[PATCH RFC v2] riscv: Enable KFENCE for riscv64

[Marco Elver]

On Mon, 24 May 2021 at 03:56, Liu Shixin <liushixin2 at huawei.com> wrote:
> On 2021/5/23 10:38, Palmer Dabbelt wrote:
> > On Fri, 14 May 2021 08:20:10 PDT (-0700), elver at google.com wrote:
> >> On Fri, 14 May 2021 at 05:11, Liu Shixin <liushixin2 at huawei.com> wrote:
> >>> Add architecture specific implementation details for KFENCE and enable
> >>> KFENCE for the riscv64 architecture. In particular, this implements the
> >>> required interface in <asm/kfence.h>.
> >>>
> >>> KFENCE requires that attributes for pages from its memory pool can
> >>> individually be set. Therefore, force the kfence pool to be mapped at
> >>> page granularity.
> >>>
> >>> I tested this patch using the testcases in kfence_test.c and all passed.
> >>>
> >>> Signed-off-by: Liu Shixin <liushixin2 at huawei.com>
> >>
> >> Acked-by: Marco Elver <elver at google.com>
> >>
> >>
> >>> ---
> >>> v1->v2: Change kmalloc() to pte_alloc_one_kernel() for allocating pte.
> >>>
> >>>  arch/riscv/Kconfig              |  1 +
> >>>  arch/riscv/include/asm/kfence.h | 51 +++++++++++++++++++++++++++++++++
> >>>  arch/riscv/mm/fault.c           | 11 ++++++-
> >>>  3 files changed, 62 insertions(+), 1 deletion(-)
> >>>  create mode 100644 arch/riscv/include/asm/kfence.h
> >>>
> >>> diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig
> >>> index c426e7d20907..000d8aba1030 100644
> >>> --- a/arch/riscv/Kconfig
> >>> +++ b/arch/riscv/Kconfig
> >>> @@ -64,6 +64,7 @@ config RISCV
> >>>         select HAVE_ARCH_JUMP_LABEL_RELATIVE
> >>>         select HAVE_ARCH_KASAN if MMU && 64BIT
> >>>         select HAVE_ARCH_KASAN_VMALLOC if MMU && 64BIT
> >>> +       select HAVE_ARCH_KFENCE if MMU && 64BIT
> >>>         select HAVE_ARCH_KGDB
> >>>         select HAVE_ARCH_KGDB_QXFER_PKT
> >>>         select HAVE_ARCH_MMAP_RND_BITS if MMU
> >>> diff --git a/arch/riscv/include/asm/kfence.h b/arch/riscv/include/asm/kfence.h
> >>> new file mode 100644
> >>> index 000000000000..c25d67e0b8ba
> >>> --- /dev/null
> >>> +++ b/arch/riscv/include/asm/kfence.h
> >>> @@ -0,0 +1,51 @@
> >>> +/* SPDX-License-Identifier: GPL-2.0 */
> >>> +
> >>> +#ifndef _ASM_RISCV_KFENCE_H
> >>> +#define _ASM_RISCV_KFENCE_H
> >>> +
> >>> +#include <linux/kfence.h>
> >>> +#include <linux/pfn.h>
> >>> +#include <asm-generic/pgalloc.h>
> >>> +#include <asm/pgtable.h>
> >>> +
> >>> +static inline bool arch_kfence_init_pool(void)
> >>> +{
> >>> +       int i;
> >>> +       unsigned long addr;
> >>> +       pte_t *pte;
> >>> +       pmd_t *pmd;
> >>> +
> >>> +       for (addr = (unsigned long)__kfence_pool; is_kfence_address((void *)addr);
> >>> +            addr += PAGE_SIZE) {
> >>> +               pte = virt_to_kpte(addr);
> >>> +               pmd = pmd_off_k(addr);
> >>> +
> >>> +               if (!pmd_leaf(*pmd) && pte_present(*pte))
> >>> +                       continue;
> >>> +
> >>> +               pte = pte_alloc_one_kernel(&init_mm);
> >>> +               for (i = 0; i < PTRS_PER_PTE; i++)
> >>> +                       set_pte(pte + i, pfn_pte(PFN_DOWN(__pa((addr & PMD_MASK) + i * PAGE_SIZE)), PAGE_KERNEL));
> >>> +
> >>> +               set_pmd(pmd, pfn_pmd(PFN_DOWN(__pa(pte)), PAGE_TABLE));
> >>> +               flush_tlb_kernel_range(addr, addr + PMD_SIZE);
> >>> +       }
> >>> +
> >>> +       return true;
> >>> +}
> >
> > I'm not fundamentally opposed to this, but the arm64 approach where pages are split at runtime when they have mis-matched permissions seems cleaner to me.  I'm not sure why x86 is doing it during init, though, as IIUC set_memory_4k() will work for both.
> >
> > Upgrading our __set_memory() with the ability to split pages (like arm64 has) seems generally useful, and would let us trivially implement the dynamic version of this.  We'll probably end up with the ability to split pages anyway, so that would be the least code in the long run.
> >
> > If there's some reason to prefer statically allocating the pages I'm fine with this, though.
> >
> As I understand，the arm64 approach does not implement dynamic splitting.
> If kfence is enabled in arch arm64, the linear map need to be forcibly mapped
> at page granularity. But x86 does not have such constraints as it only split pages
> in the kfence pool, so I think the x86 approach is better as it has less influence
> on the whole.

Correct.

I think either riscv gains set_memory_4k(), like x86, or we go with
the approach in this patch. Unless you see this is trivially
implementable, I wouldn't want to block this patch. It's better to
have something working, and then incrementally improve later if riscv
ever gets set_memory_4k().

Thanks,
-- Marco