From: Goldwyn Rodrigues <rgoldwyn@xxxxxxxx> This is an attempt to reduce the memory footprint by using a shared page(s) for shared extent(s) in the filesystem. I am hoping to start a discussion to iron out the details for implementation. Abstract If mutiple files share an extent, reads from each of the files would read individual page copies in the inode pagecache mapping, leading to waste of memory. Instead add the read pages of the filesystem to underlying device's bd_inode as opposed to file's inode mapping. The cost is that file offset to device offset conversion happens early in the read cycle. Motivation: - Reduce memory usage for shared extents - Ease DAX implementation for shared extents - Reduce Container memory utilization Implementation In the read path, pages are read and added to the block_device's inode's mapping as opposed to the inode's mapping. This is limited to reads, while write's (and read before writes) still go through inode's i_mapping. The path does check the inode's i_mapping before falling back to block device's i_mapping to read pages which may be dirty. The cost of the operation is file_to_device_offset() translation on reads. The translation should return a valid value only in case the file is CoW. This also means that page->mapping may not point to file's mapping. Questions: 1. Are there security implications for performing this for read-only pages? An alternate idea is to add a "struct fspage", which would be pointed by file's mapping and would point to the block device's page. Multiple files with shared extents have their own independent fspage pointing to the same page mapped to block device's mapping. Any security parameters, if required, would be in this structure. The advantage of this approach is it would be more flexible with respect to CoW when the page is dirtied after reads. With the current approach, a read for write is an independent operation so we can end up with two copies of the same page. This implementation got complicated too quickly. 2. Should pages be dropped after writebacks (and clone_range) to avoid duplicate copies? Limitations: 1. The filesystem have exactly one underlying device. 2. Page size should be equal to filesystem block size Goldwyn Rodrigues (5): mm: Use file parameter to determine bdi mm: Switch mapping to device mapping btrfs: Add sharedext mount option btrfs: Set s_bdev for btrfs super block btrfs: function to convert file offset to device offset fs/btrfs/ctree.h | 1 + fs/btrfs/file.c | 42 ++++++++++++++++++++++++++++++++++++++++-- fs/btrfs/super.c | 7 +++++++ include/linux/fs.h | 7 ++++++- mm/filemap.c | 34 ++++++++++++++++++++++++++-------- mm/readahead.c | 3 +++ 6 files changed, 83 insertions(+), 11 deletions(-) -- 2.33.1
