1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3 * NILFS B-tree node cache
4 *
5 * Copyright (C) 2005-2008 Nippon Telegraph and Telephone Corporation.
6 *
7 * Originally written by Seiji Kihara.
8 * Fully revised by Ryusuke Konishi for stabilization and simplification.
9 *
10 */
11
12 #include <linux/types.h>
13 #include <linux/buffer_head.h>
14 #include <linux/mm.h>
15 #include <linux/backing-dev.h>
16 #include <linux/gfp.h>
17 #include "nilfs.h"
18 #include "mdt.h"
19 #include "dat.h"
20 #include "page.h"
21 #include "btnode.h"
22
23
24 /**
25 * nilfs_init_btnc_inode - initialize B-tree node cache inode
26 * @btnc_inode: inode to be initialized
27 *
28 * nilfs_init_btnc_inode() sets up an inode for B-tree node cache.
29 */
nilfs_init_btnc_inode(struct inode * btnc_inode)30 void nilfs_init_btnc_inode(struct inode *btnc_inode)
31 {
32 struct nilfs_inode_info *ii = NILFS_I(btnc_inode);
33
34 btnc_inode->i_mode = S_IFREG;
35 ii->i_flags = 0;
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;
39 }
40
nilfs_btnode_cache_clear(struct address_space * btnc)41 void nilfs_btnode_cache_clear(struct address_space *btnc)
42 {
43 invalidate_mapping_pages(btnc, 0, -1);
44 truncate_inode_pages(btnc, 0);
45 }
46
47 struct buffer_head *
nilfs_btnode_create_block(struct address_space * btnc,__u64 blocknr)48 nilfs_btnode_create_block(struct address_space *btnc, __u64 blocknr)
49 {
50 struct inode *inode = btnc->host;
51 struct buffer_head *bh;
52
53 bh = nilfs_grab_buffer(inode, btnc, blocknr, BIT(BH_NILFS_Node));
54 if (unlikely(!bh))
55 return ERR_PTR(-ENOMEM);
56
57 if (unlikely(buffer_mapped(bh) || buffer_uptodate(bh) ||
58 buffer_dirty(bh))) {
59 /*
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).
65 */
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);
69 goto failed;
70 }
71 memset(bh->b_data, 0, i_blocksize(inode));
72 bh->b_blocknr = blocknr;
73 set_buffer_mapped(bh);
74 set_buffer_uptodate(bh);
75
76 unlock_page(bh->b_page);
77 put_page(bh->b_page);
78 return bh;
79
80 failed:
81 unlock_page(bh->b_page);
82 put_page(bh->b_page);
83 brelse(bh);
84 return ERR_PTR(-EIO);
85 }
86
nilfs_btnode_submit_block(struct address_space * btnc,__u64 blocknr,sector_t pblocknr,blk_opf_t opf,struct buffer_head ** pbh,sector_t * submit_ptr)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)
90 {
91 struct buffer_head *bh;
92 struct inode *inode = btnc->host;
93 struct page *page;
94 int err;
95
96 bh = nilfs_grab_buffer(inode, btnc, blocknr, BIT(BH_NILFS_Node));
97 if (unlikely(!bh))
98 return -ENOMEM;
99
100 err = -EEXIST; /* internal code */
101 page = bh->b_page;
102
103 if (buffer_uptodate(bh) || buffer_dirty(bh))
104 goto found;
105
106 if (pblocknr == 0) {
107 pblocknr = blocknr;
108 if (inode->i_ino != NILFS_DAT_INO) {
109 struct the_nilfs *nilfs = inode->i_sb->s_fs_info;
110
111 /* blocknr is a virtual block number */
112 err = nilfs_dat_translate(nilfs->ns_dat, blocknr,
113 &pblocknr);
114 if (unlikely(err)) {
115 brelse(bh);
116 goto out_locked;
117 }
118 }
119 }
120
121 if (opf & REQ_RAHEAD) {
122 if (pblocknr != *submit_ptr + 1 || !trylock_buffer(bh)) {
123 err = -EBUSY; /* internal code */
124 brelse(bh);
125 goto out_locked;
126 }
127 } else { /* opf == REQ_OP_READ */
128 lock_buffer(bh);
129 }
130 if (buffer_uptodate(bh)) {
131 unlock_buffer(bh);
132 err = -EEXIST; /* internal code */
133 goto found;
134 }
135 set_buffer_mapped(bh);
136 bh->b_blocknr = pblocknr; /* set block address for read */
137 bh->b_end_io = end_buffer_read_sync;
138 get_bh(bh);
139 submit_bh(opf, bh);
140 bh->b_blocknr = blocknr; /* set back to the given block address */
141 *submit_ptr = pblocknr;
142 err = 0;
143 found:
144 *pbh = bh;
145
146 out_locked:
147 unlock_page(page);
148 put_page(page);
149 return err;
150 }
151
152 /**
153 * nilfs_btnode_delete - delete B-tree node buffer
154 * @bh: buffer to be deleted
155 *
156 * nilfs_btnode_delete() invalidates the specified buffer and delete the page
157 * including the buffer if the page gets unbusy.
158 */
nilfs_btnode_delete(struct buffer_head * bh)159 void nilfs_btnode_delete(struct buffer_head *bh)
160 {
161 struct address_space *mapping;
162 struct page *page = bh->b_page;
163 pgoff_t index = page_index(page);
164 int still_dirty;
165
166 get_page(page);
167 lock_page(page);
168 wait_on_page_writeback(page);
169
170 nilfs_forget_buffer(bh);
171 still_dirty = PageDirty(page);
172 mapping = page->mapping;
173 unlock_page(page);
174 put_page(page);
175
176 if (!still_dirty && mapping)
177 invalidate_inode_pages2_range(mapping, index, index);
178 }
179
180 /**
181 * nilfs_btnode_prepare_change_key
182 * prepare to move contents of the block for old key to one of new key.
183 * the old buffer will not be removed, but might be reused for new buffer.
184 * it might return -ENOMEM because of memory allocation errors,
185 * and might return -EIO because of disk read errors.
186 */
nilfs_btnode_prepare_change_key(struct address_space * btnc,struct nilfs_btnode_chkey_ctxt * ctxt)187 int nilfs_btnode_prepare_change_key(struct address_space *btnc,
188 struct nilfs_btnode_chkey_ctxt *ctxt)
189 {
190 struct buffer_head *obh, *nbh;
191 struct inode *inode = btnc->host;
192 __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
193 int err;
194
195 if (oldkey == newkey)
196 return 0;
197
198 obh = ctxt->bh;
199 ctxt->newbh = NULL;
200
201 if (inode->i_blkbits == PAGE_SHIFT) {
202 struct page *opage = obh->b_page;
203 lock_page(opage);
204 retry:
205 /* BUG_ON(oldkey != obh->b_folio->index); */
206 if (unlikely(oldkey != opage->index))
207 NILFS_PAGE_BUG(opage,
208 "invalid oldkey %lld (newkey=%lld)",
209 (unsigned long long)oldkey,
210 (unsigned long long)newkey);
211
212 xa_lock_irq(&btnc->i_pages);
213 err = __xa_insert(&btnc->i_pages, newkey, opage, GFP_NOFS);
214 xa_unlock_irq(&btnc->i_pages);
215 /*
216 * Note: page->index will not change to newkey until
217 * nilfs_btnode_commit_change_key() will be called.
218 * To protect the page in intermediate state, the page lock
219 * is held.
220 */
221 if (!err)
222 return 0;
223 else if (err != -EBUSY)
224 goto failed_unlock;
225
226 err = invalidate_inode_pages2_range(btnc, newkey, newkey);
227 if (!err)
228 goto retry;
229 /* fallback to copy mode */
230 unlock_page(opage);
231 }
232
233 nbh = nilfs_btnode_create_block(btnc, newkey);
234 if (IS_ERR(nbh))
235 return PTR_ERR(nbh);
236
237 BUG_ON(nbh == obh);
238 ctxt->newbh = nbh;
239 return 0;
240
241 failed_unlock:
242 unlock_page(obh->b_page);
243 return err;
244 }
245
246 /**
247 * nilfs_btnode_commit_change_key
248 * commit the change_key operation prepared by prepare_change_key().
249 */
nilfs_btnode_commit_change_key(struct address_space * btnc,struct nilfs_btnode_chkey_ctxt * ctxt)250 void nilfs_btnode_commit_change_key(struct address_space *btnc,
251 struct nilfs_btnode_chkey_ctxt *ctxt)
252 {
253 struct buffer_head *obh = ctxt->bh, *nbh = ctxt->newbh;
254 __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
255 struct page *opage;
256
257 if (oldkey == newkey)
258 return;
259
260 if (nbh == NULL) { /* blocksize == pagesize */
261 opage = obh->b_page;
262 if (unlikely(oldkey != opage->index))
263 NILFS_PAGE_BUG(opage,
264 "invalid oldkey %lld (newkey=%lld)",
265 (unsigned long long)oldkey,
266 (unsigned long long)newkey);
267 mark_buffer_dirty(obh);
268
269 xa_lock_irq(&btnc->i_pages);
270 __xa_erase(&btnc->i_pages, oldkey);
271 __xa_set_mark(&btnc->i_pages, newkey, PAGECACHE_TAG_DIRTY);
272 xa_unlock_irq(&btnc->i_pages);
273
274 opage->index = obh->b_blocknr = newkey;
275 unlock_page(opage);
276 } else {
277 nilfs_copy_buffer(nbh, obh);
278 mark_buffer_dirty(nbh);
279
280 nbh->b_blocknr = newkey;
281 ctxt->bh = nbh;
282 nilfs_btnode_delete(obh); /* will decrement bh->b_count */
283 }
284 }
285
286 /**
287 * nilfs_btnode_abort_change_key
288 * abort the change_key operation prepared by prepare_change_key().
289 */
nilfs_btnode_abort_change_key(struct address_space * btnc,struct nilfs_btnode_chkey_ctxt * ctxt)290 void nilfs_btnode_abort_change_key(struct address_space *btnc,
291 struct nilfs_btnode_chkey_ctxt *ctxt)
292 {
293 struct buffer_head *nbh = ctxt->newbh;
294 __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
295
296 if (oldkey == newkey)
297 return;
298
299 if (nbh == NULL) { /* blocksize == pagesize */
300 xa_erase_irq(&btnc->i_pages, newkey);
301 unlock_page(ctxt->bh->b_page);
302 } else {
303 /*
304 * When canceling a buffer that a prepare operation has
305 * allocated to copy a node block to another location, use
306 * nilfs_btnode_delete() to initialize and release the buffer
307 * so that the buffer flags will not be in an inconsistent
308 * state when it is reallocated.
309 */
310 nilfs_btnode_delete(nbh);
311 }
312 }
313