core: fix core_mmu_map_pages()

[jenswi-linaro]

Adds missing dsb_ishst() at the end of core_mmu_map_pages() needed to
guarantee that changes to translation tables are visible.

Reported-by: Stuart Yoder stuart.yoder@arm.com
Signed-off-by: Jens Wiklander jens.wiklander@linaro.org

[jenswi-linaro]

@stuyoder does this fix #1805 ?

[etienne-lms]

@jenswi-linaro, I wonder why this issue does not occur with the pager, where pages can be loaded from the pagefault handler without dsb in some conditions (bsd only when valid mmu descriptors are updated) before handler return to normal execution ?

[jenswi-linaro]

We're doing TLB invalidations instead. Or have you been able to find a suspicious location?

[etienne-lms]

Sorry, my mistake. tee_pager_handle_fault() always ends with a tbli or dsb.

[etienne-lms]

Reviewed-by: Etienne Carriere <etienne.carriere@linaro.org>

Maybe we could add in the comments of prototypes of core_mmu_map_pages()/core_mmu_unmap_pages() that caches are synchronized against updated mapping when these routines return.

[stuyoder]

Yes, I confirmed that translate fault I see goes away with this patch applied.

[jenswi-linaro]

@stuyoder, can I add your Tested-by?

[stuyoder]

Sure, thats fine.

[jenswi-linaro]

Tags applied.

@etienne-lms, I think the interface (and implementation) need to be cleaned up a bit. It's actually one item I'd like to discuss at Connect.