Re: accel/tcg/translator.c question about translator_access

[Richard Henderson]

On 1/31/23 17:06, Sid Manning wrote:
> There is an assert in translator_access that I hit while running on a version of QEMU 
> integrated into a Virtual Platform.
>
> Since this function can return null anyway I tried the following experiment:
...
> -            assert(phys_page != -1);
> +            if(phys_page == -1) {
> +                return NULL;
> +            }
...
> which avoided the issue and the test ran to completion.  What is this assert 
> trying to catch?


One half of the instruction in ram and the other half of the instruction in 
mmio.

If the entire instruction is in mmio, then we correctly translate, but do not cache the 
result (since the io can produce different results on every access).  But if we have 
started caching the result, because we start in ram, then we will incorrectly cache the 
mmio access.

This really should never happen.  How did it occur?


r~