On Thu, Jan 26, 2023 at 1:45 PM Alexei Starovoitov
<alexei.starovoitov@xxxxxxxxx> wrote:
>
> On Tue, Jan 17, 2023 at 10:49 PM Yafang Shao <laoar.shao@xxxxxxxxx> wrote:
> > > > I just don't want to add many if-elses or switch-cases into
> > > > bpf_map_memory_footprint(), because I think it is a little ugly.
> > > > Introducing a new map ops could make it more clear. For example,
> > > > static unsigned long bpf_map_memory_footprint(const struct bpf_map *map)
> > > > {
> > > > unsigned long size;
> > > >
> > > > if (map->ops->map_mem_footprint)
> > > > return map->ops->map_mem_footprint(map);
> > > >
> > > > size = round_up(map->key_size + bpf_map_value_size(map), 8);
> > > > return round_up(map->max_entries * size, PAGE_SIZE);
> > > > }
> > >
> > > It is also ugly, because bpf_map_value_size() already has if-stmt.
> > > I prefer to keep all estimates in one place.
> > > There is no need to be 100% accurate.
> >
> > Per my investigation, it can be almost accurate with little effort.
> > Take the htab for example,
> > static unsigned long htab_mem_footprint(const struct bpf_map *map)
> > {
> > struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
> > unsigned long size = 0;
> >
> > if (!htab_is_prealloc(htab)) {
> > size += htab_elements_size(htab);
> > }
> > size += kvsize(htab->elems);
> > size += percpu_size(htab->extra_elems);
> > size += kvsize(htab->buckets);
> > size += bpf_mem_alloc_size(&htab->pcpu_ma);
> > size += bpf_mem_alloc_size(&htab->ma);
> > if (htab->use_percpu_counter)
> > size += percpu_size(htab->pcount.counters);
> > size += percpu_size(htab->map_locked[i]) * HASHTAB_MAP_LOCK_COUNT;
> > size += kvsize(htab);
> > return size;
> > }
>
> Please don't.
> Above doesn't look maintainable.
It is similar to htab_map_free(). These pointers are the pointers
which will be freed in map_free().
We just need to keep map_mem_footprint() in sync with map_free(). It
won't be a problem for maintenance.
> Look at kvsize(htab). Do you really care about hundred bytes?
> Just accept that there will be a small constant difference
> between what show_fdinfo reports and the real memory.
The point is we don't have a clear idea what the margin is.
> You cannot make it 100%.
> There is kfence that will allocate 4k though you asked kmalloc(8).
>
We already have ksize()[1], which covers the kfence.
[1]. https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/mm/slab_common.c#n1431
> > We just need to get the real memory size from the pointer instead of
> > calculating the size again.
> > For non-preallocated htab, it is a little trouble to get the element
> > size (not the unit_size), but it won't be a big deal.
>
> You'd have to convince mm folks that kvsize() is worth doing.
> I don't think it will be easy.
>
As I mentioned above, we already have ksize(), so we only need to
introduce vsize(). Per my understanding, we can simply use
vm_struct->size to get the vmalloc size, see also the patch #5 in this
patchset[2].
Andrew, Uladzislau, Christoph, do you have any comments on this newly
introduced vsize()[2] ?
[2]. https://lore.kernel.org/bpf/20230112155326.26902-6-laoar.shao@xxxxxxxxx/
> > > With a callback devs will start thinking that this is somehow
> > > a requirement to report precise memory.
> > >
> > > > > > > bpf side tracks all of its allocation. There is no need to do that
> > > > > > > in generic mm side.
> > > > > > > Exposing an aggregated single number if /proc/meminfo also looks wrong.
> > > > > >
> > > > > > Do you mean that we shouldn't expose it in /proc/meminfo ?
> > > > >
> > > > > We should not because it helps one particular use case only.
> > > > > Somebody else might want map mem info per container,
> > > > > then somebody would need it per user, etc.
> > > >
> > > > It seems we should show memcg info and user info in bpftool map show.
> > >
> > > Show memcg info? What do you have in mind?
> > >
> >
> > Each bpf map is charged to a memcg. If we know a bpf map belongs to
> > which memcg, we can know the map mem info per container.
> > Currently we can get the memcg info from the process which loads it,
> > but it can't apply to pinned-bpf-map.
> > So it would be better if we can show it in bpftool-map-show.
>
> That sounds useful.
> Have you looked at bpf iterators and how bpftool is using
> them to figure out which process loaded bpf prog and created particular map?
Yes, I have looked at it. I know what you mean. It seems we can
introduce a memcg_iter or something else to implement it.
--
Regards
Yafang
