1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* -*- mode: c; c-basic-offset: 8; -*-
3 * vim: noexpandtab sw=8 ts=8 sts=0:
7 * Node local data allocation
9 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
13 #include <linux/types.h>
14 #include <linux/slab.h>
15 #include <linux/highmem.h>
16 #include <linux/bitops.h>
18 #include <cluster/masklog.h>
23 #include "blockcheck.h"
27 #include "localalloc.h"
31 #include "ocfs2_trace.h"
33 #include "buffer_head_io.h"
35 #define OCFS2_LOCAL_ALLOC(dinode) (&((dinode)->id2.i_lab))
37 static u32 ocfs2_local_alloc_count_bits(struct ocfs2_dinode *alloc);
39 static int ocfs2_local_alloc_find_clear_bits(struct ocfs2_super *osb,
40 struct ocfs2_dinode *alloc,
42 struct ocfs2_alloc_reservation *resv);
44 static void ocfs2_clear_local_alloc(struct ocfs2_dinode *alloc);
46 static int ocfs2_sync_local_to_main(struct ocfs2_super *osb,
48 struct ocfs2_dinode *alloc,
49 struct inode *main_bm_inode,
50 struct buffer_head *main_bm_bh);
52 static int ocfs2_local_alloc_reserve_for_window(struct ocfs2_super *osb,
53 struct ocfs2_alloc_context **ac,
54 struct inode **bitmap_inode,
55 struct buffer_head **bitmap_bh);
57 static int ocfs2_local_alloc_new_window(struct ocfs2_super *osb,
59 struct ocfs2_alloc_context *ac);
61 static int ocfs2_local_alloc_slide_window(struct ocfs2_super *osb,
62 struct inode *local_alloc_inode);
65 * ocfs2_la_default_mb() - determine a default size, in megabytes of
68 * Generally, we'd like to pick as large a local alloc as
69 * possible. Performance on large workloads tends to scale
70 * proportionally to la size. In addition to that, the reservations
71 * code functions more efficiently as it can reserve more windows for
74 * Some things work against us when trying to choose a large local alloc:
76 * - We need to ensure our sizing is picked to leave enough space in
77 * group descriptors for other allocations (such as block groups,
78 * etc). Picking default sizes which are a multiple of 4 could help
79 * - block groups are allocated in 2mb and 4mb chunks.
81 * - Likewise, we don't want to starve other nodes of bits on small
82 * file systems. This can easily be taken care of by limiting our
83 * default to a reasonable size (256M) on larger cluster sizes.
85 * - Some file systems can't support very large sizes - 4k and 8k in
86 * particular are limited to less than 128 and 256 megabytes respectively.
88 * The following reference table shows group descriptor and local
89 * alloc maximums at various cluster sizes (4k blocksize)
91 * csize: 4K group: 126M la: 121M
92 * csize: 8K group: 252M la: 243M
93 * csize: 16K group: 504M la: 486M
94 * csize: 32K group: 1008M la: 972M
95 * csize: 64K group: 2016M la: 1944M
96 * csize: 128K group: 4032M la: 3888M
97 * csize: 256K group: 8064M la: 7776M
98 * csize: 512K group: 16128M la: 15552M
99 * csize: 1024K group: 32256M la: 31104M
101 #define OCFS2_LA_MAX_DEFAULT_MB 256
102 #define OCFS2_LA_OLD_DEFAULT 8
103 unsigned int ocfs2_la_default_mb(struct ocfs2_super *osb)
107 unsigned int la_max_mb;
108 unsigned int megs_per_slot;
109 struct super_block *sb = osb->sb;
111 gd_mb = ocfs2_clusters_to_megabytes(osb->sb,
112 8 * ocfs2_group_bitmap_size(sb, 0, osb->s_feature_incompat));
115 * This takes care of files systems with very small group
116 * descriptors - 512 byte blocksize at cluster sizes lower
117 * than 16K and also 1k blocksize with 4k cluster size.
119 if ((sb->s_blocksize == 512 && osb->s_clustersize <= 8192)
120 || (sb->s_blocksize == 1024 && osb->s_clustersize == 4096))
121 return OCFS2_LA_OLD_DEFAULT;
124 * Leave enough room for some block groups and make the final
125 * value we work from a multiple of 4.
133 * Keep window sizes down to a reasonable default
135 if (la_mb > OCFS2_LA_MAX_DEFAULT_MB) {
137 * Some clustersize / blocksize combinations will have
138 * given us a larger than OCFS2_LA_MAX_DEFAULT_MB
139 * default size, but get poor distribution when
140 * limited to exactly 256 megabytes.
142 * As an example, 16K clustersize at 4K blocksize
143 * gives us a cluster group size of 504M. Paring the
144 * local alloc size down to 256 however, would give us
145 * only one window and around 200MB left in the
146 * cluster group. Instead, find the first size below
147 * 256 which would give us an even distribution.
149 * Larger cluster group sizes actually work out pretty
150 * well when pared to 256, so we don't have to do this
151 * for any group that fits more than two
152 * OCFS2_LA_MAX_DEFAULT_MB windows.
154 if (gd_mb > (2 * OCFS2_LA_MAX_DEFAULT_MB))
157 unsigned int gd_mult = gd_mb;
159 while (gd_mult > 256)
160 gd_mult = gd_mult >> 1;
166 megs_per_slot = osb->osb_clusters_at_boot / osb->max_slots;
167 megs_per_slot = ocfs2_clusters_to_megabytes(osb->sb, megs_per_slot);
168 /* Too many nodes, too few disk clusters. */
169 if (megs_per_slot < la_mb)
170 la_mb = megs_per_slot;
172 /* We can't store more bits than we can in a block. */
173 la_max_mb = ocfs2_clusters_to_megabytes(osb->sb,
174 ocfs2_local_alloc_size(sb) * 8);
175 if (la_mb > la_max_mb)
181 void ocfs2_la_set_sizes(struct ocfs2_super *osb, int requested_mb)
183 struct super_block *sb = osb->sb;
184 unsigned int la_default_mb = ocfs2_la_default_mb(osb);
185 unsigned int la_max_mb;
187 la_max_mb = ocfs2_clusters_to_megabytes(sb,
188 ocfs2_local_alloc_size(sb) * 8);
190 trace_ocfs2_la_set_sizes(requested_mb, la_max_mb, la_default_mb);
192 if (requested_mb == -1) {
193 /* No user request - use defaults */
194 osb->local_alloc_default_bits =
195 ocfs2_megabytes_to_clusters(sb, la_default_mb);
196 } else if (requested_mb > la_max_mb) {
197 /* Request is too big, we give the maximum available */
198 osb->local_alloc_default_bits =
199 ocfs2_megabytes_to_clusters(sb, la_max_mb);
201 osb->local_alloc_default_bits =
202 ocfs2_megabytes_to_clusters(sb, requested_mb);
205 osb->local_alloc_bits = osb->local_alloc_default_bits;
208 static inline int ocfs2_la_state_enabled(struct ocfs2_super *osb)
210 return (osb->local_alloc_state == OCFS2_LA_THROTTLED ||
211 osb->local_alloc_state == OCFS2_LA_ENABLED);
214 void ocfs2_local_alloc_seen_free_bits(struct ocfs2_super *osb,
215 unsigned int num_clusters)
217 spin_lock(&osb->osb_lock);
218 if (osb->local_alloc_state == OCFS2_LA_DISABLED ||
219 osb->local_alloc_state == OCFS2_LA_THROTTLED)
220 if (num_clusters >= osb->local_alloc_default_bits) {
221 cancel_delayed_work(&osb->la_enable_wq);
222 osb->local_alloc_state = OCFS2_LA_ENABLED;
224 spin_unlock(&osb->osb_lock);
227 void ocfs2_la_enable_worker(struct work_struct *work)
229 struct ocfs2_super *osb =
230 container_of(work, struct ocfs2_super,
232 spin_lock(&osb->osb_lock);
233 osb->local_alloc_state = OCFS2_LA_ENABLED;
234 spin_unlock(&osb->osb_lock);
238 * Tell us whether a given allocation should use the local alloc
239 * file. Otherwise, it has to go to the main bitmap.
241 * This function does semi-dirty reads of local alloc size and state!
242 * This is ok however, as the values are re-checked once under mutex.
244 int ocfs2_alloc_should_use_local(struct ocfs2_super *osb, u64 bits)
249 spin_lock(&osb->osb_lock);
250 la_bits = osb->local_alloc_bits;
252 if (!ocfs2_la_state_enabled(osb))
255 /* la_bits should be at least twice the size (in clusters) of
256 * a new block group. We want to be sure block group
257 * allocations go through the local alloc, so allow an
258 * allocation to take up to half the bitmap. */
259 if (bits > (la_bits / 2))
264 trace_ocfs2_alloc_should_use_local(
265 (unsigned long long)bits, osb->local_alloc_state, la_bits, ret);
266 spin_unlock(&osb->osb_lock);
270 int ocfs2_load_local_alloc(struct ocfs2_super *osb)
273 struct ocfs2_dinode *alloc = NULL;
274 struct buffer_head *alloc_bh = NULL;
276 struct inode *inode = NULL;
277 struct ocfs2_local_alloc *la;
279 if (osb->local_alloc_bits == 0)
282 if (osb->local_alloc_bits >= osb->bitmap_cpg) {
283 mlog(ML_NOTICE, "Requested local alloc window %d is larger "
284 "than max possible %u. Using defaults.\n",
285 osb->local_alloc_bits, (osb->bitmap_cpg - 1));
286 osb->local_alloc_bits =
287 ocfs2_megabytes_to_clusters(osb->sb,
288 ocfs2_la_default_mb(osb));
291 /* read the alloc off disk */
292 inode = ocfs2_get_system_file_inode(osb, LOCAL_ALLOC_SYSTEM_INODE,
300 status = ocfs2_read_inode_block_full(inode, &alloc_bh,
301 OCFS2_BH_IGNORE_CACHE);
307 alloc = (struct ocfs2_dinode *) alloc_bh->b_data;
308 la = OCFS2_LOCAL_ALLOC(alloc);
310 if (!(le32_to_cpu(alloc->i_flags) &
311 (OCFS2_LOCAL_ALLOC_FL|OCFS2_BITMAP_FL))) {
312 mlog(ML_ERROR, "Invalid local alloc inode, %llu\n",
313 (unsigned long long)OCFS2_I(inode)->ip_blkno);
318 if ((la->la_size == 0) ||
319 (le16_to_cpu(la->la_size) > ocfs2_local_alloc_size(inode->i_sb))) {
320 mlog(ML_ERROR, "Local alloc size is invalid (la_size = %u)\n",
321 le16_to_cpu(la->la_size));
326 /* do a little verification. */
327 num_used = ocfs2_local_alloc_count_bits(alloc);
329 /* hopefully the local alloc has always been recovered before
332 || alloc->id1.bitmap1.i_used
333 || alloc->id1.bitmap1.i_total
335 mlog(ML_ERROR, "inconsistent detected, clean journal with"
336 " unrecovered local alloc, please run fsck.ocfs2!\n"
337 "found = %u, set = %u, taken = %u, off = %u\n",
338 num_used, le32_to_cpu(alloc->id1.bitmap1.i_used),
339 le32_to_cpu(alloc->id1.bitmap1.i_total),
340 OCFS2_LOCAL_ALLOC(alloc)->la_bm_off);
346 osb->local_alloc_bh = alloc_bh;
347 osb->local_alloc_state = OCFS2_LA_ENABLED;
354 trace_ocfs2_load_local_alloc(osb->local_alloc_bits);
362 * return any unused bits to the bitmap and write out a clean
365 * local_alloc_bh is optional. If not passed, we will simply use the
366 * one off osb. If you do pass it however, be warned that it *will* be
367 * returned brelse'd and NULL'd out.*/
368 void ocfs2_shutdown_local_alloc(struct ocfs2_super *osb)
372 struct inode *local_alloc_inode = NULL;
373 struct buffer_head *bh = NULL;
374 struct buffer_head *main_bm_bh = NULL;
375 struct inode *main_bm_inode = NULL;
376 struct ocfs2_dinode *alloc_copy = NULL;
377 struct ocfs2_dinode *alloc = NULL;
379 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_mutex.
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_mutex.
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 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)) != -1) {
868 if (bitoff == left) {
869 /* mlog(0, "bitoff (%d) == left", bitoff); */
872 /* mlog(0, "Found a zero: bitoff = %d, startoff = %d, "
873 "numfound = %d\n", bitoff, startoff, numfound);*/
875 /* Ok, we found a zero bit... is it contig. or do we
877 if (bitoff == startoff) {
878 /* we found a zero */
882 /* got a zero after some ones */
886 /* we got everything we needed */
887 if (numfound == *numbits) {
888 /* mlog(0, "Found it all!\n"); */
893 trace_ocfs2_local_alloc_find_clear_bits_search_bitmap(bitoff, numfound);
895 if (numfound == *numbits)
896 bitoff = startoff - numfound;
902 ocfs2_resv_discard(resmap, resv);
904 trace_ocfs2_local_alloc_find_clear_bits(*numbits,
905 le32_to_cpu(alloc->id1.bitmap1.i_total),
911 static void ocfs2_clear_local_alloc(struct ocfs2_dinode *alloc)
913 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
916 alloc->id1.bitmap1.i_total = 0;
917 alloc->id1.bitmap1.i_used = 0;
919 for(i = 0; i < le16_to_cpu(la->la_size); i++)
920 la->la_bitmap[i] = 0;
924 /* turn this on and uncomment below to aid debugging window shifts. */
925 static void ocfs2_verify_zero_bits(unsigned long *bitmap,
929 unsigned int tmp = count;
931 if (ocfs2_test_bit(start + tmp, bitmap)) {
932 printk("ocfs2_verify_zero_bits: start = %u, count = "
933 "%u\n", start, count);
934 printk("ocfs2_verify_zero_bits: bit %u is set!",
943 * sync the local alloc to main bitmap.
945 * assumes you've already locked the main bitmap -- the bitmap inode
946 * passed is used for caching.
948 static int ocfs2_sync_local_to_main(struct ocfs2_super *osb,
950 struct ocfs2_dinode *alloc,
951 struct inode *main_bm_inode,
952 struct buffer_head *main_bm_bh)
955 int bit_off, left, count, start;
959 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
961 trace_ocfs2_sync_local_to_main(
962 le32_to_cpu(alloc->id1.bitmap1.i_total),
963 le32_to_cpu(alloc->id1.bitmap1.i_used));
965 if (!alloc->id1.bitmap1.i_total) {
969 if (le32_to_cpu(alloc->id1.bitmap1.i_used) ==
970 le32_to_cpu(alloc->id1.bitmap1.i_total)) {
974 la_start_blk = ocfs2_clusters_to_blocks(osb->sb,
975 le32_to_cpu(la->la_bm_off));
976 bitmap = la->la_bitmap;
977 start = count = bit_off = 0;
978 left = le32_to_cpu(alloc->id1.bitmap1.i_total);
980 while ((bit_off = ocfs2_find_next_zero_bit(bitmap, left, start))
982 if ((bit_off < left) && (bit_off == start)) {
988 blkno = la_start_blk +
989 ocfs2_clusters_to_blocks(osb->sb,
992 trace_ocfs2_sync_local_to_main_free(
993 count, start - count,
994 (unsigned long long)la_start_blk,
995 (unsigned long long)blkno);
997 status = ocfs2_release_clusters(handle,
1006 if (bit_off >= left)
1009 start = bit_off + 1;
1018 enum ocfs2_la_event {
1019 OCFS2_LA_EVENT_SLIDE, /* Normal window slide. */
1020 OCFS2_LA_EVENT_FRAGMENTED, /* The global bitmap has
1021 * enough bits theoretically
1022 * free, but a contiguous
1023 * allocation could not be
1025 OCFS2_LA_EVENT_ENOSPC, /* Global bitmap doesn't have
1026 * enough bits free to satisfy
1029 #define OCFS2_LA_ENABLE_INTERVAL (30 * HZ)
1031 * Given an event, calculate the size of our next local alloc window.
1033 * This should always be called under i_mutex of the local alloc inode
1034 * so that local alloc disabling doesn't race with processes trying to
1035 * use the allocator.
1037 * Returns the state which the local alloc was left in. This value can
1038 * be ignored by some paths.
1040 static int ocfs2_recalc_la_window(struct ocfs2_super *osb,
1041 enum ocfs2_la_event event)
1046 spin_lock(&osb->osb_lock);
1047 if (osb->local_alloc_state == OCFS2_LA_DISABLED) {
1048 WARN_ON_ONCE(osb->local_alloc_state == OCFS2_LA_DISABLED);
1053 * ENOSPC and fragmentation are treated similarly for now.
1055 if (event == OCFS2_LA_EVENT_ENOSPC ||
1056 event == OCFS2_LA_EVENT_FRAGMENTED) {
1058 * We ran out of contiguous space in the primary
1059 * bitmap. Drastically reduce the number of bits used
1060 * by local alloc until we have to disable it.
1062 bits = osb->local_alloc_bits >> 1;
1063 if (bits > ocfs2_megabytes_to_clusters(osb->sb, 1)) {
1065 * By setting state to THROTTLED, we'll keep
1066 * the number of local alloc bits used down
1067 * until an event occurs which would give us
1068 * reason to assume the bitmap situation might
1071 osb->local_alloc_state = OCFS2_LA_THROTTLED;
1072 osb->local_alloc_bits = bits;
1074 osb->local_alloc_state = OCFS2_LA_DISABLED;
1076 queue_delayed_work(osb->ocfs2_wq, &osb->la_enable_wq,
1077 OCFS2_LA_ENABLE_INTERVAL);
1082 * Don't increase the size of the local alloc window until we
1083 * know we might be able to fulfill the request. Otherwise, we
1084 * risk bouncing around the global bitmap during periods of
1087 if (osb->local_alloc_state != OCFS2_LA_THROTTLED)
1088 osb->local_alloc_bits = osb->local_alloc_default_bits;
1091 state = osb->local_alloc_state;
1092 spin_unlock(&osb->osb_lock);
1097 static int ocfs2_local_alloc_reserve_for_window(struct ocfs2_super *osb,
1098 struct ocfs2_alloc_context **ac,
1099 struct inode **bitmap_inode,
1100 struct buffer_head **bitmap_bh)
1104 *ac = kzalloc(sizeof(struct ocfs2_alloc_context), GFP_KERNEL);
1112 (*ac)->ac_bits_wanted = osb->local_alloc_bits;
1113 status = ocfs2_reserve_cluster_bitmap_bits(osb, *ac);
1114 if (status == -ENOSPC) {
1115 if (ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_ENOSPC) ==
1119 ocfs2_free_ac_resource(*ac);
1120 memset(*ac, 0, sizeof(struct ocfs2_alloc_context));
1128 *bitmap_inode = (*ac)->ac_inode;
1129 igrab(*bitmap_inode);
1130 *bitmap_bh = (*ac)->ac_bh;
1134 if ((status < 0) && *ac) {
1135 ocfs2_free_alloc_context(*ac);
1145 * pass it the bitmap lock in lock_bh if you have it.
1147 static int ocfs2_local_alloc_new_window(struct ocfs2_super *osb,
1149 struct ocfs2_alloc_context *ac)
1152 u32 cluster_off, cluster_count;
1153 struct ocfs2_dinode *alloc = NULL;
1154 struct ocfs2_local_alloc *la;
1156 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
1157 la = OCFS2_LOCAL_ALLOC(alloc);
1159 trace_ocfs2_local_alloc_new_window(
1160 le32_to_cpu(alloc->id1.bitmap1.i_total),
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_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 trace_ocfs2_local_alloc_new_window_result(
1224 OCFS2_LOCAL_ALLOC(alloc)->la_bm_off,
1225 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;
1246 ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_SLIDE);
1248 /* This will lock the main bitmap for us. */
1249 status = ocfs2_local_alloc_reserve_for_window(osb,
1254 if (status != -ENOSPC)
1259 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
1260 if (IS_ERR(handle)) {
1261 status = PTR_ERR(handle);
1267 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
1269 /* We want to clear the local alloc before doing anything
1270 * else, so that if we error later during this operation,
1271 * local alloc shutdown won't try to double free main bitmap
1272 * bits. Make a copy so the sync function knows which bits to
1274 alloc_copy = kmemdup(alloc, osb->local_alloc_bh->b_size, GFP_NOFS);
1281 status = ocfs2_journal_access_di(handle,
1282 INODE_CACHE(local_alloc_inode),
1283 osb->local_alloc_bh,
1284 OCFS2_JOURNAL_ACCESS_WRITE);
1290 ocfs2_clear_local_alloc(alloc);
1291 ocfs2_journal_dirty(handle, osb->local_alloc_bh);
1293 status = ocfs2_sync_local_to_main(osb, handle, alloc_copy,
1294 main_bm_inode, main_bm_bh);
1300 status = ocfs2_local_alloc_new_window(osb, handle, ac);
1302 if (status != -ENOSPC)
1307 atomic_inc(&osb->alloc_stats.moves);
1311 ocfs2_commit_trans(osb, handle);
1315 iput(main_bm_inode);
1319 ocfs2_free_alloc_context(ac);