No kernel nhg original

[donaldsharp]

PR #21125 attempts to modify slightly how NHG data is sent to the kernel for connected/local/kernel NHG's. The approach taken there was too complicated and in order for me to figure out what was going on I had to reverse engineer it to a degree that my approach is just much simplier and frankly easier to understand. Here's my alternate approach to the problem

[greptile-apps]

Greptile Summary

This PR simplifies how NHGs for connected/local/kernel (system) routes are handled in the dataplane by letting them flow through the full dplane queue with a skip_kernel flag instead of short-circuiting early and discarding the context. The previous approach discarded the dplane context before FPM and other providers could see it; the new approach allows providers (particularly FPM) to observe and act on these NHG installs while still preventing kernel programming.

• zebra_dplane.c: Replaces the early-return that marked NHGs as installed and freed the context with a call to dplane_ctx_set_skip_kernel(), then lets the context flow normally through the provider chain.
• zebra_nhg.c: Removes the now-unnecessary special-case handling of ZEBRA_DPLANE_REQUEST_SUCCESS in zebra_nhg_install_kernel and adds UNSET_FLAG(nhe->flags, NEXTHOP_GROUP_QUEUED) when promoting an NHG from the delayed-install path to a real kernel install, which is critical for correctness.
• test_fpm_topo1.py: Adds a new test verifying that system-route NHGs are forwarded to FPM but not installed in the Linux kernel.

Confidence Score: 3/5

The C changes are correct but the new test may always time out, leaving the PR's primary behavioral claim unverified.

The core C implementation is clean and the flag-management changes in zebra_nhg.c are correct. The concern is with the test: NEXTHOP_GROUP_FPM is only written during the FPM batch resync walk (fpm_nhg_send_cb), not when individual NHG contexts flow through fpm_nl_enqueue. A connected/local/kernel NHG added after the last FPM sync will never have fpm=True in its show output, so check_nhg_sent_to_fpm will time out on every CI run.

tests/topotests/fpm_testing_topo1/test_fpm_topo1.py — the FPM presence assertion needs a different verification strategy.

Important Files Changed

Filename	Overview
zebra/zebra_dplane.c	Replaces the early-exit/ctx-discard for NEXTHOP_GROUP_INITIAL_DELAY_INSTALL with dplane_ctx_set_skip_kernel(); ctx now flows through the full provider chain.
zebra/zebra_nhg.c	Removes special SUCCESS handling from the dplane result switch (now always an error) and adds critical NEXTHOP_GROUP_QUEUED clearing when promoting a system-route NHG to a real kernel install.
tests/topotests/fpm_testing_topo1/test_fpm_topo1.py	New test for system-route NHG FPM delivery and kernel-skip; the FPM presence check relies on NEXTHOP_GROUP_FPM, which is only set during batch FPM resync, not via the individual dplane ctx path the PR introduces.

Sequence Diagram

sequenceDiagram
    participant RIB as zebra_rib (main thread)
    participant NHG as zebra_nhg
    participant DPLANE as dplane thread
    participant FPM as FPM provider
    participant KERNEL as kernel provider

    RIB->>NHG: zebra_nhg_install_kernel(nhe, CONNECT/LOCAL/KERNEL)
    Note over NHG: INITIAL_DELAY_INSTALL set skip flag-clearing block
    NHG->>DPLANE: dplane_nexthop_add(nhe)
    DPLANE->>DPLANE: dplane_ctx_set_skip_kernel(ctx)
    DPLANE->>DPLANE: dplane_update_enqueue(ctx)
    DPLANE-->>NHG: ZEBRA_DPLANE_REQUEST_QUEUED
    NHG->>NHG: SET_FLAG NEXTHOP_GROUP_QUEUED

    DPLANE->>FPM: fpm_nl_process ctx with NH_INSTALL op
    FPM->>FPM: fpm_nl_enqueue encode and send NHG to FPM socket
    FPM-->>DPLANE: enqueue to output

    DPLANE->>KERNEL: kernel provider dequeues ctx
    KERNEL->>KERNEL: dplane_ctx_is_skip_kernel true
    KERNEL->>KERNEL: set status SUCCESS skip netlink
    KERNEL-->>DPLANE: enqueue to output

    DPLANE-->>NHG: zebra_nhg_dplane_result SUCCESS
    NHG->>NHG: UNSET QUEUED REINSTALL SET INSTALLED
    NHG->>NHG: zebra_nhg_handle_install(nhe)

Loading

Prompt To Fix All With AI

Fix the following 2 code review issues. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 2
tests/topotests/fpm_testing_topo1/test_fpm_topo1.py:256-264
**FPM flag not set via dplane ctx path**

`NEXTHOP_GROUP_FPM` is only set inside `fpm_nhg_send_cb` — the batch hash-walk that runs during an FPM reconnect/resync. When an NHG install context flows through the FPM dplane provider via `fpm_nl_enqueue` (the new path this PR introduces), the flag is never written to the `nhg_hash_entry`. For an interface address added *after* the last FPM sync, `nhg_data.get("fpm", False)` will remain `False` indefinitely, causing the assertion to time out on every run. A more reliable check would either force an FPM reconnect before the assertion (triggering a full resync) or inspect the mock FPM server's received message log directly.

### Issue 2 of 2
tests/topotests/fpm_testing_topo1/test_fpm_topo1.py:258-265
**Misleading error message: `result` is a boolean, not NHG data**

`topotest.run_and_expect` returns `(success, last_function_return_value)`. Here `check_nhg_sent_to_fpm` returns `True` or `False`, so when the assertion fails the message reads `NHG data: False` instead of meaningful diagnostic information. Consider capturing the raw NHG JSON inside the check closure and including it in the assertion message.

_{Reviews (1): Last reviewed commit: "tests: Add topotest cases for sending RI..." | Re-trigger Greptile}

[GaladrielZhao]

Current SONiC fpmsyncd handles NHG via RTM_NEWNEXTHOP / RTM_DELNEXTHOP in onNextHopMsg.

Below is how each route type's NHG is processed:

• Connected routes: treated as normal unicast routes. NHG referenced by nh_id is resolved from the in-memory map and expanded into ROUTE_TABLE fields (or NEXTHOP_GROUP_TABLE for multipath). Works correctly.
• Kernel routes (proto=kernel, type=RTN_UNICAST): same path as connected routes. No special handling.
• Local routes (type=RTN_LOCAL): the route itself is intentionally skipped in onRouteMsg ("BUM routes aren't supported yet"). However, its NHG is still received via RTM_NEWNEXTHOP:
  • single NH: stored only in memory (m_nh_groups), never written to DB.
  • group NH: written to NEXTHOP_GROUP_TABLE on creation. While this creates a brief orphan entry, multipath local routes are uncommon, so the impact is nearly negligible. Even if it happens, the entry is cleaned up on RTM_DELNEXTHOP.

So from the SONiC side, having FRR send all three types of NHGs causes no functional issue and actually fixes the missing connected-route NHG bug we observed.

[GaladrielZhao]

Thanks, looks good to me.

[riw777]

looks good

[riw777]

looks like there's a memory leak :-(

r2: zebra has memory leaks:

[donaldsharp]

ci:rerun

[riw777]

current failure looks like it is in conditional advertisement ... which seems unrelated