Loading...
1// SPDX-License-Identifier: GPL-2.0+
2/*
3 * btnode.c - 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
23void nilfs_btnode_cache_clear(struct address_space *btnc)
24{
25 invalidate_mapping_pages(btnc, 0, -1);
26 truncate_inode_pages(btnc, 0);
27}
28
29struct buffer_head *
30nilfs_btnode_create_block(struct address_space *btnc, __u64 blocknr)
31{
32 struct inode *inode = NILFS_BTNC_I(btnc);
33 struct buffer_head *bh;
34
35 bh = nilfs_grab_buffer(inode, btnc, blocknr, BIT(BH_NILFS_Node));
36 if (unlikely(!bh))
37 return NULL;
38
39 if (unlikely(buffer_mapped(bh) || buffer_uptodate(bh) ||
40 buffer_dirty(bh))) {
41 brelse(bh);
42 BUG();
43 }
44 memset(bh->b_data, 0, i_blocksize(inode));
45 bh->b_bdev = inode->i_sb->s_bdev;
46 bh->b_blocknr = blocknr;
47 set_buffer_mapped(bh);
48 set_buffer_uptodate(bh);
49
50 unlock_page(bh->b_page);
51 put_page(bh->b_page);
52 return bh;
53}
54
55int nilfs_btnode_submit_block(struct address_space *btnc, __u64 blocknr,
56 sector_t pblocknr, int mode, int mode_flags,
57 struct buffer_head **pbh, sector_t *submit_ptr)
58{
59 struct buffer_head *bh;
60 struct inode *inode = NILFS_BTNC_I(btnc);
61 struct page *page;
62 int err;
63
64 bh = nilfs_grab_buffer(inode, btnc, blocknr, BIT(BH_NILFS_Node));
65 if (unlikely(!bh))
66 return -ENOMEM;
67
68 err = -EEXIST; /* internal code */
69 page = bh->b_page;
70
71 if (buffer_uptodate(bh) || buffer_dirty(bh))
72 goto found;
73
74 if (pblocknr == 0) {
75 pblocknr = blocknr;
76 if (inode->i_ino != NILFS_DAT_INO) {
77 struct the_nilfs *nilfs = inode->i_sb->s_fs_info;
78
79 /* blocknr is a virtual block number */
80 err = nilfs_dat_translate(nilfs->ns_dat, blocknr,
81 &pblocknr);
82 if (unlikely(err)) {
83 brelse(bh);
84 goto out_locked;
85 }
86 }
87 }
88
89 if (mode_flags & REQ_RAHEAD) {
90 if (pblocknr != *submit_ptr + 1 || !trylock_buffer(bh)) {
91 err = -EBUSY; /* internal code */
92 brelse(bh);
93 goto out_locked;
94 }
95 } else { /* mode == READ */
96 lock_buffer(bh);
97 }
98 if (buffer_uptodate(bh)) {
99 unlock_buffer(bh);
100 err = -EEXIST; /* internal code */
101 goto found;
102 }
103 set_buffer_mapped(bh);
104 bh->b_bdev = inode->i_sb->s_bdev;
105 bh->b_blocknr = pblocknr; /* set block address for read */
106 bh->b_end_io = end_buffer_read_sync;
107 get_bh(bh);
108 submit_bh(mode, mode_flags, bh);
109 bh->b_blocknr = blocknr; /* set back to the given block address */
110 *submit_ptr = pblocknr;
111 err = 0;
112found:
113 *pbh = bh;
114
115out_locked:
116 unlock_page(page);
117 put_page(page);
118 return err;
119}
120
121/**
122 * nilfs_btnode_delete - delete B-tree node buffer
123 * @bh: buffer to be deleted
124 *
125 * nilfs_btnode_delete() invalidates the specified buffer and delete the page
126 * including the buffer if the page gets unbusy.
127 */
128void nilfs_btnode_delete(struct buffer_head *bh)
129{
130 struct address_space *mapping;
131 struct page *page = bh->b_page;
132 pgoff_t index = page_index(page);
133 int still_dirty;
134
135 get_page(page);
136 lock_page(page);
137 wait_on_page_writeback(page);
138
139 nilfs_forget_buffer(bh);
140 still_dirty = PageDirty(page);
141 mapping = page->mapping;
142 unlock_page(page);
143 put_page(page);
144
145 if (!still_dirty && mapping)
146 invalidate_inode_pages2_range(mapping, index, index);
147}
148
149/**
150 * nilfs_btnode_prepare_change_key
151 * prepare to move contents of the block for old key to one of new key.
152 * the old buffer will not be removed, but might be reused for new buffer.
153 * it might return -ENOMEM because of memory allocation errors,
154 * and might return -EIO because of disk read errors.
155 */
156int nilfs_btnode_prepare_change_key(struct address_space *btnc,
157 struct nilfs_btnode_chkey_ctxt *ctxt)
158{
159 struct buffer_head *obh, *nbh;
160 struct inode *inode = NILFS_BTNC_I(btnc);
161 __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
162 int err;
163
164 if (oldkey == newkey)
165 return 0;
166
167 obh = ctxt->bh;
168 ctxt->newbh = NULL;
169
170 if (inode->i_blkbits == PAGE_SHIFT) {
171 struct page *opage = obh->b_page;
172 lock_page(opage);
173retry:
174 /* BUG_ON(oldkey != obh->b_page->index); */
175 if (unlikely(oldkey != opage->index))
176 NILFS_PAGE_BUG(opage,
177 "invalid oldkey %lld (newkey=%lld)",
178 (unsigned long long)oldkey,
179 (unsigned long long)newkey);
180
181 xa_lock_irq(&btnc->i_pages);
182 err = __xa_insert(&btnc->i_pages, newkey, opage, GFP_NOFS);
183 xa_unlock_irq(&btnc->i_pages);
184 /*
185 * Note: page->index will not change to newkey until
186 * nilfs_btnode_commit_change_key() will be called.
187 * To protect the page in intermediate state, the page lock
188 * is held.
189 */
190 if (!err)
191 return 0;
192 else if (err != -EBUSY)
193 goto failed_unlock;
194
195 err = invalidate_inode_pages2_range(btnc, newkey, newkey);
196 if (!err)
197 goto retry;
198 /* fallback to copy mode */
199 unlock_page(opage);
200 }
201
202 nbh = nilfs_btnode_create_block(btnc, newkey);
203 if (!nbh)
204 return -ENOMEM;
205
206 BUG_ON(nbh == obh);
207 ctxt->newbh = nbh;
208 return 0;
209
210 failed_unlock:
211 unlock_page(obh->b_page);
212 return err;
213}
214
215/**
216 * nilfs_btnode_commit_change_key
217 * commit the change_key operation prepared by prepare_change_key().
218 */
219void nilfs_btnode_commit_change_key(struct address_space *btnc,
220 struct nilfs_btnode_chkey_ctxt *ctxt)
221{
222 struct buffer_head *obh = ctxt->bh, *nbh = ctxt->newbh;
223 __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
224 struct page *opage;
225
226 if (oldkey == newkey)
227 return;
228
229 if (nbh == NULL) { /* blocksize == pagesize */
230 opage = obh->b_page;
231 if (unlikely(oldkey != opage->index))
232 NILFS_PAGE_BUG(opage,
233 "invalid oldkey %lld (newkey=%lld)",
234 (unsigned long long)oldkey,
235 (unsigned long long)newkey);
236 mark_buffer_dirty(obh);
237
238 xa_lock_irq(&btnc->i_pages);
239 __xa_erase(&btnc->i_pages, oldkey);
240 __xa_set_mark(&btnc->i_pages, newkey, PAGECACHE_TAG_DIRTY);
241 xa_unlock_irq(&btnc->i_pages);
242
243 opage->index = obh->b_blocknr = newkey;
244 unlock_page(opage);
245 } else {
246 nilfs_copy_buffer(nbh, obh);
247 mark_buffer_dirty(nbh);
248
249 nbh->b_blocknr = newkey;
250 ctxt->bh = nbh;
251 nilfs_btnode_delete(obh); /* will decrement bh->b_count */
252 }
253}
254
255/**
256 * nilfs_btnode_abort_change_key
257 * abort the change_key operation prepared by prepare_change_key().
258 */
259void nilfs_btnode_abort_change_key(struct address_space *btnc,
260 struct nilfs_btnode_chkey_ctxt *ctxt)
261{
262 struct buffer_head *nbh = ctxt->newbh;
263 __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
264
265 if (oldkey == newkey)
266 return;
267
268 if (nbh == NULL) { /* blocksize == pagesize */
269 xa_erase_irq(&btnc->i_pages, newkey);
270 unlock_page(ctxt->bh->b_page);
271 } else
272 brelse(nbh);
273}
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 */
30void 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
41void nilfs_btnode_cache_clear(struct address_space *btnc)
42{
43 invalidate_mapping_pages(btnc, 0, -1);
44 truncate_inode_pages(btnc, 0);
45}
46
47struct buffer_head *
48nilfs_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 folio_unlock(bh->b_folio);
77 folio_put(bh->b_folio);
78 return bh;
79
80failed:
81 folio_unlock(bh->b_folio);
82 folio_put(bh->b_folio);
83 brelse(bh);
84 return ERR_PTR(-EIO);
85}
86
87int 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 folio *folio;
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 folio = bh->b_folio;
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;
143found:
144 *pbh = bh;
145
146out_locked:
147 folio_unlock(folio);
148 folio_put(folio);
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 */
159void nilfs_btnode_delete(struct buffer_head *bh)
160{
161 struct address_space *mapping;
162 struct folio *folio = bh->b_folio;
163 pgoff_t index = folio->index;
164 int still_dirty;
165
166 folio_get(folio);
167 folio_lock(folio);
168 folio_wait_writeback(folio);
169
170 nilfs_forget_buffer(bh);
171 still_dirty = folio_test_dirty(folio);
172 mapping = folio->mapping;
173 folio_unlock(folio);
174 folio_put(folio);
175
176 if (!still_dirty && mapping)
177 invalidate_inode_pages2_range(mapping, index, index);
178}
179
180/**
181 * nilfs_btnode_prepare_change_key - prepare to change the search key of a
182 * b-tree node block
183 * @btnc: page cache in which the b-tree node block is buffered
184 * @ctxt: structure for exchanging context information for key change
185 *
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
191 * in the page cache.
192 *
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.
200 *
201 * Note that the current implementation does not support folio sizes larger
202 * than the page size.
203 *
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.
207 */
208int nilfs_btnode_prepare_change_key(struct address_space *btnc,
209 struct nilfs_btnode_chkey_ctxt *ctxt)
210{
211 struct buffer_head *obh, *nbh;
212 struct inode *inode = btnc->host;
213 __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
214 int err;
215
216 if (oldkey == newkey)
217 return 0;
218
219 obh = ctxt->bh;
220 ctxt->newbh = NULL;
221
222 if (inode->i_blkbits == PAGE_SHIFT) {
223 struct folio *ofolio = obh->b_folio;
224 folio_lock(ofolio);
225retry:
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);
232
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);
236 /*
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
240 * is held.
241 */
242 if (!err)
243 return 0;
244 else if (err != -EBUSY)
245 goto failed_unlock;
246
247 err = invalidate_inode_pages2_range(btnc, newkey, newkey);
248 if (!err)
249 goto retry;
250 /* fallback to copy mode */
251 folio_unlock(ofolio);
252 }
253
254 nbh = nilfs_btnode_create_block(btnc, newkey);
255 if (IS_ERR(nbh))
256 return PTR_ERR(nbh);
257
258 BUG_ON(nbh == obh);
259 ctxt->newbh = nbh;
260 return 0;
261
262 failed_unlock:
263 folio_unlock(obh->b_folio);
264 return err;
265}
266
267/**
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
272 *
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.
280 *
281 * As with nilfs_btnode_prepare_change_key(), the current implementation does
282 * not support folio sizes larger than the page size.
283 */
284void nilfs_btnode_commit_change_key(struct address_space *btnc,
285 struct nilfs_btnode_chkey_ctxt *ctxt)
286{
287 struct buffer_head *obh = ctxt->bh, *nbh = ctxt->newbh;
288 __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
289 struct folio *ofolio;
290
291 if (oldkey == newkey)
292 return;
293
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);
302
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);
307
308 ofolio->index = obh->b_blocknr = newkey;
309 folio_unlock(ofolio);
310 } else {
311 nilfs_copy_buffer(nbh, obh);
312 mark_buffer_dirty(nbh);
313
314 nbh->b_blocknr = newkey;
315 ctxt->bh = nbh;
316 nilfs_btnode_delete(obh); /* will decrement bh->b_count */
317 }
318}
319
320/**
321 * nilfs_btnode_abort_change_key - abort the change of the search key of a
322 * b-tree node block
323 * @btnc: page cache in which the b-tree node block is buffered
324 * @ctxt: structure for exchanging context information for key change
325 *
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.
331 *
332 * As with nilfs_btnode_prepare_change_key(), the current implementation does
333 * not support folio sizes larger than the page size.
334 */
335void nilfs_btnode_abort_change_key(struct address_space *btnc,
336 struct nilfs_btnode_chkey_ctxt *ctxt)
337{
338 struct buffer_head *nbh = ctxt->newbh;
339 __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
340
341 if (oldkey == newkey)
342 return;
343
344 if (nbh == NULL) { /* blocksize == pagesize */
345 xa_erase_irq(&btnc->i_pages, newkey);
346 folio_unlock(ctxt->bh->b_folio);
347 } else {
348 /*
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.
354 */
355 nilfs_btnode_delete(nbh);
356 }
357}