Re: [PATCH v3 2/2] riscv: Cleanup KASAN_VMALLOC support

From: Palmer Dabbelt
Date: Thu Apr 15 2021 - 02:16:37 EST


On Tue, 30 Mar 2021 02:47:30 PDT (-0700), geert@xxxxxxxxxxxxxx wrote:
Hi Palmer,

On Tue, Mar 30, 2021 at 7:08 AM Palmer Dabbelt <palmer@xxxxxxxxxxx> wrote:
On Sat, 13 Mar 2021 00:45:05 PST (-0800), alex@xxxxxxxx wrote:
> When KASAN vmalloc region is populated, there is no userspace process and
> the page table in use is swapper_pg_dir, so there is no need to read
> SATP. Then we can use the same scheme used by kasan_populate_p*d
> functions to go through the page table, which harmonizes the code.
>
> In addition, make use of set_pgd that goes through all unused page table
> levels, contrary to p*d_populate functions, which makes this function work
> whatever the number of page table levels.
>
> Signed-off-by: Alexandre Ghiti <alex@xxxxxxxx>
> Reviewed-by: Palmer Dabbelt <palmerdabbelt@xxxxxxxxxx>
> ---
> arch/riscv/mm/kasan_init.c | 59 ++++++++++++--------------------------
> 1 file changed, 18 insertions(+), 41 deletions(-)
>
> diff --git a/arch/riscv/mm/kasan_init.c b/arch/riscv/mm/kasan_init.c
> index 57bf4ae09361..c16178918239 100644
> --- a/arch/riscv/mm/kasan_init.c
> +++ b/arch/riscv/mm/kasan_init.c
> @@ -11,18 +11,6 @@
> #include <asm/fixmap.h>
> #include <asm/pgalloc.h>
>
> -static __init void *early_alloc(size_t size, int node)
> -{
> - void *ptr = memblock_alloc_try_nid(size, size,
> - __pa(MAX_DMA_ADDRESS), MEMBLOCK_ALLOC_ACCESSIBLE, node);
> -
> - if (!ptr)
> - panic("%pS: Failed to allocate %zu bytes align=%zx nid=%d from=%llx\n",
> - __func__, size, size, node, (u64)__pa(MAX_DMA_ADDRESS));
> -
> - return ptr;
> -}
> -
> extern pgd_t early_pg_dir[PTRS_PER_PGD];
> asmlinkage void __init kasan_early_init(void)
> {
> @@ -155,38 +143,27 @@ static void __init kasan_populate(void *start, void *end)
> memset(start, KASAN_SHADOW_INIT, end - start);
> }
>
> -void __init kasan_shallow_populate(void *start, void *end)
> +static void __init kasan_shallow_populate_pgd(unsigned long vaddr, unsigned long end)
> {
> - unsigned long vaddr = (unsigned long)start & PAGE_MASK;
> - unsigned long vend = PAGE_ALIGN((unsigned long)end);
> - unsigned long pfn;
> - int index;
> + unsigned long next;
> void *p;
> - pud_t *pud_dir, *pud_k;
> - pgd_t *pgd_dir, *pgd_k;
> - p4d_t *p4d_dir, *p4d_k;
> -
> - while (vaddr < vend) {
> - index = pgd_index(vaddr);
> - pfn = csr_read(CSR_SATP) & SATP_PPN;
> - pgd_dir = (pgd_t *)pfn_to_virt(pfn) + index;
> - pgd_k = init_mm.pgd + index;
> - pgd_dir = pgd_offset_k(vaddr);
> - set_pgd(pgd_dir, *pgd_k);
> -
> - p4d_dir = p4d_offset(pgd_dir, vaddr);
> - p4d_k = p4d_offset(pgd_k, vaddr);
> -
> - vaddr = (vaddr + PUD_SIZE) & PUD_MASK;
> - pud_dir = pud_offset(p4d_dir, vaddr);
> - pud_k = pud_offset(p4d_k, vaddr);
> -
> - if (pud_present(*pud_dir)) {
> - p = early_alloc(PAGE_SIZE, NUMA_NO_NODE);
> - pud_populate(&init_mm, pud_dir, p);
> + pgd_t *pgd_k = pgd_offset_k(vaddr);
> +
> + do {
> + next = pgd_addr_end(vaddr, end);
> + if (pgd_page_vaddr(*pgd_k) == (unsigned long)lm_alias(kasan_early_shadow_pmd)) {
> + p = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
> + set_pgd(pgd_k, pfn_pgd(PFN_DOWN(__pa(p)), PAGE_TABLE));
> }
> - vaddr += PAGE_SIZE;
> - }
> + } while (pgd_k++, vaddr = next, vaddr != end);
> +}
> +
> +static void __init kasan_shallow_populate(void *start, void *end)
> +{
> + unsigned long vaddr = (unsigned long)start & PAGE_MASK;
> + unsigned long vend = PAGE_ALIGN((unsigned long)end);
> +
> + kasan_shallow_populate_pgd(vaddr, vend);
>
> local_flush_tlb_all();
> }

Thanks, this is on for-next.

Your for-next does not include your fixes branch, hence they now conflict,
and for-next lacks the local_flush_tlb_all().

This came up before and I don't think we ever sorted out what the right thing to do is. Right now I'm keeping for-next pinned an at early RC, but fast-forwarding fixes to the latest RC every time I sent a PR. I don't have fixes merged back into for-next because I don't want those merges to show up when I send my merge window PRs.

For this one I purposefully left out the local_flush_tlb_all() whene I pulled in this patch, and was planning on fixing it up along with any other merge conflicts when I send along the PR. It does all seem like a bit of a song and dance here, though, so I'm open to suggestions as to how to run this better -- though last time I went through that exercise it seemed like everyone had their own way of doing it, they all had a different set of issues, and I was at least familiar with this flavor of craziness.

I was kind of tempted to convert for-next over into a branch that only contains merges, though, which would make it a bit easier to merge fixes in.


Gr{oetje,eeting}s,

Geert