[PATCH v2] riscv: mm: Fix alignment of phys_ram_base

[Alexandre Ghiti]

Hi Xu,

On 03/12/2024 15:49, Xu Lu wrote:
> This commit fixes the alignment of phys_ram_base in RISC-V.
>
> In sparse vmemmap model, the virtual address of vmemmap is calculated as:
> ((struct page *)VMEMMAP_START - (phys_ram_base >> PAGE_SHIFT)).
> And the struct page's va can be calculated with an offset:
> (vmemmap + (pfn)).
>
> However, when initializing struct pages, kernel actually starts from the
> first page from the same section that phys_ram_base belongs to. If the
> first page's physical address is not (phys_ram_base >> PAGE_SHIFT), then
> we get an va below VMEMMAP_START when calculating va for it's struct page.
>
> For example, if phys_ram_base starts from 0x82000000 with pfn 0x82000, the
> first page in the same section is actually pfn 0x80000. During
> init_unavailable_range(), we will initialize struct page for pfn 0x80000
> with virtual address ((struct page *)VMEMMAP_START - 0x2000), which is
> below VMEMMAP_START as well as PCI_IO_END.
>
> This commit fixes this bug by aligning phys_ram_base with SECTION_SIZE.
>
> Fixes: c3bcc65d4d2e ("riscv: Start of DRAM should at least be aligned on PMD size for the direct mapping")
> Signed-off-by: Xu Lu <luxu.kernel at bytedance.com>
> ---
>   arch/riscv/mm/init.c | 15 ++++++++++++++-
>   1 file changed, 14 insertions(+), 1 deletion(-)
>
> diff --git a/arch/riscv/mm/init.c b/arch/riscv/mm/init.c
> index 0e8c20adcd98..974cafa7c85e 100644
> --- a/arch/riscv/mm/init.c
> +++ b/arch/riscv/mm/init.c
> @@ -33,6 +33,9 @@
>   #include <asm/pgtable.h>
>   #include <asm/sections.h>
>   #include <asm/soc.h>
> +#ifdef CONFIG_SPARSEMEM
> +#include <asm/sparsemem.h>
> +#endif
>   #include <asm/tlbflush.h>
>   
>   #include "../kernel/head.h"
> @@ -59,6 +62,12 @@ EXPORT_SYMBOL(pgtable_l4_enabled);
>   EXPORT_SYMBOL(pgtable_l5_enabled);
>   #endif
>   
> +#ifdef CONFIG_SPARSEMEM
> +#define RISCV_MEMSTART_ALIGN	(1UL << SECTION_SIZE_BITS)
> +#else
> +#define RISCV_MEMSTART_ALIGN	PMD_SIZE
> +#endif
> +
>   phys_addr_t phys_ram_base __ro_after_init;
>   EXPORT_SYMBOL(phys_ram_base);
>   
> @@ -239,9 +248,13 @@ static void __init setup_bootmem(void)
>   	/*
>   	 * Make sure we align the start of the memory on a PMD boundary so that
>   	 * at worst, we map the linear mapping with PMD mappings.
> +	 *
> +	 * Also, make sure we align the start of the memory on a SECTION boundary
> +	 * when CONFIG_SPARSEMEM_VMEMMAP is enabled to ensure the correctness of
> +	 * pfn_to_page().
>   	 */
>   	if (!IS_ENABLED(CONFIG_XIP_KERNEL))
> -		phys_ram_base = memblock_start_of_DRAM() & PMD_MASK;
> +		phys_ram_base = round_down(memblock_start_of_DRAM(), RISCV_MEMSTART_ALIGN);
>   
>   	/*
>   	 * In 64-bit, any use of __va/__pa before this point is wrong as we


That's a good catch indeed. But I'm wondering if it would be more 
correct to fix the macro vmemmap instead of phys_ram_base since 
phys_ram_base is supposed to hold the real base of the system memory, 
which would be wrong with your patch. I mean something like that instead 
(or similar, I haven't tested):

#define vmemmap         ((struct page *)VMEMMAP_START - 
(round_down(memblock_start_of_DRAM(), RISCV_MEMSTART_ALIGN) >> PAGE_SHIFT))

And the fixes tag should be:

Fixes: a11dd49dcb93 ("riscv: Sparse-Memory/vmemmap out-of-bounds fix")

Thanks,

Alex