IO_URING_REGISTER(2) Linux Programmer's Manual IO_URING_REGISTER(2)

io_uring_register - register files or user buffers for asynchronous I/O

#include <linux/io_uring.h>
int io_uring_register(unsigned int fd, unsigned int opcode,
                      void *arg, unsigned int nr_args);

The io_uring_register() system call registers resources (e.g. user buffers, files, eventfd, personality, restrictions) for use in an io_uring(7) instance referenced by fd. Registering files or user buffers allows the kernel to take long term references to internal data structures or create long term mappings of application memory, greatly reducing per-I/O overhead.

fd is the file descriptor returned by a call to io_uring_setup(2). opcode can be one of:

arg points to a struct iovec array of nr_args entries. The buffers associated with the iovecs will be locked in memory and charged against the user's RLIMIT_MEMLOCK resource limit. See getrlimit(2) for more information. Additionally, there is a size limit of 1GiB per buffer. Currently, the buffers must be anonymous, non-file-backed memory, such as that returned by malloc(3) or mmap(2) with the MAP_ANONYMOUS flag set. It is expected that this limitation will be lifted in the future. Huge pages are supported as well. Note that the entire huge page will be pinned in the kernel, even if only a portion of it is used.

After a successful call, the supplied buffers are mapped into the kernel and eligible for I/O. To make use of them, the application must specify the IORING_OP_READ_FIXED or IORING_OP_WRITE_FIXED opcodes in the submission queue entry (see the struct io_uring_sqe definition in io_uring_enter(2)), and set the buf_index field to the desired buffer index. The memory range described by the submission queue entry's addr and len fields must fall within the indexed buffer.

It is perfectly valid to setup a large buffer and then only use part of it for an I/O, as long as the range is within the originally mapped region.

An application can increase or decrease the size or number of registered buffers by first unregistering the existing buffers, and then issuing a new call to io_uring_register() with the new buffers.

Note that before 5.13 registering buffers would wait for the ring to idle. If the application currently has requests in-flight, the registration will wait for those to finish before proceeding.

An application need not unregister buffers explicitly before shutting down the io_uring instance. Available since 5.1.

Register buffers for I/O. Similar to IORING_REGISTER_BUFFERS but aims to have a more extensible ABI.

arg points to a struct io_uring_rsrc_register, and nr_args should be set to the number of bytes in the structure.


struct io_uring_rsrc_register {

__u32 nr;
__u32 resv;
__u64 resv2;
__aligned_u64 data;
__aligned_u64 tags; };


The data field contains a pointer to a struct iovec array of nr entries. The tags field should either be 0, then tagging is disabled, or point to an array of nr "tags" (unsigned 64 bit integers). If a tag is zero, then tagging for this particular resource (a buffer in this case) is disabled. Otherwise, after the resource had been unregistered and it's not used anymore, a CQE will be posted with user_data set to the specified tag and all other fields zeroed.

Note that resource updates, e.g. IORING_REGISTER_BUFFERS_UPDATE, don't necessarily deallocate resources by the time it returns, but they might be held alive until all requests using it complete.

Available since 5.13.

Updates registered buffers with new ones, either turning a sparse entry into a real one, or replacing an existing entry.

arg must contain a pointer to a struct io_uring_rsrc_update2, which contains an offset on which to start the update, and an array of struct iovec. tags points to an array of tags. nr must contain the number of descriptors in the passed in arrays. See IORING_REGISTER_BUFFERS2 for the resource tagging description.


struct io_uring_rsrc_update2 {

__u32 offset;
__u32 resv;
__aligned_u64 data;
__aligned_u64 tags;
__u32 nr;
__u32 resv2; };


Available since 5.13.

This operation takes no argument, and arg must be passed as NULL. All previously registered buffers associated with the io_uring instance will be released. Available since 5.1.
Register files for I/O. arg contains a pointer to an array of nr_args file descriptors (signed 32 bit integers).

To make use of the registered files, the IOSQE_FIXED_FILE flag must be set in the flags member of the struct io_uring_sqe, and the fd member is set to the index of the file in the file descriptor array.

The file set may be sparse, meaning that the fd field in the array may be set to -1. See IORING_REGISTER_FILES_UPDATE for how to update files in place.

Note that before 5.13 registering files would wait for the ring to idle. If the application currently has requests in-flight, the registration will wait for those to finish before proceeding. See IORING_REGISTER_FILES_UPDATE for how to update an existing set without that limitation.

Files are automatically unregistered when the io_uring instance is torn down. An application needs only unregister if it wishes to register a new set of fds. Available since 5.1.

Register files for I/O. Similar to IORING_REGISTER_FILES.

arg points to a struct io_uring_rsrc_register, and nr_args should be set to the number of bytes in the structure.

The data field contains a pointer to an array of nr file descriptors (signed 32 bit integers). tags field should either be 0 or or point to an array of nr "tags" (unsigned 64 bit integers). See IORING_REGISTER_BUFFERS2 for more info on resource tagging.

Note that resource updates, e.g. IORING_REGISTER_FILES_UPDATE, don't necessarily deallocate resources, they might be held until all requests using that resource complete.

Available since 5.13.

This operation replaces existing files in the registered file set with new ones, either turning a sparse entry (one where fd is equal to -1 ) into a real one, removing an existing entry (new one is set to -1 ), or replacing an existing entry with a new existing entry.

arg must contain a pointer to a struct io_uring_files_update, which contains an offset on which to start the update, and an array of file descriptors to use for the update. nr_args must contain the number of descriptors in the passed in array. Available since 5.5.

File descriptors can be skipped if they are set to IORING_REGISTER_FILES_SKIP. Skipping an fd will not touch the file associated with the previous fd at that index. Available since 5.12.

Similar to IORING_REGISTER_FILES_UPDATE, replaces existing files in the registered file set with new ones, either turning a sparse entry (one where fd is equal to -1 ) into a real one, removing an existing entry (new one is set to -1 ), or replacing an existing entry with a new existing entry.

arg must contain a pointer to a struct io_uring_rsrc_update2, which contains an offset on which to start the update, and an array of file descriptors to use for the update stored in data. tags points to an array of tags. nr must contain the number of descriptors in the passed in arrays. See IORING_REGISTER_BUFFERS2 for the resource tagging description.

Available since 5.13.

This operation requires no argument, and arg must be passed as NULL. All previously registered files associated with the io_uring instance will be unregistered. Available since 5.1.
It's possible to use eventfd(2) to get notified of completion events on an io_uring instance. If this is desired, an eventfd file descriptor can be registered through this operation. arg must contain a pointer to the eventfd file descriptor, and nr_args must be 1. Note that while io_uring generally takes care to avoid spurious events, they can occur. Similarly, batched completions of CQEs may only trigger a single eventfd notification even if multiple CQEs are posted. The application should make no assumptions on number of events being available having a direct correlation to eventfd notifications posted. An eventfd notification must thus only be treated as a hint to check the CQ ring for completions. Available since 5.2.

An application can temporarily disable notifications, coming through the registered eventfd, by setting the IORING_CQ_EVENTFD_DISABLED bit in the flags field of the CQ ring. Available since 5.8.

This works just like IORING_REGISTER_EVENTFD , except notifications are only posted for events that complete in an async manner. This means that events that complete inline while being submitted do not trigger a notification event. The arguments supplied are the same as for IORING_REGISTER_EVENTFD. Available since 5.6.
Unregister an eventfd file descriptor to stop notifications. Since only one eventfd descriptor is currently supported, this operation takes no argument, and arg must be passed as NULL and nr_args must be zero. Available since 5.2.
This operation returns a structure, io_uring_probe, which contains information about the opcodes supported by io_uring on the running kernel. arg must contain a pointer to a struct io_uring_probe, and nr_args must contain the size of the ops array in that probe struct. The ops array is of the type io_uring_probe_op, which holds the value of the opcode and a flags field. If the flags field has IO_URING_OP_SUPPORTED set, then this opcode is supported on the running kernel. Available since 5.6.
This operation registers credentials of the running application with io_uring, and returns an id associated with these credentials. Applications wishing to share a ring between separate users/processes can pass in this credential id in the sqe personality field. If set, that particular sqe will be issued with these credentials. Must be invoked with arg set to NULL and nr_args set to zero. Available since 5.6.
This operation unregisters a previously registered personality with io_uring. nr_args must be set to the id in question, and arg must be set to NULL. Available since 5.6.
This operation enables an io_uring ring started in a disabled state (IORING_SETUP_R_DISABLED was specified in the call to io_uring_setup(2)). While the io_uring ring is disabled, submissions are not allowed and registrations are not restricted.

After the execution of this operation, the io_uring ring is enabled: submissions and registration are allowed, but they will be validated following the registered restrictions (if any). This operation takes no argument, must be invoked with arg set to NULL and nr_args set to zero. Available since 5.10.

arg points to a struct io_uring_restriction array of nr_args entries.

With an entry it is possible to allow an io_uring_register() opcode, or specify which opcode and flags of the submission queue entry are allowed, or require certain flags to be specified (these flags must be set on each submission queue entry).

All the restrictions must be submitted with a single io_uring_register() call and they are handled as an allowlist (opcodes and flags not registered, are not allowed).

Restrictions can be registered only if the io_uring ring started in a disabled state (IORING_SETUP_R_DISABLED must be specified in the call to io_uring_setup(2)).

Available since 5.10.

By default, async workers created by io_uring will inherit the CPU mask of its parent. This is usually all the CPUs in the system, unless the parent is being run with a limited set. If this isn't the desired outcome, the application may explicitly tell io_uring what CPUs the async workers may run on. arg must point to a cpu_set_t mask, and nr_args the byte size of that mask.

Available since 5.14.

Undoes a CPU mask previously set with IORING_REGISTER_IOWQ_AFF. Must not have arg or nr_args set.

Available since 5.14.

By default, io_uring limits the unbounded workers created to the maximum processor count set by RLIMIT_NPROC and the bounded workers is a function of the SQ ring size and the number of CPUs in the system. Sometimes this can be excessive (or too little, for bounded), and this command provides a way to change the count per ring (per NUMA node) instead.

arg must be set to an unsigned int pointer to an array of two values, with the values in the array being set to the maximum count of workers per NUMA node. Index 0 holds the bounded worker count, and index 1 holds the unbounded worker count. On successful return, the passed in array will contain the previous maximum valyes for each type. If the count being passed in is 0, then this command returns the current maximum values and doesn't modify the current setting. nr_args must be set to 2, as the command takes two values.

Available since 5.15.

Whenever io_uring_enter(2) is called to submit request or wait for completions, the kernel must grab a reference to the file descriptor. If the application using io_uring is threaded, the file table is marked as shared, and the reference grab and put of the file descriptor count is more expensive than it is for a non-threaded application.

Similarly to how io_uring allows registration of files, this allow registration of the ring file descriptor itself. This reduces the overhead of the io_uring_enter(2) system call.

arg must be set to an unsigned int pointer to an array of type struct io_uring_rsrc_register of nr_args number of entries. The data field of this struct must point to an io_uring file descriptor, and the offset field can be either -1 or an explicit offset desired for the registered file descriptor value. If -1 is used, then upon successful return of this system call, the field will contain the value of the registered file descriptor to be used for future io_uring_enter(2) system calls.

On successful completion of this request, the returned descriptors may be used instead of the real file descriptor for io_uring_enter(2), provided that IORING_ENTER_REGISTERED_RING is set in the flags for the system call. This flag tells the kernel that a registered descriptor is used rather than a real file descriptor.

Each thread or process using a ring must register the file descriptor directly by issuing this request.o

The maximum number of supported registered ring descriptors is currently limited to 16.

Available since 5.18.

Unregister descriptors previously registered with IORING_REGISTER_RING_FDS.

arg must be set to an unsigned int pointer to an array of type struct io_uring_rsrc_register of nr_args number of entries. Only the offset field should be set in the structure, containing the registered file descriptor offset previously returned from IORING_REGISTER_RING_FDS that the application wishes to unregister.

Note that this isn't done automatically on ring exit, if the thread or task that previously registered a ring file descriptor isn't exiting. It is recommended to manually unregister any previously registered ring descriptors if the ring is closed and the task persists. This will free up a registration slot, making it available for future use.

Available since 5.18.

On success, io_uring_register() returns 0. On error, -1 is returned, and errno is set accordingly.

The opcode field is not allowed due to registered restrictions.
One or more fds in the fd array are invalid.
IORING_REGISTER_ENABLE_RINGS or IORING_REGISTER_RESTRICTIONS was specified, but the io_uring ring is not disabled.
IORING_REGISTER_BUFFERS or IORING_REGISTER_FILES or IORING_REGISTER_RESTRICTIONS was specified, but there were already buffers, files, or restrictions registered.
buffer is outside of the process' accessible address space, or iov_len is greater than 1GiB.
IORING_REGISTER_BUFFERS or IORING_REGISTER_FILES was specified, but nr_args is 0.
IORING_REGISTER_BUFFERS was specified, but nr_args exceeds UIO_MAXIOV
IORING_UNREGISTER_BUFFERS or IORING_UNREGISTER_FILES was specified, and nr_args is non-zero or arg is non-NULL.
IORING_REGISTER_RESTRICTIONS was specified, but nr_args exceeds the maximum allowed number of restrictions or restriction opcode is invalid.
IORING_REGISTER_FILES was specified and nr_args exceeds the maximum allowed number of files in a fixed file set.
IORING_REGISTER_FILES was specified and adding nr_args file references would exceed the maximum allowed number of files the user is allowed to have according to the RLIMIT_NOFILE resource limit and the caller does not have CAP_SYS_RESOURCE capability. Note that this is a per user limit, not per process.
Insufficient kernel resources are available, or the caller had a non-zero RLIMIT_MEMLOCK soft resource limit, but tried to lock more memory than the limit permitted. This limit is not enforced if the process is privileged (CAP_IPC_LOCK).
IORING_UNREGISTER_BUFFERS or IORING_UNREGISTER_FILES was specified, but there were no buffers or files registered.
Attempt to register files or buffers on an io_uring instance that is already undergoing file or buffer registration, or is being torn down.
User buffers point to file-backed memory.
2019-01-17 Linux