dpdk: add patch removing redundant rsvd_tbl8s counter

[maxime-leroy]

Bring in a local DPDK patch that fixes a latent bug in the IPv6 FIB TRIE algorithm where rsvd_tbl8s drifts past zero on certain ADD/DEL patterns and wraps to UINT32_MAX, causing every subsequent long-prefix add to fail with -ENOSPC even though tbl8_pool_pos is far below number_tbl8s.

The bug surfaces when a covering parent prefix is deleted before its children: depth_diff at DEL time is computed against a different RIB state than at ADD time and the cumulative reservation accounting diverges. Replacing rsvd_tbl8s with the actual pool occupancy counter (tbl8_pool_pos for trie, cur_tbl8s for dir24_8) eliminates the asymmetry. The pre-allocation refusal semantics are preserved.

The patch is intended to be submitted upstream to dpdk-dev; it is carried locally until accepted so grout users get a working FIB6 under SRv6 churn.

Summary

Introduces a DPDK patch addressing an underflow bug in IPv6 FIB trie's pre-allocation reservation counter.

Problem

The rsvd_tbl8s counter in struct rte_trie_tbl (IPv6) drifts and underflows to UINT32_MAX when a covering parent prefix is deleted before its children. The counter is updated by depth_diff computed at each ADD/DEL from current RIB state via rte_rib6_get_nxt() and rte_rib6_lookup_parent(). Since the RIB is not invariant between ADD and later DEL, depth_diff at DEL time need not equal the value at ADD time. Example asymmetry:

• ADD /28: depth_diff=1, rsvd_tbl8s += 1
• ADD /48 (child of /28): depth_diff=2, rsvd_tbl8s += 2
• DEL /28 (children alive): sibling found, depth_diff=0, rsvd_tbl8s -= 0
• DEL /48: depth_diff=3, rsvd_tbl8s -= 3
• Net: rsvd_tbl8s decreased by 1 per cycle

Over repeated cycles, the counter wraps to UINT32_MAX-N, causing the pre-check to unconditionally reject ADD operations for long prefixes even when actual pool occupancy (tbl8_pool_pos) is far below number_tbl8s. Observed in production with BGP SRv6 churn under FRR/zebra restart cycles, where parent locators (/40, /48) are deleted before child SIDs (/64, /128). After ~174 asymmetric cycles, rsvd_tbl8s wrapped to 4294967122 with tbl8_pool_pos at 39 of 262144.

Solution

Replace rsvd_tbl8s with actual pool occupancy counters:

• trie: trie_modify() now compares dp->tbl8_pool_pos + depth_diff against dp->number_tbl8s. The tbl8_pool_pos counter is maintained natively by tbl8_get() and tbl8_put() and is always accurate by construction.
• dir24_8: dir24_8_modify() now compares dp->cur_tbl8s against dp->number_tbl8s. The cur_tbl8s counter was added by commit 96c3d06a3578 and already tracks occupancy in tbl8_alloc() and tbl8_cleanup_and_free(), making rsvd_tbl8s redundant. The dir24_8 increment is symmetric (binary, 0 or 1 based on whether a /25-/32 sibling exists) and exhibits no underflow bug.

Changes

• subprojects/dpdk-25.11.wrap: Added dpdk/fib6-remove-redundant-rsvd_tbl8s-counter.patch to diff_files list.
• subprojects/packagefiles/dpdk/fib6-remove-redundant-rsvd_tbl8s-counter.patch (new file): Patch removes uint32_t rsvd_tbl8s field from struct rte_trie_tbl and struct dir24_8_tbl, removes all increment/decrement code paths for the counter, and simplifies goto-based early-success paths to direct return 0.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: ae3dd4d8-5031-4ea5-8198-2f4be712c136

📥 Commits

Reviewing files that changed from the base of the PR and between 0ab251b and 4c392cc.

📒 Files selected for processing (2)

• subprojects/dpdk-25.11.wrap
• subprojects/packagefiles/dpdk/fib6-remove-redundant-rsvd_tbl8s-counter.patch

---

📝 Walkthrough

Walkthrough

This PR updates DPDK patch files to remove the rsvd_tbl8s reservation counter from both the IPv4 dir24_8 table and IPv6 trie FIB data structures. The wrap file registers a new patch file that modifies two FIB table structs and updates their corresponding modify functions to use actual tbl8 occupancy counters (cur_tbl8s for dir24_8, tbl8_pool_pos + depth_diff for trie) for capacity checks instead of maintaining reserved counters. The refactoring eliminates the increment/decrement accounting code for rsvd_tbl8s on ADD and DEL operations and simplifies early-success return paths by removing goto-based control flow.

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}