feat[needs minor version update]: replace POSIX MQ publish notification with eventfd

[k1832]

Description

Replace the POSIX MQ-based publish notification mechanism with kernel-managed eventfd.
The previous MQ path required the publisher to call mq_send() once per subscriber
in userspace (N+1 syscalls per publish including the PUBLISH ioctl). With eventfd,
the kernel module signals subscriber eventfds directly within ioctl(PUBLISH),
reducing the cost to a single syscall regardless of subscriber count.

Key changes:

• Kernel: subscriber_info stores eventfd_ctx. publish_msg collects notify_ctx
  pointers under topic_rwsem write lock, then signals outside the lock to avoid
  contention. Stack buffer (64 entries) with heap fallback for larger subscriber counts.
• Userspace: Subscriber creates eventfd and passes to kernel via ioctl(ADD_SUBSCRIBER).
  Publisher's publish_core is simplified from ~80 to ~15 lines (MQ send loop removed).
• Compatibility: agnocast_eventfd_signal wrapper for kernel < 6.8 API difference.

Detailed design, benchmark analysis, and alternative considered: see design doc (linked below).

Benchmark Results

Measured with isorc26_tmp,
10 topics, 100Hz, 3-iter mean of per-iter p50/p99, --no-rt. Test machine: 32 logical CPUs.

Publish latency (publisher wall-clock, from loan() to publish() completion)

Subs	MQ p50 / p99	eventfd p50 / p99	p50 diff
4	11.5 / 37.1 us	7.1 / 20.9 us	-38%
8	17.0 / 51.3 us	8.3 / 30.3 us	-51%
16	24.5 / 61.6 us	18.4 / 47.4 us	-25%
32	56.9 / 133.5 us	58.7 / 138.2 us	+3%
64	186.4 / 404.2 us	141.5 / 330.0 us	-24%
128	674.2 / 1084.4 us	594.4 / 1206.2 us	-12%

E2E latency (from publisher timestamp just before loan() to subscriber callback entry)

Subs	MQ p50 / p99	eventfd p50 / p99	p50 diff
4	61.3 / 210.4 us	30.2 / 183.2 us	-51%
8	36.3 / 220.0 us	26.1 / 204.7 us	-28%
16	27.8 / 119.1 us	22.4 / 124.4 us	-19%
32	45.4 / 156.4 us	55.3 / 195.1 us	+22%
64	119.6 / 521.6 us	96.6 / 412.6 us	-19%
128	387.8 / 1454.9 us	340.0 / 1803.7 us	-12%

Observations

• eventfd is clearly faster at low-to-mid N (p50 improvement up to -51%).
• Around N ≈ number of logical CPUs (32) the two approaches converge — the
  scheduler's wake_up O(N) cost dominates and the mechanism difference gets buried.
• At high N, E2E p50 is smaller than publish latency p50 (e.g. N=128: E2E=340 us <
  publish=594 us). This is because there is 1 publish-latency sample but N E2E samples
  per publish, so the E2E median points to the "middle-woken" subscriber. In the p99
  column the last-woken subscriber's scheduling latency dominates, so the ordering
  reverses (E2E p99 > publish p99 at high N).
• See design doc for details.

Alternative explored: per-topic shared eventfd

We also prototyped a per-topic shared eventfd (one eventfd shared by all subscribers
of a topic, signaled once per publish) to test whether consolidating N eventfd_signal
calls into one would improve high-N latency. Contrary to the hypothesis, the single
wake_up_poll() iterating N ep_poll_callbacks under one spin_lock showed similar
or slightly worse latency due to concentrated scheduler contention.

• Prototype branch: experiment/shared-eventfd
• Analysis: see design doc

Related links

• Design doc (TIER IV internal link): https://tier4.atlassian.net/wiki/x/1QDsMQE

How was this PR tested?

• Autoware (required)
• bash scripts/test/e2e_test_1to1.bash (required)
• bash scripts/test/e2e_test_2to2.bash (required)
• kunit tests (required when modifying the kernel module)
• bash scripts/test/run_requires_kernel_module_tests.bash (required)
• sample application
• isorc26_tmp benchmark (subs = 4, 8, 16, 32, 64, 128)

Notes for reviewers

MQ usage in other paths (bridge request and bridge delegation) remains unchanged.
Dead code from the old subscription MQ cleanup path (daemon mq_unlink loop) will be
removed in a follow-up PR.

Version Update Label (Required)

Please add exactly one of the following labels to this PR:

• need-major-update: User API breaking changes
• need-minor-update: Internal API breaking changes (heaphook/kmod/agnocastlib compatibility)
• need-patch-update: Bug fixes and other changes

Important notes:

• If you need need-major-update or need-minor-update, please include this in the PR title as well.
  • Example: fix(foo)[needs major version update]: bar or feat(baz)[needs minor version update]: qux
• After receiving approval from reviewers, add the run-build-test label. The PR can only be merged after the build tests pass.

See CONTRIBUTING.md for detailed versioning rules.

[Copilot]

Pull request overview

This PR replaces Agnocast’s POSIX message-queue–based publish notification path with per-subscriber eventfd, moving fan-out notification into the kernel ioctl(PUBLISH) path to avoid N-per-subscriber syscalls in userspace.

Changes:

• Kernel module: store eventfd_ctx per subscriber and signal subscriber eventfds during AGNOCAST_PUBLISH_MSG_CMD (with stack buffer + heap fallback for large subscriber counts).
• Userspace library: create/pass eventfd via AGNOCAST_ADD_SUBSCRIBER_CMD, update epoll handling to read/clear eventfd notifications, and simplify publisher publish path (remove MQ send loop).
• Tests/KUnit: update ioctl call signatures and remove publish subscriber-id buffer expectations tied to the old MQ design.

Reviewed changes

Copilot reviewed 32 out of 32 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
src/agnocastlib/test/integration/src/node_for_executor_test.cpp	Replace MQ-based executor-test notification with eventfd creation and write()-based triggering.
src/agnocastlib/test/integration/include/node_for_executor_test.hpp	Swap MQ receiver/sender members for an eventfds_ list.
src/agnocastlib/src/agnocast_subscription.cpp	Create eventfd for non-take subscriptions, pass it to kernel via ioctl, and track it for cleanup.
src/agnocastlib/src/agnocast_publisher.cpp	Remove userspace MQ notification loop; publishing becomes a single ioctl call.
src/agnocastlib/src/agnocast_epoll.cpp	Switch wake-up handling from mq_receive to read(eventfd) to clear notifications.
src/agnocastlib/include/agnocast/agnocast_subscription.hpp	Remove MQ subscription state; register callbacks with notify_eventfd_ and close it on destruction.
src/agnocastlib/include/agnocast/agnocast_publisher.hpp	Update publish_core signature and remove publisher-side MQ state/cleanup.
src/agnocastlib/include/agnocast/agnocast_mq.hpp	Remove MqMsgAgnocast (no longer needed for publish notification).
src/agnocastlib/include/agnocast/agnocast_ioctl.hpp	Add eventfd to add-subscriber args; remove publish subscriber-id buffer fields.
src/agnocastlib/include/agnocast/agnocast_epoll.hpp	Use notify_eventfd in epoll_ctl(ADD) for subscription readiness.
src/agnocastlib/include/agnocast/agnocast_callback_info.hpp	Replace mqdes with notify_eventfd in callback registration/info.
agnocast_kmod/agnocast_kunit/agnocast_kunit_take_msg.c	Update ioctl helper calls for new publish/add-subscriber signatures.
agnocast_kmod/agnocast_kunit/agnocast_kunit_set_ros2_subscriber_num.c	Update add-subscriber calls to pass eventfd argument.
agnocast_kmod/agnocast_kunit/agnocast_kunit_remove_subscriber.c	Update publish/add-subscriber calls for new ioctl signatures.
agnocast_kmod/agnocast_kunit/agnocast_kunit_remove_publisher.c	Update publish/add-subscriber calls for new ioctl signatures.
agnocast_kmod/agnocast_kunit/agnocast_kunit_release_sub_ref.c	Update publish/add-subscriber calls for new ioctl signatures.
agnocast_kmod/agnocast_kunit/agnocast_kunit_receive_msg.c	Update publish/add-subscriber calls for new ioctl signatures.
agnocast_kmod/agnocast_kunit/agnocast_kunit_publish_msg.h	Remove test case tied to the removed subscriber-id buffer API.
agnocast_kmod/agnocast_kunit/agnocast_kunit_publish_msg.c	Update publish/add-subscriber calls and remove subscriber-id buffer assertions/tests.
agnocast_kmod/agnocast_kunit/agnocast_kunit_get_topic_subscriber_info.c	Update add-subscriber call signature to include eventfd argument.
agnocast_kmod/agnocast_kunit/agnocast_kunit_get_topic_publisher_info.c	Update add-subscriber call signature to include eventfd argument.
agnocast_kmod/agnocast_kunit/agnocast_kunit_get_subscriber_qos.c	Update add-subscriber call signature to include eventfd argument.
agnocast_kmod/agnocast_kunit/agnocast_kunit_get_subscriber_num.c	Update add-subscriber calls to pass eventfd argument (including bridge variants).
agnocast_kmod/agnocast_kunit/agnocast_kunit_get_publisher_num.c	Update add-subscriber calls to pass eventfd argument (including bridge variants).
agnocast_kmod/agnocast_kunit/agnocast_kunit_get_node_subscriber_topics.c	Update add-subscriber calls to pass eventfd argument.
agnocast_kmod/agnocast_kunit/agnocast_kunit_do_exit.c	Update publish/add-subscriber calls for new ioctl signatures.
agnocast_kmod/agnocast_kunit/agnocast_kunit_add_subscriber.c	Update add-subscriber calls to pass eventfd argument.
agnocast_kmod/agnocast_ioctl.c	Kernel: accept eventfd, store eventfd_ctx, and signal eventfds during publish (outside topic_rwsem when possible).
agnocast_kmod/agnocast_internal.h	Add eventfd support (eventfd_ctx in subscriber_info) and compatibility wrapper for eventfd_signal API change.
agnocast_kmod/agnocast_internal.c	Ensure eventfd_ctx_put() is called during subscriber exit cleanup.
agnocast_kmod/agnocast_exit.c	Ensure eventfd_ctx_put() is called when removing all topics/subscribers on shutdown.
agnocast_kmod/agnocast.h	Kernel ABI: add eventfd to add-subscriber args; remove publish subscriber-id buffer fields; update function prototypes.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The signature of publish_core has changed (removed the opened_mqs parameter), but there is still a test-side mock implementation with the old signature (src/agnocastlib/test/unit/test_mocked_agnocast.cpp). This will cause a compile/link failure due to the mismatched declaration/definition. Update the mock to match the new publish_core signature (and remove MQ-related types/includes there).