Re: [PATCH v2 2/2] kfence: Alloc kfence_pool after system startup

[Tianchen Ding]

On 2022/3/7 10:23, Tianchen Ding wrote:
> On 2022/3/7 07:52, Marco Elver wrote:
> > On Sat, 5 Mar 2022 at 15:49, Tianchen Ding <dtcccc@xxxxxxxxxxxxxxxxx> 
> > wrote:
> > [...]
> > > +static int kfence_init_late(void)
> > > +{
> > > +       const unsigned long nr_pages = KFENCE_POOL_SIZE / PAGE_SIZE;
> > > +       struct page *pages;
> > > +
> > > +       pages = alloc_contig_pages(nr_pages, GFP_KERNEL, 
> > > first_online_node, NULL);
> >
> > > mm/kfence/core.c:836:17: error: implicit declaration of function 
> > > ‘alloc_contig_pages’ [-Werror=implicit-function-declaration]
> >
> > This doesn't build without CMA. See ifdef CONFIG_CONTIG_ALLOC in
> > gfp.h, which declares alloc_contig_pages.
> >
> > Will alloc_pages() work as you expect? If so, perhaps only use
> > alloc_contig_pages() #ifdef CONFIG_CONTIG_ALLOC.
>
> alloc_pages() will be fine. We could free "tail" pages after inited.
> Will send v3 soon.

Oh, I remember why we use alloc_contig_pages()...
alloc_pages() (or alloc_pages_exact()) only support pages less than 
MAX_ORDER (default 11). The alloc would fail when KFENCE_NUM_OBJECTS >= 512.

So the design would be:
ifndef CONFIG_CONTIG_ALLOC and KFENCE_NUM_OBJECTS exceeds MAX_ORDER, we 
do not support alloc KFENCE pool after system startup.

> > Thanks,
> > -- Marco