perf(tfork): add opt-in residual latency probes- #11

[yiying-zhang]

Summary

• add aggregate CRIU_TFORK_PROFILE=1 Phase A task-loop buckets;
• mark Phase B parent setup, restore-child phases, and every restore futex wait;
• keep normal behavior and log volume unchanged when profiling is disabled;
• allow the Phase B benchmark to pass extra clone arguments.

Motivation

Earlier profiling left 168.5 ms of the 100-process Phase A loop unassigned and
266–306 ms of Phase B after polling was replaced with futex wakeups. These
probes distinguish real work from synchronization slack without a per-task log
line.

Results

At 100 processes, the first-fork medians assign the 509.805 ms task loop mainly
to core/thread serialization (93.683 ms), page handoff (86.470 ms), parasite
infection (83.064 ms), VMA collection (73.169 ms), stop/thread/cure (40.064
ms), and MM/FS serialization (38.477 ms).

Phase B futex waits total 237.891 ms. The dominant waits are 190.512 ms for the
100-task restore tree and 44.549 ms for namespace preparation. This is useful
work/dependency time, not residual 100 ms polling quantization.

An interleaved overhead check found no measurable Phase A penalty (584.336 ms
enabled versus 588.673 ms disabled; -0.7% noise).

Validation

• remote Linux build passed;
• ten independently prepared 100-process sources completed successfully;
• enabled/disabled overhead A/B completed successfully;
• exact cleanup left no owned residual resources.