Per-CPU scripts

lunatik run [--context=softirq | hardirq] --percpu <script> creates one runtime per CPU id and dispatches a callback to the runtime of the CPU it fires on. The runtimes share the registrations a script makes once for all of them, a netfilter hook and a kprobe, and each reads its own id with lunatik.cpu().

A runtime is a CPU, not a connection. A netfilter hook runs in the packet’s own processing path (NF_HOOK from ip_rcv, ip_output and their peers), so the runtime is whichever CPU that path is on, and the packets of one connection reach several. Over loopback and veth the transmit side queues the packet to its own CPU’s backlog (__netif_rx from loopback_xmit and veth_forward_skb), which the receive softirq drains on that CPU. On a NIC it is the CPU the queue’s interrupt is bound to; with RPS, the one the queue’s map picks from the flow hash, stable while the map is; with RFS, the one where the flow’s last recvmsg ran, which follows a reader that migrates. An outbound hook reached from a sendmsg runs on the sending process’s CPU, but the same hook number also fires from the receive softirq, forwarding a packet or sending a RST for one, and from the timer softirq on a retransmission, where it is that softirq’s CPU. A kprobe reaches the runtime of the CPU the probed call ran on.

State that must see a whole flow therefore belongs in something the runtimes share, a table published in lunatik._ENV or the conntrack mark; the runtime holds what is per-CPU, a counter or a cache. An rcu.table() the script body creates is not shared: the body runs once per runtime, so each gets its own.

lunatik run creates one runtime per possible CPU and runs their bodies one after another, so a body publishes the shared table when it finds none, and the bodies after it take that one:

local lunatik = require("lunatik")
local rcu     = require("rcu")

local env = lunatik._ENV
env.flows = env.flows or rcu.table()
local flows = env.flows

The table stays in lunatik._ENV after the runtimes stop, until a script sets env.flows to nil.

A binding whose global registration the runtimes cannot share refuses a percpu runtime with not allowed in a percpu runtime, which its @raise says.