Re: Some more questions with regards to QEMU clock record and replay

[Arnabjyoti Kalita]

Thank you Pavel.

Your answers make the clock record-replay process much clearer to me now.

Best Regards,

Arnab

On Tue, Mar 2, 2021 at 12:49 PM Pavel Dovgalyuk <pavel.dovgalyuk@ispras.ru> wrote:

On 01.03.2021 20:16, Arnabjyoti Kalita wrote:
> Hello all,
>
> I am really thankful for the wonderful answers in my last post linked below-
>
> https://lists.nongnu.org/archive/html/qemu-discuss/2021-02/msg00131.html
>
> In continuation with the last post, I have a few more questions to ask -
>
> My experiment is still, mostly the same. I record clock values in KVM
> mode, and then replay the clock values in TCG mode. However, now I am
> recording and replaying all of the clock values (I was only
> recording/replaying the host clock previously). However, I do not use
> the -icount feature.
>
> - Why are clock values being replayed at checkpoints?

Timers are replayed at checkpoints to be synchronized with vCPU.
Other clock requests (e.g., caused by vCPU instruction) are replayed
immediately.

> - Can we ignore replaying at checkpoints and do a dumb replay as and
> when the clock read actually happens?

I think we can, if we need just clock synchronization.

> - Based on the documentation available, I can see that checkpoints are
> necessary for thread synchronization. Does this mean, if I do not replay
> clock values at checkpoints, the guest kernel scheduler might behave
> incorrectly during replay ?

Checkpoints are related to QEMU threads, not guest threads.
Timers are needed for virtual devices, that can generate interrupts, DMA
requests and so on. Therefore we synchronize them with vCPU to make
execution deterministic.


Pavel Dovgalyuk