Re: Questions about Disk I/O Record and Replay

[Pavel Dovgalyuk]

On 13.08.2021 08:53, Arnabjyoti Kalita wrote:
> Hello all,
>
> I am trying to use the blkreplay driver to record and replay disk
> operations. My disk record and replay mechanism is a variant of the
> inbuilt record replay mechanism in QEMU.
>
> During record, I just store the completion ids of the disk operations
> as they happen. Something like this -
>
> void replay_block_event(QEMUBH *bh, uint64_t id)
> {
>      if (replay_mode == REPLAY_MODE_RECORD) {
>          if (start_recording) {
>              replay_put_qword(id, "disk");  // writes ids to a separate
> file dedicated for disk I/O record
>          }
>          qemu_bh_schedule(bh);
>      }
> }
>
> During replay, all I do is store the disk ID into an events list and
> ask the CPU to continue executing instructions.
>
> void replay_block_event(QEMUBH *bh, uint64_t id)
> {
>     if (replay_mode == REPLAY_MODE_RECORD) {
>         /* as shown above */
>     }
>     if (replay_mode == REPLAY_MODE_PLAY) {
>         BlockEvent *event = g_malloc0(sizeof(BlockEvent));
>         event->opaque = bh;
>         event->id = id;
>         QTAILQ_INSERT_TAIL(&blk_events_list, event, blk_events);
>         qemu_cpu_kick(first_cpu);                    <- Replayed guest
> gets stuck here
>     }
>     else {
>         qemu_bh_schedule(bh);
>     }
> }
>
> I know when the disk interrupt happens and so all the event ids that
> we store will be replayed right before that.
>
> My questions are as follows -
>
> 1. Is the above approach a good idea to achieve deterministic disk I/O replay?
>
> 2. Should I also replay disk event I/O as and when they arrive? How do
> I ensure that it actually gets completed before a checkpoint? 

I do not remember the details about this.

> checkpoint in this case, is the disk interrupt.
>
> 3. How can I handle out-of-order disk I/O completions, in this
> scenario? How should I proceed?

This is handled by saving events at the checkpoint.

> I would love to get more details on these questions.
>
> Thank you very much for all your help so far.
>
> Best Regards,
> Arnabjyoti Kalita