With tun.staged.workers set, each Linux UDP listener classifies a recvmmsg
batch, hands its data packets to that many decrypt workers, and an ordered
writer commits the batch in arrival order: replay window, roaming, firewall,
control packets and tun writes go through the same helpers inline receive
uses. Six batches per routine bound memory and backpressure. Conns that cannot
hand off their buffers stay inline.
The listener's classification is speculative. Before touching replay state the
writer resolves each data packet through its own batch-local hostmap cache,
exactly where inline receive would, so data after a close in the same batch is
still delivered while data after a removal or replacement between batches is
dropped, with the same RecvError reply inline receive sends. Known replays are
marked by a precheck and skip decryption, but keep their place so that reply
stays identical when an earlier RecvError deleted their tunnel.
New API: udp.StdConn.ListenOutHandoff, and the same on TesterConn for e2e,
runs the receive loop while handing each batch's buffers to the caller in
exchange for replacements. It is found through an unexported interface in
nebula, so udp.Conn and its other implementations are unchanged.
Metrics: flow.rx.batches, and flow.stage.{rx_read,rx_decrypt,rx_write}.q<N>.busy_ns.
Tests cover both ciphers, ordering, duplicate and corrupted packets from other
addresses, window jumps, control packets mid-batch, RecvError then replay,
tunnel replacement, firewall parity, fallback, real UDP sockets, partial
batches at close, and a handshake and traffic through e2e that checks pipeline
use on Linux and the inline fallback elsewhere. No isolated receive throughput
claim is made.
Set tun.staged.workers to a positive count on Linux, with tun.use_offloads
enabled, to move encryption off each tun reader. The reader filters and
reserves counters for a whole read, hands the batch to that many encrypt
workers, and an ordered sender flushes the batches in read order. Eight slots
per routine bound the batches in flight; a full pipeline blocks the reader.
Zero workers, the default, keeps today's inline path, as do tun queues that
cannot hand off their read buffer and builds whose AES-GCM needs nonce-ordered
encryption (BoringCrypto, FIPS).
IMPORTANT: staged senders need peers with the 8192-entry replay window of
master. Released v1.11.x peers use 1024 and will drop reordered batches. The
window is not negotiated.
Readers keep today's pinning. The sender is only pinned with an explicit
tun.cpu_affinity: sender i uses entry i+1, unless that is its reader's CPU,
so a one-entry list leaves the sender unpinned rather than sharing it: two
busy transmit threads on one CPU serialise (eight TCP streams over 100 GbE
fell from 25.9 to 22.8 Gbit/s). Without a list the sender is not pinned. Placing it by queue index on a
hybrid CPU put it on an efficiency core, and one TCP stream over 100 GbE
reached 18.6-19.6 Gbit/s pinned against 20.2-20.6 unpinned.
New API, each needed across a package boundary:
- tio.Offload.SwapRxBuf hands the buffer of the last Read to the reader and
reads into the one given. It is a method on the concrete queue, found through
an unexported interface in nebula, so tio.Queue and its other
implementations are unchanged.
- batch.SendBatch.Reset drops queued packets without sending, for a sender
whose sockets are already closed; Flush now uses it.
Metrics, registered only for active pipelines: flow.tx.batches counts reads
handed to the sender, flow.stage.{tx_read,tx_encrypt,tx_send}.q<N>.busy_ns the
elapsed busy time of each stage, to find the limiting one.
Tests cover ordering and counters under backpressure with both ciphers (AES
skipped where nonce order is required), read errors with batches in flight,
shutdown, exhaustion, relays, fallback, firewall parity, and the actual reader
and send thread affinity for every placement, with pinning on and off: the
sender is pinned only to an explicit entry other than its reader's CPU.
* examples: set socket buffers and raise tun tx_queue
* examples: use the default tcp max window for socket buffers
* examples: note buffers are linux only, clarify tx_queue default
* tun: default tx_queue to 5000
* add IPv6 reject packet generation (ICMPv6 Destination Unreachable and TCP RST)
* use ICMPv6 code 1 (administratively prohibited) and cap body at 1000 bytes
* cleanup, use ICMP error code 13 for ipv4
* better docs
* cleanup
* add sshd.sandbox_dir config option
Sanitize SSH profile paths (ssh.go:514,683,719) — restrict os.Create(a[0]) to a safe directory.
Add a config option in the config file to specify the sandbox directory. For backwards compatibility, if the config is not specified, keep the current behavior.
* update default and example
* use os.TempDir() for sshd.sandbox_dir default
* split sandbox path validation into separate conditionals
Separate the combined && check in sshSanitizeFilePath into two distinct
conditionals with specific error messages: one for paths resolving to the
sandbox directory itself, and one for paths outside the sandbox.
Co-Authored-By: Claude <svc-devxp-claude@slack-corp.com>
* fix: trim leading zeros from p256 signature swap result
bigmod.Nat.Bytes() returns fixed-size 32-byte slices, but ASN.1 INTEGER
parsing strips leading zeros. This caused a flaky test failure (~1/256
chance) when the S value's high byte was zero.
Co-Authored-By: Claude <svc-devxp-claude@slack-corp.com>
---------
Co-authored-by: Claude <svc-devxp-claude@slack-corp.com>
- Accept certs signed by trusted CAs
- Username must match the cert principal if set
- Any username can be used if cert principal is empty
- Don't allow removed pubkeys/CAs to be used after reload
Ryan Huber