On Mon, Dec 4, 2023 at 12:30 AM Chengming Zhou <chengming.zhou@xxxxxxxxx> wrote:
>
> On 2023/12/1 04:35, Johannes Weiner wrote:
> > On Thu, Nov 30, 2023 at 12:07:41PM -0800, Nhat Pham wrote:
> >> On Thu, Nov 30, 2023 at 11:57 AM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> >>>
> >>> On Thu, Nov 30, 2023 at 11:40:18AM -0800, Nhat Pham wrote:
> >>>> This patch changes list_lru interface so that the caller must explicitly
> >>>> specify numa node and memcg when adding and removing objects. The old
> >>>> list_lru_add() and list_lru_del() are renamed to list_lru_add_obj() and
> >>>> list_lru_del_obj(), respectively.
> >>>
> >>> Wouldn't it be better to add list_lru_add_memcg() and
> >>> list_lru_del_memcg() and have:
> >>>
> >>> +bool list_lru_del(struct list_lru *lru, struct list_head *item)
> >>> +{
> >>> + int nid = page_to_nid(virt_to_page(item));
> >>> + struct mem_cgroup *memcg = list_lru_memcg_aware(lru) ?
> >>> + mem_cgroup_from_slab_obj(item) : NULL;
> >>> +
> >>> + return list_lru_del_memcg(lru, item, nid, memcg);
> >>> +}
> >>>
> >>> Seems like _most_ callers will want the original versions and only
> >>> a few will want the explicit memcg/nid versions. No?
> >>>
> >>
> >> I actually did something along that line in earlier iterations of this
> >> patch series (albeit with poorer naming - __list_lru_add() instead of
> >> list_lru_add_memcg()). The consensus after some back and forth was
> >> that the original list_lru_add() was not a very good design (the
> >> better one was this new version that allows for explicit numa/memcg
> >> selection). So I agreed to fix it everywhere as a prep patch.
> >>
> >> I don't have strong opinions here to be completely honest, but I do
> >> think this new API makes more sense (at the cost of quite a bit of
> >> elbow grease to fix every callsites and extra reviewing).
> >
> > Maybe I can shed some light since I was pushing for doing it this way.
> >
> > The quiet assumption that 'struct list_head *item' is (embedded in) a
> > slab object that is also charged to a cgroup is a bit much, given that
> > nothing in the name or documentation of the function points to that.
> >
> > It bit us in the THP shrinker where that list head is embedded in a
> > tailpage (virt_to_page(page) is fun to debug). And it caused some
> > confusion in this case as well, where the zswap entry is a slab object
> > but not charged (the entry descriptor is not attractive for cgroup
> > accounting, only the backing memory it points to.)
>
> Hi,
>
> I have a question, maybe I missed something since I haven't read all
> the earlier versions.
>
> IIUC, the problem here is that "zswap_entry" has different memcg and node
> than the "page", so I wonder if we can just charge "zswap_entry" to the
> same memcg of the "page".
>
> Like we can do these when allocating the "zswap_entry":
>
> old_memcg = set_active_memcg(memcg)
> kmem_cache_alloc_lru(zswap_entry_cache, lru, gfp)
> set_active_memcg(old_memcg)
>
> The good points are:
>
> 1. "zswap_entry" is charged to the memcg of "page", which is more sensible?
>
> 2. We can reuse the kmem_cache_alloc_lru() interface, which makes code simpler
> since we don't need to manage list_lru_memcg by ourselves.
>
> 3. Maybe the new list_lru_add() and list_lru_del() are not needed anymore?
> Since the "zswap_entry" is of the same memcg and node with the "page".
> But don't know if THP shrinker still need it.
>
> Thanks!
That idea was considered in earlier iterations/discussions of the
patch series as well. Charging things is not free - there is an
overhead associated with it, which is why we are usually selective
about whether to charge something. We were not super keen to do this
for zswap_entry just to plumb around the list_lru's restriction. Might
as well pay the price of extending the list_lru interface now.
If in the future, not charging the zswap entry causes a separate
isolation issue, we could revisit this decision and charge it.
Otherwise, IMHO we should just stick with this for now.
>
> >
> > Yes, for most users - at least right now - the current assumption is
> > accurate. The thinking was just that if we do have to differentiate
> > callers now anyway, we might as well make the interface a bit more
> > self-documenting and harder to misuse going forward, even if it's a
> > bit more churn now.
> >
> >
