On Sat, Jun 20, 2020 at 02:00:18PM -0700, Andrew Morton wrote:
> On Sat, 20 Jun 2020 14:47:19 -0400 Waiman Long <[email protected]> wrote:
>
> > It was found that running the LTP test on a PowerPC system could produce
> > erroneous values in /proc/meminfo, like:
> >
> > MemTotal: 531915072 kB
> > MemFree: 507962176 kB
> > MemAvailable: 1100020596352 kB
> >
> > Using bisection, the problem is tracked down to commit 9c315e4d7d8c
> > ("mm: memcg/slab: cache page number in memcg_(un)charge_slab()").
> >
> > In memcg_uncharge_slab() with a "int order" argument:
> >
> > unsigned int nr_pages = 1 << order;
> > :
> > mod_lruvec_state(lruvec, cache_vmstat_idx(s), -nr_pages);
> >
> > The mod_lruvec_state() function will eventually call the
> > __mod_zone_page_state() which accepts a long argument. Depending on
> > the compiler and how inlining is done, "-nr_pages" may be treated as
> > a negative number or a very large positive number. Apparently, it was
> > treated as a large positive number in that PowerPC system leading to
> > incorrect stat counts. This problem hasn't been seen in x86-64 yet,
> > perhaps the gcc compiler there has some slight difference in behavior.
> >
> > It is fixed by making nr_pages a signed value. For consistency, a
> > similar change is applied to memcg_charge_slab() as well.
>
> This is somewhat disturbing.
>
> > --- a/mm/slab.h
> > +++ b/mm/slab.h
> > @@ -348,7 +348,7 @@ static __always_inline int memcg_charge_slab(struct
> > page *page,
> > gfp_t gfp, int order,
> > struct kmem_cache *s)
> > {
> > - unsigned int nr_pages = 1 << order;
> > + int nr_pages = 1 << order;
> > struct mem_cgroup *memcg;
> > struct lruvec *lruvec;
> > int ret;
> > @@ -388,7 +388,7 @@ static __always_inline int memcg_charge_slab(struct
> > page *page,
> > static __always_inline void memcg_uncharge_slab(struct page *page, int
> > order,
> > struct kmem_cache *s)
> > {
> > - unsigned int nr_pages = 1 << order;
> > + int nr_pages = 1 << order;
> > struct mem_cgroup *memcg;
> > struct lruvec *lruvec;
> >
>
> I grabbed the patch, but Roman's "mm: memcg/slab: charge individual
> slab objects instead of pages"
> (http://lkml.kernel.org/r/[email protected]) deletes
> both these functions.
It looks like Waiman's patch should be backported to stable.
So if you can queue it before my series, that would be nice.
>
> It replaces the offending code with, afaict,
>
>
> static inline void memcg_slab_free_hook(struct kmem_cache *s, struct page
> *page,
> void *p)
> {
> struct obj_cgroup *objcg;
> unsigned int off;
>
> if (!memcg_kmem_enabled() || is_root_cache(s))
> return;
>
> off = obj_to_index(s, page, p);
> objcg = page_obj_cgroups(page)[off];
> page_obj_cgroups(page)[off] = NULL;
>
> obj_cgroup_uncharge(objcg, obj_full_size(s));
> mod_objcg_state(objcg, page_pgdat(page), cache_vmstat_idx(s),
> >>> -obj_full_size(s));
>
> obj_cgroup_put(objcg);
> }
>
> -obj_full_size() returns size_t so I guess that's OK.
>
>
>
> Also
>
>
> static __always_inline void uncharge_slab_page(struct page *page, int order,
> struct kmem_cache *s)
> {
> #ifdef CONFIG_MEMCG_KMEM
> if (memcg_kmem_enabled() && !is_root_cache(s)) {
> memcg_free_page_obj_cgroups(page);
> percpu_ref_put_many(&s->memcg_params.refcnt, 1 << order);
> }
> #endif
> mod_node_page_state(page_pgdat(page), cache_vmstat_idx(s),
> >>> -(PAGE_SIZE << order));
> }
>
> PAGE_SIZE is unsigned long so I guess that's OK as well.
>
>
> Still, perhaps both could be improved. Negating an unsigned scalar is
> a pretty ugly thing to do.
>
> Am I wrong in thinking that all those mod_foo() functions need careful
> review?
>
I'll take a look too.
Thanks!
