I am not 100% sure why we bail out on special mappings: all we have to do is
reliably identify anon pages, and we should be able to do that.
But they map e.g. kernel memory (at least for VM_PFNMAP, purely and by
implication really VM_IO), it wouldn't make sense for KSM to be asked to
try to merge these right?
And of course no underlying struct page to pin, no reference counting
either, so I think you'd end up in trouble potentially here wouldn't you?
And how would the CoW work?
KSM only operates on anonymous pages. It cannot de-duplicate anything
else. (therefore, only MAP_PRIVATE applies)
Anything else (no struct page, not a CoW anon folio in such a mapping)
is skipped.
Take a look at scan_get_next_rmap_item() where we do
folio = folio_walk_start(&fw, vma, ksm_scan.address, 0);
if (folio) {
if (!folio_is_zone_device(folio) &&
folio_test_anon(folio)) {
folio_get(folio);
tmp_page = fw.page;
}
folio_walk_end(&fw, vma)
}
Before I changed that code, we were using GUP. And GUP just always
refuses VM_IO|VM_PFNMAP because it cannot handle it properly.
So, assuming we could remove the VM_PFNMAP | VM_IO | VM_DONTEXPAND |
VM_MIXEDMAP constraint from vma_ksm_compatible(), could we simplify?
Well I question removing this constraint for above reasons.
At any rate, even if we _could_ this feels like a bigger change that we
should come later.
"bigger" -- it might just be removing these 4 flags from the check ;)
I'll dig a bit more.
--
Cheers,
David / dhildenb
