commit "slub: Acquire_slab() avoid loop" wrongly causes rest of partial slabs to be skipped?

commit "slub: Acquire_slab() avoid loop" wrongly causes rest of partial slabs to be skipped?

[Jann Horn]

The commit message of commit 7ced371971966 ("slub: Acquire_slab()
avoid loop") claims:

> Avoid the loop in acquire slab and simply fail if there is a conflict.
>
> This will cause the next page on the list to be considered.

However, get_partial_node() looks like this:

static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
        struct kmem_cache_cpu *c, gfp_t flags)
{
  struct page *page, *page2;
  void *object = NULL;
  unsigned int available = 0;
  int objects;
[...]
  spin_lock(&n->list_lock);
  list_for_each_entry_safe(page, page2, &n->partial, slab_list) {
    void *t;
[...]
    t = acquire_slab(s, n, page, object == NULL, &objects);
    if (!t)
      break;
[...]
  }
  spin_unlock(&n->list_lock);
  return object;
}

So actually, if the cmpxchg() fails, we'll entirely bail out of
get_partial_node() and might, if the system isn't NUMA, fall back to
allocating more memory with new_slab()? That seems to me like it might
cause fragmented slabs to slowly use more memory than they should over
time.

Should the loop in get_partial_node() be using "continue" instead of
"break" in that case, so that the rest of the partial list will be
considered as the commit message claims?

Re: commit "slub: Acquire_slab() avoid loop" wrongly causes rest of partial slabs to be skipped?

[Joonsoo Kim]

Hello, Jann.

2020년 12월 23일 (수) 오후 9:52, Jann Horn <jannh@google.com>님이 작성:
>
> The commit message of commit 7ced371971966 ("slub: Acquire_slab()
> avoid loop") claims:
>
> > Avoid the loop in acquire slab and simply fail if there is a conflict.
> >
> > This will cause the next page on the list to be considered.
>
> However, get_partial_node() looks like this:
>
> static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
>         struct kmem_cache_cpu *c, gfp_t flags)
> {
>   struct page *page, *page2;
>   void *object = NULL;
>   unsigned int available = 0;
>   int objects;
> [...]
>   spin_lock(&n->list_lock);
>   list_for_each_entry_safe(page, page2, &n->partial, slab_list) {
>     void *t;
> [...]
>     t = acquire_slab(s, n, page, object == NULL, &objects);
>     if (!t)
>       break;
> [...]
>   }
>   spin_unlock(&n->list_lock);
>   return object;
> }
>
> So actually, if the cmpxchg() fails, we'll entirely bail out of
> get_partial_node() and might, if the system isn't NUMA, fall back to
> allocating more memory with new_slab()? That seems to me like it might
> cause fragmented slabs to slowly use more memory than they should over
> time.

Nice catch! I think you are right.

> Should the loop in get_partial_node() be using "continue" instead of
> "break" in that case, so that the rest of the partial list will be
> considered as the commit message claims?

Please send the patch.

Thanks.