On Mon, 18 May 2026 10:27:00 +0200 Robin Jarry <[email protected]> wrote:
> Instead of guessing what are the proper header lengths, pass > a rte_net_hdr_lens struct to rte_net_get_ptype and use it to get the > proper header lengths/offsets in tap_verify_csum. > > This allows supporting stacked VLAN/QinQ tags and IPv6 extensions. > > Signed-off-by: Robin Jarry <[email protected]> Still needs some polishing around Ipv6 ext headers per AI review... [PATCH dpdk v5] net/tap: use offsets provided by rte_net_get_ptype The refactor is mechanically correct. rte_net_get_ptype() with a non-NULL rte_net_hdr_lens fills l2_len walking the full VLAN/QinQ stack (up to RTE_NET_VLAN_MAX_DEPTH), and for IPv4 sets l3_len = rte_ipv4_hdr_len(ip4h). Replacing the local l2_len/l3_len arithmetic with hlen->* preserves the existing bounds checks and genuinely fixes the stacked-VLAN case (the old code only handled single VLAN +4 or QinQ +8). The unnecessary cast of l4_hdr to struct rte_udp_hdr * was correctly dropped. The {0} initializer on the rte_net_hdr_lens stack variable is appropriate since rte_net_get_ptype only writes the fields it parses. Info: - The commit message says "This allows supporting stacked VLAN/QinQ tags and IPv6 extensions." The first half is true. The IPv6-extensions half is not: the else branch that catches RTE_PTYPE_L3_IPV6_EXT and returns is unchanged, and the accompanying comment "IPv6 extensions are not supported." in that branch is also unchanged. For an IPv6 packet with extension headers, rte_net_get_ptype() will set RTE_PTYPE_L3_IPV6_EXT (not RTE_PTYPE_L3_IPV6), so the else-if branch is skipped and tap_verify_csum() returns without verifying L3 or L4 checksums. The hlen now carries the correct l3_len including extensions, so this is a prerequisite for IPv6-ext support but not the feature itself. Suggest rewording to "This is a prerequisite for IPv6 extension header support" or dropping that half of the sentence. - drivers/net/tap/rte_eth_tap.c:549: minor scoping observation only (no action needed) -- the hlen struct is filled unconditionally on every received packet whereas previously rte_net_get_ptype() was called with NULL when offload check was disabled. The cost is a handful of stack writes per packet; consistent with how other PMDs handle this.
