[PATCH] RISC-V: KVM: Remove automatic I/O mapping for VM_PFNMAP

[Guo Ren]

On Mon, Oct 20, 2025 at 6:08 AM <fangyu.yu at linux.alibaba.com> wrote:
>
> From: Fangyu Yu <fangyu.yu at linux.alibaba.com>
>
> As of commit aac6db75a9fc ("vfio/pci: Use unmap_mapping_range()"),
> vm_pgoff may no longer guaranteed to hold the PFN for VM_PFNMAP
> regions. Using vma->vm_pgoff to derive the HPA here may therefore
> produce incorrect mappings.
>
> Instead, I/O mappings for such regions can be established on-demand
> during g-stage page faults, making the upfront ioremap in this path
> is unnecessary.
>
> Fixes: 9d05c1fee837 ("RISC-V: KVM: Implement stage2 page table programming")
The Fixes tag should be 'commit aac6db75a9fc ("vfio/pci: Use
unmap_mapping_range()")'.

A stable tree necessitates minimizing the "Fixes tag" interference.

We also need to
Cc: Jason Gunthorpe <jgg at nvidia.com>
Cc: Alex Williamson <alex.williamson at redhat.com>
For review.

> Signed-off-by: Fangyu Yu <fangyu.yu at linux.alibaba.com>
> ---
>  arch/riscv/kvm/mmu.c | 20 +-------------------
>  1 file changed, 1 insertion(+), 19 deletions(-)
>
> diff --git a/arch/riscv/kvm/mmu.c b/arch/riscv/kvm/mmu.c
> index 525fb5a330c0..84c04c8f0892 100644
> --- a/arch/riscv/kvm/mmu.c
> +++ b/arch/riscv/kvm/mmu.c
> @@ -197,8 +197,7 @@ int kvm_arch_prepare_memory_region(struct kvm *kvm,
>
>         /*
>          * A memory region could potentially cover multiple VMAs, and
> -        * any holes between them, so iterate over all of them to find
> -        * out if we can map any of them right now.
> +        * any holes between them, so iterate over all of them.
>          *
>          *     +--------------------------------------------+
>          * +---------------+----------------+   +----------------+
> @@ -229,32 +228,15 @@ int kvm_arch_prepare_memory_region(struct kvm *kvm,
>                 vm_end = min(reg_end, vma->vm_end);
>
>                 if (vma->vm_flags & VM_PFNMAP) {
> -                       gpa_t gpa = base_gpa + (vm_start - hva);
> -                       phys_addr_t pa;
> -
> -                       pa = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT;
> -                       pa += vm_start - vma->vm_start;
> -
>                         /* IO region dirty page logging not allowed */
>                         if (new->flags & KVM_MEM_LOG_DIRTY_PAGES) {
>                                 ret = -EINVAL;
>                                 goto out;
>                         }
> -
> -                       ret = kvm_riscv_mmu_ioremap(kvm, gpa, pa, vm_end - vm_start,
> -                                                   writable, false);
> -                       if (ret)
> -                               break;
Defering the ioremap to the g-stage page fault looks good to me, as it
simplifies the implementation here.

Acked-by: Guo Ren <guoren at kernel.org>

>                 }
>                 hva = vm_end;
>         } while (hva < reg_end);
>
> -       if (change == KVM_MR_FLAGS_ONLY)
> -               goto out;
> -
> -       if (ret)
> -               kvm_riscv_mmu_iounmap(kvm, base_gpa, size);
> -
>  out:
>         mmap_read_unlock(current->mm);
>         return ret;
> --
> 2.50.1
>


-- 
Best Regards
 Guo Ren