Filesystems

Purpose

fs/ contains both the VFS (Virtual File System) abstraction and the in-tree implementations of every supported file system. The VFS is the indirection layer that translates read, write, open, stat, mount, etc., into per-filesystem operations.

Directory layout

fs/
├── Kconfig, Kconfig.binfmt, Makefile
├── inode.c, dcache.c, namei.c, namespace.c, super.c, file.c, file_table.c
├── open.c, read_write.c, splice.c, sync.c, fcntl.c, ioctl.c, stat.c, statfs.c
├── attr.c, xattr.c, posix_acl.c
├── pipe.c, fifo.c, anon_inodes.c, eventfd.c, timerfd.c, signalfd.c
├── select.c, poll.c (in include/), epoll/  (separate? check)
├── exec.c (in fs/), binfmt_elf.c, binfmt_misc.c, binfmt_script.c, binfmt_flat.c
├── inotify/, fsnotify/, dnotify/, fanotify/  # Notification frameworks
├── locks.c, mbcache.c, drop_caches.c, fs-writeback.c, dax.c
├── ext2/ ext4/ xfs/ btrfs/ f2fs/ bcachefs/ ntfs3/ udf/ romfs/ ramfs/ cramfs/ jfs/ erofs/ orangefs/ ubifs/ jffs2/ adfs/ affs/ befs/ bfs/ qnx4/ qnx6/ omfs/ minix/ coda/ exfat/ fat/
├── proc/ sysfs/ tracefs/ debugfs/ pstore/ efivarfs/ configfs/ resctrl/ cachefiles/
├── tmpfs (under mm/shmem.c, but UAPI lives in fs)
├── nfs/ nfsd/ nfs_common/ lockd/ nls/ exportfs/      # NFS client/server
├── smb/ ksmbd/ cifs (within smb/)                    # SMB client + Samba in-kernel
├── ceph/                                              # Ceph client
├── 9p/                                                # 9P (used by virtio-9p)
├── fuse/                                              # FUSE: user-space filesystem framework
├── overlayfs/                                         # Stackable overlayfs
├── autofs/                                            # autofs auto-mounter
├── netfs/, fscache/                                   # Network-FS cache framework
├── kernfs/                                            # Backing for sysfs, cgroupfs, tracefs
├── verity/                                            # fs-verity
├── crypto/                                            # fscrypt (filesystem-level encryption)
├── notify/                                            # core fsnotify infrastructure
├── nls/                                               # NLS character set tables
├── unicode/                                           # Unicode tables for case folding (ext4, f2fs)
├── pidfs.c, mount.h, internal.h, ...                  # Misc core
└── ...

(Around 80 subdirectories total.)

Key abstractions

The VFS is built on four core objects:

Object	File	Role
struct super_block	include/linux/fs.h, implementation in fs/super.c	A mounted filesystem instance.
struct inode	include/linux/fs.h, fs/inode.c	An on-disk object (file, directory, symlink, special).
struct dentry	fs/dcache.c	A name-to-inode binding in the dcache (directory cache).
struct file	fs/file_table.c	An open file descriptor's kernel-side state.

Each is paired with an operation table:

• struct super_operations
• struct inode_operations
• struct dentry_operations
• struct file_operations
• struct address_space_operations (page cache integration)

A filesystem implements these tables and registers a struct file_system_type.

How it works

graph LR
    SC[Syscall: read/open/stat/...] --> VFS[VFS layer]
    VFS -->|"path lookup<br/>fs/namei.c"| DCACHE[dcache]
    DCACHE -->|"miss"| INODE_OPS["inode_operations<br/>per-FS"]
    INODE_OPS --> FS[("Concrete FS<br/>(ext4, btrfs, xfs, ...)")]
    VFS -->|"file ops"| FOP["file_operations<br/>per-FS"]
    FOP --> PAGE_CACHE[mm/filemap.c]
    PAGE_CACHE -->|"miss"| AOP["address_space_operations<br/>per-FS"]
    AOP --> BIO[block/bio.c]
    BIO --> BLK[Block layer]

VFS path lookup

Path resolution happens in fs/namei.c. It walks each component, consulting the dcache for cached (parent, name) -> dentry bindings. On dcache misses it calls into the filesystem's inode_operations->lookup. Special features handled here: symlinks, mountpoints (a different superblock), bind mounts, RCU-walk for read-mostly fast paths, follow-link, mount namespaces.

File I/O

read(2), write(2), pread64, pwrite64, readv, writev go through vfs_read / vfs_write in fs/read_write.c. For regular files this typically dispatches to generic_file_read_iter / generic_file_write_iter which talk to the page cache, which talks to address_space_operations to populate or flush.

Mount and namespaces

fs/namespace.c implements the mount tree, mount namespaces, and bind/move/remount. The fsopen/fsconfig/fsmount/move_mount API is the modern mount syscall set; the older mount(2) is still supported.

Notifications

fsnotify is the core in fs/notify/ with three frontends: inotify (file-descriptor-based), dnotify (legacy directory-based), and fanotify (more powerful, can be used for AV scanning, with permission events).

binfmt loaders

fs/binfmt_*.c register handlers for executable formats:

• binfmt_elf.c — ELF (the common case).
• binfmt_elf_fdpic.c — FDPIC ELF (no-MMU).
• binfmt_script.c — #! shebang.
• binfmt_misc.c — runtime-registered (qemu-user, wine).
• binfmt_flat.c — bFLT.

do_execveat_common() in fs/exec.c walks the registered formats.

Concrete filesystems

FS	Page	Notes
ext4	fs/ext4/	The traditional default for Linux distributions. Journaled.
xfs	fs/xfs/	Mature, scalable filesystem. Uses iomap and a delayed-allocation design.
btrfs	fs/btrfs/	Copy-on-write filesystem with snapshots, checksums, integrated RAID. ~40 source files.
f2fs	fs/f2fs/	Flash-friendly filesystem optimized for SSDs / eMMC.
bcachefs	fs/bcachefs/	COW filesystem with snapshots, replication, encryption, compression. Recently merged.
ntfs3	fs/ntfs3/	Modern read-write NTFS.
fuse	fs/fuse/	Forwarding to a user-space daemon over /dev/fuse.
nfs / nfsd	fs/nfs/, fs/nfsd/	NFS client and server.
smb (CIFS) / ksmbd	fs/smb/	SMB client and in-kernel SMB server.
ceph	fs/ceph/	Ceph filesystem client.
9p	fs/9p/	Plan 9 protocol; common in QEMU/virtio.
overlayfs	fs/overlayfs/	Stackable union filesystem (used by container runtimes).
tmpfs / shmem	code in mm/shmem.c	Backed by page cache; ABI surface in fs/.
proc / sysfs / tracefs / debugfs / bpf	various	Pseudo-filesystems exposing kernel state.

fscrypt and fs-verity

• fscrypt (fs/crypto/) — per-file/per-directory encryption, used by ext4, f2fs, ubifs.
• fs-verity (fs/verity/) — Merkle-tree authentication of file contents.

Network-FS cache

fs/netfs/ and fs/fscache/ provide a generic local-cache framework for network filesystems (NFS, AFS, Ceph, 9p).

Integration points

• mm/: page cache (address_space), folios, writeback, readahead.
• block/: filesystems submit bios; the block layer schedules them.
• net/: NFS, Ceph, SMB, 9p, sunrpc all use kernel sockets.
• security/: LSM hooks at path_*, inode_*, file_*.
• crypto/: fscrypt and fs-verity use the crypto API.
• kernel/: tracepoints, capability checks, signal delivery on file events.

Key source files

File	Purpose
fs/namei.c	Path-name resolution.
fs/dcache.c	The dentry cache.
fs/inode.c	Inode lifecycle, struct inode_operations glue.
fs/file.c, fs/file_table.c	Per-process file descriptor table, struct file.
fs/super.c	Superblock lifecycle, mount.
fs/namespace.c	Mount tree and mount namespaces.
fs/exec.c	execve machinery.
fs/binfmt_elf.c	ELF binary loader.
fs/read_write.c	vfs_read, vfs_write, vfs_iter_read.
fs/fs-writeback.c	Per-bdi writeback threads.
fs/buffer.c	The buffer-head layer (legacy block I/O for some FSes).
fs/iomap/	The iomap library used by xfs, gfs2, ext4 (DAX), bcachefs.

Entry points for modification

• New filesystem: create fs/myfs/, register a file_system_type, implement the four operation tables. Read Documentation/filesystems/vfs.rst and Documentation/filesystems/porting.rst.
• New VFS-level feature: modify fs/namei.c, fs/file.c, etc., touch every consumer. Send to the FS list (linux-fsdevel@vger.kernel.org).
• New filesystem syscall: rare; coordinate with linux-api@.
• New mount option: extend the new mount API in fs/fs_context.c and fs/fs_parser.c.

Related pages

• Block layer — where filesystem bios go.
• Memory management — page cache.
• Networking — kernel sockets used by network FSes.
• Security — LSM hooks on path/inode/file.
• Crypto — fscrypt, fs-verity.