Requesting Attendance: Trond Myklebust, James Lentini and Anshul Madan Proposal: We propose to discuss the design and implementation issues involved in introducing a new copy() system call into the Linux kernel. The goal of this session would be to determine the required functionality and obtain a general consensus on the necessary VFS interface changes. We believe that the LSF Summit offers a unique opportunity to discuss this topic and gain consensus with developers from several different file system communities. We are implementing the NFS copy offload operations and associated system call(s). We would contribute to the discussions on API capabilities, design issues, and other VFS issues. We would provide insights into different design trade-offs in the VFS layer, and also report on the progress of our implementation. Keywords: File systems Background: There is an emerging trend for file systems to support intelligent copies. OCFS2 supports intelligent copies via a reflink ioctl(2). At the same time, a proposal has been made for the NFS protocol to support intelligent copy operations, see IETF draft draft-lentini-nfsv4-server-side-copy. We expect other file systems to add support for intelligent copy capabilities in the future. To take full advantage of these capabilities, the Linux kernel's system call interface must be expanded to allow userspace applications to use this feature in a uniform way. This topic has been discussed both on Linux development mailing lists [1] and at previous LSF summits. Joel Becker originally proposed a syscall called reflink() that was specific to copying an inode to a copy-on-write snapshot. Since then, the need for a generic copy interface that encompassed the reflink semantics as well as less restrictive ones has been recognized based on discussion between Linus Torvalds and Joel. Motivation: Intelligent copies have performance advantages over a traditional read/write copy. With data deduplication, fragments of data can be shared between files. Thus using conventional syscalls like read() and write() for doing a copy may lead to the unnecessary data transfers to and from the file system. Using a copy() syscall, the file system can make use of it's intelligence and thus avoid unnecessary data transfers. The IT trend toward virtualization via hypervisors has created an emerging use case to copy a virtual disk over NFS. The use of a copy operation will save network bandwidth on the client and server, and intra-server file copy has the potential to avoid all physical data copies, if the file system is deduplicated as described above. Implementing a VFS inode operation which has a generic interface applicable to all types of file systems would allow applications to access this functionality in a uniform way. The VFS copy() inode operation and corresponding interface should encompass the semantics of various file system copy() implementations, such as support for asynchronous copies, copy status, and copy abort calls. --- [1] Relevant posts: http://www.mail-archive.com/ocfs2-devel@xxxxxxxxxxxxxx/msg04347.html http://www.mail-archive.com/ocfs2-devel@xxxxxxxxxxxxxx/msg04395.html -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
