Skip to content

dpdk: preserve kernel fallback for dynamic rx flows#204

Merged
dleshchev merged 2 commits into
mainfrom
dpdk-dynamic-rx-kernel-fallback
Jun 25, 2026
Merged

dpdk: preserve kernel fallback for dynamic rx flows#204
dleshchev merged 2 commits into
mainfrom
dpdk-dynamic-rx-kernel-fallback

Conversation

@dleshchev

@dleshchev dleshchev commented Jun 25, 2026 •
edited
Loading

Copy link
Copy Markdown
Collaborator

Summary

  • install the flow-isolation send-to-kernel fallback before programming dynamic RX flows
  • track fallback installation per {port, group} so repeated dynamic flow updates do not duplicate the rule
  • clear fallback tracking when programmed flows are destroyed

Fixes #205.

Validation

  • git diff --check origin/main...HEAD
  • IMAGE_TAG=daqiri:dynamic-rx-kernel-fallback BASE_TARGET=dpdk DAQIRI_ENGINE="dpdk ibverbs" scripts/build-container.sh

@greptile-apps

greptile-apps Bot commented Jun 25, 2026 •
edited
Loading

Copy link
Copy Markdown
Contributor

Greptile Summary

This PR fixes a correctness gap where flow_isolation: true queues-only configs would black-hole unmatched control-plane traffic (ARP, SSH, etc.) once the first dynamic RX flow was added, because the send-to-kernel fallback was only installed for static startup flows. The fix installs the fallback (and the group-0 ETH jump) immediately before the first dynamic rule on each {port, group} pair, with deduplication so repeated flow additions don't create duplicate fallback rules.

  • ensure_dynamic_flow_isolation_fallback_locked is called inside add_rx_flow_async and add_rx_flows_async (both under flow_lock_) and also enforces the invariant that standard IPv4/UDP and flex-item flows cannot mix groups on one interface.
  • send_to_kernel_fallback_installed_ tracks installations per {port, group} and is cleared in destroy_programmed_flows, symmetric with the fallback flows being tracked in programmed_flows_; the ETH jump set (eth_jump_installed_) is intentionally not cleared because those flows are long-lived and managed separately.
  • docs/api-reference/configuration.md is updated to describe the new dynamic-flow fallback behavior and the single-group batch constraint.

Confidence Score: 5/5

Safe to merge — focused bug fix with correct deduplication, lifecycle symmetry, and no new data-plane race conditions.

The implementation is correct: fallback installation is idempotent, the conflict check for mixed flow groups is sound, ordering (fallback before match rules) is right, and the clear-on-destroy symmetry between send_to_kernel_fallback_installed_ and eth_jump_installed_ matches their respective lifetimes. Both commits are DCO-signed and follow the required commit title format.

docs/tutorials/configuration-walkthrough.md (note 19) and AGENTS.md (Current limitations) still describe the send-to-kernel fallback as init-time only and should be updated.

Important Files Changed

Filename Overview
src/engines/dpdk/daqiri_dpdk_engine.cpp Adds ensure_dynamic_flow_isolation_fallback_locked to install the send-to-kernel fallback and ETH jump rule before the first dynamic RX flow is programmed; deduplicates via send_to_kernel_fallback_installed_ and clears tracking on destroy_programmed_flows. Logic is correct: conflict detection, ordering, deduplication, and the clear-on-destroy symmetry with eth_jump_installed_ all hold.
src/engines/dpdk/daqiri_dpdk_engine.h Adds send_to_kernel_fallback_installed_ member set and declares ensure_dynamic_flow_isolation_fallback_locked; straightforward additions that mirror the existing eth_jump_installed_ pattern.
docs/api-reference/configuration.md Updates the flow_isolation description to reflect that the first dynamic RX flow now installs the send-to-kernel fallback, and adds a note that mixing standard and flex-item flow classes is not supported within a dynamic batch. Change is accurate and consistent with the implementation.

Reviews (3): Last reviewed commit: "#205 - Support dynamic flex-item isolati..." | Re-trigger Greptile

@dleshchev dleshchev mentioned this pull request Jun 25, 2026
Closed
@dleshchev
#205 - Preserve kernel fallback for dynamic RX flows
cc4109e 
Install the flow-isolation send-to-kernel fallback before programming
dynamic RX flows so control-plane traffic (ARP, SSH, application control
planes) keeps reaching the Linux kernel. Previously the fallback was
installed only for static startup flows, so a queues-only config with
flow_isolation: true black-holed all unmatched/kernel traffic once the
first dynamic flow was added.

- ensure_dynamic_flow_isolation_fallback_locked() installs the ETH jump
  rule and send-to-kernel fallback under flow_lock_ in add_rx_flow_async
  and add_rx_flows_async.
- Track fallback installation per {port, group} so repeated dynamic flow
  updates do not duplicate the rule.
- Clear fallback tracking when programmed flows are destroyed.
- Document the dynamic-flow fallback behavior in configuration.md.

Signed-off-by: Denis Leshchev <dleshchev@nvidia.com>
@dleshchev dleshchev force-pushed the dpdk-dynamic-rx-kernel-fallback branch from 995a7c7 to cc4109e Compare June 25, 2026 11:05
@dleshchev
#205 - Support dynamic flex-item isolation fallback
e635ee3 
Signed-off-by: Denis Leshchev <dleshchev@nvidia.com>
@dleshchev dleshchev merged commit defd3ad into main Jun 25, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

[BUG] Dynamic RX flows under flow_isolation drop kernel/control traffic (missing send-to-kernel fallback)

1 participant

@dleshchev