Enable I/O polling mode for file descriptors that support it. Instead of relying on interrupts, the kernel polls for completions. This reduces latency for high-performance storage devices (NVMe, etc.) but requires:

• Files opened with O_DIRECT
• Hardware and drivers that support polling
• The application to call io_uring_enter(2) to reap completions (busy-polling)
• Storage device configuration for polling support

IOPOLL rings cannot use IRQ-driven completion; the application must poll. Only request types that support polling may be issued on an IOPOLL ring. This mode is commonly used for scenarios that purely do polled I/O on storage devices like NVMe.

Using IOPOLL generally requires storage device setup. For NVMe devices, the kernel parameter nvme.poll_queues=X must be set, where X is the number of completion queues on the NVMe device to set aside for polling operations.

Create a kernel thread that polls the submission queue. Eliminates the need for system calls to submit I/O. See io_uring_sqpoll(7) for details.

Pin the SQPOLL thread to a specific CPU. Requires IORING_SETUP_SQPOLL. The CPU is specified in sq_thread_cpu of struct io_uring_params.

Enable hybrid polling mode. Instead of pure busy-polling, the kernel uses an adaptive approach that may sleep briefly, reducing CPU usage while still providing low latency. This is a middle ground between interrupt-driven and pure polling modes.

Disable interrupting the application for completion processing. Normally, the kernel signals the application when completions are ready, which can interrupt system calls. With this flag, completions are only processed when the application returns to userspace from any system call, not just io_uring-related ones. This means completions may be processed after read(2), write(2), poll(2), or any other syscall returns.

This improves performance by eliminating asynchronous interrupts but requires the application to regularly enter the kernel to process completions. Recommended for most applications that have an event loop.

When completions are pending, set IORING_SQ_TASKRUN in the SQ ring flags. This allows applications to check if there is completion work to process without making a system call. Typically used with IORING_SETUP_COOP_TASKRUN.

Defer completion task work to when the application explicitly enters the kernel via io_uring_enter(2). Unlike IORING_SETUP_COOP_TASKRUN, completions are only processed during io_uring-related syscalls, not on return from arbitrary syscalls. This provides the tightest and most predictable control over when completion processing occurs, as well as optimal cache behavior since work runs in the application's context.

This flag should be considered the default mode for applications setting up a ring. It requires IORING_SETUP_SINGLE_ISSUER and a ring created per-thread. The application must regularly call io_uring_enter(2) (via io_uring_submit(3), io_uring_wait_cqe(3), or similar) to process deferred work; failing to do so will stall completions.

Some features require this flag:

• Ring resizing (io_uring_register_resize_rings(3))
• Zero-copy receive (IORING_OP_RECV_ZC)

Hint that only one task will submit requests to this ring. Enables internal optimizations including reduced locking overhead. The first task to submit a request becomes the designated submitter; others attempting to submit will get -EEXIST.

Each thread or task having its own ring is the idiomatic use case for io_uring. Sharing a ring between multiple threads or tasks is discouraged as it requires additional synchronization and prevents many optimizations. Applications should create a ring per thread rather than sharing rings.

Override the default completion queue size. By default, the CQ has twice as many entries as the SQ. Set cq_entries in struct io_uring_params to specify a custom CQ size. Must be a power of 2.

Larger CQ sizes are useful when the application may submit many requests before processing completions, avoiding CQ overflow.

Clamp the SQ and CQ sizes to the maximum allowed values instead of returning -EINVAL if the requested sizes are too large. Useful when the application wants the largest possible rings without querying limits.

Use 128-byte SQEs instead of the default 64 bytes. Required for some operations that need extra space, such as IORING_OP_URING_CMD passthrough commands.

Use 32-byte CQEs instead of the default 16 bytes. Required for operations that return extra data, such as some passthrough commands or when using IORING_OP_MSG_RING.

Do not create the SQ array. The SQ array is a level of indirection that allows SQEs to be submitted in a different order than they appear in the ring. Most applications submit SQEs in order and do not need this. This flag saves memory and is required for some modes like IORING_SETUP_REGISTERED_FD_ONLY.

Use non-circular submission queue mode. The kernel ignores the SQ head and tail pointers and instead fetches SQEs starting from index 0 on each submit. The application places all SQEs at the beginning of the ring before calling io_uring_enter(2), and the sq_entries parameter determines how many SQEs are submitted.

Requires IORING_SETUP_NO_SQARRAY. Not compatible with IORING_SETUP_SQPOLL.

This mode keeps SQEs hot in cache by always accessing the same memory locations at the start of the ring, improving performance for workloads that submit small batches frequently.

Allow the ring to return a mix of 16-byte and 32-byte CQEs, controlled per-request. When a request needs a 32-byte CQE, it sets IOSQE_BIG_CQE in its flags. Otherwise, a 16-byte CQE is used. Requires IORING_SETUP_CQE32.

This is useful when certain operations require 32-byte CQEs (such as some passthrough commands) but most operations do not. Using mixed mode instead of IORING_SETUP_CQE32 alone provides efficiency benefits in terms of memory bandwidth and usage, since the smaller 16-byte CQEs are used for operations that do not need the extra space.

Allow the ring to accept a mix of 64-byte and 128-byte SQEs. When a request needs a 128-byte SQE, it sets IOSQE_BIG_SQE in its flags. Requires IORING_SETUP_SQE128.

This is useful when certain operations require 128-byte SQEs (such as IORING_OP_URING_CMD) but most operations do not. Using mixed mode instead of IORING_SETUP_SQE128 alone provides efficiency benefits in terms of memory bandwidth and usage, since the smaller 64-byte SQEs are used for operations that do not need the extra space.

The application provides its own memory for the rings instead of the kernel allocating and the application mmap'ing it. The application fills in sq_off.user_addr, cq_off.user_addr, and sq_sqes.user_addr in struct io_uring_params with addresses of application-allocated memory.

This is useful for placing rings in specific memory (huge pages, shared memory, etc.) or for creating rings without mmap.

The ring file descriptor is not installed in the process's file descriptor table. Instead, a "registered ring" index is returned in ring_fd that can be used with io_uring_enter(2) when IORING_ENTER_REGISTERED_RING is set. This reduces per-operation overhead.

Requires IORING_SETUP_NO_SQARRAY. The application must use io_uring_register_ring_fd(3) to use the ring or access it via the registered index.

Create the ring in a disabled state. The ring will not accept submissions until it is enabled via io_uring_enable_rings(3). This is useful when setting up restrictions or registered resources before allowing I/O. See io_uring_register_restrictions(3).

Continue processing submissions even if one fails. Normally, if an SQE fails during submission (not execution), subsequent SQEs in the same submit call are not processed. With this flag, all SQEs are processed regardless of earlier failures.

The failed SQE still generates a CQE with the error; this flag only affects whether subsequent SQEs are submitted. This is probably the behavior most applications expect, since CQEs are generated for failed submissions anyway and the application must handle them regardless.

Share the async worker thread pool with another ring. Set wq_fd in struct io_uring_params to the file descriptor of the ring to share with. This reduces resource usage when an application uses multiple rings.

When combined with IORING_SETUP_SQPOLL, the SQPOLL thread is also shared.

        IORING_SETUP_DEFER_TASKRUN |
        IORING_SETUP_COOP_TASKRUN

This combination provides the best latency and throughput for applications where each thread has its own ring and processes completions in a dedicated event loop.

        IORING_SETUP_SINGLE_ISSUER |
        IORING_SETUP_DEFER_TASKRUN

For NVMe or other devices that support polling, this eliminates interrupt overhead. Combined with DEFER_TASKRUN for optimal completion handling.

        IORING_SETUP_SQ_AFF

For workloads that benefit from eliminating submission syscall overhead. See io_uring_sqpoll(7).

/* First ring */
p1.flags = IORING_SETUP_SQPOLL;
/* Subsequent rings */
p2.flags = IORING_SETUP_SQPOLL | IORING_SETUP_ATTACH_WQ;
p2.wq_fd = ring1_fd;

Reduces kernel thread and workqueue overhead when using multiple rings.