[dbernheisel]

A host that calls back into Lua from an Elixir function can't stop just that callback when it runs long. :max_instructions is per evaluation, and killing the process throws away the whole state. Lua.set_deadline/2 gives running Lua a System.monotonic_time(:millisecond) deadline. Past it, Lua raises a catchable "deadline exceeded" at the next loop back-edge or call, the same places :max_instructions is checked. It's what a lua_sethook count hook gives a host in reference Lua.

• the clock is read every 256 ticks, and only while a deadline is set
• a pcall that traps the error is interrupted again on its next tick, so the error reaches the host
• Lua.deadline/1 returns the current deadline, so a host can give a nested call a shorter one and restore it after

Benchmarks

feat/vm-gc (f5d18a5) vs this branch (105bb9c), lua (chunk) median, quick mode. Base and branch ran interleaved per workload, two rounds averaged (oop five rounds). Apple M4 Max, OTP 29, Elixir 1.20.0. Other work was running on the machine (load 5–11), so a few percent either way is noise.

Workload	Base	Branch	Change
fibonacci	456.3 ms	456.6 ms	+0.1%
closures	411.7 μs	411.8 μs	+0.0%
oop	114.2 μs	116.5 μs	+2.0%
metamethods: self-call method dispatch	511.9 μs	495.4 μs	-3.2%
metamethods: 3-level __index chain	208.8 μs	216.3 μs	+3.6%
metamethods: arithmetic/relational	2.26 ms	2.15 ms	-4.9%
pcall success path	247.6 μs	238.9 μs	-3.5%
pcall raise + catch	823.4 μs	802.5 μs	-2.5%
varargs + multiple returns	3.58 ms	3.45 ms	-3.6%
table.concat	35.9 μs	34.2 μs	-4.6%
string.format	85.8 μs	83.4 μs	-2.8%
table build	17.8 μs	17.2 μs	-3.5%
table sort	23.3 μs	23.1 μs	-0.7%
table iterate/sum	24.9 μs	25.2 μs	+1.0%
table map + reduce	49.9 μs	50.0 μs	+0.3%
table pairs (hash)	85.5 μs	83.6 μs	-2.2%

Cost when a deadline is set, on this branch: 4M loop iterations in a compiled function, nine runs each.

Limit	Min	Median
none	245 ms	251 ms
deadline set, not reached	259 ms	266 ms
max_instructions: 1_000_000_000	338 ms	345 ms