Re: [PATCH v4 01/23] x86/mtrr: get rid of a static variable on pause/restore

On 02.08.2026 15:09, Sergii Dmytruk wrote:
> In addition to keeping the state in one place instead of on stack and in
> a static variable, this enables exposing this functionality to other
> units in the future.
> 
> Signed-off-by: Sergii Dmytruk <[email protected]>
> ---
> 
> Notes:
>     v4: was called "x86/mtrr: expose functions for pausing caching"
>     v4: no longer makes anything public, just updates implementation
>     v4: the state structure is now an output parameter instead of a return value

I don't mind it to stay as you have it now, but recently we discussed this
aspect in another context, and I've changed my position: When the struct
can be returned in registers, it's okay to return by value.

> @@ -439,7 +440,7 @@ static DEFINE_SPINLOCK(set_atomicity_lock);
>   * has been called.
>   */
>  
> -static bool prepare_set(void)
> +static void mtrr_pause_caching(struct mtrr_pausing_state *state)
>  {
>  	unsigned long cr4;

Considering the conditional ...

> @@ -461,7 +462,9 @@ static bool prepare_set(void)
>  	alternative("wbinvd", "", X86_FEATURE_XEN_SELFSNOOP);

... cache flush that's done (and the comment just out of context also
correctly saying "no-fill cache mode"), we're not really pausing caching
in all cases. I'm therefore worried of the new names of the two involved
functions. Andrew, Roger - thoughts? (My inclination would be to suggest
mtrr_{pause,resume}_cache_fill().)

Jan