[PATCH] arm64: kaslr: Fix kaslr end boundary of virt addr

[Ard Biesheuvel]

On 28 November 2017 at 20:41, Ard Biesheuvel <ard.biesheuvel at linaro.org> wrote:
> On 21 November 2017 at 03:44, Chen Feng <puck.chen at hisilicon.com> wrote:
>> With kaslr and kasan enable both, I got the follow issue.
>>
>> [   16.130523s]kasan: reg->base = 100000000, phys_end =1c0000000,start = ffffffff40000000, end = ffffffc000000000
>> [   16.142517s]___alloc_bootmem_nopanic:257
>> [   16.148284s]__alloc_memory_core_early:63, addr = 197fc7fc0
>> [   16.155670s]__alloc_memory_core_early:65, virt = ffffffffd7fc7fc0
>> [   16.163635s]__alloc_memory_core_early:67, toshow = ffffff8ffaff8ff8
>> [   16.171783s]__alloc_memory_core_early:69, show_phy = ffffffe2649f8ff8
>> [   16.180145s]Unable to handle kernel paging request at virtual address ffffff8ffaff8ff8
>> [   16.189971s]pgd = ffffffad9c507000
>> [   16.195220s][ffffff8ffaff8ff8] *pgd=0000000197fc8003, *pud=0000000197fc8003
>>
>> *reg->base = 100000000, phys_end =1c0000000,start = ffffffff40000000, end = ffffffc000000000*
>>
>> memstart_addr 0
>> ARM64_MEMSTART_ALIGN 0x40000000
>> memstart_offset_seed 0xffc7
>> PHYS_OFFSET = 0 - memstart_addr = 0 - 3E40000000 = FFFFFFC1C0000000
>>
>> reg->base = 0x100000000  -> 0xffffffff40000000
>> phys_end  = 0x1c0000000  -> 0xffffffc000000000  This is confused, end less than start.
>>
>
> This looks a bit weird because we add the PAGE_OFFSET, but it simply
> wraps at the top of the address space.
>
> So this code in kasan_init()
>
> void *start = (void *)__phys_to_virt(reg->base);
> void *end = (void *)__phys_to_virt(reg->base + reg->size);
>
> if (start >= end)
>      break;
>
> is essentially incorrect, because it translates an address that is
> strictly outside of the current memblock region. If the KASLR code
> happens to map DRAM all the way at the top of the linear region (which
> is what occurs in your case), end - 1 is the last valid address.
>
> So I think the minimal correct fix would be
>
> diff --git a/arch/arm64/mm/kasan_init.c b/arch/arm64/mm/kasan_init.c
> index acba49fb5aac..3214aa9d90be 100644
> --- a/arch/arm64/mm/kasan_init.c
> +++ b/arch/arm64/mm/kasan_init.c
> @@ -216,7 +216,7 @@ void __init kasan_init(void)
>
>         for_each_memblock(memory, reg) {
>                 void *start = (void *)__phys_to_virt(reg->base);
> -               void *end = (void *)__phys_to_virt(reg->base + reg->size);
> +               void *end = start + reg->size;
>
>                 if (start >= end)
>                         break;
>

This does not actually fix the problem, end will be 0x0 which is still
< start. We'll need a '-1' in there somewhere to really fix it.

> given that mappings in the linear region are congruent with the
> underlying physical regions (unless I am missing something wrt special
> start/end values in memblock, but in that case, they should not be p2v
> translated before the evaluation)
>
> However, since having DRAM at the very top appears to break other things as well
>
> vmemmap : 0xffffffbf00000000 - 0xffffffc000000000   (     4 GB maximum)
>           0xffffffbfff000000 - 0xffffffbf00000000   (17592186040336 MB actual)
> memory  : 0xffffffffc0000000 - 0x               0   (  1024 MB)
>
> I will leave it to Will and/or Catalin to decide whether they prefer
> to follow your approach instead, and prevent KASLR from mapping DRAM
> all the way at the top of the address space. Otherwise, we'll need to
> track down all problematic uses of __phys_to_virt() et al, because
> there will surely be more.
>

I guess it takes quite a bit of discipline to deal with overflow in
all 'end' calculations, although I suppose this is more common on
32-bit architectures.