On Thu, Sep 17, 2020 at 11:14 AM Peter Xu <peterx@xxxxxxxxxx> wrote:
>
> In my humble opinion, the real solution is still to use MADV_DONTFORK properly
> so we should never share the DMA pages with others when we know the fact.
Is this all just because somebody does a fork() after doing page pinning?
If so, I feel this should be trivially fixed in copy_one_pte().
That's where we currently do
/*
* If it's a COW mapping, write protect it both
* in the parent and the child
*/
if (is_cow_mapping(vm_flags) && pte_write(pte)) {
ptep_set_wrprotect(src_mm, addr, src_pte);
pte = pte_wrprotect(pte);
}
and I feel that that is where we could just change the code to do a
COW event for pinned pages (and *not* mark the parent write protected,
since the parent page now isn't a COW page).
Because if that's the case that Jason is hitting, then I feel that
really is the correct fix: make sure that the pinning action is
meaningful.
As mentioned, I really think the whole (and only) point of page
pinning is that it should keep the page locked in the page tables. And
by "locked" I mean exactly that: not just present, but writable.
And then the "we never COW a pinned page" comes not from the COW code
doing magic, but by it simply never becoming non-writable - because
the page table entry is locked!
Linus
