[PATCH v2] riscv: Use PUD/P4D/PGD pages for the linear mapping

[Rob Herring]

On Mon, Dec 12, 2022 at 8:31 AM Alexandre Ghiti <alexghiti at rivosinc.com> wrote:
>
> During the early page table creation, we used to set the mapping for
> PAGE_OFFSET to the kernel load address: but the kernel load address is
> always offseted by PMD_SIZE which makes it impossible to use PUD/P4D/PGD
> pages as this physical address is not aligned on PUD/P4D/PGD size (whereas
> PAGE_OFFSET is).
>
> But actually we don't have to establish this mapping (ie set va_pa_offset)
> that early in the boot process because:
>
> - first, setup_vm installs a temporary kernel mapping and among other
>   things, discovers the system memory,
> - then, setup_vm_final creates the final kernel mapping and takes
>   advantage of the discovered system memory to create the linear
>   mapping.
>
> During the first phase, we don't know the start of the system memory and
> then until the second phase is finished, we can't use the linear mapping at
> all and phys_to_virt/virt_to_phys translations must not be used because it
> would result in a different translation from the 'real' one once the final
> mapping is installed.
>
> So here we simply delay the initialization of va_pa_offset to after the
> system memory discovery. But to make sure noone uses the linear mapping
> before, we add some guard in the DEBUG_VIRTUAL config.
>
> Finally we can use PUD/P4D/PGD hugepages when possible, which will result
> in a better TLB utilization.
>
> Note that we rely on the firmware to protect itself using PMP.
>
> Signed-off-by: Alexandre Ghiti <alexghiti at rivosinc.com>
> ---
>
> v2:
> - Add a comment on why RISCV64 does not need to set initrd_start/end that
>   early in the boot process, as asked by Rob
>
> Note that this patch is rebased on top of:
> [PATCH v1 1/1] riscv: mm: call best_map_size many times during linear-mapping
>
>  arch/riscv/include/asm/page.h | 16 ++++++++++++++++
>  arch/riscv/mm/init.c          | 25 +++++++++++++++++++------
>  arch/riscv/mm/physaddr.c      | 16 ++++++++++++++++
>  drivers/of/fdt.c              |  7 ++++++-
>  4 files changed, 57 insertions(+), 7 deletions(-)

[...]

> diff --git a/drivers/of/fdt.c b/drivers/of/fdt.c
> index 7b571a631639..012554445054 100644
> --- a/drivers/of/fdt.c
> +++ b/drivers/of/fdt.c
> @@ -895,8 +895,13 @@ static void __early_init_dt_declare_initrd(unsigned long start,
>          * enabled since __va() is called too early. ARM64 does make use
>          * of phys_initrd_start/phys_initrd_size so we can skip this
>          * conversion.
> +        * On RISCV64, the usage of __va() before the linear mapping exists
> +        * is wrong and RISCV64 rightly calls reserve_initrd_mem after it is
> +        * available where it actually resets the translation that is done
> +        * here and re-computes it.

No, I want a single comment that covers both cases (and the next arch
we add here). Something like this:

__va() is not yet available this early on some platforms. In that
case, the platform uses phys_initrd_start/phys_initrd_size instead and
does the VA conversion itself.

Rob