Hi, I wrote a test for migration auto converge and found out a strange thing: 1. Enable auto converge 2. Set max-bandwidth 1Gb/s 3. Set downtime-limit 1ms 4. Run standard test (just writes a byte per page) 5. Wait for converge 6. It's converged with 99% throttle percentage 7. The result downtime was about 300-600ms <<<<
It's much higher than expected 1ms. I figured out that cpu_throttle_thread() function sleeps for 100ms+ for high throttle percentage (>=95%) in VCPU thread. And it sleeps even after a cpu kick. I tried to fix it by using timedwait for ms part of sleep. E.g timedwait(halt_cond, 1ms) + usleep(500). But I'm not sure about using timedwait function here with qemu_global_mutex. The original function uses qemu_mutex_unlock_iothread + qemu_mutex_lock_iothread It differs from locking/unlocking (inside timedwait) qemu_global_mutex because of using qemu_bql_mutex_lock_func function which could be anything. This is why the series is RFC. What do you think? Thanks! Yury Kotov (2): qemu-thread: Add qemu_cond_timedwait cpus: Fix throttling during vm_stop cpus.c | 27 +++++++++++++++++++-------- include/qemu/thread.h | 12 ++++++++++++ util/qemu-thread-posix.c | 40 ++++++++++++++++++++++++++++------------ util/qemu-thread-win32.c | 16 ++++++++++++++++ 4 files changed, 75 insertions(+), 20 deletions(-) -- 2.22.0
