|
HTML rendering created 2025-09-06 by Michael Kerrisk, author of The Linux Programming Interface. For details of in-depth Linux/UNIX system programming training courses that I teach, look here. Hosting by jambit GmbH. |
|
|
NAME | SYNOPSIS | DESCRIPTION | CONFORMING TO | EXAMPLES | SEE ALSO | COLOPHON |
|
|
|
io_uring(7) Linux Programmer's Manual io_uring(7)
io_uring - Asynchronous I/O facility
#include <linux/io_uring.h>
io_uring is a Linux-specific API for asynchronous I/O. It allows
the user to submit one or more I/O requests, which are processed
asynchronously without blocking the calling process. io_uring
gets its name from ring buffers which are shared between user
space and kernel space. This arrangement allows for efficient I/O,
while avoiding the overhead of copying buffers between them, where
possible. This interface makes io_uring different from other UNIX
I/O APIs, wherein, rather than just communicate between kernel and
user space with system calls, ring buffers are used as the main
mode of communication. This arrangement has various performance
benefits which are discussed in a separate section below. This
man page uses the terms shared buffers, shared ring buffers and
queues interchangeably.
The general programming model you need to follow for io_uring is
outlined below
• Set up shared buffers with io_uring_setup(2) and mmap(2),
mapping into user space shared buffers for the submission
queue (SQ) and the completion queue (CQ). You place I/O
requests you want to make on the SQ, while the kernel
places the results of those operations on the CQ.
• For every I/O request you need to make (like to read a
file, write a file, accept a socket connection, etc), you
create a submission queue entry, or SQE, describe the I/O
operation you need to get done and add it to the tail of
the submission queue (SQ). Each I/O operation is, in
essence, the equivalent of a system call you would have
made otherwise, if you were not using io_uring. For
instance, a SQE with the opcode set to IORING_OP_READ will
request a read operation to be issued that is similar to
the read(2) system call. Refer to the opcode documentation
in io_uring_enter(2) for all supported opcodes and their
properties. You can add more than one SQE to the queue
depending on the number of operations you want to request.
• After you add one or more SQEs, you need to call
io_uring_enter(2) to tell the kernel to dequeue your I/O
requests off the SQ and begin processing them.
• For each SQE you submit, once it is done processing the
request, the kernel places a completion queue event or CQE
at the tail of the completion queue or CQ. The kernel
places exactly one matching CQE in the CQ for every SQE you
submit on the SQ. After you retrieve a CQE, minimally, you
might be interested in checking the res field of the CQE
structure, which corresponds to the return value of the
system call's equivalent, had you used it directly without
using io_uring. Given that io_uring is an async interface,
errno is never used for passing back error information.
Instead, res will contain what the equivalent system call
would have returned in case of success, and in case of
error res will contain -errno. For example, if the normal
read system call would have returned -1 and set errno to
EINVAL, then res would contain -EINVAL. If the normal
system call would have returned a read size of 1024, then
res would contain 1024.
• Optionally, io_uring_enter(2) can also wait for a specified
number of requests to be processed by the kernel before it
returns. If you specified a certain number of completions
to wait for, the kernel would have placed at least those
many number of CQEs on the CQ, which you can then readily
read, right after the return from io_uring_enter(2).
• It is important to remember that I/O requests submitted to
the kernel can complete in any order. It is not necessary
for the kernel to process one request after another, in the
order you placed them. Given that the interface is a ring,
the requests are attempted in order, however that doesn't
imply any sort of ordering on their execution or
completion. When more than one request is in flight, it is
not possible to determine which one will execute or
complete first. When you dequeue CQEs off the CQ, you
should always check which submitted request it corresponds
to. The most common method for doing so is utilizing the
user_data field in the request, which is passed back on the
completion side.
• Concretely, for operations where strict ordering is
required, such as for sends and receives on a stream-
oriented TCP socket, it is generally unsafe to have more
than one outstanding send, or more than one outstanding
receive (the two directions are independent) on a given
socket at a time, as the kernel may reorder their execution
if poll arming or other background kernel activities are
involved. However, io_uring provides various facilities to
enable applications to efficiently pipeline their
operations safely. If the requests are submitted in a
single batch, the application may use IOSQE_IO_LINK to
enforce an execution order in the kernel. Otherwise,
io_uring provides advanced features like multi shot and
send/receive bundles to allow applications to provide more
data in fewer, more efficient trips to the kernel. Even if
these features are used, applications must still ensure
they do not overlap different sends or different receives
on a given file.
Adding to and reading from the queues:
• You add SQEs to the tail of the SQ. The kernel reads SQEs
off the head of the queue.
• The kernel adds CQEs to the tail of the CQ. You read CQEs
off the head of the queue.
It should be noted that depending on the configuration io_uring's
behavior can deviate from the behavior outlined above, like not
posting a CQE for every SQE when setting IOSQE_CQE_SKIP_SUCCESS in
the SQE or posting multiple CQEs for a single SQE for multi shot
operations or requiring an io_uring_enter(2) syscall to make the
kernel begin processing newly added SQEs when using submission
queue polling.
Submission queue polling
One of the goals of io_uring is to provide a means for efficient
I/O. To this end, io_uring supports a polling mode that lets you
avoid the call to io_uring_enter(2), which you use to inform the
kernel that you have queued SQEs on to the SQ. With SQ Polling,
io_uring starts a kernel thread that polls the submission queue
for any I/O requests you submit by adding SQEs. With SQ Polling
enabled, there is no need for you to call io_uring_enter(2),
letting you avoid the overhead of system calls. A designated
kernel thread dequeues SQEs off the SQ as you add them and
dispatches them for asynchronous processing.
Setting up io_uring
The main steps in setting up io_uring consist of mapping in the
shared buffers with mmap(2) calls. In the example program
included in this man page, the function app_setup_uring() sets up
io_uring with a QUEUE_DEPTH deep submission queue. Pay attention
to the 2 mmap(2) calls that set up the shared submission and
completion queues. If your kernel is older than version 5.4,
three mmap(2) calls are required.
Submitting I/O requests
The process of submitting a request consists of describing the I/O
operation you need to get done using an io_uring_sqe structure
instance. These details describe the equivalent system call and
its parameters. Because the range of I/O operations Linux
supports are very varied and the io_uring_sqe structure needs to
be able to describe them, it has several fields, some packed into
unions for space efficiency. Here is a simplified version of
struct io_uring_sqe with some of the most often used fields:
struct io_uring_sqe {
__u8 opcode; /* type of operation for this sqe */
__s32 fd; /* file descriptor to do IO on */
__u64 off; /* offset into file */
__u64 addr; /* pointer to buffer or iovecs */
__u32 len; /* buffer size or number of iovecs */
__u64 user_data; /* data to be passed back at completion time */
__u8 flags; /* IOSQE_ flags */
...
};
Here is struct io_uring_sqe in full:
struct io_uring_sqe {
__u8 opcode; /* type of operation for this sqe */
__u8 flags; /* IOSQE_ flags */
__u16 ioprio; /* ioprio for the request */
__s32 fd; /* file descriptor to do IO on */
union {
__u64 off; /* offset into file */
__u64 addr2;
struct {
__u32 cmd_op;
__u32 __pad1;
};
};
union {
__u64 addr; /* pointer to buffer or iovecs */
__u64 splice_off_in;
struct {
__u32 level;
__u32 optname;
};
};
__u32 len; /* buffer size or number of iovecs */
union {
__kernel_rwf_t rw_flags;
__u32 fsync_flags;
__u16 poll_events; /* compatibility */
__u32 poll32_events; /* word-reversed for BE */
__u32 sync_range_flags;
__u32 msg_flags;
__u32 timeout_flags;
__u32 accept_flags;
__u32 cancel_flags;
__u32 open_flags;
__u32 statx_flags;
__u32 fadvise_advice;
__u32 splice_flags;
__u32 rename_flags;
__u32 unlink_flags;
__u32 hardlink_flags;
__u32 xattr_flags;
__u32 msg_ring_flags;
__u32 uring_cmd_flags;
__u32 waitid_flags;
__u32 futex_flags;
__u32 install_fd_flags;
__u32 nop_flags;
};
__u64 user_data; /* data to be passed back at completion time */
/* pack this to avoid bogus arm OABI complaints */
union {
/* index into fixed buffers, if used */
__u16 buf_index;
/* for grouped buffer selection */
__u16 buf_group;
} __attribute__((packed));
/* personality to use, if used */
__u16 personality;
union {
__s32 splice_fd_in;
__u32 file_index;
__u32 optlen;
struct {
__u16 addr_len;
__u16 __pad3[1];
};
};
union {
struct {
__u64 addr3;
__u64 __pad2[1];
};
__u64 optval;
/*
* If the ring is initialized with IORING_SETUP_SQE128, then
* this field is used for 80 bytes of arbitrary command data
*/
__u8 cmd[0];
};
};
To submit an I/O request to io_uring, you need to acquire a
submission queue entry (SQE) from the submission queue (SQ), fill
it up with details of the operation you want to submit and call
io_uring_enter(2). There are helper functions of the form
io_uring_prep_X to enable proper setup of the SQE. If you want to
avoid calling io_uring_enter(2), you have the option of setting up
Submission Queue Polling.
SQEs are added to the tail of the submission queue. The kernel
picks up SQEs off the head of the SQ. The general algorithm to
get the next available SQE and update the tail is as follows.
struct io_uring_sqe *sqe;
unsigned tail, index;
tail = *sqring->tail;
index = tail & (*sqring->ring_mask);
sqe = &sqring->sqes[index];
/* fill up details about this I/O request */
describe_io(sqe);
/* fill the sqe index into the SQ ring array */
sqring->array[index] = index;
tail++;
atomic_store_explicit(sqring->tail, tail, memory_order_release);
To get the index of an entry, the application must mask the
current tail index with the size mask of the ring. This holds
true for both SQs and CQs. Once the SQE is acquired, the
necessary fields are filled in, describing the request. While the
CQ ring directly indexes the shared array of CQEs, the submission
side has an indirection array between them. The submission side
ring buffer is an index into this array, which in turn contains
the index into the SQEs.
The following code snippet demonstrates how a read operation, an
equivalent of a preadv2(2) system call is described by filling up
an SQE with the necessary parameters.
struct iovec iovecs[16];
...
sqe->opcode = IORING_OP_READV;
sqe->fd = fd;
sqe->addr = (unsigned long) iovecs;
sqe->len = 16;
sqe->off = offset;
sqe->flags = 0;
Memory ordering
Modern compilers and CPUs freely reorder reads and writes
without affecting the program's outcome to optimize
performance. Some aspects of this need to be kept in mind
on SMP systems since io_uring involves buffers shared
between kernel and user space. These buffers are both
visible and modifiable from kernel and user space. As
heads and tails belonging to these shared buffers are
updated by kernel and user space, changes need to be
coherently visible on either side, irrespective of whether
a CPU switch took place after the kernel-user mode switch
happened. We use memory barriers to enforce this
coherency. Being significantly large subjects on their
own, memory barriers are out of scope for further
discussion on this man page. For more information on
modern memory models the reader may refer to the
Documentation/memory-barriers.txt in the kernel tree or to
the documentation of the formal C11 or kernel memory model.
Letting the kernel know about I/O submissions
Once you place one or more SQEs on to the SQ, you need to
let the kernel know that you've done so. You can do this
by calling the io_uring_enter(2) system call. This system
call is also capable of waiting for a specified count of
events to complete. This way, you can be sure to find
completion events in the completion queue without having to
poll it for events later.
Reading completion events
Similar to the submission queue (SQ), the completion queue (CQ) is
a shared buffer between the kernel and user space. Whereas you
placed submission queue entries on the tail of the SQ and the
kernel read off the head, when it comes to the CQ, the kernel
places completion queue events or CQEs on the tail of the CQ and
you read off its head.
Submission is flexible (and thus a bit more complicated) since it
needs to be able to encode different types of system calls that
take various parameters. Completion, on the other hand is simpler
since we're looking only for a return value back from the kernel.
This is easily understood by looking at the completion queue event
structure, struct io_uring_cqe:
struct io_uring_cqe {
__u64 user_data; /* sqe->data submission passed back */
__s32 res; /* result code for this event */
__u32 flags;
};
Here, user_data is custom data that is passed unchanged from
submission to completion. That is, from SQEs to CQEs. This field
can be used to set context, uniquely identifying submissions that
got completed. Given that I/O requests can complete in any order,
this field can be used to correlate a submission with a
completion. res is the result from the system call that was
performed as part of the submission; its return value.
The flags field carries request-specific information. As of the
6.12 kernel, the following flags are defined:
IORING_CQE_F_BUFFER
If set, the upper 16 bits of the flags field carries the
buffer ID that was chosen for this request. The request
must have been issued with IOSQE_BUFFER_SELECT set, and
used with a request type that supports buffer selection.
Additionally, buffers must have been provided upfront
either via the IORING_OP_PROVIDE_BUFFERS or the
IORING_REGISTER_PBUF_RING methods.
IORING_CQE_F_MORE
If set, the application should expect more completions from
the request. This is used for requests that can generate
multiple completions, such as multi-shot requests, receive,
or accept.
IORING_CQE_F_SOCK_NONEMPTY
If set, upon receiving the data from the socket in the
current request, the socket still had data left on
completion of this request.
IORING_CQE_F_NOTIF
Set for notification CQEs, as seen with the zero-copy
networking send and receive support.
IORING_CQE_F_BUF_MORE
If set, the buffer ID set in the completion will get more
completions. This means that the provided buffer has been
partially consumed and there's more buffer space left, and
hence the application should expect more completions with
this buffer ID. Each completion will continue where the
previous one left off. This can only happen if the provided
buffer ring has been setup with IOU_PBUF_RING_INC to allow
for incremental / partial consumption of buffers.
The general sequence to read completion events off the completion
queue is as follows:
unsigned head;
head = *cqring->head;
if (head != atomic_load_acquire(cqring->tail)) {
struct io_uring_cqe *cqe;
unsigned index;
index = head & (cqring->mask);
cqe = &cqring->cqes[index];
/* process completed CQE */
process_cqe(cqe);
/* CQE consumption complete */
head++;
}
atomic_store_explicit(cqring->head, head, memory_order_release);
It helps to be reminded that the kernel adds CQEs to the tail of
the CQ, while you need to dequeue them off the head. To get the
index of an entry at the head, the application must mask the
current head index with the size mask of the ring. Once the CQE
has been consumed or processed, the head needs to be updated to
reflect the consumption of the CQE. Attention should be paid to
the read and write barriers to ensure successful read and update
of the head.
io_uring performance
Because of the shared ring buffers between kernel and user space,
io_uring can be a zero-copy system. Copying buffers to and from
becomes necessary when system calls that transfer data between
kernel and user space are involved. But since the bulk of the
communication in io_uring is via buffers shared between the kernel
and user space, this huge performance overhead is completely
avoided.
While system calls may not seem like a significant overhead, in
high performance applications, making a lot of them will begin to
matter. While workarounds the operating system has in place to
deal with Spectre and Meltdown are ideally best done away with,
unfortunately, some of these workarounds are around the system
call interface, making system calls not as cheap as before on
affected hardware. While newer hardware should not need these
workarounds, hardware with these vulnerabilities can be expected
to be in the wild for a long time. While using synchronous
programming interfaces or even when using asynchronous programming
interfaces under Linux, there is at least one system call involved
in the submission of each request. In io_uring, on the other
hand, you can batch several requests in one go, simply by queueing
up multiple SQEs, each describing an I/O operation you want and
make a single call to io_uring_enter(2). This is possible due to
io_uring's shared buffers based design.
While this batching in itself can avoid the overhead associated
with potentially multiple and frequent system calls, you can
reduce even this overhead further with Submission Queue Polling,
by having the kernel poll and pick up your SQEs for processing as
you add them to the submission queue. This avoids the
io_uring_enter(2) call you need to make to tell the kernel to pick
SQEs up. For high-performance applications, this means even fewer
system call overheads.
io_uring is Linux-specific.
The following example uses io_uring to copy stdin to stdout.
Using shell redirection, you should be able to copy files with
this example. Because it uses a queue depth of only one, this
example processes I/O requests one after the other. It is
purposefully kept this way to aid understanding. In real-world
scenarios however, you'll want to have a larger queue depth to
parallelize I/O request processing so as to gain the kind of
performance benefits io_uring provides with its asynchronous
processing of requests.
#include <stdio.h>
#include <stdlib.h>
#include <sys/stat.h>
#include <sys/ioctl.h>
#include <sys/syscall.h>
#include <sys/mman.h>
#include <sys/uio.h>
#include <linux/fs.h>
#include <fcntl.h>
#include <unistd.h>
#include <string.h>
#include <stdatomic.h>
#include <linux/io_uring.h>
#define QUEUE_DEPTH 1
#define BLOCK_SZ 1024
/* Macros for barriers needed by io_uring */
#define io_uring_smp_store_release(p, v) \
atomic_store_explicit((_Atomic typeof(*(p)) *)(p), (v), \
memory_order_release)
#define io_uring_smp_load_acquire(p) \
atomic_load_explicit((_Atomic typeof(*(p)) *)(p), \
memory_order_acquire)
int ring_fd;
unsigned *sring_tail, *sring_mask, *sring_array,
*cring_head, *cring_tail, *cring_mask;
struct io_uring_sqe *sqes;
struct io_uring_cqe *cqes;
char buff[BLOCK_SZ];
off_t offset;
/*
* System call wrappers provided since glibc does not yet
* provide wrappers for io_uring system calls.
* */
int io_uring_setup(unsigned entries, struct io_uring_params *p)
{
int ret;
ret = syscall(__NR_io_uring_setup, entries, p);
return (ret < 0) ? -errno : ret;
}
int io_uring_enter(int ring_fd, unsigned int to_submit,
unsigned int min_complete, unsigned int flags)
{
int ret;
ret = syscall(__NR_io_uring_enter, ring_fd, to_submit,
min_complete, flags, NULL, 0);
return (ret < 0) ? -errno : ret;
}
int app_setup_uring(void) {
struct io_uring_params p;
void *sq_ptr, *cq_ptr;
/* See io_uring_setup(2) for io_uring_params.flags you can set */
memset(&p, 0, sizeof(p));
ring_fd = io_uring_setup(QUEUE_DEPTH, &p);
if (ring_fd < 0) {
perror("io_uring_setup");
return 1;
}
/*
* io_uring communication happens via 2 shared kernel-user space ring
* buffers, which can be jointly mapped with a single mmap() call in
* kernels >= 5.4.
*/
int sring_sz = p.sq_off.array + p.sq_entries * sizeof(unsigned);
int cring_sz = p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe);
/* Rather than check for kernel version, the recommended way is to
* check the features field of the io_uring_params structure, which is a
* bitmask. If IORING_FEAT_SINGLE_MMAP is set, we can do away with the
* second mmap() call to map in the completion ring separately.
*/
if (p.features & IORING_FEAT_SINGLE_MMAP) {
if (cring_sz > sring_sz)
sring_sz = cring_sz;
cring_sz = sring_sz;
}
/* Map in the submission and completion queue ring buffers.
* Kernels < 5.4 only map in the submission queue, though.
*/
sq_ptr = mmap(0, sring_sz, PROT_READ | PROT_WRITE,
MAP_SHARED | MAP_POPULATE,
ring_fd, IORING_OFF_SQ_RING);
if (sq_ptr == MAP_FAILED) {
perror("mmap");
return 1;
}
if (p.features & IORING_FEAT_SINGLE_MMAP) {
cq_ptr = sq_ptr;
} else {
/* Map in the completion queue ring buffer in older kernels separately */
cq_ptr = mmap(0, cring_sz, PROT_READ | PROT_WRITE,
MAP_SHARED | MAP_POPULATE,
ring_fd, IORING_OFF_CQ_RING);
if (cq_ptr == MAP_FAILED) {
perror("mmap");
return 1;
}
}
/* Save useful fields for later easy reference */
sring_tail = sq_ptr + p.sq_off.tail;
sring_mask = sq_ptr + p.sq_off.ring_mask;
sring_array = sq_ptr + p.sq_off.array;
/* Map in the submission queue entries array */
sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),
PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE,
ring_fd, IORING_OFF_SQES);
if (sqes == MAP_FAILED) {
perror("mmap");
return 1;
}
/* Save useful fields for later easy reference */
cring_head = cq_ptr + p.cq_off.head;
cring_tail = cq_ptr + p.cq_off.tail;
cring_mask = cq_ptr + p.cq_off.ring_mask;
cqes = cq_ptr + p.cq_off.cqes;
return 0;
}
/*
* Read from completion queue.
* In this function, we read completion events from the completion queue.
* We dequeue the CQE, update and head and return the result of the operation.
* */
int read_from_cq() {
struct io_uring_cqe *cqe;
unsigned head;
/* Read barrier */
head = io_uring_smp_load_acquire(cring_head);
/*
* Remember, this is a ring buffer. If head == tail, it means that the
* buffer is empty.
* */
if (head == *cring_tail)
return -1;
/* Get the entry */
cqe = &cqes[head & (*cring_mask)];
if (cqe->res < 0)
fprintf(stderr, "Error: %s\n", strerror(abs(cqe->res)));
head++;
/* Write barrier so that update to the head are made visible */
io_uring_smp_store_release(cring_head, head);
return cqe->res;
}
/*
* Submit a read or a write request to the submission queue.
* */
int submit_to_sq(int fd, int op) {
unsigned index, tail;
/* Add our submission queue entry to the tail of the SQE ring buffer */
tail = *sring_tail;
index = tail & *sring_mask;
struct io_uring_sqe *sqe = &sqes[index];
/* Fill in the parameters required for the read or write operation */
sqe->opcode = op;
sqe->fd = fd;
sqe->addr = (unsigned long) buff;
if (op == IORING_OP_READ) {
memset(buff, 0, sizeof(buff));
sqe->len = BLOCK_SZ;
}
else {
sqe->len = strlen(buff);
}
sqe->off = offset;
sring_array[index] = index;
tail++;
/* Update the tail */
io_uring_smp_store_release(sring_tail, tail);
/*
* Tell the kernel we have submitted events with the io_uring_enter()
* system call. We also pass in the IORING_ENTER_GETEVENTS flag which
* causes the io_uring_enter() call to wait until min_complete
* (the 3rd param) events complete.
* */
int ret = io_uring_enter(ring_fd, 1,1,
IORING_ENTER_GETEVENTS);
if(ret < 0) {
perror("io_uring_enter");
return -1;
}
return ret;
}
int main(int argc, char *argv[]) {
int res;
/* Setup io_uring for use */
if(app_setup_uring()) {
fprintf(stderr, "Unable to setup uring!\n");
return 1;
}
/*
* A while loop that reads from stdin and writes to stdout.
* Breaks on EOF.
*/
while (1) {
/* Initiate read from stdin and wait for it to complete */
submit_to_sq(STDIN_FILENO, IORING_OP_READ);
/* Read completion queue entry */
res = read_from_cq();
if (res > 0) {
/* Read successful. Write to stdout. */
submit_to_sq(STDOUT_FILENO, IORING_OP_WRITE);
read_from_cq();
} else if (res == 0) {
/* reached EOF */
break;
}
else if (res < 0) {
/* Error reading file */
fprintf(stderr, "Error: %s\n", strerror(abs(res)));
break;
}
offset += res;
}
return 0;
}
io_uring_enter(2) io_uring_register(2) io_uring_setup(2)
This page is part of the liburing (A library for io_uring)
project. Information about the project can be found at
⟨https://github.com/axboe/liburing⟩. If you have a bug report for
this manual page, send it to io-uring@vger.kernel.org. This page
was obtained from the project's upstream Git repository
⟨https://github.com/axboe/liburing⟩ on 2025-08-11. (At that time,
the date of the most recent commit that was found in the
repository was 2025-08-02.) If you discover any rendering
problems in this HTML version of the page, or you believe there is
a better or more up-to-date source for the page, or you have
corrections or improvements to the information in this COLOPHON
(which is not part of the original manual page), send a mail to
man-pages@man7.org
Linux 2020-07-26 io_uring(7)
Pages that refer to this page: io_uring_register(2), systemd.exec(5)
|
HTML rendering created 2025-09-06 by Michael Kerrisk, author of The Linux Programming Interface. For details of in-depth Linux/UNIX system programming training courses that I teach, look here. Hosting by jambit GmbH. |
|