Hi Ricardo,
On Sat, Nov 12, 2022 at 08:17:06AM +0000, Ricardo Koller wrote:
[...]
> +/**
> + * kvm_pgtable_stage2_split() - Split a range of huge pages into leaf PTEs pointing
> + * to PAGE_SIZE guest pages.
> + * @pgt: Page-table structure initialised by kvm_pgtable_stage2_init*().
> + * @addr: Intermediate physical address from which to split.
> + * @size: Size of the range.
> + * @mc: Cache of pre-allocated and zeroed memory from which to allocate
> + * page-table pages.
> + *
> + * @addr and the end (@addr + @size) are effectively aligned down and up to
> + * the top level huge-page block size. This is an exampe using 1GB
> + * huge-pages and 4KB granules.
> + *
> + * [---input range---]
> + * : :
> + * [--1G block pte--][--1G block pte--][--1G block pte--][--1G block pte--]
> + * : :
> + * [--2MB--][--2MB--][--2MB--][--2MB--]
> + * : :
> + * [ ][ ][:][ ][ ][ ][ ][ ][:][ ][ ][ ]
> + * : :
> + *
> + * Return: 0 on success, negative error code on failure. Note that
> + * kvm_pgtable_stage2_split() is best effort: it tries to break as many
> + * blocks in the input range as allowed by the size of the memcache. It
> + * will fail it wasn't able to break any block.
> + */
> +int kvm_pgtable_stage2_split(struct kvm_pgtable *pgt, u64 addr, u64 size, void *mc);
> +
> /**
> * kvm_pgtable_walk() - Walk a page-table.
> * @pgt: Page-table structure initialised by kvm_pgtable_*_init().
> diff --git a/arch/arm64/kvm/hyp/pgtable.c b/arch/arm64/kvm/hyp/pgtable.c
> index d1f309128118..9c42eff6d42e 100644
> --- a/arch/arm64/kvm/hyp/pgtable.c
> +++ b/arch/arm64/kvm/hyp/pgtable.c
> @@ -1267,6 +1267,80 @@ static int stage2_create_removed(kvm_pte_t *ptep, u64 phys, u32 level,
> return __kvm_pgtable_visit(&data, mm_ops, ptep, level);
> }
>
> +struct stage2_split_data {
> + struct kvm_s2_mmu *mmu;
> + void *memcache;
> + struct kvm_pgtable_mm_ops *mm_ops;
You can also get at mm_ops through kvm_pgtable_visit_ctx
> +};
> +
> +static int stage2_split_walker(const struct kvm_pgtable_visit_ctx *ctx,
> + enum kvm_pgtable_walk_flags visit)
> +{
> + struct stage2_split_data *data = ctx->arg;
> + struct kvm_pgtable_mm_ops *mm_ops = data->mm_ops;
> + kvm_pte_t pte = ctx->old, attr, new;
> + enum kvm_pgtable_prot prot;
> + void *mc = data->memcache;
> + u32 level = ctx->level;
> + u64 phys;
> +
> + if (WARN_ON_ONCE(kvm_pgtable_walk_shared(ctx)))
> + return -EINVAL;
> +
> + /* Nothing to split at the last level */
> + if (level == KVM_PGTABLE_MAX_LEVELS - 1)
> + return 0;
> +
> + /* We only split valid block mappings */
> + if (!kvm_pte_valid(pte) || kvm_pte_table(pte, ctx->level))
> + return 0;
> +
> + phys = kvm_pte_to_phys(pte);
> + prot = kvm_pgtable_stage2_pte_prot(pte);
> + stage2_set_prot_attr(data->mmu->pgt, prot, &attr);
> +
> + /*
> + * Eager page splitting is best-effort, so we can ignore the error.
> + * The returned PTE (new) will be valid even if this call returns
> + * error: new will be a single (big) block PTE. The only issue is
> + * that it will affect dirty logging performance, as the huge-pages
> + * will have to be split on fault, and so we WARN.
> + */
> + WARN_ON(stage2_create_removed(&new, phys, level, attr, mc, mm_ops));
I don't believe we should warn in this case, at least not
unconditionally. ENOMEM is an expected outcome, for example.
Additionally, I believe you'll want to bail out at this point to avoid
installing a potentially garbage PTE as well.
> + stage2_put_pte(ctx, data->mmu, mm_ops);
Ah, I see why you've relaxed the WARN in patch 1 now.
I would recommend you follow the break-before-make pattern and use the
helpers here as well. stage2_try_break_pte() will demote the store to
WRITE_ONCE() if called from a non-shared context.
Then the WARN will behave as expected in stage2_make_pte().
> + /*
> + * Note, the contents of the page table are guaranteed to be made
> + * visible before the new PTE is assigned because
> + * stage2_make__pte() writes the PTE using smp_store_release().
typo: stage2_make_pte()
--
Thanks,
Oliver
