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>
41 #include "localalloc.h"
46 #include "buffer_head_io.h"
48 #define OCFS2_LOCAL_ALLOC(dinode) (&((dinode)->id2.i_lab))
50 static inline int ocfs2_local_alloc_window_bits(struct ocfs2_super *osb);
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,
58 static void ocfs2_clear_local_alloc(struct ocfs2_dinode *alloc);
60 static int ocfs2_sync_local_to_main(struct ocfs2_super *osb,
61 struct ocfs2_journal_handle *handle,
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_journal_handle *handle,
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,
73 struct ocfs2_journal_handle *handle,
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 * Determine how large our local alloc window should be, in bits.
82 * These values (and the behavior in ocfs2_alloc_should_use_local) have
83 * been chosen so that most allocations, including new block groups go
84 * through local alloc.
86 static inline int ocfs2_local_alloc_window_bits(struct ocfs2_super *osb)
88 BUG_ON(osb->s_clustersize_bits < 12);
90 return 2048 >> (osb->s_clustersize_bits - 12);
94 * Tell us whether a given allocation should use the local alloc
95 * file. Otherwise, it has to go to the main bitmap.
97 int ocfs2_alloc_should_use_local(struct ocfs2_super *osb, u64 bits)
99 int la_bits = ocfs2_local_alloc_window_bits(osb);
101 if (osb->local_alloc_state != OCFS2_LA_ENABLED)
104 /* la_bits should be at least twice the size (in clusters) of
105 * a new block group. We want to be sure block group
106 * allocations go through the local alloc, so allow an
107 * allocation to take up to half the bitmap. */
108 if (bits > (la_bits / 2))
114 int ocfs2_load_local_alloc(struct ocfs2_super *osb)
117 struct ocfs2_dinode *alloc = NULL;
118 struct buffer_head *alloc_bh = NULL;
120 struct inode *inode = NULL;
121 struct ocfs2_local_alloc *la;
125 /* read the alloc off disk */
126 inode = ocfs2_get_system_file_inode(osb, LOCAL_ALLOC_SYSTEM_INODE,
134 status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno,
135 &alloc_bh, 0, inode);
141 alloc = (struct ocfs2_dinode *) alloc_bh->b_data;
142 la = OCFS2_LOCAL_ALLOC(alloc);
144 if (!(le32_to_cpu(alloc->i_flags) &
145 (OCFS2_LOCAL_ALLOC_FL|OCFS2_BITMAP_FL))) {
146 mlog(ML_ERROR, "Invalid local alloc inode, %llu\n",
147 (unsigned long long)OCFS2_I(inode)->ip_blkno);
152 if ((la->la_size == 0) ||
153 (le16_to_cpu(la->la_size) > ocfs2_local_alloc_size(inode->i_sb))) {
154 mlog(ML_ERROR, "Local alloc size is invalid (la_size = %u)\n",
155 le16_to_cpu(la->la_size));
160 /* do a little verification. */
161 num_used = ocfs2_local_alloc_count_bits(alloc);
163 /* hopefully the local alloc has always been recovered before
166 || alloc->id1.bitmap1.i_used
167 || alloc->id1.bitmap1.i_total
169 mlog(ML_ERROR, "Local alloc hasn't been recovered!\n"
170 "found = %u, set = %u, taken = %u, off = %u\n",
171 num_used, le32_to_cpu(alloc->id1.bitmap1.i_used),
172 le32_to_cpu(alloc->id1.bitmap1.i_total),
173 OCFS2_LOCAL_ALLOC(alloc)->la_bm_off);
175 osb->local_alloc_bh = alloc_bh;
176 osb->local_alloc_state = OCFS2_LA_ENABLED;
190 * return any unused bits to the bitmap and write out a clean
193 * local_alloc_bh is optional. If not passed, we will simply use the
194 * one off osb. If you do pass it however, be warned that it *will* be
195 * returned brelse'd and NULL'd out.*/
196 void ocfs2_shutdown_local_alloc(struct ocfs2_super *osb)
199 struct ocfs2_journal_handle *handle;
200 struct inode *local_alloc_inode = NULL;
201 struct buffer_head *bh = NULL;
202 struct buffer_head *main_bm_bh = NULL;
203 struct inode *main_bm_inode = NULL;
204 struct ocfs2_dinode *alloc_copy = NULL;
205 struct ocfs2_dinode *alloc = NULL;
209 if (osb->local_alloc_state == OCFS2_LA_UNUSED)
213 ocfs2_get_system_file_inode(osb,
214 LOCAL_ALLOC_SYSTEM_INODE,
216 if (!local_alloc_inode) {
222 osb->local_alloc_state = OCFS2_LA_DISABLED;
224 main_bm_inode = ocfs2_get_system_file_inode(osb,
225 GLOBAL_BITMAP_SYSTEM_INODE,
227 if (!main_bm_inode) {
233 mutex_lock(&main_bm_inode->i_mutex);
235 status = ocfs2_meta_lock(main_bm_inode, NULL, &main_bm_bh, 1);
241 /* WINDOW_MOVE_CREDITS is a bit heavy... */
242 handle = ocfs2_start_trans(osb, NULL, OCFS2_WINDOW_MOVE_CREDITS);
243 if (IS_ERR(handle)) {
244 mlog_errno(PTR_ERR(handle));
249 bh = osb->local_alloc_bh;
250 alloc = (struct ocfs2_dinode *) bh->b_data;
252 alloc_copy = kmalloc(bh->b_size, GFP_KERNEL);
257 memcpy(alloc_copy, alloc, bh->b_size);
259 status = ocfs2_journal_access(handle, local_alloc_inode, bh,
260 OCFS2_JOURNAL_ACCESS_WRITE);
266 ocfs2_clear_local_alloc(alloc);
268 status = ocfs2_journal_dirty(handle, bh);
275 osb->local_alloc_bh = NULL;
276 osb->local_alloc_state = OCFS2_LA_UNUSED;
278 status = ocfs2_sync_local_to_main(osb, handle, alloc_copy,
279 main_bm_inode, main_bm_bh);
284 ocfs2_commit_trans(handle);
290 ocfs2_meta_unlock(main_bm_inode, 1);
293 mutex_unlock(&main_bm_inode->i_mutex);
297 if (local_alloc_inode)
298 iput(local_alloc_inode);
307 * We want to free the bitmap bits outside of any recovery context as
308 * we'll need a cluster lock to do so, but we must clear the local
309 * alloc before giving up the recovered nodes journal. To solve this,
310 * we kmalloc a copy of the local alloc before it's change for the
311 * caller to process with ocfs2_complete_local_alloc_recovery
313 int ocfs2_begin_local_alloc_recovery(struct ocfs2_super *osb,
315 struct ocfs2_dinode **alloc_copy)
318 struct buffer_head *alloc_bh = NULL;
319 struct inode *inode = NULL;
320 struct ocfs2_dinode *alloc;
322 mlog_entry("(slot_num = %d)\n", slot_num);
326 inode = ocfs2_get_system_file_inode(osb,
327 LOCAL_ALLOC_SYSTEM_INODE,
335 mutex_lock(&inode->i_mutex);
337 status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno,
338 &alloc_bh, 0, inode);
344 *alloc_copy = kmalloc(alloc_bh->b_size, GFP_KERNEL);
345 if (!(*alloc_copy)) {
349 memcpy((*alloc_copy), alloc_bh->b_data, alloc_bh->b_size);
351 alloc = (struct ocfs2_dinode *) alloc_bh->b_data;
352 ocfs2_clear_local_alloc(alloc);
354 status = ocfs2_write_block(osb, alloc_bh, inode);
359 if ((status < 0) && (*alloc_copy)) {
368 mutex_unlock(&inode->i_mutex);
377 * Step 2: By now, we've completed the journal recovery, we've stamped
378 * a clean local alloc on disk and dropped the node out of the
379 * recovery map. Dlm locks will no longer stall, so lets clear out the
382 int ocfs2_complete_local_alloc_recovery(struct ocfs2_super *osb,
383 struct ocfs2_dinode *alloc)
386 struct ocfs2_journal_handle *handle;
387 struct buffer_head *main_bm_bh = NULL;
388 struct inode *main_bm_inode;
392 main_bm_inode = ocfs2_get_system_file_inode(osb,
393 GLOBAL_BITMAP_SYSTEM_INODE,
395 if (!main_bm_inode) {
401 mutex_lock(&main_bm_inode->i_mutex);
403 status = ocfs2_meta_lock(main_bm_inode, NULL, &main_bm_bh, 1);
409 handle = ocfs2_start_trans(osb, NULL, OCFS2_WINDOW_MOVE_CREDITS);
410 if (IS_ERR(handle)) {
411 status = PTR_ERR(handle);
417 /* we want the bitmap change to be recorded on disk asap */
418 handle->k_handle->h_sync = 1;
420 status = ocfs2_sync_local_to_main(osb, handle, alloc,
421 main_bm_inode, main_bm_bh);
425 ocfs2_commit_trans(handle);
428 ocfs2_meta_unlock(main_bm_inode, 1);
431 mutex_unlock(&main_bm_inode->i_mutex);
444 * make sure we've got at least bitswanted contiguous bits in the
445 * local alloc. You lose them when you drop i_mutex.
447 * We will add ourselves to the transaction passed in, but may start
448 * our own in order to shift windows.
450 int ocfs2_reserve_local_alloc_bits(struct ocfs2_super *osb,
451 struct ocfs2_journal_handle *passed_handle,
453 struct ocfs2_alloc_context *ac)
456 struct ocfs2_dinode *alloc;
457 struct inode *local_alloc_inode;
458 unsigned int free_bits;
462 BUG_ON(!passed_handle);
464 BUG_ON(passed_handle->k_handle);
467 ocfs2_get_system_file_inode(osb,
468 LOCAL_ALLOC_SYSTEM_INODE,
470 if (!local_alloc_inode) {
475 ocfs2_handle_add_inode(passed_handle, local_alloc_inode);
477 if (osb->local_alloc_state != OCFS2_LA_ENABLED) {
482 if (bits_wanted > ocfs2_local_alloc_window_bits(osb)) {
483 mlog(0, "Asking for more than my max window size!\n");
488 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
490 if (le32_to_cpu(alloc->id1.bitmap1.i_used) !=
491 ocfs2_local_alloc_count_bits(alloc)) {
492 ocfs2_error(osb->sb, "local alloc inode %llu says it has "
493 "%u free bits, but a count shows %u",
494 (unsigned long long)le64_to_cpu(alloc->i_blkno),
495 le32_to_cpu(alloc->id1.bitmap1.i_used),
496 ocfs2_local_alloc_count_bits(alloc));
501 free_bits = le32_to_cpu(alloc->id1.bitmap1.i_total) -
502 le32_to_cpu(alloc->id1.bitmap1.i_used);
503 if (bits_wanted > free_bits) {
504 /* uhoh, window change time. */
506 ocfs2_local_alloc_slide_window(osb, local_alloc_inode);
508 if (status != -ENOSPC)
514 ac->ac_inode = igrab(local_alloc_inode);
515 get_bh(osb->local_alloc_bh);
516 ac->ac_bh = osb->local_alloc_bh;
517 ac->ac_which = OCFS2_AC_USE_LOCAL;
520 if (local_alloc_inode)
521 iput(local_alloc_inode);
527 int ocfs2_claim_local_alloc_bits(struct ocfs2_super *osb,
528 struct ocfs2_journal_handle *handle,
529 struct ocfs2_alloc_context *ac,
535 struct inode *local_alloc_inode;
538 struct ocfs2_dinode *alloc;
539 struct ocfs2_local_alloc *la;
542 BUG_ON(ac->ac_which != OCFS2_AC_USE_LOCAL);
544 bits_wanted = ac->ac_bits_wanted - ac->ac_bits_given;
545 local_alloc_inode = ac->ac_inode;
546 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
547 la = OCFS2_LOCAL_ALLOC(alloc);
549 start = ocfs2_local_alloc_find_clear_bits(osb, alloc, bits_wanted);
551 /* TODO: Shouldn't we just BUG here? */
557 bitmap = la->la_bitmap;
558 *bit_off = le32_to_cpu(la->la_bm_off) + start;
559 /* local alloc is always contiguous by nature -- we never
560 * delete bits from it! */
561 *num_bits = bits_wanted;
563 status = ocfs2_journal_access(handle, local_alloc_inode,
565 OCFS2_JOURNAL_ACCESS_WRITE);
572 ocfs2_set_bit(start++, bitmap);
574 alloc->id1.bitmap1.i_used = cpu_to_le32(*num_bits +
575 le32_to_cpu(alloc->id1.bitmap1.i_used));
577 status = ocfs2_journal_dirty(handle, osb->local_alloc_bh);
589 static u32 ocfs2_local_alloc_count_bits(struct ocfs2_dinode *alloc)
594 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
598 buffer = la->la_bitmap;
599 for (i = 0; i < le16_to_cpu(la->la_size); i++)
600 count += hweight8(buffer[i]);
606 static int ocfs2_local_alloc_find_clear_bits(struct ocfs2_super *osb,
607 struct ocfs2_dinode *alloc,
610 int numfound, bitoff, left, startoff, lastzero;
613 mlog_entry("(numbits wanted = %u)\n", numbits);
615 if (!alloc->id1.bitmap1.i_total) {
616 mlog(0, "No bits in my window!\n");
621 bitmap = OCFS2_LOCAL_ALLOC(alloc)->la_bitmap;
623 numfound = bitoff = startoff = 0;
625 left = le32_to_cpu(alloc->id1.bitmap1.i_total);
626 while ((bitoff = ocfs2_find_next_zero_bit(bitmap, left, startoff)) != -1) {
627 if (bitoff == left) {
628 /* mlog(0, "bitoff (%d) == left", bitoff); */
631 /* mlog(0, "Found a zero: bitoff = %d, startoff = %d, "
632 "numfound = %d\n", bitoff, startoff, numfound);*/
634 /* Ok, we found a zero bit... is it contig. or do we
636 if (bitoff == startoff) {
637 /* we found a zero */
641 /* got a zero after some ones */
645 /* we got everything we needed */
646 if (numfound == numbits) {
647 /* mlog(0, "Found it all!\n"); */
652 mlog(0, "Exiting loop, bitoff = %d, numfound = %d\n", bitoff,
655 if (numfound == numbits)
656 bitoff = startoff - numfound;
665 static void ocfs2_clear_local_alloc(struct ocfs2_dinode *alloc)
667 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
671 alloc->id1.bitmap1.i_total = 0;
672 alloc->id1.bitmap1.i_used = 0;
674 for(i = 0; i < le16_to_cpu(la->la_size); i++)
675 la->la_bitmap[i] = 0;
681 /* turn this on and uncomment below to aid debugging window shifts. */
682 static void ocfs2_verify_zero_bits(unsigned long *bitmap,
686 unsigned int tmp = count;
688 if (ocfs2_test_bit(start + tmp, bitmap)) {
689 printk("ocfs2_verify_zero_bits: start = %u, count = "
690 "%u\n", start, count);
691 printk("ocfs2_verify_zero_bits: bit %u is set!",
700 * sync the local alloc to main bitmap.
702 * assumes you've already locked the main bitmap -- the bitmap inode
703 * passed is used for caching.
705 static int ocfs2_sync_local_to_main(struct ocfs2_super *osb,
706 struct ocfs2_journal_handle *handle,
707 struct ocfs2_dinode *alloc,
708 struct inode *main_bm_inode,
709 struct buffer_head *main_bm_bh)
712 int bit_off, left, count, start;
716 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
718 mlog_entry("total = %u, COUNT = %u, used = %u\n",
719 le32_to_cpu(alloc->id1.bitmap1.i_total),
720 ocfs2_local_alloc_count_bits(alloc),
721 le32_to_cpu(alloc->id1.bitmap1.i_used));
723 if (!alloc->id1.bitmap1.i_total) {
724 mlog(0, "nothing to sync!\n");
728 if (le32_to_cpu(alloc->id1.bitmap1.i_used) ==
729 le32_to_cpu(alloc->id1.bitmap1.i_total)) {
730 mlog(0, "all bits were taken!\n");
734 la_start_blk = ocfs2_clusters_to_blocks(osb->sb,
735 le32_to_cpu(la->la_bm_off));
736 bitmap = la->la_bitmap;
737 start = count = bit_off = 0;
738 left = le32_to_cpu(alloc->id1.bitmap1.i_total);
740 while ((bit_off = ocfs2_find_next_zero_bit(bitmap, left, start))
742 if ((bit_off < left) && (bit_off == start)) {
748 blkno = la_start_blk +
749 ocfs2_clusters_to_blocks(osb->sb,
752 mlog(0, "freeing %u bits starting at local alloc bit "
753 "%u (la_start_blk = %llu, blkno = %llu)\n",
754 count, start - count,
755 (unsigned long long)la_start_blk,
756 (unsigned long long)blkno);
758 status = ocfs2_free_clusters(handle, main_bm_inode,
759 main_bm_bh, blkno, count);
776 static int ocfs2_local_alloc_reserve_for_window(struct ocfs2_super *osb,
777 struct ocfs2_journal_handle *handle,
778 struct ocfs2_alloc_context **ac,
779 struct inode **bitmap_inode,
780 struct buffer_head **bitmap_bh)
784 *ac = kcalloc(1, sizeof(struct ocfs2_alloc_context), GFP_KERNEL);
791 (*ac)->ac_handle = handle;
792 (*ac)->ac_bits_wanted = ocfs2_local_alloc_window_bits(osb);
794 status = ocfs2_reserve_cluster_bitmap_bits(osb, *ac);
796 if (status != -ENOSPC)
801 *bitmap_inode = (*ac)->ac_inode;
802 igrab(*bitmap_inode);
803 *bitmap_bh = (*ac)->ac_bh;
807 if ((status < 0) && *ac) {
808 ocfs2_free_alloc_context(*ac);
817 * pass it the bitmap lock in lock_bh if you have it.
819 static int ocfs2_local_alloc_new_window(struct ocfs2_super *osb,
820 struct ocfs2_journal_handle *handle,
821 struct ocfs2_alloc_context *ac)
824 u32 cluster_off, cluster_count;
825 struct ocfs2_dinode *alloc = NULL;
826 struct ocfs2_local_alloc *la;
830 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
831 la = OCFS2_LOCAL_ALLOC(alloc);
833 if (alloc->id1.bitmap1.i_total)
834 mlog(0, "asking me to alloc a new window over a non-empty "
837 mlog(0, "Allocating %u clusters for a new window.\n",
838 ocfs2_local_alloc_window_bits(osb));
840 /* Instruct the allocation code to try the most recently used
841 * cluster group. We'll re-record the group used this pass
843 ac->ac_last_group = osb->la_last_gd;
845 /* we used the generic suballoc reserve function, but we set
846 * everything up nicely, so there's no reason why we can't use
847 * the more specific cluster api to claim bits. */
848 status = ocfs2_claim_clusters(osb, handle, ac,
849 ocfs2_local_alloc_window_bits(osb),
850 &cluster_off, &cluster_count);
852 if (status != -ENOSPC)
857 osb->la_last_gd = ac->ac_last_group;
859 la->la_bm_off = cpu_to_le32(cluster_off);
860 alloc->id1.bitmap1.i_total = cpu_to_le32(cluster_count);
861 /* just in case... In the future when we find space ourselves,
862 * we don't have to get all contiguous -- but we'll have to
863 * set all previously used bits in bitmap and update
864 * la_bits_set before setting the bits in the main bitmap. */
865 alloc->id1.bitmap1.i_used = 0;
866 memset(OCFS2_LOCAL_ALLOC(alloc)->la_bitmap, 0,
867 le16_to_cpu(la->la_size));
869 mlog(0, "New window allocated:\n");
870 mlog(0, "window la_bm_off = %u\n",
871 OCFS2_LOCAL_ALLOC(alloc)->la_bm_off);
872 mlog(0, "window bits = %u\n", le32_to_cpu(alloc->id1.bitmap1.i_total));
879 /* Note that we do *NOT* lock the local alloc inode here as
880 * it's been locked already for us. */
881 static int ocfs2_local_alloc_slide_window(struct ocfs2_super *osb,
882 struct inode *local_alloc_inode)
885 struct buffer_head *main_bm_bh = NULL;
886 struct inode *main_bm_inode = NULL;
887 struct ocfs2_journal_handle *handle = NULL;
888 struct ocfs2_dinode *alloc;
889 struct ocfs2_dinode *alloc_copy = NULL;
890 struct ocfs2_alloc_context *ac = NULL;
894 handle = ocfs2_alloc_handle(osb);
901 /* This will lock the main bitmap for us. */
902 status = ocfs2_local_alloc_reserve_for_window(osb,
908 if (status != -ENOSPC)
913 handle = ocfs2_start_trans(osb, handle, OCFS2_WINDOW_MOVE_CREDITS);
914 if (IS_ERR(handle)) {
915 status = PTR_ERR(handle);
921 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
923 /* We want to clear the local alloc before doing anything
924 * else, so that if we error later during this operation,
925 * local alloc shutdown won't try to double free main bitmap
926 * bits. Make a copy so the sync function knows which bits to
928 alloc_copy = kmalloc(osb->local_alloc_bh->b_size, GFP_KERNEL);
934 memcpy(alloc_copy, alloc, osb->local_alloc_bh->b_size);
936 status = ocfs2_journal_access(handle, local_alloc_inode,
938 OCFS2_JOURNAL_ACCESS_WRITE);
944 ocfs2_clear_local_alloc(alloc);
946 status = ocfs2_journal_dirty(handle, osb->local_alloc_bh);
952 status = ocfs2_sync_local_to_main(osb, handle, alloc_copy,
953 main_bm_inode, main_bm_bh);
959 status = ocfs2_local_alloc_new_window(osb, handle, ac);
961 if (status != -ENOSPC)
966 atomic_inc(&osb->alloc_stats.moves);
971 ocfs2_commit_trans(handle);
983 ocfs2_free_alloc_context(ac);