la_bitmap, 0,
1197 le16_to_cpu(la->la_size));
1199 ocfs2_resmap_restart(&osb->osb_la_resmap, cluster_count,
1200 OCFS2_LOCAL_ALLOC(alloc)->la_bitmap);
1202 trace_ocfs2_local_alloc_new_window_result(
1203 OCFS2_LOCAL_ALLOC(alloc)->la_bm_off,
1204 le32_to_cpu(alloc->id1.bitmap1.i_total));
1212 /* Note that we do *NOT* lock the local alloc inode here as
1213 * it's been locked already for us. */
1214 static int ocfs2_local_alloc_slide_window(struct ocfs2_super *osb,
1215 struct inode *local_alloc_inode)
1218 struct buffer_head *main_bm_bh = NULL;
1219 struct inode *main_bm_inode = NULL;
1220 handle_t *handle = NULL;
1221 struct ocfs2_dinode *alloc;
1222 struct ocfs2_dinode *alloc_copy = NULL;
1223 struct ocfs2_alloc_context *ac = NULL;
1225 ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_SLIDE);
1227 /* This will lock the main bitmap for us. */
1228 status = ocfs2_local_alloc_reserve_for_window(osb,
1233 if (status != -ENOSPC)
1238 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
1239 if (IS_ERR(handle)) {
1240 status = PTR_ERR(handle);
1246 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
1248 /* We want to clear the local alloc before doing anything
1249 * else, so that if we error later during this operation,
1250 * local alloc shutdown won't try to double free main bitmap
1251 * bits. Make a copy so the sync function knows which bits to
1253 alloc_copy = kmalloc(osb->local_alloc_bh->b_size, GFP_NOFS);
1259 memcpy(alloc_copy, alloc, osb->local_alloc_bh->b_size);
1261 status = ocfs2_journal_access_di(handle,
1262 INODE_CACHE(local_alloc_inode),
1263 osb->local_alloc_bh,
1264 OCFS2_JOURNAL_ACCESS_WRITE);
1270 ocfs2_clear_local_alloc(alloc);
1271 ocfs2_journal_dirty(handle, osb->local_alloc_bh);
1273 status = ocfs2_sync_local_to_main(osb, handle, alloc_copy,
1274 main_bm_inode, main_bm_bh);
1280 status = ocfs2_local_alloc_new_window(osb, handle, ac);
1282 if (status != -ENOSPC)
1287 atomic_inc(&osb->alloc_stats.moves);
1291 ocfs2_commit_trans(osb, handle);
1296 iput(main_bm_inode);
1302 ocfs2_free_alloc_context(ac);