Re: [PATCH v1 09/16] kvm: arm/arm64: Delay stage2 page table allocation

[Christoffer Dall]

On Tue, Jan 09, 2018 at 07:04:04PM +0000, Suzuki K Poulose wrote:
> We allocate the entry level page tables for stage2 when the
> VM is created. This doesn't give us the flexibility of configuring
> the physical address space size for a VM. In order to allow
> the VM to choose the required size, we delay the allocation of
> stage2 entry level tables until we really try to map something.
> 
> This could be either when the VM creates a memory range or when
> it tries to map a device memory. So we add in a hook to these
> two places to make sure the tables are allocated. We use
> kvm->slots_lock to serialize the allocation entry point, since
> we add hooks to the arch specific call back with the mutex held.
> 
> Cc: Marc Zyngier <marc.zyngier@xxxxxxx>
> Cc: Christoffer Dall <cdall@xxxxxxxxxx>
> Signed-off-by: Suzuki K Poulose <suzuki.poulose@xxxxxxx>
> ---
>  virt/kvm/arm/arm.c | 18 ++++++----------
>  virt/kvm/arm/mmu.c | 61 +++++++++++++++++++++++++++++++++++++++++++++---------
>  2 files changed, 57 insertions(+), 22 deletions(-)
> 
> diff --git a/virt/kvm/arm/arm.c b/virt/kvm/arm/arm.c
> index 19b720ddedce..d06f00566664 100644
> --- a/virt/kvm/arm/arm.c
> +++ b/virt/kvm/arm/arm.c
> @@ -127,13 +127,13 @@ int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
>  	for_each_possible_cpu(cpu)
>  		*per_cpu_ptr(kvm->arch.last_vcpu_ran, cpu) = -1;
>  
> -	ret = kvm_alloc_stage2_pgd(kvm);
> -	if (ret)
> -		goto out_fail_alloc;
> -
>  	ret = create_hyp_mappings(kvm, kvm + 1, PAGE_HYP);
> -	if (ret)
> -		goto out_free_stage2_pgd;
> +	if (ret) {
> +		free_percpu(kvm->arch.last_vcpu_ran);
> +		kvm->arch.last_vcpu_ran = NULL;
> +		return ret;
> +	}
> +
>  
>  	kvm_vgic_early_init(kvm);
>  
> @@ -145,12 +145,6 @@ int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
>  				kvm_vgic_get_max_vcpus() : KVM_MAX_VCPUS;
>  
>  	return ret;
> -out_free_stage2_pgd:
> -	kvm_free_stage2_pgd(kvm);
> -out_fail_alloc:
> -	free_percpu(kvm->arch.last_vcpu_ran);
> -	kvm->arch.last_vcpu_ran = NULL;
> -	return ret;
>  }
>  
>  bool kvm_arch_has_vcpu_debugfs(void)
> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c
> index c94c61ac38b9..257f2a8ccfc7 100644
> --- a/virt/kvm/arm/mmu.c
> +++ b/virt/kvm/arm/mmu.c
> @@ -1011,15 +1011,39 @@ static int stage2_pmdp_test_and_clear_young(pmd_t *pmd)
>  	return stage2_ptep_test_and_clear_young((pte_t *)pmd);
>  }
>  
> -/**
> - * kvm_phys_addr_ioremap - map a device range to guest IPA
> - *
> - * @kvm:	The KVM pointer
> - * @guest_ipa:	The IPA at which to insert the mapping
> - * @pa:		The physical address of the device
> - * @size:	The size of the mapping
> +/*
> + * Finalise the stage2 page table layout. Must be called with kvm->slots_lock
> + * held.
>   */
> -int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
> +static int __kvm_init_stage2_table(struct kvm *kvm)
> +{
> +	/* Double check if somebody has already allocated it */

dubious comment: Either leave it out or explain that we need to check
again with the mutex held.

> +	if (likely(kvm->arch.pgd))
> +		return 0;
> +	return kvm_alloc_stage2_pgd(kvm);
> +}
> +
> +static int kvm_init_stage2_table(struct kvm *kvm)
> +{
> +	int rc;
> +
> +	/*
> +	 * Once allocated, the stage2 entry level tables are only
> +	 * freed when the KVM instance is destroyed. So, if we see
> +	 * something valid here, that guarantees that we have
> +	 * done the one time allocation and it is something valid
> +	 * and won't go away until the last reference to the KVM
> +	 * is gone.
> +	 */

Really not sure if this comment adds something beyond what's described
by the code already?

Thanks,
-Christoffer

> +	if (likely(kvm->arch.pgd))
> +		return 0;
> +	mutex_lock(&kvm->slots_lock);
> +	rc = __kvm_init_stage2_table(kvm);
> +	mutex_unlock(&kvm->slots_lock);
> +	return rc;
> +}
> +
> +static int __kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
>  			  phys_addr_t pa, unsigned long size, bool writable)
>  {
>  	phys_addr_t addr, end;
> @@ -1055,6 +1079,23 @@ int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
>  	return ret;
>  }
>  
> +/**
> + * kvm_phys_addr_ioremap - map a device range to guest IPA.
> + * Acquires kvm->slots_lock for making sure that the stage2 is initialized.
> + *
> + * @kvm:	The KVM pointer
> + * @guest_ipa:	The IPA at which to insert the mapping
> + * @pa:		The physical address of the device
> + * @size:	The size of the mapping
> + */
> +int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
> +			  phys_addr_t pa, unsigned long size, bool writable)
> +{
> +	if (unlikely(kvm_init_stage2_table(kvm)))
> +		return -ENOMEM;
> +	return __kvm_phys_addr_ioremap(kvm, guest_ipa, pa, size, writable);
> +}
> +
>  static bool transparent_hugepage_adjust(kvm_pfn_t *pfnp, phys_addr_t *ipap)
>  {
>  	kvm_pfn_t pfn = *pfnp;
> @@ -1912,7 +1953,7 @@ int kvm_arch_prepare_memory_region(struct kvm *kvm,
>  				goto out;
>  			}
>  
> -			ret = kvm_phys_addr_ioremap(kvm, gpa, pa,
> +			ret = __kvm_phys_addr_ioremap(kvm, gpa, pa,
>  						    vm_end - vm_start,
>  						    writable);
>  			if (ret)
> @@ -1943,7 +1984,7 @@ void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
>  int kvm_arch_create_memslot(struct kvm *kvm, struct kvm_memory_slot *slot,
>  			    unsigned long npages)
>  {
> -	return 0;
> +	return __kvm_init_stage2_table(kvm);
>  }
>  
>  void kvm_arch_memslots_updated(struct kvm *kvm, struct kvm_memslots *slots)
> -- 
> 2.13.6
>