idle-pause: wake for an interrupt masked by a raised spl

[slaapliedje]

An idle-park build can hang IRIX for good. We hit it on an IRIX 6.5.22 Indy (R4400, a nightly with lightning,jitv2,idle-pause) after some hours of light use: keyboard, serial console and telnet all dead, the MIPS-CPU thread at 0%. The monitor showed the CPU parked:

stopped pc=ffffffff880e3a98
CP0 Status: 0x00004001 (IM:_6______ K:K IE)

IE with IM6 alone is IRIX's splhi: on IP22, IP6 is the bus error, the one line left unmasked at high spl. The code at the PC is a loop that raises the spl, calls a function, drops it again (splx) and tests a flag an interrupt handler clears. Most of each pass is at high spl, so the detector's batch-end samples land there, the state repeats, and park waits for ip & im. The timer's interrupt then arrives masked. The guest would take it a few instructions later at splx, but a parked CPU never gets there, and nothing raises IP6. stop / step / cont on the monitor freed it: the first step went straight to the exception vector.

This is the sibling of idle-pause-must-not-park-with-interrupts-masked.md, with IM partly clear rather than zero.

The change

src/cpu/idle_park.rs. The mask update samples is the guest's at that moment, not the loop's, so:

• update declines to park while any line is pending, masked or not: the guest is about to take it.
• park returns when any line becomes pending, masked or not.

"Nothing pending, wait for anything" can always be satisfied by the guest's own timer, whatever spl the sample caught. A line held pending and masked for a long time would keep the CPU from parking (it spins, as without idle-pause), never from running. The IM == 0 guard stays.

Also: rules/perf/idle-pause-masked-interrupt-must-wake.md with the details, a pointer to it from the IM == 0 rule, and a CHANGELOG entry.

Tests

Each new test hangs until its deadline (and fails) without the fix:

• update_declines_to_park_while_a_masked_interrupt_is_pending (idle_park.rs).
• a_masked_interrupt_ends_the_park (idle_park.rs): parked with only IM6 set, IP4 raised from another thread.
• idle_park_wakes_for_an_interrupt_masked_by_a_raised_spl (mips_exec_test.rs): the spl loop as guest code, run through step_int, 1000-step batches, update and park as MipsCpu::start does; an interrupt handler sets the flag; IP4 is raised from another thread once the CPU has parked; the loop must finish.

Tests that park now take PARK_TEST_LOCK. There is one PARKER slot, so two tests parking at once wake each other's threads. They also wait for parked() rather than a guessed sleep, so the lock is held only briefly.

• cargo test --release --features idle-pause --lib: 1056 passed.
• --features idle-pause,jitv2 --lib: 1530 passed. One existing test, an_interrupt_ends_the_park_without_waiting_out_the_slice, measures wake latency with a 300 µs limit while the whole suite runs, and goes over it now and then. That rate is unchanged by this PR: 1 run in 25 with this change, and 1 in 25 on main, run alternately on the same machine.
• IRIX 6.5.22 on the Indy, a build with the nightly's features plus this change, no_idle = false:
  • Boots to login.
  • Idles at about 15% of a host core, against 95%+ with no_idle = true, so parking still pays.
  • A two-hour soak, a round every 20 s of disk (tar of /usr/include, find), NFS (cksum of a host file) and network (ping) work with the guest otherwise idle: all 347 rounds answered. Before the fix the same guest hung within hours.

Found while using iris as the test machine for an IRIX port of the Manx browser. Thanks for the emulator.

🤖 Generated with Claude Code

https://claude.ai/code/session_01APKCDaLfEv8mFqUmDtAruT

[techomancer]

That's why I like my laptop to sound like turbo jet! I't makes it autentic fast fan sgi experience ;-)