Re: [PATCH 0/9] accel/tcg: Fix page_set_flags and related [#1528]

[Joel Stanley]

On Mon, 6 Mar 2023 at 02:14, Richard Henderson
<richard.henderson@linaro.org> wrote:
>
> The primary issue is that of overflow, where "end" for the last
> page of the 32-bit address space overflows to 0.  The fix is to
> use "last" instead, which can always be represented.
>
> This requires that we adjust reserved_va as well, because of
>
> -/*
> - * There are a number of places where we assign reserved_va to a variable
> - * of type abi_ulong and expect it to fit.  Avoid the last page.
> - */
> -#   define MAX_RESERVED_VA  (0xfffffffful & TARGET_PAGE_MASK)
>
> and the related
>
> -        /*
> -         * reserved_va must be aligned with the host page size
> -         * as it is used with mmap()
> -         */
> -        reserved_va = local_max_va & qemu_host_page_mask;
>
> whereby we avoided the final (host | guest) page of the address space
> because of said overflow.  With the change in representation, we can
> always use UINT32_MAX as the end of the 32-bit address space.
>
> This was observable on ppc64le (or any other 64k page host) not being
> able to load any arm32 binary, because the COMMPAGE goes at 0xffff0000,
> which violated that last host page problem above.
>
> The issue is resolved in patch 4, but the rest clean up other interfaces
> with the same issue.  I'm not touching any interfaces that use start+len
> instead of start+end.

Thanks for looking at this Richard. I gave it a spin on a ppc64le host
and it resolved the assert.

Tested-by: Joel Stanley <joel@jms.id.au>

Cheers,

Joel