[PATCH 9/9] iosys_map: embed the is_iomem bit into the pointer.

[Dave Airlie]

On Fri, 23 May 2025 at 01:10, Lucas De Marchi <lucas.demarchi at intel.com> wrote:
>
> On Thu, May 22, 2025 at 04:52:18PM +1000, Dave Airlie wrote:
> >From: Dave Airlie <airlied at redhat.com>
> >
> >This reduces this struct from 16 to 8 bytes, and it gets embedded
> >into a lot of things.
> >
> >Signed-off-by: Dave Airlie <airlied at redhat.com>
>
> Replied too early on cover. Chatting with Michal Wajdeczko today, this
> may break things as we then can't byte-address anymore. It seems
> particularly dangerous when using the iosys_map_wr/iosys_map_rd as
> there's nothing preventing an unaligned address and we increment the map
> with the sizeof() of a struct that could be __packed. Example: in
> xe_guc_ads.c we use it to write packed structs like guc_gt_system_info.
> In this particular case it doesn't give unaligned address, but we should
> probably then protect iosys_map from doing the wrong thing.
>
> So, if we are keeping this patch, we should probably protect
> initially-unaligned addresses and the iosys_map_incr() call?

oh interesting, my thoughts was of course nobody would want to use
this for < 32-byte aligned ptrs :-)

but I forgot about using the incr for stuff, I do wonder if the incr
could be modelled on a base + offset, as I do think for a lot of stuff
we'd want to retain the original vaddr for unmapping or other things,

I'll play around a bit more next week with at least protecting against bad uses.

Dave.