Chunkfs README Val Henson <firstname.lastname@example.org> Summary ------- Chunkfs is an experimental local file system designed to quickly recover from file system corruption. Each file system is divided up into many small chunks, each of which can be checked and repaired with very few references to other chunks. In most cases, only a small part of the file system must be checked and repaired before it can be brought back online, requiring minutes instead of hours of downtime to recover from a file system error. Status ------ Chunkfs development began in February 2007. Growing a file into two chunks is supported but not much else. See the project web site for the current status: http://chunkfs.org License ------- Chunkfs is licensed under the GNU General Public License version 2. Chunks is released as a patch against the Linux kernel, which contains a copy of the GPLv2. Funding ------- Development of chunkfs was funded by: Intel EMC Centera VAH Consulting How it works ------------ Each file system is divided up into many small chunks. Each chunk is marked as dirty or clean. Things that make a chunk dirty are in-progress writes to metadata (creat(), chmod(), extending a file, etc.), I/O errors reported by the disk, and any data integrity errors observed by the file system code (bad checksums, wrong magic number, etc.). At mount time, each dirty chunk is checked with fsck and repaired if necessary, with limited references to other chunks. Clean chunks are not checked unless specifically requested. The metadata inside a chunk is structured so that nearly all references are within the chunk only. No block pointers or hard links may cross chunk boundaries. Only one kind of reference crosses the chunk boundary: inode continuations. When an inode needs to grow out of a chunk (either its data has outgrown the free space or we want to link to an inode outside its chunk), we allocate a new inode in the appropriate chunk and link the two inodes together with forward and back pointers, creating a continuation. Logically, the two inodes are parts of a single file or directory. When checking a chunk containing a continuation inode, the forward and back pointers allow us to quickly find the relevant information in any other chunks, without reading all the metadata in the entire chunk. Each chunk also keeps a bitmap of all inodes with continuations, as in some circumstances all continuations must be checked. For more information, see the documentation section on the project web site.