[JackDoan]

maybe don't merge until 1.12 is cut

[cloudflare-workers-and-pages]

Deploying docs-nebula with    Cloudflare Pages

Latest commit:	1a15505
Status:	✅  Deploy successful!
Preview URL:	https://d2a43285.docs-nebula.pages.dev
Branch Preview URL:	https://tun-offloads.docs-nebula.pages.dev

View logs

[jasikpark]

Checked this against slackhq/nebula@abff1a4. Accurate throughout, including the batch memory math and the platform caveats. One thing worth adding:

listen.batch: 1 silently disables GRO. udp_linux.go:80 is the only prepareGRO call site and reads if s.Batch > 1 && s.Offloads, with nothing logged when it's skipped. Since the new listen.batch paragraph invites lowering the value to save memory, worth noting that 1 turns receive-side offload off entirely. The gate is GRO-only, GSO still applies at batch: 1.

Two optional wording nits on listen.udp_offloads, both about "probed when the socket is created, and Nebula falls back to standard sends and receives":

• The send-side fallback is still sendmmsg with one entry per datagram, so "unbatched sends" is more precise than "standard sends".
• Fallback isn't only at creation. An EIO on a superpacket disables GSO at runtime and replays the batch, since the setsockopt probe can't see per-route limitations.

[JackDoan]

Superseded by #390