Net: the Temple rests on the Sabbath, the CIA burns its own CPU

God rested on the seventh day. The Temple's network stack never rested at all.

The sin

An idle Temple on the VDS, with nobody knocking, ate 100% of a CPU. The hosting company billed us for the sin of gluttony, and the glowing ones in Langley would have been proud.

The scheduler already has a HLT (Kernel/Sched.HC), but Seth reaches it only after a full round in which no task was ready. Every wait in the network stack was a while (...) Yield; loop, and a task that Yields is ready again at once. So Seth never had a quiet round and never reached HLT:

• NetHandler staring at an empty frame fifo
• accept() on the Confessional (:23) and HtServ (:80)
• recvfrom() on the DNS oracle (:53)
• the TCP waits in connect, send, recv and close. Close could spin up to TCP_CLOSE_TIMEOUT for every connection a port scanner abandoned. The CIA scans, the Temple pays.

TASKf_IDLE did not help. It is bookkeeping for the CPU meter, it puts nobody to sleep.

The repentance

• Tcp.HC, Udp.HC: every wait loop now does Sleep(1) instead of hosting company does.
• NetHandlerTask.HC: the handler lies down on wake_jiffy (at most 10 ms) before it looks in the fifo, behind a locked XCHG (LXchgI64). A frame that lands in between finds it already lying down, so no wakeup is lost.
• NetFifo.HC: the NIC's IRQ calls the handler by name, the way God called Samuel in the temple. It zeroes wake_jiffy in NetFifoPushCopy(), so frames still go up the stack at once.

No kernel changes. Adam recompiles at boot, a reboot after tools/update_vds.py is enough.

Testimony

QEMU, idle live CD, pcnet, 2 vCPU, TCG on a Mac:

	Host CPU at idle
before	~98–100% of a core
after	~27–30% of a core

What is left is mostly TCG emulating a 1 kHz timer and drawing the screen. Under KVM on the VDS it should be lower, not measured there yet. The Confessional answered three commands (accept, recv, send, close) in 0.16 s.

Not tested: HtServ over HTTP (it is off on the live CD unless HtServEnable;) and the UDP side (NTP, DNS oracle) on their own.

What remains

Sleep(1) is still polling: one look per tick per waiting task. It costs almost nothing only because TempleOS ticks at 1 kHz anyway. The righteous way is wait channels, as in BSD tsleep()/wakeup(): the socket remembers its waiting task and NetHandler wakes it when data arrives. That is a later MR.

Be still, and know that I am God. Psalm 46:10.