PHP: rfc:io_ring

This page is read only. You can view the source, but not change it. Ask your administrator if you think this is wrong.
====== PHP RFC: Ring API ======
  * Version: 0.1
  * Date: 2026-09-29
  * Author: Jakub Zelenka, bukka@php.net
  * Status: Draft
  * First Published at: https://wiki.php.net/rfc/io_ring


===== Introduction =====

The [[rfc:io_hooks|IO hooks and operations RFC]] lets a provider, typically an event loop or a Fiber
scheduler, execute the blocking operations of PHP's stream layer, so that a Fiber suspends inside
''fread()'' or ''curl_exec()'' instead of blocking the process. It ships one executor for those operations,
''Io\Poll\OperationQueue'', built on the Poll API. That executor can tell when a socket or a pipe is ready and
let the engine perform the read, but it can do nothing about operations that have no notion of readiness, such as a
read of a regular file, an ''fsync()'' or a host name lookup. Under such a provider those still block the whole
process, exactly as today.

This RFC adds the second executor, the **Ring**: an operation queue that performs operations itself, on the
platform's native asynchronous IO where there is one and on a thread pool where there is none. Under a
provider on the Ring, ''file_get_contents()'' on a slow disk, a lookup against a slow resolver and an
''fsync()'' suspend the calling Fiber the way a socket read does.

==== Readiness and completion ====

There are two ways an operating system can help a program do many IO operations at once. A **readiness**
interface (''select'', ''poll'', epoll, kqueue) answers the question "which of these descriptors can I operate
on right now without blocking?", and the program then performs the operation itself. A **completion**
interface (io_uring on Linux, IO completion ports on Windows) takes the whole operation, "read 8 KiB from
this descriptor into this buffer", performs it in the background, and reports when it is done.

The completion model covers everything the readiness model does, and also what it cannot. Regular files
are always "ready" and always block, and so are library calls that are not descriptor operations at all, such
as a name lookup. It also batches, so that under load many operations become a single system call. io_uring is the
Linux form of this model and has been the direction of Linux IO since kernel 5.1. This RFC is designed around
it, and the other backends are its portable form.

==== The library ====

The Ring is implemented on top of [[https://github.com/libior/ior|ior]], a C library with an API shaped like
liburing and three backends: io_uring on Linux, IO completion ports on Windows, and a thread pool on the
other POSIX platforms, which is also the fallback on Linux where io_uring exists but cannot be used (a
kernel older than 5.19, ''io_uring_disabled'', a seccomp profile such as Docker's default, a memory lock
limit). ior is bundled
with PHP the way pcre2 and other libraries are, so the Ring is available on every platform PHP builds on and
needs no external dependency.

libuv, the usual answer to portable asynchronous IO, is deliberately not used. It is a full event loop with
its own callback style and its own threads, it is readiness first with a thread pool bolted on for files, and
a design that inherits it inherits that model. The Poll API plus ior's thread pool cover everything libuv
would, and the completion model here is io_uring's.

===== Terminology =====

The terms of the hooks RFC (provider, operation, completion, operation queue, capability, registration)
apply here unchanged. The Ring adds a few of its own.

**Backend.** The mechanism the Ring runs on, which is io_uring, IO completion ports (IOCP) or the thread pool. ior
picks one when the Ring is created and the choice is reported, not requested.

**Submission and completion queue.** The two halves of a ring. Operations are placed in the submission queue
and handed to the backend in a batch. The backend places their outcomes in the completion queue, from which
the provider reaps them. The **depth** is how many entries fit in one batch. It does not bound how many
operations may be in flight.

**Offloading.** Performing an operation off the calling thread, in the kernel or on a worker thread, so the
Fiber that asked for it can be suspended meanwhile. The Ring offloads everything it is handed. The Poll
queue offloads nothing and only waits for readiness.

**Multishot.** A single submission that produces many completions. One multishot accept on a listening
socket reports every incoming connection, one multishot poll on a descriptor reports every change of its
readiness. The Ring uses them to serve repeated waits without a submission per wait.

**Notification handle.** A Poll API handle owned by the Ring that becomes readable whenever a completion is
posted, so that a program which waits on a Poll context can be woken by the Ring without waiting on the Ring
itself.

**Orphaned operation.** An operation whose caller went away, because its Fiber was destroyed while
suspended, while the backend still holds its buffer. The Ring keeps it until the backend is done and drops
the result.

===== Proposal =====

==== The Engine ====

''Io\Ring\Engine'' is an ''Io\OperationQueue'', the interface the hooks RFC defines, with the same ''submit()'',
''cancel()'', ''waitCompletions()'' and ''countPending()''. A provider written against the interface works on it
unchanged. The minimal scheduler of the hooks RFC becomes a completion based scheduler by constructing it
with ''new Io\Ring\Engine()'' instead of ''new Io\Poll\OperationQueue()''.

Where the Poll queue turns an operation into a watcher and completes it as ''Ready'' for the engine to
perform, the Ring performs it and completes it as ''Done'':

^ Operation ^ On the Ring ^
| ''Poll'', ''Timer'', ''Any'' | a poll or timeout submission. An Any is its members submitted together, and the Ring cancels the losers when one completes |
| ''Read'', ''Write'' | performed by the backend: the kernel's worker pool for a regular file on io_uring, an overlapped read on IOCP, a pool thread on the thread backend |
| ''Recv'', ''Send'', ''Accept'', ''Connect'' | performed by the backend, inline during the submit when the socket is ready on the native backends |
| ''GetAddrInfo'', ''GetNameInfo'', ''Fsync'' | a job on ior's worker pool, which every backend has for such calls, the native ones included |
| ''WaitPid'', ''SigWait'' | native on every backend |

It is named ''Engine'' rather than ''OperationQueue'', as ''Random\Engine'' is. It implements the queue
interface, but it also owns the ring, chooses the backend, executes the operations itself and exposes a
notification handle, where ''Io\Poll\OperationQueue'' only turns operations into watchers.

==== Backend selection ====

The backend is not selectable from PHP. ior picks the best one for the platform and ''getBackend()'' reports
which. The test suite can force a backend so that every backend the platform has is exercised, but that
is a testing facility and not part of the API.

==== Capabilities ====

The hooks RFC lets a provider report what it can do beyond waiting for readiness. A Ring can serve more than
the Poll queue, and ''Engine'' reports it in two sets. ''getHookCapabilities()'' is what a provider on the Ring
should report by default, which is ''EdgeRegistrations'' where the backend's multishot poll reports edges, which is
io_uring and the thread pool but not IOCP, and ''DirectAccept''. ''getSupportedHookCapabilities()'' is what the
backend can serve for a provider that opts in, which adds ''Files'' on every backend, and ''DirectData'' on
the native backends, where a receive on a ready socket is performed inline during the submit.

The default is deliberately not everything the Ring can do, because the measurements below show that the
rest loses in the common case. ''DirectData'' hands every data operation to the provider before the engine
tries the system call. On io_uring that is one system call either way, but for a provider written in PHP it
adds a round trip through userland per operation, which loses whenever the data is already there. It fits a
provider written in C. ''Files'' hands regular file reads and writes to the Ring. A cached read is a memory
copy, and the Ring turns it into a submission, a completion and a Fiber switch, so it pays off only for
storage known to be slow, such as network file systems, or an ''fsync()'' heavy workload.

''DirectAccept'' is in the default because the Ring serves it from a multishot accept on the listening socket.
Accepted connections wait in a buffer, and an accept that finds one completes at submit, so the accepting
Fiber takes a burst of connections without a pass of the scheduler per connection. A direct accept without
that took one connection per pass and made the last connection of a burst wait for every pass before it,
measured as a p99 of about a second at a thousand connections. With the multishot accept it is 23 ms, in
line with the Poll queue.

==== Embedding a Ring in a Poll based loop ====

A loop that wants to stay on its own ''Io\Poll\Context'' and still take file operations off the thread does
not have to move to the Ring. ''Engine::getHandle()'' returns the notification handle, an ''Io\Poll\NotifyHandle''
that is raised for every posted completion, so the Ring can be added to the context with ''Event::Notify''
like any other handle, on every platform. When it fires, the loop calls ''waitCompletions()'' with a zero
timeout until it returns an empty array. The first call clears the handle, and what is left unreaped after
that is not announced again. The handle is the Ring's to signal, so ''notify()'' throws on it.

==== Cancellation and abandoned operations ====

On the Poll queue giving up on an operation is immediate. The watcher is removed and the operation is over.
On the Ring an operation may have handed a buffer to the kernel or a worker thread, and the buffer stays in
use until the backend reports the operation finished, cancelled or not. The Ring hides this from the
provider. After ''cancel()'' returns, the provider is never handed a completion for the operation, and the Ring
consumes the backend's completion itself. A Fiber destroyed while suspended in an operation leaves the
operation orphaned. The Ring keeps a reference on the stream so the buffer stays valid, cancels the request,
and finishes it silently later. The stream stays frozen until then and bytes already received are dropped.
A provider needs no ''finally'' for any of this, and a Ring destroyed with operations in flight cancels them
all and waits for each before it exits.

Every input the backend touches (addresses, host names, signal sets) is copied into the Ring's own record
at submit, and outputs reach the operation only when its completion is delivered, so a cancelled or orphaned
operation never writes into memory that is gone. Only the stream's own read buffer is handed to the backend
directly, so that the buffered read path makes no copies.

==== Fork ====

A Ring inherited by a forked child is refused, not reused. The io_uring instance belongs to the parent and
the thread pool's workers do not exist in the child. Every ''Engine'' method throws ''RingException'' there, and
the child closes its copies of the Ring's descriptors right after ''pcntl_fork()'' without touching the
parent's Ring. ''pcntl_fork()'' itself throws while any operation is in flight, as the hooks RFC specifies.

==== Shared listeners ====

A multishot accept takes connections off the kernel's backlog early. For a listening socket shared by several
processes, workers forked after ''listen()'', that pulls connections into one process where an idle sibling
cannot take them. The stream context option ''accept_multishot'' under ''socket'' keeps such a listener out of
it, so its accepts stay one at a time. Sockets bound with ''SO_REUSEPORT'' are separate per process and need
nothing.

<code php>
<?php

$ctx = stream_context_create(['socket' => ['accept_multishot' => false]]);
$server = stream_socket_server('tcp://0.0.0.0:8080', $errno, $errstr,
    STREAM_SERVER_BIND | STREAM_SERVER_LISTEN, $ctx);
</code>

==== Windows ====

IO completion ports can only complete operations on files opened for overlapped IO, a mode that cannot be
set after the fact and that PHP's file functions never used. The plain file wrapper therefore gets an
overlapped path of its own, used for files opened while a provider with the ''Files'' capability is
installed. Everything else keeps the C runtime path and is read synchronously. Only sockets can be polled on
IOCP, so a Poll operation on anything else completes as ''Unsupported'' and the engine falls back as the hooks
RFC specifies, and the Ring does not offer ''EdgeRegistrations'' there, since IOCP's multishot poll reports
persisting readiness rather than changes.

==== Poll or Ring ====

The two queues are not a fallback and a goal. Each is the better executor for a class of workloads, and a
provider picks by what it runs and where.

The Poll queue is the better choice where io_uring is unavailable or unwanted, which includes containers
under Docker's default seccomp profile and distributions that ship io_uring disabled. There ior falls back
to the thread pool, and a handoff to a worker and back per socket operation is worse than epoll with the
syscall-first attempt. It is also the better choice where the process should have no threads, which
matters for CLI scripts and FPM workers. For socket-bound workloads it is on par, since readiness with the
syscall-first attempt is one system call per operation, the same as io_uring's inline completion, and it
needs none of the buffer lifetime machinery. An existing loop is easier to adapt to it, since Revolt, AMPHP
and ReactPHP are readiness loops. And where maturity matters it is the conservative choice, since the Poll
API sits on mechanisms that are decades old and the Ring is new code over a new library.

The Ring is the better choice where regular files matter, since a file read has no readiness form and only
a completion executor can take it off the calling thread. The same holds where blocking library calls
matter, since name lookups and ''fsync()'' have a home only in a work pool. On Windows the Ring is the
native model, since IOCP is what the platform offers and WSAPoll does not scale. And under load a ring turns
many operations into one system call, where a readiness loop pays one per operation.

The two combine. A Poll based loop on Unix embeds a Ring for file and DNS operations through its
notification handle, and a Ring based loop serves every readiness consumer through poll operations.

==== Measurements ====

A benchmark project (release builds on Linux 6.8, a Fiber scheduler over each queue, CPU pinning, repeated
runs with significance tests) compared the queues. Throughput is relative to the Poll queue. "syscall-first" is
the Ring's default, "Direct" has ''DirectData'' and ''DirectAccept'' with one-shot accepts:

^ Scenario ^ Ring io_uring, Direct ^ Ring io_uring, syscall-first ^ Ring threads ^
| Ping-pong over a socket pair | -14% | +8.5% | -91% |
| Unbuffered 64 byte reads, unix / tcp | -69% / -88% | +5.5% / -0.5% | +3% / -5% |
| Pipes, 64 byte / 64 KiB | +56% / +36% | +57% / +32% | -81% / -82% |
| Echo, 100 / 1000 / 10000 connections | -17% / +3% / -16% | -2% / -1.5% / +1% | -92% / -88% / -63% |
| HTTP server, 100 connections | -8.5% | -5.5% | -86% |
| Cached regular files with ''Files'' | -54% | (''Files'' off: as Poll) | -55% |
| DNS lookups | -44% | -45% | -12% |
| 1000 Fibers sleeping 1 ms | -63% | | -80% |
| ''stream_select()'' over 400 streams | +24% | -2% | -78% |
| ''fsync()'' with concurrent writers | | +264% | |

The defaults follow from these numbers. Syscall-first io_uring is on par with the Poll queue for sockets
and ahead on pipes, ''fsync()'' and large selects. Direct operations lose whenever data is usually there, so
they are opt-in. Cached file reads lose on the Ring, so ''Files'' is opt-in. The thread backend is far behind
on sockets, which is what the Poll queue is for where io_uring is unavailable. The thread backend column
predates a rework of that backend in ior. Measured again after it, syscall-first on threads is 50 to 85%
behind the Poll queue on sockets.

===== API =====

<code php>
<?php

namespace Io\Ring;

/** Which backend ior chose. Informational: there is no way to request one. */
enum Backend
{
    case IoUring;
    case Iocp;
    case Threads;
}

/**
 * The Ring: an operation queue that executes every operation itself, on
 * io_uring, IO completion ports or ior's thread pool. Operations reach it
 * from a provider's run(); a provider on it is a completion based scheduler.
 */
final class Engine implements \Io\OperationQueue
{
    /**
     * Submission queue depth; 0 is the default (256). It sizes the batches
     * handed to the backend and does not bound the operations in flight.
     */
    public function __construct(int $entries = 0) {}

    public function getBackend(): Backend {}

    /**
     * The notification handle: raised for every posted completion, so a loop
     * on its own Poll context can embed the Ring by adding it with
     * Event::Notify and, when it fires, calling waitCompletions() with a zero
     * timeout until it returns an empty array. Owned by the Ring and the
     * Ring's to signal: notify() throws on it.
     */
    public function getHandle(): \Io\Poll\NotifyHandle {}

    /* Io\OperationQueue */

    public function submit(\Io\Operation $op, mixed $data = null): ?\Io\Completion {}

    public function cancel(\Io\Operation $op): void {}

    public function add(\Io\Registration $registration): void {}

    public function remove(\Io\Registration $registration): void {}

    /** @return list<\Io\Completion> */
    public function waitCompletions(?\Time\Duration $timeout = null, ?int $max = null): array {}

    public function countPending(): int {}

    /**
     * What a provider on this Ring should report by default: EdgeRegistrations
     * where the backend reports edges, and DirectAccept.
     * @return list<\Io\Hooks\Capability>
     */
    public function getHookCapabilities(): array {}

    /**
     * What this backend can serve for a provider that opts in: additionally
     * Files on every backend and DirectData on the native ones.
     * @return list<\Io\Hooks\Capability>
     */
    public function getSupportedHookCapabilities(): array {}
}

/** Also thrown by any use of a Ring in a forked child. */
class RingException extends \Io\IoException {}

/** Submit, cancel and wait failures, with the errno in the code. */
class FailedRingOperationException extends RingException {}
</code>

The classes are final, not serializable and have strict properties.

The operations the Ring executes are always created by the engine. A user-created read on a stream would
bypass the stream's buffer and filters, and the only correct asynchronous read on a stream is one that fills
that buffer, which is what ''fread()'' under a provider already is. Ring features for userland, such as
operations on ''Socket'' objects or dedicated file handles with caller-supplied buffers, would layer on top of
this class later with operation classes of their own.

===== Usage Examples =====

==== A scheduler on the Ring ====

The ''Scheduler'' class of the hooks RFC, unchanged, on a Ring:

<code php>
<?php

$scheduler = new Scheduler(new \Io\Ring\Engine());
\Io\Hooks\set_hooks($scheduler);

$scheduler->spawn(function () {
    echo strlen(file_get_contents('/var/log/big.log')), " bytes\n";   // Read: a worker reads it
});
$scheduler->spawn(function () {
    $c = stream_socket_client('tcp://example.org:80');                // lookup on the pool, then Connect
    fwrite($c, "GET / HTTP/1.0\r\nHost: example.org\r\n\r\n");
    echo strlen(stream_get_contents($c)), " bytes\n";
});
$scheduler->spawn(function () {
    $f = fopen('/data/journal.bin', 'a');
    fwrite($f, str_repeat('x', 1 << 20));
    fflush($f);                                                       // Fsync on the pool
    echo "journal written\n";
});

$scheduler->loop();
</code>

With the Ring's default capabilities the file read and the ''fsync()'' still run synchronously, since ''Files''
is opt-in. A provider that knows its storage is slow opts in:

<code php>
<?php

final class FileOffloadingScheduler extends Scheduler
{
    public function __construct(private \Io\Ring\Engine $ring)
    {
        parent::__construct($ring);
    }

    public function getCapabilities(): array
    {
        return [...$this->ring->getHookCapabilities(), \Io\Hooks\Capability::Files];
    }
}
</code>

==== A Poll loop embedding a Ring ====

A provider on its own Poll context that sends only the operations without a readiness form to a Ring and
reaps the Ring when its notification handle fires. Socket operations stay on the context. The base class is
a provider that answers every operation with one-shot watchers on its context, as the hooks RFC describes
and its test suite contains. Only the parts that concern the Ring are shown.

<code php>
<?php

final class MixedProvider extends ContextProvider
{
    private \Io\Ring\Engine $ring;

    public function __construct()
    {
        parent::__construct();
        $this->ring = new \Io\Ring\Engine();
        $this->context->add($this->ring->getHandle(), [\Io\Poll\Event::Notify], $this->ring);
    }

    public function getCapabilities(): array
    {
        return [\Io\Hooks\Capability::Files];          // file operations come here
    }

    public function run(\Io\Operation $op): \Io\Completion
    {
        if ($op instanceof \Io\Operation\Read || $op instanceof \Io\Operation\Write
                || $op instanceof \Io\Operation\Fsync || $op instanceof \Io\Operation\GetAddrInfo) {
            $this->ring->submit($op, \Fiber::getCurrent());
            $this->waiting++;
            return \Fiber::suspend();                  // resumed with the Completion
        }
        return parent::run($op);
    }

    /* In loop(), when the watcher whose data is the Ring fires: reap until empty */
    private function onRingReadable(): void
    {
        while ($completions = $this->ring->waitCompletions(\Time\Duration::fromSeconds(0))) {
            foreach ($completions as $c) {
                $this->waiting--;
                $this->ready[] = [$c->getData(), $c];
            }
        }
    }
}
</code>

==== Checking the backend ====

<code php>
<?php

$ring = new \Io\Ring\Engine();
echo "backend: ", $ring->getBackend()->name, "\n";
foreach ($ring->getSupportedHookCapabilities() as $capability) {
    echo "can serve: ", $capability->name, "\n";
}
</code>

===== Backward Incompatible Changes =====

None. The RFC adds the ''Io\Ring'' namespace and the ''accept_multishot'' stream context option, which is read
only under a provider with ''DirectAccept''. Bundling ior adds a library to the PHP build, as pcre2 and
others are.

===== Proposed PHP Version(s) =====

Next minor version, PHP 8.7, together with or after the IO hooks RFC, which it requires.

===== RFC Impact =====

==== To SAPIs ====

A Ring on the thread pool backend brings worker threads into the process, which is a consideration for FPM
workers and CLI scripts that fork. Nothing creates a Ring unless a script does.

==== To Existing Extensions ====

On Windows, extensions that cast a stream to a ''FILE*'' or a descriptor get it from a synchronous reopen of
the same file when the stream was opened for overlapped IO, positioned where the stream is, and work as
before.

==== To Opcache ====

None.

==== New Constants ====

None.

==== php.ini Defaults ====

None.

===== Open Issues =====

  * **Shared listeners.** Whether the multishot accept should be opt-in per listener rather than opt-out, since a provider cannot tell a listener shared with forked workers from a private one.
  * **Windows plain files.** Whether files should be opened for overlapped IO regardless of the installed provider, now that the test suite passes with it forced on and the cost is measured (a large sequential read a fifth slower, a write a tenth), or only under a ''Files'' provider as now, so that a file opened while nobody offloads costs nothing.

===== Unaffected PHP Functionality =====

Everything that does not construct an ''Io\Ring\Engine''. A provider on ''Io\Poll\OperationQueue'' behaves as
the hooks RFC specifies.

===== Future Scope =====

  * Operations created by userland, on unbuffered resources such as ''Socket'' objects or dedicated file handles, with caller buffers and vectored IO.
  * ''EdgeRegistrations'' on IOCP, once ior's IOCP backend reports readiness changes rather than persisting readiness.
  * Overlapped ''proc_open()'' pipes on Windows, so that pipe reads and writes can be completed by the Ring.

===== Proposed Voting Choices =====

As per the voting RFC, a yes/no vote with a 2/3 majority is needed for this proposal to be accepted.

<doodle title="Should the Ring API be added to PHP?" auth="bukka" voteType="single" closed="true">
   * Yes
   * No
   * Abstain
</doodle>

The vote started on 2026-XX-XX at XX:XX UTC and ends on 2026-XX-XX at XX:XX UTC.

===== Patches and Tests =====

A working implementation is on the ''io_hooks_poc'' branch at https://github.com/bukka/php-src/tree/io_hooks_poc.
The hooks test suite in ''ext/standard/tests/streams/hooks'' runs every scenario on the Ring as well, on both
Linux backends, and CI runs the hooks, poll, TLS and curl tests on the Ring on Linux, macOS and Windows.
ior is at https://github.com/libior/ior.

===== Implementation =====

After the project is implemented, this section will contain:
  - the version(s) it was merged to
  - a link to the git commit(s)

===== References =====

  * [[rfc:io_hooks|PHP RFC: IO hooks and operations]]
  * [[rfc:poll_api_additions|PHP RFC: Polling API additions]]
  * ior: https://github.com/libior/ior
  * io_uring: https://kernel.dk/io_uring.pdf
  * liburing: https://github.com/axboe/liburing
  * Windows IO completion ports: https://learn.microsoft.com/en-us/windows/win32/fileio/i-o-completion-ports

===== Changelog =====

  * 0.1 - Initial draft
Page Tools
