Re: [PATCH 4/5] xen/xsm: Improve fallback handling in xsm_fixup_ops()

[Jan Beulich <jbeulich@xxxxxxxx>]

On 05.11.2021 14:55, Andrew Cooper wrote:
> +void __init xsm_fixup_ops(struct xsm_ops *ops)
> +{
> +    /*
> +     * We make some simplifying assumptions about struct xsm_ops; that it is
> +     * made exclusively of function pointers to non-init text.
> +     *
> +     * This allows us to walk over struct xsm_ops as if it were an array of
> +     * unsigned longs.
> +     */
> +    unsigned long *dst = _p(ops);
> +    unsigned long *src = _p(&dummy_ops);

I'm afraid I consider this an abuse of _p(): It hides casting when
that would better not be hidden (and there's then also a pointless
step through "unsigned long" in the casting). I suppose this is
also why "src" didn't end up "const unsigned long *" - with spelled
out casts the casting away of const might have been more noticable.

> +    for ( ; dst < (unsigned long *)(ops + 1); src++, dst++ )
> +    {
> +        /*
> +         * If you encounter this BUG(), then you've most likely added a new
> +         * XSM hook but failed to provide the default implementation in
> +         * dummy_ops.
> +         *
> +         * If not, then perhaps a function pointer to an init function, or
> +         * something which isn't a function pointer at all.
> +         */
> +        BUG_ON(!is_kernel_text(*src));

Just as a remark, not a request to change anything: A cause of this
triggering may also be is_kernel_text() not covering all text
sections. Some of what recently we've been talking about informally
may lead to new text section variants appearing, and whether those
would sensibly end up inside [_stext,_etext) is uncertain.

Jan