Re: [PATCH v2 13/16] arm64: Implement ioremap/iounmap hooks calling into KVM's MMIO guard

[Andrew Jones <drjones@redhat.com>]

On Mon, Oct 04, 2021 at 06:48:46PM +0100, Marc Zyngier wrote:
> Implement the previously defined ioremap/iounmap hooks for arm64,
> calling into KVM's MMIO guard if available.
> 
> Signed-off-by: Marc Zyngier <maz@kernel.org>
> ---
>  arch/arm64/mm/ioremap.c | 112 ++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 112 insertions(+)
> 
> diff --git a/arch/arm64/mm/ioremap.c b/arch/arm64/mm/ioremap.c
> index b7c81dacabf0..5334cbdc9f64 100644
> --- a/arch/arm64/mm/ioremap.c
> +++ b/arch/arm64/mm/ioremap.c
> @@ -9,13 +9,125 @@
>   * Copyright (C) 2012 ARM Ltd.
>   */
>  
> +#define pr_fmt(fmt)	"ioremap: " fmt
> +
>  #include <linux/export.h>
>  #include <linux/mm.h>
>  #include <linux/vmalloc.h>
> +#include <linux/slab.h>
>  #include <linux/io.h>
> +#include <linux/arm-smccc.h>
>  
>  #include <asm/fixmap.h>
>  #include <asm/tlbflush.h>
> +#include <asm/hypervisor.h>
> +
> +struct ioremap_guard_ref {
> +	refcount_t	count;
> +};
> +
> +static DEFINE_STATIC_KEY_FALSE(ioremap_guard_key);
> +static DEFINE_XARRAY(ioremap_guard_array);
> +static DEFINE_MUTEX(ioremap_guard_lock);
> +
> +void ioremap_phys_range_hook(phys_addr_t phys_addr, size_t size, pgprot_t prot)
> +{
> +	if (!static_branch_unlikely(&ioremap_guard_key))
> +		return;
> +
> +	if (pfn_valid(__phys_to_pfn(phys_addr)))
> +		return;
> +
> +	mutex_lock(&ioremap_guard_lock);
> +
> +	while (size) {
> +		u64 pfn = phys_addr >> PAGE_SHIFT;
> +		struct ioremap_guard_ref *ref;
> +		struct arm_smccc_res res;
> +
> +		ref = xa_load(&ioremap_guard_array, pfn);
> +		if (ref) {
> +			refcount_inc(&ref->count);
> +			goto next;
> +		}
> +
> +		/*
> +		 * It is acceptable for the allocation to fail, specially
> +		 * if trying to ioremap something very early on, like with
> +		 * earlycon, which happens long before kmem_cache_init.
> +		 * This page will be permanently accessible, similar to a
> +		 * saturated refcount.
> +		 */
> +		ref = kzalloc(sizeof(*ref), GFP_KERNEL);
> +		if (ref) {
> +			refcount_set(&ref->count, 1);
> +			if (xa_err(xa_store(&ioremap_guard_array, pfn, ref,
> +					    GFP_KERNEL))) {
> +				kfree(ref);
> +				ref = NULL;
> +			}
> +		}
> +
> +		arm_smccc_1_1_hvc(ARM_SMCCC_VENDOR_HYP_KVM_MMIO_GUARD_MAP_FUNC_ID,
> +				  phys_addr, prot, &res);

OK, I see this follows the document and passes prot in x2, even though the
hypercall implementation doesn't look at it [yet].

> +		if (res.a0 != SMCCC_RET_SUCCESS) {
> +			pr_warn_ratelimited("Failed to register %llx\n",
> +					    phys_addr);
> +			xa_erase(&ioremap_guard_array, pfn);
> +			kfree(ref);
> +			goto out;
> +		}
> +
> +	next:
> +		size -= PAGE_SIZE;
> +		phys_addr += PAGE_SIZE;

Looks like we're assuming the guard granule to be PAGE_SIZE here. Looking
ahead at the next patch, I see it must be PAGE_SIZE, because if the info
hypercall doesn't have a matching value, then mmio guarding doesn't
happen at all. Maybe it should be documented that for this feature the
host and guest must have matching page sizes.

> +	}
> +out:
> +	mutex_unlock(&ioremap_guard_lock);
> +}
> +
> +void iounmap_phys_range_hook(phys_addr_t phys_addr, size_t size)
> +{
> +	if (!static_branch_unlikely(&ioremap_guard_key))
> +		return;
> +
> +	VM_BUG_ON(phys_addr & ~PAGE_MASK || size & ~PAGE_MASK);
> +
> +	mutex_lock(&ioremap_guard_lock);
> +
> +	while (size) {
> +		u64 pfn = phys_addr >> PAGE_SHIFT;
> +		struct ioremap_guard_ref *ref;
> +		struct arm_smccc_res res;
> +
> +		ref = xa_load(&ioremap_guard_array, pfn);
> +		if (!ref) {
> +			pr_warn_ratelimited("%llx not tracked, left mapped\n",
> +					    phys_addr);
> +			goto next;
> +		}
> +
> +		if (!refcount_dec_and_test(&ref->count))
> +			goto next;
> +
> +		xa_erase(&ioremap_guard_array, pfn);
> +		kfree(ref);
> +
> +		arm_smccc_1_1_hvc(ARM_SMCCC_VENDOR_HYP_KVM_MMIO_GUARD_UNMAP_FUNC_ID,
> +				  phys_addr, &res);
> +		if (res.a0 != SMCCC_RET_SUCCESS) {
> +			pr_warn_ratelimited("Failed to unregister %llx\n",
> +					    phys_addr);
> +			goto out;
> +		}
> +
> +	next:
> +		size -= PAGE_SIZE;
> +		phys_addr += PAGE_SIZE;
> +	}
> +out:
> +	mutex_unlock(&ioremap_guard_lock);
> +}
>  
>  static void __iomem *__ioremap_caller(phys_addr_t phys_addr, size_t size,
>  				      pgprot_t prot, void *caller)
> -- 
> 2.30.2
>

Thanks,
drew