1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * Block/Cluster mapping functions
8 * Copyright (C) 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, version 2, as published by the Free Software Foundation.
14 * This program is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * General Public License for more details.
19 * You should have received a copy of the GNU General Public
20 * License along with this program; if not, write to the
21 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
22 * Boston, MA 021110-1307, USA.
26 #include <linux/init.h>
27 #include <linux/types.h>
28 #include <linux/fiemap.h>
30 #define MLOG_MASK_PREFIX ML_EXTENT_MAP
31 #include <cluster/masklog.h>
37 #include "extent_map.h"
42 #include "buffer_head_io.h"
45 * The extent caching implementation is intentionally trivial.
47 * We only cache a small number of extents stored directly on the
48 * inode, so linear order operations are acceptable. If we ever want
49 * to increase the size of the extent map, then these algorithms must
53 void ocfs2_extent_map_init(struct inode *inode)
55 struct ocfs2_inode_info *oi = OCFS2_I(inode);
57 oi->ip_extent_map.em_num_items = 0;
58 INIT_LIST_HEAD(&oi->ip_extent_map.em_list);
61 static void __ocfs2_extent_map_lookup(struct ocfs2_extent_map *em,
63 struct ocfs2_extent_map_item **ret_emi)
66 struct ocfs2_extent_map_item *emi;
70 list_for_each_entry(emi, &em->em_list, ei_list) {
71 range = emi->ei_cpos + emi->ei_clusters;
73 if (cpos >= emi->ei_cpos && cpos < range) {
74 list_move(&emi->ei_list, &em->em_list);
82 static int ocfs2_extent_map_lookup(struct inode *inode, unsigned int cpos,
83 unsigned int *phys, unsigned int *len,
87 struct ocfs2_inode_info *oi = OCFS2_I(inode);
88 struct ocfs2_extent_map_item *emi;
90 spin_lock(&oi->ip_lock);
92 __ocfs2_extent_map_lookup(&oi->ip_extent_map, cpos, &emi);
94 coff = cpos - emi->ei_cpos;
95 *phys = emi->ei_phys + coff;
97 *len = emi->ei_clusters - coff;
99 *flags = emi->ei_flags;
102 spin_unlock(&oi->ip_lock);
111 * Forget about all clusters equal to or greater than cpos.
113 void ocfs2_extent_map_trunc(struct inode *inode, unsigned int cpos)
115 struct ocfs2_extent_map_item *emi, *n;
116 struct ocfs2_inode_info *oi = OCFS2_I(inode);
117 struct ocfs2_extent_map *em = &oi->ip_extent_map;
121 spin_lock(&oi->ip_lock);
122 list_for_each_entry_safe(emi, n, &em->em_list, ei_list) {
123 if (emi->ei_cpos >= cpos) {
124 /* Full truncate of this record. */
125 list_move(&emi->ei_list, &tmp_list);
126 BUG_ON(em->em_num_items == 0);
131 range = emi->ei_cpos + emi->ei_clusters;
133 /* Partial truncate */
134 emi->ei_clusters = cpos - emi->ei_cpos;
137 spin_unlock(&oi->ip_lock);
139 list_for_each_entry_safe(emi, n, &tmp_list, ei_list) {
140 list_del(&emi->ei_list);
146 * Is any part of emi2 contained within emi1
148 static int ocfs2_ei_is_contained(struct ocfs2_extent_map_item *emi1,
149 struct ocfs2_extent_map_item *emi2)
151 unsigned int range1, range2;
154 * Check if logical start of emi2 is inside emi1
156 range1 = emi1->ei_cpos + emi1->ei_clusters;
157 if (emi2->ei_cpos >= emi1->ei_cpos && emi2->ei_cpos < range1)
161 * Check if logical end of emi2 is inside emi1
163 range2 = emi2->ei_cpos + emi2->ei_clusters;
164 if (range2 > emi1->ei_cpos && range2 <= range1)
170 static void ocfs2_copy_emi_fields(struct ocfs2_extent_map_item *dest,
171 struct ocfs2_extent_map_item *src)
173 dest->ei_cpos = src->ei_cpos;
174 dest->ei_phys = src->ei_phys;
175 dest->ei_clusters = src->ei_clusters;
176 dest->ei_flags = src->ei_flags;
180 * Try to merge emi with ins. Returns 1 if merge succeeds, zero
183 static int ocfs2_try_to_merge_extent_map(struct ocfs2_extent_map_item *emi,
184 struct ocfs2_extent_map_item *ins)
187 * Handle contiguousness
189 if (ins->ei_phys == (emi->ei_phys + emi->ei_clusters) &&
190 ins->ei_cpos == (emi->ei_cpos + emi->ei_clusters) &&
191 ins->ei_flags == emi->ei_flags) {
192 emi->ei_clusters += ins->ei_clusters;
194 } else if ((ins->ei_phys + ins->ei_clusters) == emi->ei_phys &&
195 (ins->ei_cpos + ins->ei_clusters) == emi->ei_phys &&
196 ins->ei_flags == emi->ei_flags) {
197 emi->ei_phys = ins->ei_phys;
198 emi->ei_cpos = ins->ei_cpos;
199 emi->ei_clusters += ins->ei_clusters;
204 * Overlapping extents - this shouldn't happen unless we've
205 * split an extent to change it's flags. That is exceedingly
206 * rare, so there's no sense in trying to optimize it yet.
208 if (ocfs2_ei_is_contained(emi, ins) ||
209 ocfs2_ei_is_contained(ins, emi)) {
210 ocfs2_copy_emi_fields(emi, ins);
214 /* No merge was possible. */
219 * In order to reduce complexity on the caller, this insert function
220 * is intentionally liberal in what it will accept.
222 * The only rule is that the truncate call *must* be used whenever
223 * records have been deleted. This avoids inserting overlapping
224 * records with different physical mappings.
226 void ocfs2_extent_map_insert_rec(struct inode *inode,
227 struct ocfs2_extent_rec *rec)
229 struct ocfs2_inode_info *oi = OCFS2_I(inode);
230 struct ocfs2_extent_map *em = &oi->ip_extent_map;
231 struct ocfs2_extent_map_item *emi, *new_emi = NULL;
232 struct ocfs2_extent_map_item ins;
234 ins.ei_cpos = le32_to_cpu(rec->e_cpos);
235 ins.ei_phys = ocfs2_blocks_to_clusters(inode->i_sb,
236 le64_to_cpu(rec->e_blkno));
237 ins.ei_clusters = le16_to_cpu(rec->e_leaf_clusters);
238 ins.ei_flags = rec->e_flags;
241 spin_lock(&oi->ip_lock);
243 list_for_each_entry(emi, &em->em_list, ei_list) {
244 if (ocfs2_try_to_merge_extent_map(emi, &ins)) {
245 list_move(&emi->ei_list, &em->em_list);
246 spin_unlock(&oi->ip_lock);
252 * No item could be merged.
254 * Either allocate and add a new item, or overwrite the last recently
258 if (em->em_num_items < OCFS2_MAX_EXTENT_MAP_ITEMS) {
259 if (new_emi == NULL) {
260 spin_unlock(&oi->ip_lock);
262 new_emi = kmalloc(sizeof(*new_emi), GFP_NOFS);
269 ocfs2_copy_emi_fields(new_emi, &ins);
270 list_add(&new_emi->ei_list, &em->em_list);
274 BUG_ON(list_empty(&em->em_list) || em->em_num_items == 0);
275 emi = list_entry(em->em_list.prev,
276 struct ocfs2_extent_map_item, ei_list);
277 list_move(&emi->ei_list, &em->em_list);
278 ocfs2_copy_emi_fields(emi, &ins);
281 spin_unlock(&oi->ip_lock);
288 static int ocfs2_last_eb_is_empty(struct inode *inode,
289 struct ocfs2_dinode *di)
292 u64 last_eb_blk = le64_to_cpu(di->i_last_eb_blk);
293 struct buffer_head *eb_bh = NULL;
294 struct ocfs2_extent_block *eb;
295 struct ocfs2_extent_list *el;
297 ret = ocfs2_read_extent_block(INODE_CACHE(inode), last_eb_blk, &eb_bh);
303 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
306 if (el->l_tree_depth) {
307 ocfs2_error(inode->i_sb,
308 "Inode %lu has non zero tree depth in "
309 "leaf block %llu\n", inode->i_ino,
310 (unsigned long long)eb_bh->b_blocknr);
315 next_free = le16_to_cpu(el->l_next_free_rec);
317 if (next_free == 0 ||
318 (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0])))
327 * Return the 1st index within el which contains an extent start
328 * larger than v_cluster.
330 static int ocfs2_search_for_hole_index(struct ocfs2_extent_list *el,
334 struct ocfs2_extent_rec *rec;
336 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
337 rec = &el->l_recs[i];
339 if (v_cluster < le32_to_cpu(rec->e_cpos))
347 * Figure out the size of a hole which starts at v_cluster within the given
350 * If there is no more allocation past v_cluster, we return the maximum
351 * cluster size minus v_cluster.
353 * If we have in-inode extents, then el points to the dinode list and
354 * eb_bh is NULL. Otherwise, eb_bh should point to the extent block
357 int ocfs2_figure_hole_clusters(struct ocfs2_caching_info *ci,
358 struct ocfs2_extent_list *el,
359 struct buffer_head *eb_bh,
364 struct buffer_head *next_eb_bh = NULL;
365 struct ocfs2_extent_block *eb, *next_eb;
367 i = ocfs2_search_for_hole_index(el, v_cluster);
369 if (i == le16_to_cpu(el->l_next_free_rec) && eb_bh) {
370 eb = (struct ocfs2_extent_block *)eb_bh->b_data;
373 * Check the next leaf for any extents.
376 if (le64_to_cpu(eb->h_next_leaf_blk) == 0ULL)
377 goto no_more_extents;
379 ret = ocfs2_read_extent_block(ci,
380 le64_to_cpu(eb->h_next_leaf_blk),
387 next_eb = (struct ocfs2_extent_block *)next_eb_bh->b_data;
388 el = &next_eb->h_list;
389 i = ocfs2_search_for_hole_index(el, v_cluster);
393 if (i == le16_to_cpu(el->l_next_free_rec)) {
395 * We're at the end of our existing allocation. Just
396 * return the maximum number of clusters we could
399 *num_clusters = UINT_MAX - v_cluster;
401 *num_clusters = le32_to_cpu(el->l_recs[i].e_cpos) - v_cluster;
410 static int ocfs2_get_clusters_nocache(struct inode *inode,
411 struct buffer_head *di_bh,
412 u32 v_cluster, unsigned int *hole_len,
413 struct ocfs2_extent_rec *ret_rec,
414 unsigned int *is_last)
416 int i, ret, tree_height, len;
417 struct ocfs2_dinode *di;
418 struct ocfs2_extent_block *uninitialized_var(eb);
419 struct ocfs2_extent_list *el;
420 struct ocfs2_extent_rec *rec;
421 struct buffer_head *eb_bh = NULL;
423 memset(ret_rec, 0, sizeof(*ret_rec));
427 di = (struct ocfs2_dinode *) di_bh->b_data;
428 el = &di->id2.i_list;
429 tree_height = le16_to_cpu(el->l_tree_depth);
431 if (tree_height > 0) {
432 ret = ocfs2_find_leaf(INODE_CACHE(inode), el, v_cluster,
439 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
442 if (el->l_tree_depth) {
443 ocfs2_error(inode->i_sb,
444 "Inode %lu has non zero tree depth in "
445 "leaf block %llu\n", inode->i_ino,
446 (unsigned long long)eb_bh->b_blocknr);
452 i = ocfs2_search_extent_list(el, v_cluster);
455 * Holes can be larger than the maximum size of an
456 * extent, so we return their lengths in a seperate
460 ret = ocfs2_figure_hole_clusters(INODE_CACHE(inode),
473 rec = &el->l_recs[i];
475 BUG_ON(v_cluster < le32_to_cpu(rec->e_cpos));
478 ocfs2_error(inode->i_sb, "Inode %lu has bad extent "
479 "record (%u, %u, 0)", inode->i_ino,
480 le32_to_cpu(rec->e_cpos),
481 ocfs2_rec_clusters(el, rec));
489 * Checking for last extent is potentially expensive - we
490 * might have to look at the next leaf over to see if it's
493 * The first two checks are to see whether the caller even
494 * cares for this information, and if the extent is at least
495 * the last in it's list.
497 * If those hold true, then the extent is last if any of the
498 * additional conditions hold true:
499 * - Extent list is in-inode
500 * - Extent list is right-most
501 * - Extent list is 2nd to rightmost, with empty right-most
504 if (i == (le16_to_cpu(el->l_next_free_rec) - 1)) {
505 if (tree_height == 0)
507 else if (eb->h_blkno == di->i_last_eb_blk)
509 else if (eb->h_next_leaf_blk == di->i_last_eb_blk) {
510 ret = ocfs2_last_eb_is_empty(inode, di);
528 static void ocfs2_relative_extent_offsets(struct super_block *sb,
530 struct ocfs2_extent_rec *rec,
531 u32 *p_cluster, u32 *num_clusters)
534 u32 coff = v_cluster - le32_to_cpu(rec->e_cpos);
536 *p_cluster = ocfs2_blocks_to_clusters(sb, le64_to_cpu(rec->e_blkno));
537 *p_cluster = *p_cluster + coff;
540 *num_clusters = le16_to_cpu(rec->e_leaf_clusters) - coff;
543 int ocfs2_xattr_get_clusters(struct inode *inode, u32 v_cluster,
544 u32 *p_cluster, u32 *num_clusters,
545 struct ocfs2_extent_list *el,
546 unsigned int *extent_flags)
549 struct buffer_head *eb_bh = NULL;
550 struct ocfs2_extent_block *eb;
551 struct ocfs2_extent_rec *rec;
554 if (el->l_tree_depth) {
555 ret = ocfs2_find_leaf(INODE_CACHE(inode), el, v_cluster,
562 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
565 if (el->l_tree_depth) {
566 ocfs2_error(inode->i_sb,
567 "Inode %lu has non zero tree depth in "
568 "xattr leaf block %llu\n", inode->i_ino,
569 (unsigned long long)eb_bh->b_blocknr);
575 i = ocfs2_search_extent_list(el, v_cluster);
581 rec = &el->l_recs[i];
582 BUG_ON(v_cluster < le32_to_cpu(rec->e_cpos));
585 ocfs2_error(inode->i_sb, "Inode %lu has bad extent "
586 "record (%u, %u, 0) in xattr", inode->i_ino,
587 le32_to_cpu(rec->e_cpos),
588 ocfs2_rec_clusters(el, rec));
592 coff = v_cluster - le32_to_cpu(rec->e_cpos);
593 *p_cluster = ocfs2_blocks_to_clusters(inode->i_sb,
594 le64_to_cpu(rec->e_blkno));
595 *p_cluster = *p_cluster + coff;
597 *num_clusters = ocfs2_rec_clusters(el, rec) - coff;
600 *extent_flags = rec->e_flags;
608 int ocfs2_get_clusters(struct inode *inode, u32 v_cluster,
609 u32 *p_cluster, u32 *num_clusters,
610 unsigned int *extent_flags)
613 unsigned int uninitialized_var(hole_len), flags = 0;
614 struct buffer_head *di_bh = NULL;
615 struct ocfs2_extent_rec rec;
617 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
623 ret = ocfs2_extent_map_lookup(inode, v_cluster, p_cluster,
624 num_clusters, extent_flags);
628 ret = ocfs2_read_inode_block(inode, &di_bh);
634 ret = ocfs2_get_clusters_nocache(inode, di_bh, v_cluster, &hole_len,
641 if (rec.e_blkno == 0ULL) {
643 * A hole was found. Return some canned values that
644 * callers can key on. If asked for, num_clusters will
645 * be populated with the size of the hole.
649 *num_clusters = hole_len;
652 ocfs2_relative_extent_offsets(inode->i_sb, v_cluster, &rec,
653 p_cluster, num_clusters);
656 ocfs2_extent_map_insert_rec(inode, &rec);
660 *extent_flags = flags;
668 * This expects alloc_sem to be held. The allocation cannot change at
669 * all while the map is in the process of being updated.
671 int ocfs2_extent_map_get_blocks(struct inode *inode, u64 v_blkno, u64 *p_blkno,
672 u64 *ret_count, unsigned int *extent_flags)
675 int bpc = ocfs2_clusters_to_blocks(inode->i_sb, 1);
676 u32 cpos, num_clusters, p_cluster;
679 cpos = ocfs2_blocks_to_clusters(inode->i_sb, v_blkno);
681 ret = ocfs2_get_clusters(inode, cpos, &p_cluster, &num_clusters,
689 * p_cluster == 0 indicates a hole.
692 boff = ocfs2_clusters_to_blocks(inode->i_sb, p_cluster);
693 boff += (v_blkno & (u64)(bpc - 1));
699 *ret_count = ocfs2_clusters_to_blocks(inode->i_sb, num_clusters);
700 *ret_count -= v_blkno & (u64)(bpc - 1);
708 * The ocfs2_fiemap_inline() may be a little bit misleading, since
709 * it not only handles the fiemap for inlined files, but also deals
710 * with the fast symlink, cause they have no difference for extent
713 static int ocfs2_fiemap_inline(struct inode *inode, struct buffer_head *di_bh,
714 struct fiemap_extent_info *fieinfo,
718 unsigned int id_count;
719 struct ocfs2_dinode *di;
721 u32 flags = FIEMAP_EXTENT_DATA_INLINE|FIEMAP_EXTENT_LAST;
722 struct ocfs2_inode_info *oi = OCFS2_I(inode);
724 di = (struct ocfs2_dinode *)di_bh->b_data;
725 if (ocfs2_inode_is_fast_symlink(inode))
726 id_count = ocfs2_fast_symlink_chars(inode->i_sb);
728 id_count = le16_to_cpu(di->id2.i_data.id_count);
730 if (map_start < id_count) {
731 phys = oi->ip_blkno << inode->i_sb->s_blocksize_bits;
732 if (ocfs2_inode_is_fast_symlink(inode))
733 phys += offsetof(struct ocfs2_dinode, id2.i_symlink);
735 phys += offsetof(struct ocfs2_dinode,
738 ret = fiemap_fill_next_extent(fieinfo, 0, phys, id_count,
747 #define OCFS2_FIEMAP_FLAGS (FIEMAP_FLAG_SYNC)
749 int ocfs2_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
750 u64 map_start, u64 map_len)
753 u32 mapping_end, cpos;
754 unsigned int hole_size;
755 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
756 u64 len_bytes, phys_bytes, virt_bytes;
757 struct buffer_head *di_bh = NULL;
758 struct ocfs2_extent_rec rec;
760 ret = fiemap_check_flags(fieinfo, OCFS2_FIEMAP_FLAGS);
764 ret = ocfs2_inode_lock(inode, &di_bh, 0);
770 down_read(&OCFS2_I(inode)->ip_alloc_sem);
773 * Handle inline-data and fast symlink separately.
775 if ((OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
776 ocfs2_inode_is_fast_symlink(inode)) {
777 ret = ocfs2_fiemap_inline(inode, di_bh, fieinfo, map_start);
781 cpos = map_start >> osb->s_clustersize_bits;
782 mapping_end = ocfs2_clusters_for_bytes(inode->i_sb,
783 map_start + map_len);
786 while (cpos < mapping_end && !is_last) {
789 ret = ocfs2_get_clusters_nocache(inode, di_bh, cpos,
790 &hole_size, &rec, &is_last);
796 if (rec.e_blkno == 0ULL) {
802 if (rec.e_flags & OCFS2_EXT_UNWRITTEN)
803 fe_flags |= FIEMAP_EXTENT_UNWRITTEN;
804 if (rec.e_flags & OCFS2_EXT_REFCOUNTED)
805 fe_flags |= FIEMAP_EXTENT_SHARED;
807 fe_flags |= FIEMAP_EXTENT_LAST;
808 len_bytes = (u64)le16_to_cpu(rec.e_leaf_clusters) << osb->s_clustersize_bits;
809 phys_bytes = le64_to_cpu(rec.e_blkno) << osb->sb->s_blocksize_bits;
810 virt_bytes = (u64)le32_to_cpu(rec.e_cpos) << osb->s_clustersize_bits;
812 ret = fiemap_fill_next_extent(fieinfo, virt_bytes, phys_bytes,
813 len_bytes, fe_flags);
817 cpos = le32_to_cpu(rec.e_cpos)+ le16_to_cpu(rec.e_leaf_clusters);
826 up_read(&OCFS2_I(inode)->ip_alloc_sem);
828 ocfs2_inode_unlock(inode, 0);
834 int ocfs2_read_virt_blocks(struct inode *inode, u64 v_block, int nr,
835 struct buffer_head *bhs[], int flags,
836 int (*validate)(struct super_block *sb,
837 struct buffer_head *bh))
840 u64 p_block, p_count;
841 int i, count, done = 0;
843 mlog_entry("(inode = %p, v_block = %llu, nr = %d, bhs = %p, "
844 "flags = %x, validate = %p)\n",
845 inode, (unsigned long long)v_block, nr, bhs, flags,
848 if (((v_block + nr - 1) << inode->i_sb->s_blocksize_bits) >=
849 i_size_read(inode)) {
850 BUG_ON(!(flags & OCFS2_BH_READAHEAD));
855 down_read(&OCFS2_I(inode)->ip_alloc_sem);
856 rc = ocfs2_extent_map_get_blocks(inode, v_block + done,
857 &p_block, &p_count, NULL);
858 up_read(&OCFS2_I(inode)->ip_alloc_sem);
867 "Inode #%llu contains a hole at offset %llu\n",
868 (unsigned long long)OCFS2_I(inode)->ip_blkno,
869 (unsigned long long)(v_block + done) <<
870 inode->i_sb->s_blocksize_bits);
879 * If the caller passed us bhs, they should have come
880 * from a previous readahead call to this function. Thus,
881 * they should have the right b_blocknr.
883 for (i = 0; i < count; i++) {
886 BUG_ON(bhs[done + i]->b_blocknr != (p_block + i));
889 rc = ocfs2_read_blocks(INODE_CACHE(inode), p_block, count,
890 bhs + done, flags, validate);