Hi, Jann,
On Mon, Sep 21, 2020 at 11:55:06PM +0200, Jann Horn wrote:
> On Mon, Sep 21, 2020 at 11:20 PM Peter Xu <peterx@xxxxxxxxxx> wrote:
> > This patch is greatly inspired by the discussions on the list from Linus, Jason
> > Gunthorpe and others [1].
> >
> > It allows copy_pte_range() to do early cow if the pages were pinned on the
> > source mm. Currently we don't have an accurate way to know whether a page is
> > pinned or not. The only thing we have is page_maybe_dma_pinned(). However
> > that's good enough for now. Especially, with the newly added mm->has_pinned
> > flag to make sure we won't affect processes that never pinned any pages.
>
> To clarify: This patch only handles pin_user_pages() callers and
> doesn't try to address other GUP users, right? E.g. if task A uses
> process_vm_write() on task B while task B is going through fork(),
> that can still race in such a way that the written data only shows up
> in the child and not in B, right?
I saw that process_vm_write() is using pin_user_pages_remote(), so I think
after this patch applied the data will only be written to B but not the child.
Because when B fork() with these temp pinned pages, it will copy the pages
rather than write-protect them any more. IIUC the child could still have
partial data, but at last (after unpinned) B should always have the complete
data set.
>
> I dislike the whole pin_user_pages() concept because (as far as I
> understand) it fundamentally tries to fix a problem in the subset of
> cases that are more likely to occur in practice (long-term pins
> overlapping with things like writeback), and ignores the rarer cases
> ("short-term" GUP).
John/Jason or others may be better on commenting on this one. From my own
understanding, I thought it was the right thing to do so that we'll always
guarantee process B gets the whole data. From that pov this patch should make
sense even for short term gups. But maybe I've missed something.
--
Peter Xu
