btnode.c

  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 one of the following negative error codes on
205 * failure:
206 * * %-EIO	- I/O error (metadata corruption).
207 * * %-ENOMEM	- Insufficient memory available.
208 */
209int nilfs_btnode_prepare_change_key(struct address_space *btnc,
210				    struct nilfs_btnode_chkey_ctxt *ctxt)
211{
212	struct buffer_head *obh, *nbh;
213	struct inode *inode = btnc->host;
214	__u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
215	int err;
216
217	if (oldkey == newkey)
218		return 0;
219
220	obh = ctxt->bh;
221	ctxt->newbh = NULL;
222
223	if (inode->i_blkbits == PAGE_SHIFT) {
224		struct folio *ofolio = obh->b_folio;
225		folio_lock(ofolio);
226retry:
227		/* BUG_ON(oldkey != obh->b_folio->index); */
228		if (unlikely(oldkey != ofolio->index))
229			NILFS_FOLIO_BUG(ofolio,
230				       "invalid oldkey %lld (newkey=%lld)",
231				       (unsigned long long)oldkey,
232				       (unsigned long long)newkey);
233
234		xa_lock_irq(&btnc->i_pages);
235		err = __xa_insert(&btnc->i_pages, newkey, ofolio, GFP_NOFS);
236		xa_unlock_irq(&btnc->i_pages);
237		/*
238		 * Note: folio->index will not change to newkey until
239		 * nilfs_btnode_commit_change_key() will be called.
240		 * To protect the folio in intermediate state, the folio lock
241		 * is held.
242		 */
243		if (!err)
244			return 0;
245		else if (err != -EBUSY)
246			goto failed_unlock;
247
248		err = invalidate_inode_pages2_range(btnc, newkey, newkey);
249		if (!err)
250			goto retry;
251		/* fallback to copy mode */
252		folio_unlock(ofolio);
253	}
254
255	nbh = nilfs_btnode_create_block(btnc, newkey);
256	if (IS_ERR(nbh))
257		return PTR_ERR(nbh);
258
259	BUG_ON(nbh == obh);
260	ctxt->newbh = nbh;
261	return 0;
262
263 failed_unlock:
264	folio_unlock(obh->b_folio);
265	return err;
266}
267
268/**
269 * nilfs_btnode_commit_change_key - commit the change of the search key of
270 *                                  a b-tree node block
271 * @btnc: page cache in which the b-tree node block is buffered
272 * @ctxt: structure for exchanging context information for key change
273 *
274 * nilfs_btnode_commit_change_key() executes the key change based on the
275 * context @ctxt prepared by nilfs_btnode_prepare_change_key().  If no valid
276 * block buffer is prepared in "newbh" of @ctxt (i.e., a full folio move),
277 * this function removes the folio from the old index and completes the move.
278 * Otherwise, it copies the block data and inherited flag states of "oldbh"
279 * to "newbh" and clears the "oldbh" from the cache.  In either case, the
280 * relocated buffer is marked as dirty.
281 *
282 * As with nilfs_btnode_prepare_change_key(), the current implementation does
283 * not support folio sizes larger than the page size.
284 */
285void nilfs_btnode_commit_change_key(struct address_space *btnc,
286				    struct nilfs_btnode_chkey_ctxt *ctxt)
287{
288	struct buffer_head *obh = ctxt->bh, *nbh = ctxt->newbh;
289	__u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
290	struct folio *ofolio;
291
292	if (oldkey == newkey)
293		return;
294
295	if (nbh == NULL) {	/* blocksize == pagesize */
296		ofolio = obh->b_folio;
297		if (unlikely(oldkey != ofolio->index))
298			NILFS_FOLIO_BUG(ofolio,
299				       "invalid oldkey %lld (newkey=%lld)",
300				       (unsigned long long)oldkey,
301				       (unsigned long long)newkey);
302		mark_buffer_dirty(obh);
303
304		xa_lock_irq(&btnc->i_pages);
305		__xa_erase(&btnc->i_pages, oldkey);
306		__xa_set_mark(&btnc->i_pages, newkey, PAGECACHE_TAG_DIRTY);
307		xa_unlock_irq(&btnc->i_pages);
308
309		ofolio->index = obh->b_blocknr = newkey;
310		folio_unlock(ofolio);
311	} else {
312		nilfs_copy_buffer(nbh, obh);
313		mark_buffer_dirty(nbh);
314
315		nbh->b_blocknr = newkey;
316		ctxt->bh = nbh;
317		nilfs_btnode_delete(obh); /* will decrement bh->b_count */
318	}
319}
320
321/**
322 * nilfs_btnode_abort_change_key - abort the change of the search key of a
323 *                                 b-tree node block
324 * @btnc: page cache in which the b-tree node block is buffered
325 * @ctxt: structure for exchanging context information for key change
326 *
327 * nilfs_btnode_abort_change_key() cancels the key change associated with the
328 * context @ctxt prepared via nilfs_btnode_prepare_change_key() and performs
329 * any necessary cleanup.  If no valid block buffer is prepared in "newbh" of
330 * @ctxt, this function removes the folio from the destination index and aborts
331 * the move.  Otherwise, it clears "newbh" from the cache.
332 *
333 * As with nilfs_btnode_prepare_change_key(), the current implementation does
334 * not support folio sizes larger than the page size.
335 */
336void nilfs_btnode_abort_change_key(struct address_space *btnc,
337				   struct nilfs_btnode_chkey_ctxt *ctxt)
338{
339	struct buffer_head *nbh = ctxt->newbh;
340	__u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
341
342	if (oldkey == newkey)
343		return;
344
345	if (nbh == NULL) {	/* blocksize == pagesize */
346		xa_erase_irq(&btnc->i_pages, newkey);
347		folio_unlock(ctxt->bh->b_folio);
348	} else {
349		/*
350		 * When canceling a buffer that a prepare operation has
351		 * allocated to copy a node block to another location, use
352		 * nilfs_btnode_delete() to initialize and release the buffer
353		 * so that the buffer flags will not be in an inconsistent
354		 * state when it is reallocated.
355		 */
356		nilfs_btnode_delete(nbh);
357	}
358}