[RFCv2] arm64: support HAVE_ARCH_RARE_WRITE and HAVE_ARCH_RARE_WRITE_MEMCPY

[Ard Biesheuvel]

On 30 March 2017 at 15:39, Hoeun Ryu <hoeun.ryu at gmail.com> wrote:
>  This patch might be a part of Kees Cook's rare_write infrastructure series
> for [1] for arm64 architecture.
>
>  This implementation is based on Mark Rutland's suggestions [2], which is
> that a special userspace mm that maps only __start/end_rodata as RW permi-
> ssion is prepared during early boot time (paging_init) and __arch_rare_-
> write_begin() switches to the mm [2].
>
>  rare_write_mm address space is added for the special purpose and a page
> global directory is also prepared for it. The mm remaps __start_rodata ~
> __end_rodata to rodata_rw_alias_start which starts from TASK_SIZE_64 / 4
> + kaslr_offset().
>
>  It passes LKDTM's rare write test.
>
> [1] : http://www.openwall.com/lists/kernel-hardening/2017/02/27/5
> [2] : https://lkml.org/lkml/2017/3/29/704
>
> Signed-off-by: Hoeun Ryu <hoeun.ryu at gmail.com>
> ---
>  arch/arm64/Kconfig               |   2 +
>  arch/arm64/include/asm/pgtable.h |   4 ++
>  arch/arm64/mm/mmu.c              | 101 +++++++++++++++++++++++++++++++++++++++
>  3 files changed, 107 insertions(+)
>
> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> index f2b0b52..6e2c592 100644
> --- a/arch/arm64/Kconfig
> +++ b/arch/arm64/Kconfig
> @@ -102,6 +102,8 @@ config ARM64
>         select HAVE_SYSCALL_TRACEPOINTS
>         select HAVE_KPROBES
>         select HAVE_KRETPROBES
> +       select HAVE_ARCH_RARE_WRITE
> +       select HAVE_ARCH_RARE_WRITE_MEMCPY
>         select IOMMU_DMA if IOMMU_SUPPORT
>         select IRQ_DOMAIN
>         select IRQ_FORCED_THREADING
> diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
> index c213fdbd0..1514933 100644
> --- a/arch/arm64/include/asm/pgtable.h
> +++ b/arch/arm64/include/asm/pgtable.h
> @@ -741,6 +741,10 @@ static inline void update_mmu_cache(struct vm_area_struct *vma,
>  #define kc_vaddr_to_offset(v)  ((v) & ~VA_START)
>  #define kc_offset_to_vaddr(o)  ((o) | VA_START)
>
> +unsigned long __arch_rare_write_begin(void);
> +unsigned long __arch_rare_write_end(void);
> +void __arch_rare_write_memcpy(void *dst, const void *src, __kernel_size_t len);
> +

If these hook into a generic framework, shouldn't these already be
declared in a generic header file?

>  #endif /* !__ASSEMBLY__ */
>
>  #endif /* __ASM_PGTABLE_H */
> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
> index 91502e3..86b25c9 100644
> --- a/arch/arm64/mm/mmu.c
> +++ b/arch/arm64/mm/mmu.c
> @@ -570,6 +570,105 @@ static void __init map_kernel(pgd_t *pgd)
>         kasan_copy_shadow(pgd);
>  }
>
> +struct mm_struct rare_write_mm = {

static please

> +       .mm_rb          = RB_ROOT,
> +       .mm_users       = ATOMIC_INIT(2),
> +       .mm_count       = ATOMIC_INIT(1),
> +       .mmap_sem       = __RWSEM_INITIALIZER(rare_write_mm.mmap_sem),
> +       .page_table_lock= __SPIN_LOCK_UNLOCKED(rare_write_mm.page_table_lock),
> +       .mmlist         = LIST_HEAD_INIT(rare_write_mm.mmlist),
> +};
> +
> +#ifdef CONFIG_ARM64_PTDUMP_DEBUGFS
> +#include <asm/ptdump.h>
> +
> +static struct ptdump_info rare_write_ptdump_info = {
> +       .mm             = &rare_write_mm,
> +       .markers        = (struct addr_marker[]){
> +               { 0,            "rare-write start" },
> +               { TASK_SIZE_64, "rare-write end" }
> +       },
> +       .base_addr      = 0,
> +};
> +
> +static int __init ptdump_init(void)
> +{
> +       return ptdump_debugfs_register(&rare_write_ptdump_info,
> +                                      "rare_write_page_tables");
> +}
> +device_initcall(ptdump_init);
> +
> +#endif
> +
> +__always_inline unsigned long __arch_rare_write_begin(void)

These functions are not called from the same compilation unit, so what
do you think __always_inline buys you here?

> +{
> +       struct mm_struct *mm = &rare_write_mm;
> +
> +       preempt_disable();
> +
> +       __switch_mm(mm);
> +
> +       if (system_uses_ttbr0_pan()) {
> +               update_saved_ttbr0(current, mm);
> +               cpu_switch_mm(mm->pgd, mm);
> +       }
> +
> +       return 0;
> +}
> +
> +__always_inline unsigned long __arch_rare_write_end(void)
> +{
> +       struct mm_struct *mm = current->active_mm;
> +
> +       __switch_mm(mm);
> +
> +       if (system_uses_ttbr0_pan()) {
> +               cpu_set_reserved_ttbr0();
> +               if (mm != &init_mm)
> +                       update_saved_ttbr0(current, mm);
> +       }
> +
> +       preempt_enable_no_resched();
> +
> +       return 0;
> +}
> +
> +static unsigned long rodata_rw_alias_start __ro_after_init = TASK_SIZE_64 / 4;
> +
> +__always_inline void __arch_rare_write_memcpy(void *dst, const void *src,
> +                                             __kernel_size_t len)
> +{
> +       unsigned long __dst = (unsigned long)dst;
> +
> +       __dst -= (unsigned long)__start_rodata;
> +       __dst += rodata_rw_alias_start;
> +
> +       memcpy((void *)__dst, src, len);
> +}
> +
> +void __init rare_write_init(void)

static please

> +{
> +       phys_addr_t pgd_phys = early_pgtable_alloc();
> +       pgd_t *pgd = (pgd_t *)__phys_to_virt(pgd_phys);
> +       phys_addr_t pa_start = __pa_symbol(__start_rodata);
> +       unsigned long size = __end_rodata - __start_rodata;
> +
> +       BUG_ON(!PAGE_ALIGNED(pa_start));
> +       BUG_ON(!PAGE_ALIGNED(size));
> +
> +       rodata_rw_alias_start += kaslr_offset();
> +

This places the rodata alias at a fixed offset from the original,
right, even under KASLR? Given that they are completely independent, I
think you can use a random offset here rather than the same offset we
use for vmlinux.


> +       BUG_ON(rodata_rw_alias_start + size > TASK_SIZE_64);
> +
> +       rare_write_mm.pgd = pgd;

For correctness, you should add a call to mm_init_cpumask() here,
although it doesn't matter in practice on arm64.

> +       init_new_context(NULL, &rare_write_mm);
> +
> +       __create_pgd_mapping(pgd,
> +                            pa_start, rodata_rw_alias_start, size,
> +                            __pgprot(pgprot_val(PAGE_KERNEL) | PTE_NG),
> +                            early_pgtable_alloc, debug_pagealloc_enabled());

This needs to be updated after the changes that are now queued in the
arm64 tree for v4.12

BTW you can allow page/cont mappings in all cases, no need for the
debug_pagealloc_enabled() check

> +}
> +
>  /*
>   * paging_init() sets up the page tables, initialises the zone memory
>   * maps and sets up the zero page.
> @@ -603,6 +702,8 @@ void __init paging_init(void)
>          */
>         memblock_free(__pa_symbol(swapper_pg_dir) + PAGE_SIZE,
>                       SWAPPER_DIR_SIZE - PAGE_SIZE);
> +
> +       rare_write_init();
>  }
>
>  /*
> --
> 2.7.4
>