> > +/* hpage must be locked, and mmap_lock must be held in write */
> > +static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr,
> > + pmd_t *pmdp, struct page *hpage)
> > +{
> > + struct vm_fault vmf = {
> > + .vma = vma,
> > + .address = addr,
> > + .flags = 0,
>
> It means you are going to install a readonly PMD. It is fine for
> readonly fs, but it may be not performant for shmem since the
> following write will trigger wp fault then the PMD will be split. So
> it should be ok to install writable PMD for shmem.
>
> But anyway this is not a fatal problem, the patch looks good to me.
> Reviewed-by: Yang Shi <shy828301@xxxxxxxxx>
>
Thanks, Yang!
I haven't worked out the details for writable shmem, but I'm hoping
FAULT_FLAG_WRITE is enough here.
I'm also now wondering if FAULT_FLAG_REMOTE is needed here for
process_madvise(MADV_COLLAPSE) when called on behalf of another
process; though that would also mean it's needed in
__collapse_huge_page_swapin() as well (even when running as
khugepaged). I can't say I fully understand 1b2ee1266ea6 ("mm/core: Do
not enforce PKEY permissions on remote mm access"), but if the comment
in include/linux/mm_types.h is to be trusted, then this would be the
case.
