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"
47 #include "buffer_head_io.h"
49 #define OCFS2_LOCAL_ALLOC(dinode) (&((dinode)->id2.i_lab))
51 static u32 ocfs2_local_alloc_count_bits(struct ocfs2_dinode *alloc);
53 static int ocfs2_local_alloc_find_clear_bits(struct ocfs2_super *osb,
54 struct ocfs2_dinode *alloc,
56 struct ocfs2_alloc_reservation *resv);
58 static void ocfs2_clear_local_alloc(struct ocfs2_dinode *alloc);
60 static int ocfs2_sync_local_to_main(struct ocfs2_super *osb,
62 struct ocfs2_dinode *alloc,
63 struct inode *main_bm_inode,
64 struct buffer_head *main_bm_bh);
66 static int ocfs2_local_alloc_reserve_for_window(struct ocfs2_super *osb,
67 struct ocfs2_alloc_context **ac,
68 struct inode **bitmap_inode,
69 struct buffer_head **bitmap_bh);
71 static int ocfs2_local_alloc_new_window(struct ocfs2_super *osb,
73 struct ocfs2_alloc_context *ac);
75 static int ocfs2_local_alloc_slide_window(struct ocfs2_super *osb,
76 struct inode *local_alloc_inode);
79 * ocfs2_la_default_mb() - determine a default size, in megabytes of
82 * Generally, we'd like to pick as large a local alloc as
83 * possible. Performance on large workloads tends to scale
84 * proportionally to la size. In addition to that, the reservations
85 * code functions more efficiently as it can reserve more windows for
88 * Some things work against us when trying to choose a large local alloc:
90 * - We need to ensure our sizing is picked to leave enough space in
91 * group descriptors for other allocations (such as block groups,
92 * etc). Picking default sizes which are a multiple of 4 could help
93 * - block groups are allocated in 2mb and 4mb chunks.
95 * - Likewise, we don't want to starve other nodes of bits on small
96 * file systems. This can easily be taken care of by limiting our
97 * default to a reasonable size (256M) on larger cluster sizes.
99 * - Some file systems can't support very large sizes - 4k and 8k in
100 * particular are limited to less than 128 and 256 megabytes respectively.
102 * The following reference table shows group descriptor and local
103 * alloc maximums at various cluster sizes (4k blocksize)
105 * csize: 4K group: 126M la: 121M
106 * csize: 8K group: 252M la: 243M
107 * csize: 16K group: 504M la: 486M
108 * csize: 32K group: 1008M la: 972M
109 * csize: 64K group: 2016M la: 1944M
110 * csize: 128K group: 4032M la: 3888M
111 * csize: 256K group: 8064M la: 7776M
112 * csize: 512K group: 16128M la: 15552M
113 * csize: 1024K group: 32256M la: 31104M
115 #define OCFS2_LA_MAX_DEFAULT_MB 256
116 #define OCFS2_LA_OLD_DEFAULT 8
117 unsigned int ocfs2_la_default_mb(struct ocfs2_super *osb)
121 unsigned int la_max_mb;
122 unsigned int megs_per_slot;
123 struct super_block *sb = osb->sb;
125 gd_mb = ocfs2_clusters_to_megabytes(osb->sb,
126 8 * ocfs2_group_bitmap_size(sb, 0, osb->s_feature_incompat));
129 * This takes care of files systems with very small group
130 * descriptors - 512 byte blocksize at cluster sizes lower
131 * than 16K and also 1k blocksize with 4k cluster size.
133 if ((sb->s_blocksize == 512 && osb->s_clustersize <= 8192)
134 || (sb->s_blocksize == 1024 && osb->s_clustersize == 4096))
135 return OCFS2_LA_OLD_DEFAULT;
138 * Leave enough room for some block groups and make the final
139 * value we work from a multiple of 4.
147 * Keep window sizes down to a reasonable default
149 if (la_mb > OCFS2_LA_MAX_DEFAULT_MB) {
151 * Some clustersize / blocksize combinations will have
152 * given us a larger than OCFS2_LA_MAX_DEFAULT_MB
153 * default size, but get poor distribution when
154 * limited to exactly 256 megabytes.
156 * As an example, 16K clustersize at 4K blocksize
157 * gives us a cluster group size of 504M. Paring the
158 * local alloc size down to 256 however, would give us
159 * only one window and around 200MB left in the
160 * cluster group. Instead, find the first size below
161 * 256 which would give us an even distribution.
163 * Larger cluster group sizes actually work out pretty
164 * well when pared to 256, so we don't have to do this
165 * for any group that fits more than two
166 * OCFS2_LA_MAX_DEFAULT_MB windows.
168 if (gd_mb > (2 * OCFS2_LA_MAX_DEFAULT_MB))
171 unsigned int gd_mult = gd_mb;
173 while (gd_mult > 256)
174 gd_mult = gd_mult >> 1;
180 megs_per_slot = osb->osb_clusters_at_boot / osb->max_slots;
181 megs_per_slot = ocfs2_clusters_to_megabytes(osb->sb, megs_per_slot);
182 /* Too many nodes, too few disk clusters. */
183 if (megs_per_slot < la_mb)
184 la_mb = megs_per_slot;
186 /* We can't store more bits than we can in a block. */
187 la_max_mb = ocfs2_clusters_to_megabytes(osb->sb,
188 ocfs2_local_alloc_size(sb) * 8);
189 if (la_mb > la_max_mb)
195 void ocfs2_la_set_sizes(struct ocfs2_super *osb, int requested_mb)
197 struct super_block *sb = osb->sb;
198 unsigned int la_default_mb = ocfs2_la_default_mb(osb);
199 unsigned int la_max_mb;
201 la_max_mb = ocfs2_clusters_to_megabytes(sb,
202 ocfs2_local_alloc_size(sb) * 8);
204 mlog(0, "requested: %dM, max: %uM, default: %uM\n",
205 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 mlog(0, "state=%d, bits=%llu, la_bits=%d, ret=%d\n",
280 osb->local_alloc_state, (unsigned long long)bits, 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;
296 if (osb->local_alloc_bits == 0)
299 if (osb->local_alloc_bits >= osb->bitmap_cpg) {
300 mlog(ML_NOTICE, "Requested local alloc window %d is larger "
301 "than max possible %u. Using defaults.\n",
302 osb->local_alloc_bits, (osb->bitmap_cpg - 1));
303 osb->local_alloc_bits =
304 ocfs2_megabytes_to_clusters(osb->sb,
305 ocfs2_la_default_mb(osb));
308 /* read the alloc off disk */
309 inode = ocfs2_get_system_file_inode(osb, LOCAL_ALLOC_SYSTEM_INODE,
317 status = ocfs2_read_inode_block_full(inode, &alloc_bh,
318 OCFS2_BH_IGNORE_CACHE);
324 alloc = (struct ocfs2_dinode *) alloc_bh->b_data;
325 la = OCFS2_LOCAL_ALLOC(alloc);
327 if (!(le32_to_cpu(alloc->i_flags) &
328 (OCFS2_LOCAL_ALLOC_FL|OCFS2_BITMAP_FL))) {
329 mlog(ML_ERROR, "Invalid local alloc inode, %llu\n",
330 (unsigned long long)OCFS2_I(inode)->ip_blkno);
335 if ((la->la_size == 0) ||
336 (le16_to_cpu(la->la_size) > ocfs2_local_alloc_size(inode->i_sb))) {
337 mlog(ML_ERROR, "Local alloc size is invalid (la_size = %u)\n",
338 le16_to_cpu(la->la_size));
343 /* do a little verification. */
344 num_used = ocfs2_local_alloc_count_bits(alloc);
346 /* hopefully the local alloc has always been recovered before
349 || alloc->id1.bitmap1.i_used
350 || alloc->id1.bitmap1.i_total
352 mlog(ML_ERROR, "Local alloc hasn't been recovered!\n"
353 "found = %u, set = %u, taken = %u, off = %u\n",
354 num_used, le32_to_cpu(alloc->id1.bitmap1.i_used),
355 le32_to_cpu(alloc->id1.bitmap1.i_total),
356 OCFS2_LOCAL_ALLOC(alloc)->la_bm_off);
358 osb->local_alloc_bh = alloc_bh;
359 osb->local_alloc_state = OCFS2_LA_ENABLED;
367 mlog(0, "Local alloc window bits = %d\n", osb->local_alloc_bits);
374 * return any unused bits to the bitmap and write out a clean
377 * local_alloc_bh is optional. If not passed, we will simply use the
378 * one off osb. If you do pass it however, be warned that it *will* be
379 * returned brelse'd and NULL'd out.*/
380 void ocfs2_shutdown_local_alloc(struct ocfs2_super *osb)
384 struct inode *local_alloc_inode = NULL;
385 struct buffer_head *bh = NULL;
386 struct buffer_head *main_bm_bh = NULL;
387 struct inode *main_bm_inode = NULL;
388 struct ocfs2_dinode *alloc_copy = NULL;
389 struct ocfs2_dinode *alloc = NULL;
393 cancel_delayed_work(&osb->la_enable_wq);
394 flush_workqueue(ocfs2_wq);
396 if (osb->local_alloc_state == OCFS2_LA_UNUSED)
400 ocfs2_get_system_file_inode(osb,
401 LOCAL_ALLOC_SYSTEM_INODE,
403 if (!local_alloc_inode) {
409 osb->local_alloc_state = OCFS2_LA_DISABLED;
411 ocfs2_resmap_uninit(&osb->osb_la_resmap);
413 main_bm_inode = ocfs2_get_system_file_inode(osb,
414 GLOBAL_BITMAP_SYSTEM_INODE,
416 if (!main_bm_inode) {
422 mutex_lock(&main_bm_inode->i_mutex);
424 status = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 1);
430 /* WINDOW_MOVE_CREDITS is a bit heavy... */
431 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
432 if (IS_ERR(handle)) {
433 mlog_errno(PTR_ERR(handle));
438 bh = osb->local_alloc_bh;
439 alloc = (struct ocfs2_dinode *) bh->b_data;
441 alloc_copy = kmalloc(bh->b_size, GFP_NOFS);
446 memcpy(alloc_copy, alloc, bh->b_size);
448 status = ocfs2_journal_access_di(handle, INODE_CACHE(local_alloc_inode),
449 bh, OCFS2_JOURNAL_ACCESS_WRITE);
455 ocfs2_clear_local_alloc(alloc);
456 ocfs2_journal_dirty(handle, bh);
459 osb->local_alloc_bh = NULL;
460 osb->local_alloc_state = OCFS2_LA_UNUSED;
462 status = ocfs2_sync_local_to_main(osb, handle, alloc_copy,
463 main_bm_inode, main_bm_bh);
468 ocfs2_commit_trans(osb, handle);
473 ocfs2_inode_unlock(main_bm_inode, 1);
476 mutex_unlock(&main_bm_inode->i_mutex);
480 if (local_alloc_inode)
481 iput(local_alloc_inode);
490 * We want to free the bitmap bits outside of any recovery context as
491 * we'll need a cluster lock to do so, but we must clear the local
492 * alloc before giving up the recovered nodes journal. To solve this,
493 * we kmalloc a copy of the local alloc before it's change for the
494 * caller to process with ocfs2_complete_local_alloc_recovery
496 int ocfs2_begin_local_alloc_recovery(struct ocfs2_super *osb,
498 struct ocfs2_dinode **alloc_copy)
501 struct buffer_head *alloc_bh = NULL;
502 struct inode *inode = NULL;
503 struct ocfs2_dinode *alloc;
505 mlog_entry("(slot_num = %d)\n", slot_num);
509 inode = ocfs2_get_system_file_inode(osb,
510 LOCAL_ALLOC_SYSTEM_INODE,
518 mutex_lock(&inode->i_mutex);
520 status = ocfs2_read_inode_block_full(inode, &alloc_bh,
521 OCFS2_BH_IGNORE_CACHE);
527 *alloc_copy = kmalloc(alloc_bh->b_size, GFP_KERNEL);
528 if (!(*alloc_copy)) {
532 memcpy((*alloc_copy), alloc_bh->b_data, alloc_bh->b_size);
534 alloc = (struct ocfs2_dinode *) alloc_bh->b_data;
535 ocfs2_clear_local_alloc(alloc);
537 ocfs2_compute_meta_ecc(osb->sb, alloc_bh->b_data, &alloc->i_check);
538 status = ocfs2_write_block(osb, alloc_bh, INODE_CACHE(inode));
543 if ((status < 0) && (*alloc_copy)) {
551 mutex_unlock(&inode->i_mutex);
560 * Step 2: By now, we've completed the journal recovery, we've stamped
561 * a clean local alloc on disk and dropped the node out of the
562 * recovery map. Dlm locks will no longer stall, so lets clear out the
565 int ocfs2_complete_local_alloc_recovery(struct ocfs2_super *osb,
566 struct ocfs2_dinode *alloc)
570 struct buffer_head *main_bm_bh = NULL;
571 struct inode *main_bm_inode;
575 main_bm_inode = ocfs2_get_system_file_inode(osb,
576 GLOBAL_BITMAP_SYSTEM_INODE,
578 if (!main_bm_inode) {
584 mutex_lock(&main_bm_inode->i_mutex);
586 status = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 1);
592 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
593 if (IS_ERR(handle)) {
594 status = PTR_ERR(handle);
600 /* we want the bitmap change to be recorded on disk asap */
603 status = ocfs2_sync_local_to_main(osb, handle, alloc,
604 main_bm_inode, main_bm_bh);
608 ocfs2_commit_trans(osb, handle);
611 ocfs2_inode_unlock(main_bm_inode, 1);
614 mutex_unlock(&main_bm_inode->i_mutex);
622 ocfs2_init_steal_slots(osb);
628 * make sure we've got at least bits_wanted contiguous bits in the
629 * local alloc. You lose them when you drop i_mutex.
631 * We will add ourselves to the transaction passed in, but may start
632 * our own in order to shift windows.
634 int ocfs2_reserve_local_alloc_bits(struct ocfs2_super *osb,
636 struct ocfs2_alloc_context *ac)
639 struct ocfs2_dinode *alloc;
640 struct inode *local_alloc_inode;
641 unsigned int free_bits;
648 ocfs2_get_system_file_inode(osb,
649 LOCAL_ALLOC_SYSTEM_INODE,
651 if (!local_alloc_inode) {
657 mutex_lock(&local_alloc_inode->i_mutex);
660 * We must double check state and allocator bits because
661 * another process may have changed them while holding i_mutex.
663 spin_lock(&osb->osb_lock);
664 if (!ocfs2_la_state_enabled(osb) ||
665 (bits_wanted > osb->local_alloc_bits)) {
666 spin_unlock(&osb->osb_lock);
670 spin_unlock(&osb->osb_lock);
672 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
674 #ifdef CONFIG_OCFS2_DEBUG_FS
675 if (le32_to_cpu(alloc->id1.bitmap1.i_used) !=
676 ocfs2_local_alloc_count_bits(alloc)) {
677 ocfs2_error(osb->sb, "local alloc inode %llu says it has "
678 "%u free bits, but a count shows %u",
679 (unsigned long long)le64_to_cpu(alloc->i_blkno),
680 le32_to_cpu(alloc->id1.bitmap1.i_used),
681 ocfs2_local_alloc_count_bits(alloc));
687 free_bits = le32_to_cpu(alloc->id1.bitmap1.i_total) -
688 le32_to_cpu(alloc->id1.bitmap1.i_used);
689 if (bits_wanted > free_bits) {
690 /* uhoh, window change time. */
692 ocfs2_local_alloc_slide_window(osb, local_alloc_inode);
694 if (status != -ENOSPC)
700 * Under certain conditions, the window slide code
701 * might have reduced the number of bits available or
702 * disabled the the local alloc entirely. Re-check
703 * here and return -ENOSPC if necessary.
706 if (!ocfs2_la_state_enabled(osb))
709 free_bits = le32_to_cpu(alloc->id1.bitmap1.i_total) -
710 le32_to_cpu(alloc->id1.bitmap1.i_used);
711 if (bits_wanted > free_bits)
715 if (ac->ac_max_block)
716 mlog(0, "Calling in_range for max block %llu\n",
717 (unsigned long long)ac->ac_max_block);
719 ac->ac_inode = local_alloc_inode;
720 /* We should never use localalloc from another slot */
721 ac->ac_alloc_slot = osb->slot_num;
722 ac->ac_which = OCFS2_AC_USE_LOCAL;
723 get_bh(osb->local_alloc_bh);
724 ac->ac_bh = osb->local_alloc_bh;
727 if (status < 0 && local_alloc_inode) {
728 mutex_unlock(&local_alloc_inode->i_mutex);
729 iput(local_alloc_inode);
732 mlog(0, "bits=%d, slot=%d, ret=%d\n", bits_wanted, osb->slot_num,
739 int ocfs2_claim_local_alloc_bits(struct ocfs2_super *osb,
741 struct ocfs2_alloc_context *ac,
747 struct inode *local_alloc_inode;
749 struct ocfs2_dinode *alloc;
750 struct ocfs2_local_alloc *la;
753 BUG_ON(ac->ac_which != OCFS2_AC_USE_LOCAL);
755 local_alloc_inode = ac->ac_inode;
756 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
757 la = OCFS2_LOCAL_ALLOC(alloc);
759 start = ocfs2_local_alloc_find_clear_bits(osb, alloc, &bits_wanted,
762 /* TODO: Shouldn't we just BUG here? */
768 bitmap = la->la_bitmap;
769 *bit_off = le32_to_cpu(la->la_bm_off) + start;
770 *num_bits = bits_wanted;
772 status = ocfs2_journal_access_di(handle,
773 INODE_CACHE(local_alloc_inode),
775 OCFS2_JOURNAL_ACCESS_WRITE);
781 ocfs2_resmap_claimed_bits(&osb->osb_la_resmap, ac->ac_resv, start,
785 ocfs2_set_bit(start++, bitmap);
787 le32_add_cpu(&alloc->id1.bitmap1.i_used, *num_bits);
788 ocfs2_journal_dirty(handle, osb->local_alloc_bh);
795 static u32 ocfs2_local_alloc_count_bits(struct ocfs2_dinode *alloc)
800 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
804 buffer = la->la_bitmap;
805 for (i = 0; i < le16_to_cpu(la->la_size); i++)
806 count += hweight8(buffer[i]);
812 static int ocfs2_local_alloc_find_clear_bits(struct ocfs2_super *osb,
813 struct ocfs2_dinode *alloc,
815 struct ocfs2_alloc_reservation *resv)
817 int numfound, bitoff, left, startoff, lastzero;
819 struct ocfs2_alloc_reservation r;
821 struct ocfs2_reservation_map *resmap = &osb->osb_la_resmap;
823 mlog_entry("(numbits wanted = %u)\n", *numbits);
825 if (!alloc->id1.bitmap1.i_total) {
826 mlog(0, "No bits in my window!\n");
833 ocfs2_resv_init_once(&r);
834 ocfs2_resv_set_type(&r, OCFS2_RESV_FLAG_TMP);
839 if (ocfs2_resmap_resv_bits(resmap, resv, &bitoff, &numfound) == 0) {
840 if (numfound < *numbits)
846 * Code error. While reservations are enabled, local
847 * allocation should _always_ go through them.
849 BUG_ON(osb->osb_resv_level != 0);
852 * Reservations are disabled. Handle this the old way.
855 bitmap = OCFS2_LOCAL_ALLOC(alloc)->la_bitmap;
857 numfound = bitoff = startoff = 0;
859 left = le32_to_cpu(alloc->id1.bitmap1.i_total);
860 while ((bitoff = ocfs2_find_next_zero_bit(bitmap, left, startoff)) != -1) {
861 if (bitoff == left) {
862 /* mlog(0, "bitoff (%d) == left", bitoff); */
865 /* mlog(0, "Found a zero: bitoff = %d, startoff = %d, "
866 "numfound = %d\n", bitoff, startoff, numfound);*/
868 /* Ok, we found a zero bit... is it contig. or do we
870 if (bitoff == startoff) {
871 /* we found a zero */
875 /* got a zero after some ones */
879 /* we got everything we needed */
880 if (numfound == *numbits) {
881 /* mlog(0, "Found it all!\n"); */
886 mlog(0, "Exiting loop, bitoff = %d, numfound = %d\n", bitoff,
889 if (numfound == *numbits)
890 bitoff = startoff - numfound;
896 ocfs2_resv_discard(resmap, resv);
902 static void ocfs2_clear_local_alloc(struct ocfs2_dinode *alloc)
904 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
908 alloc->id1.bitmap1.i_total = 0;
909 alloc->id1.bitmap1.i_used = 0;
911 for(i = 0; i < le16_to_cpu(la->la_size); i++)
912 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 mlog_entry("total = %u, used = %u\n",
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) {
960 mlog(0, "nothing to sync!\n");
964 if (le32_to_cpu(alloc->id1.bitmap1.i_used) ==
965 le32_to_cpu(alloc->id1.bitmap1.i_total)) {
966 mlog(0, "all bits were taken!\n");
970 la_start_blk = ocfs2_clusters_to_blocks(osb->sb,
971 le32_to_cpu(la->la_bm_off));
972 bitmap = la->la_bitmap;
973 start = count = bit_off = 0;
974 left = le32_to_cpu(alloc->id1.bitmap1.i_total);
976 while ((bit_off = ocfs2_find_next_zero_bit(bitmap, left, start))
978 if ((bit_off < left) && (bit_off == start)) {
984 blkno = la_start_blk +
985 ocfs2_clusters_to_blocks(osb->sb,
988 mlog(0, "freeing %u bits starting at local alloc bit "
989 "%u (la_start_blk = %llu, blkno = %llu)\n",
990 count, start - count,
991 (unsigned long long)la_start_blk,
992 (unsigned long long)blkno);
994 status = ocfs2_release_clusters(handle,
1003 if (bit_off >= left)
1006 start = bit_off + 1;
1014 enum ocfs2_la_event {
1015 OCFS2_LA_EVENT_SLIDE, /* Normal window slide. */
1016 OCFS2_LA_EVENT_FRAGMENTED, /* The global bitmap has
1017 * enough bits theoretically
1018 * free, but a contiguous
1019 * allocation could not be
1021 OCFS2_LA_EVENT_ENOSPC, /* Global bitmap doesn't have
1022 * enough bits free to satisfy
1025 #define OCFS2_LA_ENABLE_INTERVAL (30 * HZ)
1027 * Given an event, calculate the size of our next local alloc window.
1029 * This should always be called under i_mutex of the local alloc inode
1030 * so that local alloc disabling doesn't race with processes trying to
1031 * use the allocator.
1033 * Returns the state which the local alloc was left in. This value can
1034 * be ignored by some paths.
1036 static int ocfs2_recalc_la_window(struct ocfs2_super *osb,
1037 enum ocfs2_la_event event)
1042 spin_lock(&osb->osb_lock);
1043 if (osb->local_alloc_state == OCFS2_LA_DISABLED) {
1044 WARN_ON_ONCE(osb->local_alloc_state == OCFS2_LA_DISABLED);
1049 * ENOSPC and fragmentation are treated similarly for now.
1051 if (event == OCFS2_LA_EVENT_ENOSPC ||
1052 event == OCFS2_LA_EVENT_FRAGMENTED) {
1054 * We ran out of contiguous space in the primary
1055 * bitmap. Drastically reduce the number of bits used
1056 * by local alloc until we have to disable it.
1058 bits = osb->local_alloc_bits >> 1;
1059 if (bits > ocfs2_megabytes_to_clusters(osb->sb, 1)) {
1061 * By setting state to THROTTLED, we'll keep
1062 * the number of local alloc bits used down
1063 * until an event occurs which would give us
1064 * reason to assume the bitmap situation might
1067 osb->local_alloc_state = OCFS2_LA_THROTTLED;
1068 osb->local_alloc_bits = bits;
1070 osb->local_alloc_state = OCFS2_LA_DISABLED;
1072 queue_delayed_work(ocfs2_wq, &osb->la_enable_wq,
1073 OCFS2_LA_ENABLE_INTERVAL);
1078 * Don't increase the size of the local alloc window until we
1079 * know we might be able to fulfill the request. Otherwise, we
1080 * risk bouncing around the global bitmap during periods of
1083 if (osb->local_alloc_state != OCFS2_LA_THROTTLED)
1084 osb->local_alloc_bits = osb->local_alloc_default_bits;
1087 state = osb->local_alloc_state;
1088 spin_unlock(&osb->osb_lock);
1093 static int ocfs2_local_alloc_reserve_for_window(struct ocfs2_super *osb,
1094 struct ocfs2_alloc_context **ac,
1095 struct inode **bitmap_inode,
1096 struct buffer_head **bitmap_bh)
1100 *ac = kzalloc(sizeof(struct ocfs2_alloc_context), GFP_KERNEL);
1108 (*ac)->ac_bits_wanted = osb->local_alloc_default_bits;
1109 status = ocfs2_reserve_cluster_bitmap_bits(osb, *ac);
1110 if (status == -ENOSPC) {
1111 if (ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_ENOSPC) ==
1115 ocfs2_free_ac_resource(*ac);
1116 memset(*ac, 0, sizeof(struct ocfs2_alloc_context));
1124 *bitmap_inode = (*ac)->ac_inode;
1125 igrab(*bitmap_inode);
1126 *bitmap_bh = (*ac)->ac_bh;
1130 if ((status < 0) && *ac) {
1131 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;
1153 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
1154 la = OCFS2_LOCAL_ALLOC(alloc);
1156 if (alloc->id1.bitmap1.i_total)
1157 mlog(0, "asking me to alloc a new window over a non-empty "
1160 mlog(0, "Allocating %u clusters for a new window.\n",
1161 osb->local_alloc_bits);
1163 /* Instruct the allocation code to try the most recently used
1164 * cluster group. We'll re-record the group used this pass
1166 ac->ac_last_group = osb->la_last_gd;
1168 /* we used the generic suballoc reserve function, but we set
1169 * everything up nicely, so there's no reason why we can't use
1170 * the more specific cluster api to claim bits. */
1171 status = ocfs2_claim_clusters(handle, ac, osb->local_alloc_bits,
1172 &cluster_off, &cluster_count);
1173 if (status == -ENOSPC) {
1176 * Note: We could also try syncing the journal here to
1177 * allow use of any free bits which the current
1178 * transaction can't give us access to. --Mark
1180 if (ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_FRAGMENTED) ==
1184 ac->ac_bits_wanted = osb->local_alloc_default_bits;
1185 status = ocfs2_claim_clusters(handle, ac,
1186 osb->local_alloc_bits,
1189 if (status == -ENOSPC)
1192 * We only shrunk the *minimum* number of in our
1193 * request - it's entirely possible that the allocator
1194 * might give us more than we asked for.
1197 spin_lock(&osb->osb_lock);
1198 osb->local_alloc_bits = cluster_count;
1199 spin_unlock(&osb->osb_lock);
1203 if (status != -ENOSPC)
1208 osb->la_last_gd = ac->ac_last_group;
1210 la->la_bm_off = cpu_to_le32(cluster_off);
1211 alloc->id1.bitmap1.i_total = cpu_to_le32(cluster_count);
1212 /* just in case... In the future when we find space ourselves,
1213 * we don't have to get all contiguous -- but we'll have to
1214 * set all previously used bits in bitmap and update
1215 * la_bits_set before setting the bits in the main bitmap. */
1216 alloc->id1.bitmap1.i_used = 0;
1217 memset(OCFS2_LOCAL_ALLOC(alloc)->la_bitmap, 0,
1218 le16_to_cpu(la->la_size));
1220 ocfs2_resmap_restart(&osb->osb_la_resmap, cluster_count,
1221 OCFS2_LOCAL_ALLOC(alloc)->la_bitmap);
1223 mlog(0, "New window allocated:\n");
1224 mlog(0, "window la_bm_off = %u\n",
1225 OCFS2_LOCAL_ALLOC(alloc)->la_bm_off);
1226 mlog(0, "window bits = %u\n", le32_to_cpu(alloc->id1.bitmap1.i_total));
1233 /* Note that we do *NOT* lock the local alloc inode here as
1234 * it's been locked already for us. */
1235 static int ocfs2_local_alloc_slide_window(struct ocfs2_super *osb,
1236 struct inode *local_alloc_inode)
1239 struct buffer_head *main_bm_bh = NULL;
1240 struct inode *main_bm_inode = NULL;
1241 handle_t *handle = NULL;
1242 struct ocfs2_dinode *alloc;
1243 struct ocfs2_dinode *alloc_copy = NULL;
1244 struct ocfs2_alloc_context *ac = NULL;
1248 ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_SLIDE);
1250 /* This will lock the main bitmap for us. */
1251 status = ocfs2_local_alloc_reserve_for_window(osb,
1256 if (status != -ENOSPC)
1261 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
1262 if (IS_ERR(handle)) {
1263 status = PTR_ERR(handle);
1269 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
1271 /* We want to clear the local alloc before doing anything
1272 * else, so that if we error later during this operation,
1273 * local alloc shutdown won't try to double free main bitmap
1274 * bits. Make a copy so the sync function knows which bits to
1276 alloc_copy = kmalloc(osb->local_alloc_bh->b_size, GFP_NOFS);
1282 memcpy(alloc_copy, alloc, osb->local_alloc_bh->b_size);
1284 status = ocfs2_journal_access_di(handle,
1285 INODE_CACHE(local_alloc_inode),
1286 osb->local_alloc_bh,
1287 OCFS2_JOURNAL_ACCESS_WRITE);
1293 ocfs2_clear_local_alloc(alloc);
1294 ocfs2_journal_dirty(handle, osb->local_alloc_bh);
1296 status = ocfs2_sync_local_to_main(osb, handle, alloc_copy,
1297 main_bm_inode, main_bm_bh);
1303 status = ocfs2_local_alloc_new_window(osb, handle, ac);
1305 if (status != -ENOSPC)
1310 atomic_inc(&osb->alloc_stats.moves);
1314 ocfs2_commit_trans(osb, handle);
1319 iput(main_bm_inode);
1325 ocfs2_free_alloc_context(ac);