1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Copyright (c) 2023-2024 Oracle. All Rights Reserved.
9 #include "xfs_buf_mem.h"
10 #include "xfs_trace.h"
11 #include <linux/shmem_fs.h>
12 #include "xfs_log_format.h"
13 #include "xfs_trans.h"
14 #include "xfs_buf_item.h"
15 #include "xfs_error.h"
18 * Buffer Cache for In-Memory Files
19 * ================================
21 * Online fsck wants to create ephemeral ordered recordsets. The existing
22 * btree infrastructure can do this, but we need the buffer cache to target
23 * memory instead of block devices.
25 * When CONFIG_TMPFS=y, shmemfs is enough of a filesystem to meet those
26 * requirements. Therefore, the xmbuf mechanism uses an unlinked shmem file to
27 * store our staging data. This file is not installed in the file descriptor
28 * table so that user programs cannot access the data, which means that the
29 * xmbuf must be freed with xmbuf_destroy.
31 * xmbufs assume that the caller will handle all required concurrency
32 * management; standard vfs locks (freezer and inode) are not taken. Reads
33 * and writes are satisfied directly from the page cache.
35 * The only supported block size is PAGE_SIZE, and we cannot use highmem.
39 * shmem files used to back an in-memory buffer cache must not be exposed to
40 * userspace. Upper layers must coordinate access to the one handle returned
41 * by the constructor, so establish a separate lock class for xmbufs to avoid
44 static struct lock_class_key xmbuf_i_mutex_key;
47 * Allocate a buffer cache target for a memory-backed file and set up the
54 struct xfs_buftarg **btpp)
58 struct xfs_buftarg *btp;
61 btp = kzalloc(struct_size(btp, bt_cache, 1), GFP_KERNEL);
65 file = shmem_kernel_file_setup(descr, 0, 0);
67 error = PTR_ERR(file);
70 inode = file_inode(file);
72 /* private file, private locking */
73 lockdep_set_class(&inode->i_rwsem, &xmbuf_i_mutex_key);
76 * We don't want to bother with kmapping data during repair, so don't
77 * allow highmem pages to back this mapping.
79 mapping_set_gfp_mask(inode->i_mapping, GFP_KERNEL);
81 /* ensure all writes are below EOF to avoid pagecache zeroing */
82 i_size_write(inode, inode->i_sb->s_maxbytes);
84 error = xfs_buf_cache_init(btp->bt_cache);
88 /* Initialize buffer target */
90 btp->bt_dev = (dev_t)-1U;
91 btp->bt_bdev = NULL; /* in-memory buftargs have no bdev */
93 btp->bt_meta_sectorsize = XMBUF_BLOCKSIZE;
94 btp->bt_meta_sectormask = XMBUF_BLOCKSIZE - 1;
96 error = xfs_init_buftarg(btp, XMBUF_BLOCKSIZE, descr);
100 trace_xmbuf_create(btp);
106 xfs_buf_cache_destroy(btp->bt_cache);
114 /* Free a buffer cache target for a memory-backed buffer cache. */
117 struct xfs_buftarg *btp)
119 ASSERT(xfs_buftarg_is_mem(btp));
120 ASSERT(percpu_counter_sum(&btp->bt_io_count) == 0);
122 trace_xmbuf_free(btp);
124 xfs_destroy_buftarg(btp);
125 xfs_buf_cache_destroy(btp->bt_cache);
130 /* Directly map a shmem page into the buffer cache. */
135 struct inode *inode = file_inode(bp->b_target->bt_file);
136 struct folio *folio = NULL;
138 loff_t pos = BBTOB(xfs_buf_daddr(bp));
141 ASSERT(xfs_buftarg_is_mem(bp->b_target));
143 if (bp->b_map_count != 1)
145 if (BBTOB(bp->b_length) != XMBUF_BLOCKSIZE)
147 if (offset_in_page(pos) != 0) {
148 ASSERT(offset_in_page(pos));
152 error = shmem_get_folio(inode, pos >> PAGE_SHIFT, &folio, SGP_CACHE);
156 if (filemap_check_wb_err(inode->i_mapping, 0)) {
162 page = folio_file_page(folio, pos >> PAGE_SHIFT);
165 * Mark the page dirty so that it won't be reclaimed once we drop the
166 * (potentially last) reference in xmbuf_unmap_page.
168 set_page_dirty(page);
171 bp->b_addr = page_address(page);
172 bp->b_pages = bp->b_page_array;
173 bp->b_pages[0] = page;
174 bp->b_page_count = 1;
178 /* Unmap a shmem page that was mapped into the buffer cache. */
183 struct page *page = bp->b_pages[0];
185 ASSERT(xfs_buftarg_is_mem(bp->b_target));
190 bp->b_pages[0] = NULL;
192 bp->b_page_count = 0;
195 /* Is this a valid daddr within the buftarg? */
198 struct xfs_buftarg *btp,
201 struct inode *inode = file_inode(btp->bt_file);
203 ASSERT(xfs_buftarg_is_mem(btp));
205 return daddr < (inode->i_sb->s_maxbytes >> BBSHIFT);
208 /* Discard the page backing this buffer. */
213 struct inode *inode = file_inode(bp->b_target->bt_file);
216 ASSERT(xfs_buftarg_is_mem(bp->b_target));
218 pos = BBTOB(xfs_buf_daddr(bp));
219 shmem_truncate_range(inode, pos, pos + BBTOB(bp->b_length) - 1);
223 * Finalize a buffer -- discard the backing page if it's stale, or run the
224 * write verifier to detect problems.
233 if (bp->b_flags & XBF_STALE) {
239 * Although this btree is ephemeral, validate the buffer structure so
240 * that we can detect memory corruption errors and software bugs.
242 fa = bp->b_ops->verify_struct(bp);
244 error = -EFSCORRUPTED;
245 xfs_verifier_error(bp, error, fa);
252 * Detach this xmbuf buffer from the transaction by any means necessary.
253 * All buffers are direct-mapped, so they do not need bwrite.
257 struct xfs_trans *tp,
260 struct xfs_buf_log_item *bli = bp->b_log_item;
264 bli->bli_flags &= ~(XFS_BLI_DIRTY | XFS_BLI_ORDERED |
265 XFS_BLI_LOGGED | XFS_BLI_STALE);
266 clear_bit(XFS_LI_DIRTY, &bli->bli_item.li_flags);
268 while (bp->b_log_item != NULL)
269 xfs_trans_bdetach(tp, bp);