Why use registered files?

For every I/O operation that uses a file descriptor, the kernel must:

• Look up the file descriptor in the process's file descriptor table
• Take a reference to the file to ensure it remains valid during the operation
• Release the reference when the operation completes

For applications performing many I/O operations, especially on threaded applications where the file table is shared (making reference counting more expensive), these overheads accumulate. By registering files once, the reference is held for the lifetime of the registration, and operations can use the file directly without per-operation lookups or reference counting.

Registered files are most beneficial for applications that:

• Perform many I/O operations on the same set of files
• Are multi-threaded (where file table operations are more expensive)
• Need the lowest possible per-I/O overhead

Registering files

Files are registered using io_uring_register_files(3) or io_uring_register_files_tags(3). The files are described using an array of file descriptors:

int fds[3];
fds[0] = open("file1", O_RDONLY);
fds[1] = open("file2", O_RDONLY);
fds[2] = open("file3", O_WRONLY | O_CREAT, 0644);
ret = io_uring_register_files(ring, fds, 3);

Once registered, the original file descriptors can be closed if desired. The kernel holds its own references to the underlying files.

Using registered files

To use a registered file in an I/O operation, set the IOSQE_FIXED_FILE flag in the SQE's flags field, and use the index into the registered file array (not the original file descriptor) in the fd field:

struct io_uring_sqe *sqe = io_uring_get_sqe(ring);
io_uring_prep_read(sqe, 0, buf, len, offset);  /* index 0, not fd */
io_uring_sqe_set_flags(sqe, IOSQE_FIXED_FILE);

The index is 0-based into the array passed to io_uring_register_files(3).

Sparse file registration

The file array can be sparse, meaning some slots can be empty. Empty slots are indicated by setting the file descriptor to -1. Applications can create a fully sparse table using io_uring_register_files_sparse(3) and fill in slots later:

/* Create sparse table with 100 slots */
ret = io_uring_register_files_sparse(ring, 100);
/* Later, fill in slot 5 */
int fd = open("file", O_RDONLY);
ret = io_uring_register_files_update(ring, 5, &fd, 1);

Updating registered files

Registered files can be updated using io_uring_register_files_update(3) or io_uring_register_files_update_tag(3). This can:

• Replace an existing file with a new one
• Fill in an empty slot
• Remove a file by setting the descriptor to -1

To skip updating certain slots while updating others, use the special value IORING_REGISTER_FILES_SKIP.

int fds[3];
fds[0] = new_fd;                        /* replace slot 0 */
fds[1] = IORING_REGISTER_FILES_SKIP;    /* leave slot 1 unchanged */
fds[2] = -1;                            /* remove slot 2 */
ret = io_uring_register_files_update(ring, 0, fds, 3);

Updates do not require the ring to be idle on kernels 5.13 and later. On older kernels, updates would wait for in-flight operations to complete.

File tagging

When using io_uring_register_files_tags(3) or io_uring_register_files_update_tag(3), each file can be associated with a tag value. When a file is unregistered (either explicitly or by replacement), and there are no more in-flight operations using that file, a completion queue entry is posted with user_data set to the tag value and all other fields zeroed.

This notification mechanism allows applications to know when it is safe to perform cleanup actions associated with the file.

Direct file descriptors

Some io_uring operations can allocate file descriptors directly into the registered file table, avoiding the regular file descriptor table entirely. This is done by setting the file_index field in the SQE (using io_uring_sqe_set_target_fixed_file(3)) to the desired slot, or using IORING_FILE_INDEX_ALLOC to have io_uring allocate the next available slot.

Operations that support direct descriptors include:

• IORING_OP_OPENAT / IORING_OP_OPENAT2
• IORING_OP_ACCEPT
• IORING_OP_SOCKET
• IORING_OP_PIPE

When using IORING_FILE_INDEX_ALLOC, the application should use io_uring_register_file_alloc_range(3) to specify which range of the file table should be used for allocations.

/* Reserve slots 50-99 for dynamic allocation */
io_uring_register_file_alloc_range(ring, 50, 50);
/* Accept with direct descriptor allocation */
struct io_uring_sqe *sqe = io_uring_get_sqe(ring);
io_uring_prep_accept_direct(sqe, listen_fd, NULL, NULL, 0,
                            IORING_FILE_INDEX_ALLOC);

The allocated slot index is returned in the CQE res field on success.

Closing direct descriptors

Direct descriptors (files that exist only in the registered file table) can be closed using IORING_OP_CLOSE with the IOSQE_FIXED_FILE flag set, or by updating the slot to -1 using io_uring_register_files_update(3).

Unregistering files

Files are unregistered using io_uring_unregister_files(3). This releases all registered files. Files are also automatically unregistered when the io_uring instance is destroyed.

Applications do not need to explicitly unregister files before shutting down the ring.