Re: [PATCH] mm: fix possible OOB in numa_rebuild_large_mapping()

[David Hildenbrand <david@redhat.com>]

On 07.06.24 12:32, Kefeng Wang wrote:
> The large folio is mapped with folio size aligned virtual address during
> the pagefault, eg, 'addr = ALIGN_DOWN(vmf->address, nr_pages * PAGE_SIZE)'
> in do_anonymous_page(), but after the mremap(), the virtual address only
> require PAGE_SIZE aligned, also pte is moved to new in move_page_tables(),
> then traverse the new pte in numa_rebuild_large_mapping() will hint the
> following issue,
> 
>     Unable to handle kernel paging request at virtual address 00000a80c021a788
>     Mem abort info:
>       ESR = 0x0000000096000004
>       EC = 0x25: DABT (current EL), IL = 32 bits
>       SET = 0, FnV = 0
>       EA = 0, S1PTW = 0
>       FSC = 0x04: level 0 translation fault
>     Data abort info:
>       ISV = 0, ISS = 0x00000004, ISS2 = 0x00000000
>       CM = 0, WnR = 0, TnD = 0, TagAccess = 0
>       GCS = 0, Overlay = 0, DirtyBit = 0, Xs = 0
>     user pgtable: 4k pages, 48-bit VAs, pgdp=00002040341a6000
>     [00000a80c021a788] pgd=0000000000000000, p4d=0000000000000000
>     Internal error: Oops: 0000000096000004 [#1] SMP
>     ...
>     CPU: 76 PID: 15187 Comm: git Kdump: loaded Tainted: G        W          6.10.0-rc2+ #209
>     Hardware name: Huawei TaiShan 2280 V2/BC82AMDD, BIOS 1.79 08/21/2021
>     pstate: 60400009 (nZCv daif +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
>     pc : numa_rebuild_large_mapping+0x338/0x638
>     lr : numa_rebuild_large_mapping+0x320/0x638
>     sp : ffff8000b41c3b00
>     x29: ffff8000b41c3b30 x28: ffff8000812a0000 x27: 00000000000a8000
>     x26: 00000000000000a8 x25: 0010000000000001 x24: ffff20401c7170f0
>     x23: 0000ffff33a1e000 x22: 0000ffff33a76000 x21: ffff20400869eca0
>     x20: 0000ffff33976000 x19: 00000000000000a8 x18: ffffffffffffffff
>     x17: 0000000000000000 x16: 0000000000000020 x15: ffff8000b41c36a8
>     x14: 0000000000000000 x13: 205d373831353154 x12: 5b5d333331363732
>     x11: 000000000011ff78 x10: 000000000011ff10 x9 : ffff800080273f30
>     x8 : 000000320400869e x7 : c0000000ffffd87f x6 : 00000000001e6ba8
>     x5 : ffff206f3fb5af88 x4 : 0000000000000000 x3 : 0000000000000000
>     x2 : 0000000000000000 x1 : fffffdffc0000000 x0 : 00000a80c021a780
>     Call trace:
>      numa_rebuild_large_mapping+0x338/0x638
>      do_numa_page+0x3e4/0x4e0
>      handle_pte_fault+0x1bc/0x238
>      __handle_mm_fault+0x20c/0x400
>      handle_mm_fault+0xa8/0x288
>      do_page_fault+0x124/0x498
>      do_translation_fault+0x54/0x80
>      do_mem_abort+0x4c/0xa8
>      el0_da+0x40/0x110
>      el0t_64_sync_handler+0xe4/0x158
>      el0t_64_sync+0x188/0x190
> 
> Fix it by correct the start and end, which may lead to only rebuild part
> of large mapping in one numa page fault, there is no issue since other part
> could rebuild by another pagefault.
> 
> Fixes: d2136d749d76 ("mm: support multi-size THP numa balancing")
> Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>
> ---
>   mm/memory.c | 24 +++++++++++++++---------
>   1 file changed, 15 insertions(+), 9 deletions(-)
> 
> diff --git a/mm/memory.c b/mm/memory.c
> index db9130488231..0ad57b6485ca 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -5223,15 +5223,21 @@ static void numa_rebuild_single_mapping(struct vm_fault *vmf, struct vm_area_str
>   	update_mmu_cache_range(vmf, vma, fault_addr, fault_pte, 1);
>   }
>   
> -static void numa_rebuild_large_mapping(struct vm_fault *vmf, struct vm_area_struct *vma,
> -				       struct folio *folio, pte_t fault_pte,
> -				       bool ignore_writable, bool pte_write_upgrade)
> +static void numa_rebuild_large_mapping(struct vm_fault *vmf,
> +		struct vm_area_struct *vma, struct folio *folio, int nr_pages,
> +		pte_t fault_pte, bool ignore_writable, bool pte_write_upgrade)
>   {
>   	int nr = pte_pfn(fault_pte) - folio_pfn(folio);
> -	unsigned long start = max(vmf->address - nr * PAGE_SIZE, vma->vm_start);
> -	unsigned long end = min(vmf->address + (folio_nr_pages(folio) - nr) * PAGE_SIZE, vma->vm_end);
> -	pte_t *start_ptep = vmf->pte - (vmf->address - start) / PAGE_SIZE;
> -	unsigned long addr;
> +	unsigned long folio_size = nr_pages * PAGE_SIZE;

Just re-read folio_nr_pages() here, it's cheap. Or even better, use 
folio_size();

> +	unsigned long addr = vmf->address;
> +	unsigned long start, end, align_addr;
> +	pte_t *start_ptep;
> +
> +	align_addr = ALIGN_DOWN(addr, folio_size);
> +	start = max3(addr - nr * PAGE_SIZE, align_addr, vma->vm_start);
> +	end = min3(addr + (nr_pages - nr) * PAGE_SIZE, align_addr + folio_size,

Please avoid mixing nr_pages and folio_size.

> +		   vma->vm_end);
> +	start_ptep = vmf->pte - (addr - start) / PAGE_SIZE;
			    writable);

I am not able to convince myself that the old code could not have 
resulted in a vmf->pte that underflows the page table. Am I correct?

Now align_addr would make sure that we are always within one page table 
(as long as our folio size does not exceed PMD size :) ).

Can we use PMD_SIZE instead for that, something like the following?


/* Stay within the VMA and within the page table. */
pt_start = ALIGN_DOWN(addr, PMD_SIZE);
start = max3(addr - nr << PAGE_SHIFT, pt_start, vma->vm_start);
end = min3(addr + folio_size - nr << PAGE_SHIFT,
            pt_start + PMD_SIZE, vma->vm_end);

start_ptep = vmf->pte - (addr - start) >> PAGE_SHIFT;

-- 
Cheers,

David / dhildenb