On Wed, Nov 30, 2022 at 04:13:23PM -0800, Hugh Dickins wrote:
> On Wed, 30 Nov 2022, Johannes Weiner wrote:
> >
> > Hm, I think the below should work for swap pages. Do you see anything
> > obviously wrong with it, or scenarios I haven't considered?
> >
>
> I think you're overcomplicating it, with the __swap_count(ent) business,
> and consequent unnecessarily detailed comments on the serialization.
>
> Page/folio lock prevents a !page_mapped(page) becoming a page_mapped(page),
> whether it's in swap cache or in file cache; it does not stop the sharing
> count going further up, or down even to 0, but we just don't need to worry
> about that sharing count - the MC_TARGET_PAGE case does not reject pages
> with mapcount > 1, so why complicate the swap or file case in that way?
>
> (Yes, it can be argued that all such sharing should be rejected; but we
> didn't come here to argue improvements to memcg charge moving semantics:
> just to minimize its effect on rmap, before it is fully deprecated.)
>
> Or am I missing the point of why you add that complication?
No, it just seemed odd to move shared swap *unless* it's partially
faulted. But you're right, it's probably not worth the hassle. I'll
cut this down to the page_mapped() check.
The struggle of writing code for Schroedinger's User...
> > @@ -5637,6 +5645,46 @@ static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
>
> Don't forget to trylock the page in the device_private case before this.
Yep, thanks!
> > * we call find_get_page() with swapper_space directly.
> > */
> > page = find_get_page(swap_address_space(ent), swp_offset(ent));
> > +
> > + /*
> > + * Don't move shared charges. This isn't just for saner move
> > + * semantics, it also ensures that page_mapped() is stable for
> > + * the accounting in mem_cgroup_mapcount().
>
> mem_cgroup_mapcount()??
mem_cgroup_move_account() of course! Will fix.
> > + * We have to serialize against the following paths: fork
> > + * (which may copy a page map or a swap pte), fault (which may
> > + * change a swap pte into a page map), unmap (which may cause
> > + * a page map or a swap pte to disappear), and reclaim (which
> > + * may change a page map into a swap pte).
> > + *
> > + * - Without swapcache, we only want to move the charge if
> > + * there are no other swap ptes. With the pte lock, the
> > + * swapcount is stable against all of the above scenarios
> > + * when it's 1 (our pte), which is the case we care about.
> > + *
> > + * - When there is a page in swapcache, we only want to move
> > + * charges when neither the page nor the swap entry are
> > + * mapped elsewhere. The pte lock prevents our pte from
> > + * being forked or unmapped. The page lock will stop faults
> > + * against, and reclaim of, the swapcache page. So if the
> > + * page isn't mapped, and the swap count is 1 (our pte), the
> > + * test results are stable and the charge is exclusive.
... and edit this down accordingly.
> > + */
> > + if (!page && __swap_count(ent) != 1)
> > + return NULL;
> > +
> > + if (page) {
> > + if (!trylock_page(page)) {
> > + put_page(page);
> > + return NULL;
> > + }
> > + if (page_mapped(page) || __swap_count(ent) != 1) {
> > + unlock_page(page);
> > + put_page(page);
> > + return NULL;
> > + }
> > + }
> > +
> > entry->val = ent.val;
> >
> > return page;
>
> Looks right, without the __swap_count() additions and swap count comments.
>
> And similar code in mc_handle_file_pte() - or are you saying that only
> swap should be handled this way? I would disagree.
Right, same rules apply there. I only pasted the swap one to make sure
we get aligned on the basic strategy.
> And matching trylock in mc_handle_present_pte() (and get_mctgt_type_thp()),
> instead of in mem_cgroup_move_account().
Yes.
> I haven't checked to see where the page then needs to be unlocked,
> probably some new places.
Yes, the callers of get_mctgt_type*() need to unlock (if target is
passed and the page is returned). It looks straight-forward, they
already have to do put_page().
> And I don't know what will be best for the preliminary precharge pass:
> doesn't really want the page lock at all, but it may be unnecessary
> complication to avoid taking it then unlocking it in that pass.
We could make it conditional on target, which precharge doesn't pass,
but I agree it's likely not worth optimizing that code at this point.
Thanks for taking a look, Hugh, that's excellent input.
I'll finish this patch, rebase the rmap patch on it, and add a new one
to issue a deprecation warning in mem_cgroup_move_charge_write().
Johannes
