Re: [PATCH 09/11] replay: stop us hanging in rr_wait_io_event

[Pavel Dovgalyuk]

On 05.12.2023 23:41, Alex Bennée wrote:
> A lot of the hang I see are when we end up spinning in
> rr_wait_io_event for an event that will never come in playback. As a
> new check functions which can see if we are in PLAY mode and kick us
> us the wait function so the event can be processed.
>
> This fixes most of the failures in replay_kernel.py

Is there an effect for console QEMU only?
I've tested this patch on Windows7 boot scenario and replay speed has 
not changed.

> Fixes: https://gitlab.com/qemu-project/qemu/-/issues/2013
> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
> Cc: Pavel Dovgalyuk <pavel.dovgaluk@ispras.ru>
> ---
>   include/sysemu/replay.h      |  5 +++++
>   accel/tcg/tcg-accel-ops-rr.c |  2 +-
>   replay/replay.c              | 24 ++++++++++++++++++++++++
>   3 files changed, 30 insertions(+), 1 deletion(-)
>
> diff --git a/include/sysemu/replay.h b/include/sysemu/replay.h
> index 08aae5869f..83995ae4bd 100644
> --- a/include/sysemu/replay.h
> +++ b/include/sysemu/replay.h
> @@ -70,6 +70,11 @@ int replay_get_instructions(void);
>   /*! Updates instructions counter in replay mode. */
>   void replay_account_executed_instructions(void);
>   
> +/**
> + * replay_can_wait: check if we should pause for wait-io
> + */
> +bool replay_can_wait(void);
> +
>   /* Processing clocks and other time sources */
>   
>   /*! Save the specified clock */
> diff --git a/accel/tcg/tcg-accel-ops-rr.c b/accel/tcg/tcg-accel-ops-rr.c
> index 611932f3c3..825e35b3dc 100644
> --- a/accel/tcg/tcg-accel-ops-rr.c
> +++ b/accel/tcg/tcg-accel-ops-rr.c
> @@ -109,7 +109,7 @@ static void rr_wait_io_event(void)
>   {
>       CPUState *cpu;
>   
> -    while (all_cpu_threads_idle()) {
> +    while (all_cpu_threads_idle() && replay_can_wait()) {
>           rr_stop_kick_timer();
>           qemu_cond_wait_iothread(first_cpu->halt_cond);
>       }
> diff --git a/replay/replay.c b/replay/replay.c
> index e83c01285c..042a6a9636 100644
> --- a/replay/replay.c
> +++ b/replay/replay.c
> @@ -347,6 +347,30 @@ void replay_start(void)
>       replay_enable_events();
>   }
>   
> +/*
> + * For none/record the answer is yes.
> + */
> +bool replay_can_wait(void)
> +{
> +    if (replay_mode == REPLAY_MODE_PLAY) {
> +        /*
> +         * For playback we shouldn't ever be at a point we wait. If
> +         * the instruction count has reached zero and we have an
> +         * unconsumed event we should go around again and consume it.
> +         */
> +        if (replay_state.instruction_count == 0 && 
> replay_state.has_unread_data) {
> +            return false;
> +        } else {
> +            fprintf(stderr, "Error: Invalid replay state\n");
> +            fprintf(stderr,"instruction_count = %d, has = %d, event_kind = 
> %d\n",
> +                    replay_state.instruction_count, 
> replay_state.has_unread_data, replay_state.data_kind);
> +            abort();
> +        }
> +    }
> +    return true;
> +}
> +
> +
>   void replay_finish(void)
>   {
>       if (replay_mode == REPLAY_MODE_NONE) {