1 // SPDX-License-Identifier: GPL-2.0+
3 * NILFS B-tree node cache
5 * Copyright (C) 2005-2008 Nippon Telegraph and Telephone Corporation.
7 * Originally written by Seiji Kihara.
8 * Fully revised by Ryusuke Konishi for stabilization and simplification.
12 #include <linux/types.h>
13 #include <linux/buffer_head.h>
15 #include <linux/backing-dev.h>
16 #include <linux/gfp.h>
25 * nilfs_init_btnc_inode - initialize B-tree node cache inode
26 * @btnc_inode: inode to be initialized
28 * nilfs_init_btnc_inode() sets up an inode for B-tree node cache.
30 void nilfs_init_btnc_inode(struct inode *btnc_inode)
32 struct nilfs_inode_info *ii = NILFS_I(btnc_inode);
34 btnc_inode->i_mode = S_IFREG;
36 memset(&ii->i_bmap_data, 0, sizeof(struct nilfs_bmap));
37 mapping_set_gfp_mask(btnc_inode->i_mapping, GFP_NOFS);
38 btnc_inode->i_mapping->a_ops = &nilfs_buffer_cache_aops;
41 void nilfs_btnode_cache_clear(struct address_space *btnc)
43 invalidate_mapping_pages(btnc, 0, -1);
44 truncate_inode_pages(btnc, 0);
48 nilfs_btnode_create_block(struct address_space *btnc, __u64 blocknr)
50 struct inode *inode = btnc->host;
51 struct buffer_head *bh;
53 bh = nilfs_grab_buffer(inode, btnc, blocknr, BIT(BH_NILFS_Node));
55 return ERR_PTR(-ENOMEM);
57 if (unlikely(buffer_mapped(bh) || buffer_uptodate(bh) ||
60 * The block buffer at the specified new address was already
61 * in use. This can happen if it is a virtual block number
62 * and has been reallocated due to corruption of the bitmap
63 * used to manage its allocation state (if not, the buffer
64 * clearing of an abandoned b-tree node is missing somewhere).
66 nilfs_error(inode->i_sb,
67 "state inconsistency probably due to duplicate use of b-tree node block address %llu (ino=%lu)",
68 (unsigned long long)blocknr, inode->i_ino);
71 memset(bh->b_data, 0, i_blocksize(inode));
72 bh->b_blocknr = blocknr;
73 set_buffer_mapped(bh);
74 set_buffer_uptodate(bh);
76 folio_unlock(bh->b_folio);
77 folio_put(bh->b_folio);
81 folio_unlock(bh->b_folio);
82 folio_put(bh->b_folio);
87 int nilfs_btnode_submit_block(struct address_space *btnc, __u64 blocknr,
88 sector_t pblocknr, blk_opf_t opf,
89 struct buffer_head **pbh, sector_t *submit_ptr)
91 struct buffer_head *bh;
92 struct inode *inode = btnc->host;
96 bh = nilfs_grab_buffer(inode, btnc, blocknr, BIT(BH_NILFS_Node));
100 err = -EEXIST; /* internal code */
103 if (buffer_uptodate(bh) || buffer_dirty(bh))
108 if (inode->i_ino != NILFS_DAT_INO) {
109 struct the_nilfs *nilfs = inode->i_sb->s_fs_info;
111 /* blocknr is a virtual block number */
112 err = nilfs_dat_translate(nilfs->ns_dat, blocknr,
121 if (opf & REQ_RAHEAD) {
122 if (pblocknr != *submit_ptr + 1 || !trylock_buffer(bh)) {
123 err = -EBUSY; /* internal code */
127 } else { /* opf == REQ_OP_READ */
130 if (buffer_uptodate(bh)) {
132 err = -EEXIST; /* internal code */
135 set_buffer_mapped(bh);
136 bh->b_blocknr = pblocknr; /* set block address for read */
137 bh->b_end_io = end_buffer_read_sync;
140 bh->b_blocknr = blocknr; /* set back to the given block address */
141 *submit_ptr = pblocknr;
153 * nilfs_btnode_delete - delete B-tree node buffer
154 * @bh: buffer to be deleted
156 * nilfs_btnode_delete() invalidates the specified buffer and delete the page
157 * including the buffer if the page gets unbusy.
159 void nilfs_btnode_delete(struct buffer_head *bh)
161 struct address_space *mapping;
162 struct folio *folio = bh->b_folio;
163 pgoff_t index = folio->index;
168 folio_wait_writeback(folio);
170 nilfs_forget_buffer(bh);
171 still_dirty = folio_test_dirty(folio);
172 mapping = folio->mapping;
176 if (!still_dirty && mapping)
177 invalidate_inode_pages2_range(mapping, index, index);
181 * nilfs_btnode_prepare_change_key - prepare to change the search key of a
183 * @btnc: page cache in which the b-tree node block is buffered
184 * @ctxt: structure for exchanging context information for key change
186 * nilfs_btnode_prepare_change_key() prepares to move the contents of the
187 * b-tree node block of the old key given in the "oldkey" member of @ctxt to
188 * the position of the new key given in the "newkey" member of @ctxt in the
189 * page cache @btnc. Here, the key of the block is an index in units of
190 * blocks, and if the page and block sizes match, it matches the page index
193 * If the page size and block size match, this function attempts to move the
194 * entire folio, and in preparation for this, inserts the original folio into
195 * the new index of the cache. If this insertion fails or if the page size
196 * and block size are different, it falls back to a copy preparation using
197 * nilfs_btnode_create_block(), inserts a new block at the position
198 * corresponding to "newkey", and stores the buffer head pointer in the
199 * "newbh" member of @ctxt.
201 * Note that the current implementation does not support folio sizes larger
202 * than the page size.
204 * Return: 0 on success, or the following negative error code on failure.
205 * * %-EIO - I/O error (metadata corruption).
206 * * %-ENOMEM - Insufficient memory available.
208 int nilfs_btnode_prepare_change_key(struct address_space *btnc,
209 struct nilfs_btnode_chkey_ctxt *ctxt)
211 struct buffer_head *obh, *nbh;
212 struct inode *inode = btnc->host;
213 __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
216 if (oldkey == newkey)
222 if (inode->i_blkbits == PAGE_SHIFT) {
223 struct folio *ofolio = obh->b_folio;
226 /* BUG_ON(oldkey != obh->b_folio->index); */
227 if (unlikely(oldkey != ofolio->index))
228 NILFS_FOLIO_BUG(ofolio,
229 "invalid oldkey %lld (newkey=%lld)",
230 (unsigned long long)oldkey,
231 (unsigned long long)newkey);
233 xa_lock_irq(&btnc->i_pages);
234 err = __xa_insert(&btnc->i_pages, newkey, ofolio, GFP_NOFS);
235 xa_unlock_irq(&btnc->i_pages);
237 * Note: folio->index will not change to newkey until
238 * nilfs_btnode_commit_change_key() will be called.
239 * To protect the folio in intermediate state, the folio lock
244 else if (err != -EBUSY)
247 err = invalidate_inode_pages2_range(btnc, newkey, newkey);
250 /* fallback to copy mode */
251 folio_unlock(ofolio);
254 nbh = nilfs_btnode_create_block(btnc, newkey);
263 folio_unlock(obh->b_folio);
268 * nilfs_btnode_commit_change_key - commit the change of the search key of
269 * a b-tree node block
270 * @btnc: page cache in which the b-tree node block is buffered
271 * @ctxt: structure for exchanging context information for key change
273 * nilfs_btnode_commit_change_key() executes the key change based on the
274 * context @ctxt prepared by nilfs_btnode_prepare_change_key(). If no valid
275 * block buffer is prepared in "newbh" of @ctxt (i.e., a full folio move),
276 * this function removes the folio from the old index and completes the move.
277 * Otherwise, it copies the block data and inherited flag states of "oldbh"
278 * to "newbh" and clears the "oldbh" from the cache. In either case, the
279 * relocated buffer is marked as dirty.
281 * As with nilfs_btnode_prepare_change_key(), the current implementation does
282 * not support folio sizes larger than the page size.
284 void nilfs_btnode_commit_change_key(struct address_space *btnc,
285 struct nilfs_btnode_chkey_ctxt *ctxt)
287 struct buffer_head *obh = ctxt->bh, *nbh = ctxt->newbh;
288 __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
289 struct folio *ofolio;
291 if (oldkey == newkey)
294 if (nbh == NULL) { /* blocksize == pagesize */
295 ofolio = obh->b_folio;
296 if (unlikely(oldkey != ofolio->index))
297 NILFS_FOLIO_BUG(ofolio,
298 "invalid oldkey %lld (newkey=%lld)",
299 (unsigned long long)oldkey,
300 (unsigned long long)newkey);
301 mark_buffer_dirty(obh);
303 xa_lock_irq(&btnc->i_pages);
304 __xa_erase(&btnc->i_pages, oldkey);
305 __xa_set_mark(&btnc->i_pages, newkey, PAGECACHE_TAG_DIRTY);
306 xa_unlock_irq(&btnc->i_pages);
308 ofolio->index = obh->b_blocknr = newkey;
309 folio_unlock(ofolio);
311 nilfs_copy_buffer(nbh, obh);
312 mark_buffer_dirty(nbh);
314 nbh->b_blocknr = newkey;
316 nilfs_btnode_delete(obh); /* will decrement bh->b_count */
321 * nilfs_btnode_abort_change_key - abort the change of the search key of a
323 * @btnc: page cache in which the b-tree node block is buffered
324 * @ctxt: structure for exchanging context information for key change
326 * nilfs_btnode_abort_change_key() cancels the key change associated with the
327 * context @ctxt prepared via nilfs_btnode_prepare_change_key() and performs
328 * any necessary cleanup. If no valid block buffer is prepared in "newbh" of
329 * @ctxt, this function removes the folio from the destination index and aborts
330 * the move. Otherwise, it clears "newbh" from the cache.
332 * As with nilfs_btnode_prepare_change_key(), the current implementation does
333 * not support folio sizes larger than the page size.
335 void nilfs_btnode_abort_change_key(struct address_space *btnc,
336 struct nilfs_btnode_chkey_ctxt *ctxt)
338 struct buffer_head *nbh = ctxt->newbh;
339 __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
341 if (oldkey == newkey)
344 if (nbh == NULL) { /* blocksize == pagesize */
345 xa_erase_irq(&btnc->i_pages, newkey);
346 folio_unlock(ctxt->bh->b_folio);
349 * When canceling a buffer that a prepare operation has
350 * allocated to copy a node block to another location, use
351 * nilfs_btnode_delete() to initialize and release the buffer
352 * so that the buffer flags will not be in an inconsistent
353 * state when it is reallocated.
355 nilfs_btnode_delete(nbh);