io_uring: expose multishot recv via IoUringOptions#44670
Draft
aburan28 wants to merge 4 commits intoenvoyproxy:mainfrom
Draft
io_uring: expose multishot recv via IoUringOptions#44670aburan28 wants to merge 4 commits intoenvoyproxy:mainfrom
aburan28 wants to merge 4 commits intoenvoyproxy:mainfrom
Conversation
This is a no-behavior-change preparation step for multishot recv. The ``CompletionCb`` callback type now takes a ``uint32_t flags`` argument that carries the raw ``cqe->flags`` value from the kernel. For multishot completions a follow-up change will inspect: * ``IORING_CQE_F_BUFFER`` — a buffer was selected from a buf-ring; the buffer ID is encoded in the upper bits. * ``IORING_CQE_F_MORE`` — the SQE will produce further completions. The worker callback ignores ``flags`` for now. Injected completions are defined to always carry ``flags == 0``. All ``forEveryCompletion`` callers (worker, impl tests) updated. ``IoUringSocket::on*`` virtual methods are intentionally unchanged in this commit; only ``onRead`` will need flags, in the multishot recv change. Signed-off-by: Adam Buran <a.buran28@gmail.com> Signed-off-by: Adam Buran <aburan28@gmail.com>
Adds the kernel-managed buffer ring lifecycle and the ``recv`` multishot opcode to ``IoUringImpl``. This is the plumbing layer for switching the io_uring socket read path off the per-read ``readv`` allocation; the worker change comes in a follow-up PR. New ``IoUring`` virtuals: * ``setupBufRing(group_id, count, buf_size)`` — register a buffer ring with the kernel. The buffers live in a single contiguous allocation owned by ``IoUringImpl``. Validates that ``count`` is a non-zero power of two and rejects double-setup. Falls back to ``IoUringResult::Failed`` on kernels that lack ``IORING_REGISTER_PBUF_RING`` (< 5.19). * ``prepareRecvMultishot(fd, group_id, user_data)`` — submits a recv with ``IOSQE_BUFFER_SELECT`` so the kernel pulls a buffer from the ring. The same SQE may produce multiple completions, signalled by ``IORING_CQE_F_MORE`` in ``cqe->flags``. * ``getBufferForBid(group_id, bid)`` — look up the storage backing a particular kernel-selected buffer; the consumer reads up to ``cqe->res`` bytes and then recycles. * ``recycleBuffer(group_id, bid)`` — return a consumed buffer to the ring so the kernel can reuse it. For now only one buf-ring is supported per ``IoUring`` instance. Test: * ``SetupBufRingValidatesInputs`` — exercises the rejection paths (bad count, bad buf_size, double-setup). * ``MultishotRecvDeliversBuffersAndStaysArmed`` — end-to-end with a real socketpair and a real ring: arm a multishot recv, write twice, verify both completions deliver buffers, the bid is in range, the data matches, and the SQE stays armed (F_MORE set on the first completion). Skips when the kernel lacks buf-ring support. Signed-off-by: Adam Buran <a.buran28@gmail.com> Signed-off-by: Adam Buran <aburan28@gmail.com>
When the worker is configured with multishot recv enabled and the kernel/liburing successfully sets up a buf-ring (5.19+), the ``IoUringServerSocket`` read path replaces the per-read ``readv`` SQE + ``uint8_t[]`` allocation with a single ``IORING_OP_RECV`` multishot SQE that pulls buffers from the kernel- managed ring. Each completion delivers one kernel-selected buffer; the ``BufferFragment`` wrapping it recycles the buffer back to the ring on release. Mechanics: * New ``Request::RequestType::RecvMultishot`` distinguishes the multishot SQE from a plain ``Read``. The worker's completion dispatch routes both to ``onRead`` but holds onto the ``Request*`` while ``IORING_CQE_F_MORE`` is set (the kernel reuses the same user_data for further completions on the same SQE). * ``IoUringSocket::onRead`` gains a ``uint32_t flags`` argument carrying the raw ``cqe->flags``. The buffer ID is in the upper bits when ``IORING_CQE_F_BUFFER`` is set; ``F_MORE`` indicates the SQE is still armed. * ``IoUringServerSocket::onRead`` only clears ``read_req_`` when the SQE has terminated. While armed, the bottom-of-function ``submitReadRequest`` short-circuits because ``read_req_`` is still non-null. When ``F_MORE`` clears, ``read_req_`` is freed and a new multishot SQE is submitted. * ``IoUringWorkerImpl::makeMultishotBufferFragment`` wraps the kernel buffer with a release callback that calls ``recycleBuffer`` — back-pressure / buffer return is driven by the upper-layer drain. * On older kernels ``setupBufRing`` returns ``Failed`` and the worker silently falls back to the existing ``readv`` path, so the feature is safe to ship gated behind a config flag. The worker constructor gains two new defaulted args (``enable_multishot_recv``, ``multishot_recv_buffer_count``) so all existing call sites continue to compile unchanged. Tests: * ``MultishotRecvSetupAndSubmit`` — buf-ring setup + first submit picks the multishot path and produces a ``RecvMultishot`` request. * ``MultishotRecvFallbackOnUnsupportedKernel`` — when ``setupBufRing`` fails, the worker falls back to ``prepareReadv``. * ``MultishotRecvDeliversBufferAndStaysArmed`` — completion with ``F_BUFFER | F_MORE`` delivers the buffer to the upper layer and does not re-arm the SQE; the buffer is recycled when the upper layer drains. * ``MultishotRecvReArmOnFMoreClear`` — completion with ``F_BUFFER`` but no ``F_MORE`` triggers a fresh ``prepareRecvMultishot`` to re-arm. The proto / factory wiring to actually expose this option is in a follow-up change. Signed-off-by: Adam Buran <a.buran28@gmail.com> Signed-off-by: Adam Buran <aburan28@gmail.com>
Wires the per-worker multishot recv plumbing through the proto config and factory so it can be enabled at runtime. * New ``IoUringOptions.enable_multishot_recv`` (BoolValue, default false) and ``IoUringOptions.multishot_recv_buffer_count`` (UInt32Value, default 256, must be a power of two). * ``IoUringWorkerFactoryImpl`` constructor takes the two new args and threads them through to ``IoUringWorkerImpl``. * ``socket_interface_impl.cc`` reads the proto fields and passes them to the factory. * Changelog entry under ``new_features`` describes the option and the kernel/liburing requirements (5.19+ for the buf-ring registration, 6.0+ for multishot recv itself); older kernels silently fall back to ``readv``. Each multishot recv buffer is sized to ``read_buffer_size``, so the total per-worker memory is ``multishot_recv_buffer_count * read_buffer_size`` bytes (default 256 * 8192 = 2 MiB). Signed-off-by: Adam Buran <a.buran28@gmail.com> Signed-off-by: Adam Buran <aburan28@gmail.com>
aburan28
requested a deployment
to
external-contributors
GitHub Actions
Waiting
|
Hi @aburan28, welcome and thank you for your contribution. We will try to review your Pull Request as quickly as possible. In the meantime, please take a look at the contribution guidelines if you have not done so already. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Wires the multishot recv read path through the proto config and the bootstrap extension factory so it can be enabled at runtime.
Depends on:
CompletionCbflags arg)prepareRecvMultishotAPI)Config surface
Each buffer is sized to
read_buffer_size, so the per-worker memory cost ismultishot_recv_buffer_count * read_buffer_sizebytes (default 256 * 8192 = 2 MiB per worker).Compatibility
Requires kernel 5.19+ for
IORING_REGISTER_PBUF_RINGand 6.0+ for multishot recv. On older kernels Envoy logs a warning and falls back to the existingreadvpath, so this option is safe to enable on a heterogeneous fleet.The factory constructor gains two new positional arguments. All call sites in source/test have been updated. There is no on-the-wire proto change beyond two new optional fields, so existing configs continue to work.
Test plan
IoUringWorkerFactoryImplTest.Basicupdated for the new constructor args.IoUringSocketHandleImplintegration test updated for the new constructor args.