Skip to content

PIT evaluate drops sub-100ns time, so the PIT runs fast and same-instant reads differ #1065

Closed

Description

@q5438722

Summary

PitDevice::evaluate counts whole 838ns ticks since last and advances last by them. VmTime has 100ns resolution, so wrapping_add truncates the advance, and the next evaluation counts the dropped 0–99ns again. Every PIO access to the PIT or port 0x61 is an evaluation, so the PIT runs fast, and faster the more often the guest touches it. Two reads at the same VM time can also return different counts.

Repro

pit_sub_100ns_leak.txt

Add the attached pit_sub_100ns_leak.rs to vm/devices/chipset/tests/ and run:

cargo test -p chipset --test pit_sub_100ns_leak

Both tests program counter 0 in mode 2 with count 0x50, then advance VM time by 2.5µs (2 ticks + 824ns):

  • reads_at_the_same_instant_agree latches the count twice at the same VM time.
  • restored_device_reads_the_same_count saves the device, restores it into a new device, and latches at the same VM time.

Expected

Both reads return the same count, and the restored device reads the same count as the original.

Actual

thread 'reads_at_the_same_instant_agree' panicked at vm/devices/chipset/tests/pit_sub_100ns_leak.rs:62:9:
assertion `left == right` failed: counter moved between two reads at the same VM time
  left: 79
 right: 78

The first evaluation counts 2 ticks (1676ns) but advances last by only 1600ns. The second evaluation then sees 900ns and counts a third tick. The restore test fails the same way (78 vs. 79).

Root cause

pit/mod.rs:541-545:

let delta = now.checked_sub(self.last).unwrap_or(Duration::ZERO);
let ticks = delta.as_nanos() as u64 / NANOS_PER_TICK;
self.last = self
    .last
    .wrapping_add(Duration::from_nanos(ticks * NANOS_PER_TICK)); // truncated to 100ns

The comment above this code explains why last must not be set to now: the timers would run late. Truncating to 100ns makes them run fast instead.

Impact

  • The PIT runs fast by up to 99ns per evaluation, so the drift depends on how often the guest accesses the PIT. For example, a guest that busy-polls the counter or port 0x61 to calibrate a clock sees a fast PIT.
  • PIT reads are not deterministic for a given VM time, which affects save/restore and replay.

Suggested fix

Carry the remainder below 100ns in a new field, last_frac_ns, and include it in evaluate and arm_wakeup. The field is saved as a new optional #[mesh(3)] field; without it, a restored device would count the remainder again. Old saved states restore with 0, and restore rejects values >= 100.

fix.patch

An alternative is to count ticks from a fixed origin instead of accumulating last.

Metadata

Metadata

Labels

No labels
No labels

Type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions