Re: [PATCH 2/2] x86/mm/KASLR: Adjust the vmemmap size according to paging mode

[Baoquan He <bhe@redhat.com>]

On 09/03/18 at 01:26pm, Kirill A. Shutemov wrote:
> On Mon, Sep 03, 2018 at 03:47:18PM +0800, Baoquan He wrote:
> > On 09/02/18 at 11:52pm, Kirill A. Shutemov wrote:
> > > On Thu, Aug 30, 2018 at 11:25:12PM +0800, Baoquan He wrote:
> > > > Hi Kirill,
> > > > 
> > > > I made a new version according to your suggestion, just a little
> > > > different, I didn't make 1TB as default, just calculate with the actual
> > > > size, then align up to 1TB boundary.  Just found kcore is printing more
> > > > entries than before, I thought it's caused by my code, later got it was
> > > > touchde by other people.
> > > > 
> > > > Any comment about this? I can change accordingly.
> > > 
> > > Looks good to me.
> > > 
> > > But there's corner case when struct page is unreasonably large and
> > > vmemmap_size will be way to large. We probably have to report an error if
> > > we cannot fit vmemmap properly into virtual memory layout.
> > 
> > Hmm, sizeof(struct page) can't exceed one whole page surely, otherwise
> > system bootup can't go over vmemmap initlization. Except of this, we may
> > need think about the virtual memory layout which vmemmap can be allowed
> > to occupy.
> > 
> > If KASAN enabled, KASLR disabled,
> > 4-level 1TB + 1TB hole (2TB)
> > 5-level 512TB + 2034TB hole (2.5PB)
> > 
> > If KASAN disabled, KASLR enabled,
> > 4-level 1TB + 1TB hole + 16TB  (18TB)
> > 5-level 512TB + 2034TB hole + 8PB (10.5PB)
> > 
> > So, as you can see, if add check in memory KASLR code, we should only
> > consider KASLR enabled case. We possibly don't need to worry about
> > 5-level case since the size 10.5PB is even bigger than the maximum
> > physical RAM mapping size. For 4-level, 18TB align to multiples of 2, it
> > will be 32 times of the current 1TB, then we usually assume 64 as the
> > default value of sizeof(struct page), then 64*32 == 1024. So we can add
> > check like this, what do you think? Or any other idea?
> 
> Looks reasonable to me.
> 
> But I would have the BUILD_BUG_ON() in generic code. If you struct page is
> more than 1/4 of PAGE_SIZE something is horribly broken.

Just the 1/4 of PAGE_SIZE is based on analysis of KASLR case. If
non-KASLR case, it may not be that value. Not sure if it's OK to put it
in generic code, and haven't thought of a good place, maybe in
setup_arch(), just at the beginning?

> 
> > diff --git a/arch/x86/mm/kaslr.c b/arch/x86/mm/kaslr.c
> > index 1db8e166455e..776ec759a87c 100644
> > --- a/arch/x86/mm/kaslr.c
> > +++ b/arch/x86/mm/kaslr.c
> > @@ -90,6 +90,7 @@ void __init kernel_randomize_memory(void)
> >         BUILD_BUG_ON(vaddr_start >= vaddr_end);
> >         BUILD_BUG_ON(vaddr_end != CPU_ENTRY_AREA_BASE);
> >         BUILD_BUG_ON(vaddr_end > __START_KERNEL_map);
> > +       BUILD_BUG_ON(sizeof(struct page ) > PAGE_SIZE/4);
> 
> Nitpick: redundant space before ')'.
> 
> >  
> >         if (!kaslr_memory_enabled())
> >                 return;
> > 
> > 
> > For 5-level paging mode, we
> > may not need to worry about that. Since KASAN 
> > 
> > ***4-level***
> > ffffea0000000000 - ffffeaffffffffff (=40 bits) virtual memory map (1TB)
> > ... unused hole ...
> > ffffec0000000000 - fffffbffffffffff (=44 bits) kasan shadow memory (16TB)
> > ... unused hole ...
> > 
> > 
> > 
> > ***5-level***
> > ffd4000000000000 - ffd5ffffffffffff (=49 bits) virtual memory map (512TB)
> > ... unused hole ...
> > ffdf000000000000 - fffffc0000000000 (=53 bits) kasan shadow memory (8PB) 
> > 
> > > 
> > > -- 
> > >  Kirill A. Shutemov
> 
> -- 
>  Kirill A. Shutemov