Before commit 93e72b3c612adcaca1 ("squashfs: migrate from ll_rw_block
usage to BIO"), compressed blocks read by squashfs were cached in the
page cache, but that is not the case after that commit. That has lead
to squashfs having to re-read a lot of sectors from disk/flash.
For example, the first sectors of every metadata block need to be read
twice from the disk. Once partially to read the length, and a
second time to read the block itself. Also, in linear reads of large
files, the last sectors of one data block are re-read from disk when
reading the next data block, since the compressed blocks are of variable
sizes and not aligned to device blocks. This extra I/O results in a
degrade in read performance of, for example, ~16% in one scenario on my
ARM platform using squashfs with dm-verity and NAND.
Since the decompressed data is cached in the page cache or squashfs'
internal metadata and fragment caches, caching _all_ compressed pages
would lead to a lot of double caching and is undesirable. But make the
code cache any disk blocks which were only partially requested, since
these are the ones likely to include data which is needed by other file
system blocks. This restores read performance in my test scenario.
The compressed block caching is only applied when the disk block size is
equal to the page size, to avoid having to deal with caching sub-page
reads.
Signed-off-by: Vincent Whitchurch <[email protected]>
---
Changes in v3:
- Drop Philippe Liard <[email protected]> from Cc since the address is
bouncing.
- Tweak formatting of long argument list.
- Clone the bio so that the bio_vec can be reused.
- Link to v2: https://lore.kernel.org/r/[email protected]
Changes in v2:
- Do not remove static from squashfs_bio_read()
- Link to v1: https://lore.kernel.org/r/[email protected]
---
fs/squashfs/block.c | 115 +++++++++++++++++++++++++++++++++++++++++--
fs/squashfs/squashfs_fs_sb.h | 1 +
fs/squashfs/super.c | 12 +++++
3 files changed, 125 insertions(+), 3 deletions(-)
diff --git a/fs/squashfs/block.c b/fs/squashfs/block.c
index bed3bb8b27fa..2b7f4c6fa77c 100644
--- a/fs/squashfs/block.c
+++ b/fs/squashfs/block.c
@@ -76,10 +76,100 @@ static int copy_bio_to_actor(struct bio *bio,
return copied_bytes;
}
+static int squashfs_bio_read_cached(struct bio *fullbio,
+ struct address_space *cache_mapping, u64 index, int length,
+ u64 read_start, u64 read_end, int page_count)
+{
+ struct page *head_to_cache = NULL, *tail_to_cache = NULL;
+ struct block_device *bdev = fullbio->bi_bdev;
+ int start_idx = 0, end_idx = 0;
+ struct bvec_iter_all iter_all;
+ struct bio *bio = NULL;
+ struct bio_vec *bv;
+ int idx = 0;
+ int err = 0;
+
+ bio_for_each_segment_all(bv, fullbio, iter_all) {
+ struct page *page = bv->bv_page;
+
+ if (page->mapping == cache_mapping && PageUptodate(page)) {
+ idx++;
+ continue;
+ }
+
+ /*
+ * We only use this when the device block size is the same as
+ * the page size, so read_start and read_end cover full pages.
+ *
+ * Compare these to the original required index and length to
+ * only cache pages which were requested partially, since these
+ * are the ones which are likely to be needed when reading
+ * adjacent blocks.
+ */
+ if (idx == 0 && index != read_start)
+ head_to_cache = page;
+ else if (idx == page_count - 1 && index + length != read_end)
+ tail_to_cache = page;
+
+ if (!bio || idx != end_idx) {
+ struct bio *new = bio_alloc_clone(bdev, fullbio,
+ GFP_NOIO, &fs_bio_set);
+
+ if (bio) {
+ bio_trim(bio, start_idx * PAGE_SECTORS,
+ (end_idx - start_idx) * PAGE_SECTORS);
+ bio_chain(bio, new);
+ submit_bio(bio);
+ }
+
+ bio = new;
+ start_idx = idx;
+ }
+
+ idx++;
+ end_idx = idx;
+ }
+
+ if (bio) {
+ bio_trim(bio, start_idx * PAGE_SECTORS,
+ (end_idx - start_idx) * PAGE_SECTORS);
+ err = submit_bio_wait(bio);
+ bio_put(bio);
+ }
+
+ if (err)
+ return err;
+
+ if (head_to_cache) {
+ int ret = add_to_page_cache_lru(head_to_cache, cache_mapping,
+ read_start, GFP_NOIO);
+
+ if (!ret) {
+ SetPageUptodate(head_to_cache);
+ unlock_page(head_to_cache);
+ }
+
+ }
+
+ if (tail_to_cache) {
+ int ret = add_to_page_cache_lru(tail_to_cache, cache_mapping,
+ read_end - PAGE_SIZE, GFP_NOIO);
+
+ if (!ret) {
+ SetPageUptodate(tail_to_cache);
+ unlock_page(tail_to_cache);
+ }
+ }
+
+ return 0;
+}
+
static int squashfs_bio_read(struct super_block *sb, u64 index, int length,
struct bio **biop, int *block_offset)
{
struct squashfs_sb_info *msblk = sb->s_fs_info;
+ struct inode *cache_inode = msblk->cache_inode;
+ struct address_space *cache_mapping = cache_inode ? cache_inode->i_mapping : NULL;
const u64 read_start = round_down(index, msblk->devblksize);
const sector_t block = read_start >> msblk->devblksize_log2;
const u64 read_end = round_up(index + length, msblk->devblksize);
@@ -99,13 +189,27 @@ static int squashfs_bio_read(struct super_block *sb, u64 index, int length,
for (i = 0; i < page_count; ++i) {
unsigned int len =
min_t(unsigned int, PAGE_SIZE - offset, total_len);
- struct page *page = alloc_page(GFP_NOIO);
+ struct page *page = NULL;
+
+ if (cache_mapping)
+ page = find_get_page(cache_mapping,
+ read_start + i * PAGE_SIZE);
+ if (!page)
+ page = alloc_page(GFP_NOIO);
if (!page) {
error = -ENOMEM;
goto out_free_bio;
}
- if (!bio_add_page(bio, page, len, offset)) {
+
+ if (cache_mapping) {
+ /*
+ * Use the __ version to avoid merging since we need
+ * each page to be separate when we check for and avoid
+ * cached pages.
+ */
+ __bio_add_page(bio, page, len, offset);
+ } else if (!bio_add_page(bio, page, len, offset)) {
error = -EIO;
goto out_free_bio;
}
@@ -113,7 +217,12 @@ static int squashfs_bio_read(struct super_block *sb, u64 index, int length,
total_len -= len;
}
- error = submit_bio_wait(bio);
+ if (cache_mapping)
+ error = squashfs_bio_read_cached(bio, cache_mapping, index,
+ length, read_start, read_end,
+ page_count);
+ else
+ error = submit_bio_wait(bio);
if (error)
goto out_free_bio;
diff --git a/fs/squashfs/squashfs_fs_sb.h b/fs/squashfs/squashfs_fs_sb.h
index 72f6f4b37863..dfee65845d48 100644
--- a/fs/squashfs/squashfs_fs_sb.h
+++ b/fs/squashfs/squashfs_fs_sb.h
@@ -47,6 +47,7 @@ struct squashfs_sb_info {
struct squashfs_cache *block_cache;
struct squashfs_cache *fragment_cache;
struct squashfs_cache *read_page;
+ struct inode *cache_inode;
int next_meta_index;
__le64 *id_table;
__le64 *fragment_index;
diff --git a/fs/squashfs/super.c b/fs/squashfs/super.c
index e090fae48e68..64d6bc95950b 100644
--- a/fs/squashfs/super.c
+++ b/fs/squashfs/super.c
@@ -329,6 +329,16 @@ static int squashfs_fill_super(struct super_block *sb, struct fs_context *fc)
goto failed_mount;
}
+ if (msblk->devblksize == PAGE_SIZE) {
+ msblk->cache_inode = new_inode(sb);
+ if (msblk->cache_inode == NULL)
+ goto failed_mount;
+
+ set_nlink(msblk->cache_inode, 1);
+ msblk->cache_inode->i_size = OFFSET_MAX;
+ mapping_set_gfp_mask(msblk->cache_inode->i_mapping, GFP_NOFS);
+ }
+
msblk->stream = squashfs_decompressor_setup(sb, flags);
if (IS_ERR(msblk->stream)) {
err = PTR_ERR(msblk->stream);
@@ -454,6 +464,7 @@ static int squashfs_fill_super(struct super_block *sb, struct fs_context *fc)
squashfs_cache_delete(msblk->block_cache);
squashfs_cache_delete(msblk->fragment_cache);
squashfs_cache_delete(msblk->read_page);
+ iput(msblk->cache_inode);
msblk->thread_ops->destroy(msblk);
kfree(msblk->inode_lookup_table);
kfree(msblk->fragment_index);
@@ -572,6 +583,7 @@ static void squashfs_put_super(struct super_block *sb)
squashfs_cache_delete(sbi->block_cache);
squashfs_cache_delete(sbi->fragment_cache);
squashfs_cache_delete(sbi->read_page);
+ iput(sbi->cache_inode);
sbi->thread_ops->destroy(sbi);
kfree(sbi->id_table);
kfree(sbi->fragment_index);
---
base-commit: 457391b0380335d5e9a5babdec90ac53928b23b4
change-id: 20230510-squashfs-cache-7a3b9e7355c1
Best regards,
--
Vincent Whitchurch <[email protected]>
On 16/05/2023 09:22, Vincent Whitchurch wrote:
> Before commit 93e72b3c612adcaca1 ("squashfs: migrate from ll_rw_block
> usage to BIO"), compressed blocks read by squashfs were cached in the
> page cache, but that is not the case after that commit. That has lead
> to squashfs having to re-read a lot of sectors from disk/flash.
>
> For example, the first sectors of every metadata block need to be read
> twice from the disk. Once partially to read the length, and a
> second time to read the block itself. Also, in linear reads of large
> files, the last sectors of one data block are re-read from disk when
> reading the next data block, since the compressed blocks are of variable
> sizes and not aligned to device blocks. This extra I/O results in a
> degrade in read performance of, for example, ~16% in one scenario on my
> ARM platform using squashfs with dm-verity and NAND.
>
> Since the decompressed data is cached in the page cache or squashfs'
> internal metadata and fragment caches, caching _all_ compressed pages
> would lead to a lot of double caching and is undesirable. But make the
> code cache any disk blocks which were only partially requested, since
> these are the ones likely to include data which is needed by other file
> system blocks. This restores read performance in my test scenario.
>
> The compressed block caching is only applied when the disk block size is
> equal to the page size, to avoid having to deal with caching sub-page
> reads.
>
> Signed-off-by: Vincent Whitchurch <[email protected]>
Reviewed-by: Phillip Lougher <[email protected]>
This looks generally good, but I have two more nitpicks (sorry for
beeing annoying):
> + struct inode *cache_inode = msblk->cache_inode;
> + struct address_space *cache_mapping = cache_inode ? cache_inode->i_mapping : NULL;
>
> + struct page *page = NULL;
> +
> + if (cache_mapping)
> + page = find_get_page(cache_mapping,
> + read_start + i * PAGE_SIZE);
Given we only ever use the cache inode mapping, I'd suggest to store
that in the super_block and the just use mapping->host for the iput
on unmount.
> + if (cache_mapping) {
> + /*
> + * Use the __ version to avoid merging since we need
> + * each page to be separate when we check for and avoid
> + * cached pages.
> + */
> + __bio_add_page(bio, page, len, offset);
> + } else if (!bio_add_page(bio, page, len, offset)) {
> error = -EIO;
> goto out_free_bio;
And given that random alloc_page pages are basically never mergable
we should be able to just use __bio_add_page here unconditionally.