Re: [PATCH v2 09/12] accel/tcg: Introduce EXCP_TB_FLUSH

[Richard Henderson <richard.henderson@linaro.org>]

On 9/23/25 00:10, Paolo Bonzini wrote:
> On 9/23/25 04:39, Richard Henderson wrote:
>> We are going to disallow tb_flush from within the context
>> of a running cpu.  Introduce a tcg-internal exception to
>> return out of the cpu run loop and perform the flush there.
>>
>> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> 
> One small difference is that different CPUs can fail tcg_tb_alloc() at the same time and 
> flush multiple times.
> 
> I think the check on the generation count should remain.  Instead of introducing 
> EXCP_TB_FLUSH, you can keep the guts of tb_flush() as
> 
> void tb_queue_flush(CPUState *cpu)
> {
>      unsigned tb_flush_count = qatomic_read(&tb_ctx.tb_flush_count);
>      async_safe_run_on_cpu(cpu, do_tb_flush,
>                            RUN_ON_CPU_HOST_INT(tb_flush_count));
> }
> 
> With the unconditional async_safe_run_on_cpu() hidden behind a function, this patch goes 
> away while the next one survives as
> 
>       if (unlikely(!tb)) {
> -        /* flush must be done */
> -        tb_flush(cpu);
>           mmap_unlock();
> -        /* Make the execution loop process the flush as soon as possible.  */
> -        cpu->exception_index = EXCP_INTERRUPT;
> +        tb_queue_flush(cpu);

You have a point.  It's not even that unlikely a scenario.


r~