[PATCH v2] riscv: Only extend kernel reservation if mapped read-only

[Alex Ghiti]

Le 4/29/21 à 11:05 AM, Geert Uytterhoeven a écrit :
> When the kernel mapping was moved outside of the linear mapping, the
> kernel memory reservation was increased, to take into account mapping
> granularity.  However, this is done unconditionally, regardless of
> whether the kernel memory is mapped read-only or not.
> 
> If this extension is not needed, up to 2 MiB may be lost, which has a
> big impact on e.g. Canaan K210 (64-bit nommu) platforms with only 8 MiB
> of RAM.
> 
> Reclaim the lost memory by only extending the reserved region when
> needed, i.e. depending on a simplified version of the conditional logic
> around the call to protect_kernel_linear_mapping_text_rodata().
> 
> Fixes: 2bfc6cd81bd17e43 ("riscv: Move kernel mapping outside of linear mapping")
> Signed-off-by: Geert Uytterhoeven <geert+renesas at glider.be>
> ---
> v2:
>    - Simplify the conditional, as STRICT_KERNEL_RWX depends on
>      MMU && !XIP_KERNEL.
> 
> Only tested on K210 (SiPeed MAIX BiT):
> 
>      -Memory: 5852K/8192K available (1344K kernel code, 147K rwdata, 272K rodata, 106K init, 72K bss, 2340K reserved, 0K cma-reserved)
>      +Memory: 5948K/8192K available (1344K kernel code, 147K rwdata, 272K rodata, 106K init, 72K bss, 2244K reserved, 0K cma-reserved)
> 
> Yes, I was lucky, as only 96 KiB was lost ;-)
> ---
>   arch/riscv/mm/init.c | 9 +++++++--
>   1 file changed, 7 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/riscv/mm/init.c b/arch/riscv/mm/init.c
> index 788eb222deacf994..3ebc0f5d2b73b42b 100644
> --- a/arch/riscv/mm/init.c
> +++ b/arch/riscv/mm/init.c
> @@ -136,11 +136,16 @@ void __init setup_bootmem(void)
>   
>   	/*
>   	 * Reserve from the start of the kernel to the end of the kernel
> -	 * and make sure we align the reservation on PMD_SIZE since we will
> +	 */
> +#if defined(CONFIG_64BIT) && defined(CONFIG_STRICT_KERNEL_RWX)
> +	/*
> +	 * Make sure we align the reservation on PMD_SIZE since we will
>   	 * map the kernel in the linear mapping as read-only: we do not want
>   	 * any allocation to happen between _end and the next pmd aligned page.
>   	 */
> -	memblock_reserve(vmlinux_start, (vmlinux_end - vmlinux_start + PMD_SIZE - 1) & PMD_MASK);
> +	vmlinux_end = (vmlinux_end + PMD_SIZE - 1) & PMD_MASK;
> +#endif
> +	memblock_reserve(vmlinux_start, vmlinux_end - vmlinux_start);
>   
>   	/*
>   	 * memblock allocator is not aware of the fact that last 4K bytes of
> 

I tested this on the following configs:

- rv32_defconfig (build and valid on qemu)
- defconfig (with and without CONFIG_STRICT_KERNEL_RWX) (build and valid 
on qemu)
- xip kernel (build and valid on qemu)
- nommu_k210_defconfig (build only)

so you can add:

Tested-by: Alexandre Ghiti <alex at ghiti.fr>

Thank you again for that,

Alex