KVM: arm64: Instantiate pKVM hypervisor VM and vCPU structures from EL1
authorFuad Tabba <tabba@google.com>
Thu, 10 Nov 2022 19:02:46 +0000 (19:02 +0000)
committerMarc Zyngier <maz@kernel.org>
Fri, 11 Nov 2022 17:16:24 +0000 (17:16 +0000)
With the pKVM hypervisor at EL2 now offering hypercalls to the host for
creating and destroying VM and vCPU structures, plumb these in to the
existing arm64 KVM backend to ensure that the hypervisor data structures
are allocated and initialised on first vCPU run for a pKVM guest.

In the host, 'struct kvm_protected_vm' is introduced to hold the handle
of the pKVM VM instance as well as to track references to the memory
donated to the hypervisor so that it can be freed back to the host
allocator following VM teardown. The stage-2 page-table, hypervisor VM
and vCPU structures are allocated separately so as to avoid the need for
a large physically-contiguous allocation in the host at run-time.

Tested-by: Vincent Donnefort <vdonnefort@google.com>
Signed-off-by: Fuad Tabba <tabba@google.com>
Signed-off-by: Will Deacon <will@kernel.org>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20221110190259.26861-14-will@kernel.org
arch/arm64/include/asm/kvm_host.h
arch/arm64/include/asm/kvm_pkvm.h
arch/arm64/kvm/arm.c
arch/arm64/kvm/hyp/hyp-constants.c
arch/arm64/kvm/hyp/nvhe/pkvm.c
arch/arm64/kvm/pkvm.c

index d3dd7ab9c79eaed3fa3561653cab056f8493b58d..467393e7331f10b5ca250b5503a2b8cfff968a13 100644 (file)
@@ -117,6 +117,16 @@ struct kvm_smccc_features {
 
 typedef unsigned int pkvm_handle_t;
 
+struct kvm_protected_vm {
+       pkvm_handle_t handle;
+
+       struct {
+               void *pgd;
+               void *vm;
+               void *vcpus[KVM_MAX_VCPUS];
+       } hyp_donations;
+};
+
 struct kvm_arch {
        struct kvm_s2_mmu mmu;
 
@@ -170,10 +180,10 @@ struct kvm_arch {
        struct kvm_smccc_features smccc_feat;
 
        /*
-        * For an untrusted host VM, 'pkvm_handle' is used to lookup
+        * For an untrusted host VM, 'pkvm.handle' is used to lookup
         * the associated pKVM instance in the hypervisor.
         */
-       pkvm_handle_t pkvm_handle;
+       struct kvm_protected_vm pkvm;
 };
 
 struct kvm_vcpu_fault_info {
index f4e3133d65500b0469b9a68b985d862d96b21e63..01129b0d4c689e46c149b6d91dd5b3bce99f64d6 100644 (file)
 
 #define HYP_MEMBLOCK_REGIONS 128
 
+int pkvm_init_host_vm(struct kvm *kvm);
+int pkvm_create_hyp_vm(struct kvm *kvm);
+void pkvm_destroy_hyp_vm(struct kvm *kvm);
+
 extern struct memblock_region kvm_nvhe_sym(hyp_memory)[];
 extern unsigned int kvm_nvhe_sym(hyp_memblock_nr);
 
index 94d33e296e10c57c34008af96d7a9adcb72d99ef..30d6fc5d3a93d8cad5b2419593f3f8ad04f8aac8 100644 (file)
@@ -37,6 +37,7 @@
 #include <asm/kvm_arm.h>
 #include <asm/kvm_asm.h>
 #include <asm/kvm_mmu.h>
+#include <asm/kvm_pkvm.h>
 #include <asm/kvm_emulate.h>
 #include <asm/sections.h>
 
@@ -150,6 +151,10 @@ int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
        if (ret)
                goto out_free_stage2_pgd;
 
+       ret = pkvm_init_host_vm(kvm);
+       if (ret)
+               goto out_free_stage2_pgd;
+
        if (!zalloc_cpumask_var(&kvm->arch.supported_cpus, GFP_KERNEL)) {
                ret = -ENOMEM;
                goto out_free_stage2_pgd;
@@ -187,6 +192,9 @@ void kvm_arch_destroy_vm(struct kvm *kvm)
 
        kvm_vgic_destroy(kvm);
 
+       if (is_protected_kvm_enabled())
+               pkvm_destroy_hyp_vm(kvm);
+
        kvm_destroy_vcpus(kvm);
 
        kvm_unshare_hyp(kvm, kvm + 1);
@@ -569,6 +577,12 @@ int kvm_arch_vcpu_run_pid_change(struct kvm_vcpu *vcpu)
        if (ret)
                return ret;
 
+       if (is_protected_kvm_enabled()) {
+               ret = pkvm_create_hyp_vm(kvm);
+               if (ret)
+                       return ret;
+       }
+
        if (!irqchip_in_kernel(kvm)) {
                /*
                 * Tell the rest of the code that there are userspace irqchip
index b3742a6691e87450865eefb61f0018ee1db17158..b257a3b4bfc5c6c7c9fdcba16b1d7c09bc4d76fd 100644 (file)
@@ -2,9 +2,12 @@
 
 #include <linux/kbuild.h>
 #include <nvhe/memory.h>
+#include <nvhe/pkvm.h>
 
 int main(void)
 {
        DEFINE(STRUCT_HYP_PAGE_SIZE,    sizeof(struct hyp_page));
+       DEFINE(PKVM_HYP_VM_SIZE,        sizeof(struct pkvm_hyp_vm));
+       DEFINE(PKVM_HYP_VCPU_SIZE,      sizeof(struct pkvm_hyp_vcpu));
        return 0;
 }
index 135c9a095eca2eafa3c61987aac7eece08af2178..2c73c4640e4ddc3d8e65cdeb5a4adbb657c4f2f5 100644 (file)
@@ -324,7 +324,7 @@ static pkvm_handle_t insert_vm_table_entry(struct kvm *host_kvm,
        if (idx < 0)
                return idx;
 
-       hyp_vm->kvm.arch.pkvm_handle = idx_to_vm_handle(idx);
+       hyp_vm->kvm.arch.pkvm.handle = idx_to_vm_handle(idx);
 
        /* VMID 0 is reserved for the host */
        atomic64_set(&mmu->vmid.id, idx + 1);
@@ -333,7 +333,7 @@ static pkvm_handle_t insert_vm_table_entry(struct kvm *host_kvm,
        mmu->pgt = &hyp_vm->pgt;
 
        vm_table[idx] = hyp_vm;
-       return hyp_vm->kvm.arch.pkvm_handle;
+       return hyp_vm->kvm.arch.pkvm.handle;
 }
 
 /*
@@ -458,10 +458,10 @@ int __pkvm_init_vm(struct kvm *host_kvm, unsigned long vm_hva,
                goto err_remove_vm_table_entry;
        hyp_spin_unlock(&vm_table_lock);
 
-       return hyp_vm->kvm.arch.pkvm_handle;
+       return hyp_vm->kvm.arch.pkvm.handle;
 
 err_remove_vm_table_entry:
-       remove_vm_table_entry(hyp_vm->kvm.arch.pkvm_handle);
+       remove_vm_table_entry(hyp_vm->kvm.arch.pkvm.handle);
 err_unlock:
        hyp_spin_unlock(&vm_table_lock);
 err_remove_mappings:
@@ -528,6 +528,7 @@ int __pkvm_teardown_vm(pkvm_handle_t handle)
 {
        struct pkvm_hyp_vm *hyp_vm;
        struct kvm *host_kvm;
+       unsigned int idx;
        size_t vm_size;
        int err;
 
@@ -553,6 +554,12 @@ int __pkvm_teardown_vm(pkvm_handle_t handle)
        unpin_host_vcpus(hyp_vm->vcpus, hyp_vm->nr_vcpus);
 
        /* Return the metadata pages to the host */
+       for (idx = 0; idx < hyp_vm->nr_vcpus; ++idx) {
+               struct pkvm_hyp_vcpu *hyp_vcpu = hyp_vm->vcpus[idx];
+
+               unmap_donated_memory(hyp_vcpu, sizeof(*hyp_vcpu));
+       }
+
        host_kvm = hyp_vm->host_kvm;
        vm_size = pkvm_get_hyp_vm_size(hyp_vm->kvm.created_vcpus);
        unmap_donated_memory(hyp_vm, vm_size);
index 71493136e59c200dd30ab7dd1e6170ef5d143ca1..8c443b915e439dfa925b4643cadd19ad3a0efbae 100644 (file)
@@ -6,6 +6,7 @@
 
 #include <linux/kvm_host.h>
 #include <linux/memblock.h>
+#include <linux/mutex.h>
 #include <linux/sort.h>
 
 #include <asm/kvm_pkvm.h>
@@ -94,3 +95,140 @@ void __init kvm_hyp_reserve(void)
        kvm_info("Reserved %lld MiB at 0x%llx\n", hyp_mem_size >> 20,
                 hyp_mem_base);
 }
+
+/*
+ * Allocates and donates memory for hypervisor VM structs at EL2.
+ *
+ * Allocates space for the VM state, which includes the hyp vm as well as
+ * the hyp vcpus.
+ *
+ * Stores an opaque handler in the kvm struct for future reference.
+ *
+ * Return 0 on success, negative error code on failure.
+ */
+static int __pkvm_create_hyp_vm(struct kvm *host_kvm)
+{
+       size_t pgd_sz, hyp_vm_sz, hyp_vcpu_sz;
+       struct kvm_vcpu *host_vcpu;
+       pkvm_handle_t handle;
+       void *pgd, *hyp_vm;
+       unsigned long idx;
+       int ret;
+
+       if (host_kvm->created_vcpus < 1)
+               return -EINVAL;
+
+       pgd_sz = kvm_pgtable_stage2_pgd_size(host_kvm->arch.vtcr);
+
+       /*
+        * The PGD pages will be reclaimed using a hyp_memcache which implies
+        * page granularity. So, use alloc_pages_exact() to get individual
+        * refcounts.
+        */
+       pgd = alloc_pages_exact(pgd_sz, GFP_KERNEL_ACCOUNT);
+       if (!pgd)
+               return -ENOMEM;
+
+       /* Allocate memory to donate to hyp for vm and vcpu pointers. */
+       hyp_vm_sz = PAGE_ALIGN(size_add(PKVM_HYP_VM_SIZE,
+                                       size_mul(sizeof(void *),
+                                                host_kvm->created_vcpus)));
+       hyp_vm = alloc_pages_exact(hyp_vm_sz, GFP_KERNEL_ACCOUNT);
+       if (!hyp_vm) {
+               ret = -ENOMEM;
+               goto free_pgd;
+       }
+
+       /* Donate the VM memory to hyp and let hyp initialize it. */
+       ret = kvm_call_hyp_nvhe(__pkvm_init_vm, host_kvm, hyp_vm, pgd);
+       if (ret < 0)
+               goto free_vm;
+
+       handle = ret;
+
+       host_kvm->arch.pkvm.handle = handle;
+       host_kvm->arch.pkvm.hyp_donations.pgd = pgd;
+       host_kvm->arch.pkvm.hyp_donations.vm = hyp_vm;
+
+       /* Donate memory for the vcpus at hyp and initialize it. */
+       hyp_vcpu_sz = PAGE_ALIGN(PKVM_HYP_VCPU_SIZE);
+       kvm_for_each_vcpu(idx, host_vcpu, host_kvm) {
+               void *hyp_vcpu;
+
+               /* Indexing of the vcpus to be sequential starting at 0. */
+               if (WARN_ON(host_vcpu->vcpu_idx != idx)) {
+                       ret = -EINVAL;
+                       goto destroy_vm;
+               }
+
+               hyp_vcpu = alloc_pages_exact(hyp_vcpu_sz, GFP_KERNEL_ACCOUNT);
+               if (!hyp_vcpu) {
+                       ret = -ENOMEM;
+                       goto destroy_vm;
+               }
+
+               host_kvm->arch.pkvm.hyp_donations.vcpus[idx] = hyp_vcpu;
+
+               ret = kvm_call_hyp_nvhe(__pkvm_init_vcpu, handle, host_vcpu,
+                                       hyp_vcpu);
+               if (ret)
+                       goto destroy_vm;
+       }
+
+       return 0;
+
+destroy_vm:
+       pkvm_destroy_hyp_vm(host_kvm);
+       return ret;
+free_vm:
+       free_pages_exact(hyp_vm, hyp_vm_sz);
+free_pgd:
+       free_pages_exact(pgd, pgd_sz);
+       return ret;
+}
+
+int pkvm_create_hyp_vm(struct kvm *host_kvm)
+{
+       int ret = 0;
+
+       mutex_lock(&host_kvm->lock);
+       if (!host_kvm->arch.pkvm.handle)
+               ret = __pkvm_create_hyp_vm(host_kvm);
+       mutex_unlock(&host_kvm->lock);
+
+       return ret;
+}
+
+void pkvm_destroy_hyp_vm(struct kvm *host_kvm)
+{
+       unsigned long idx, nr_vcpus = host_kvm->created_vcpus;
+       size_t pgd_sz, hyp_vm_sz;
+
+       if (host_kvm->arch.pkvm.handle)
+               WARN_ON(kvm_call_hyp_nvhe(__pkvm_teardown_vm,
+                                         host_kvm->arch.pkvm.handle));
+
+       host_kvm->arch.pkvm.handle = 0;
+
+       for (idx = 0; idx < nr_vcpus; ++idx) {
+               void *hyp_vcpu = host_kvm->arch.pkvm.hyp_donations.vcpus[idx];
+
+               if (!hyp_vcpu)
+                       break;
+
+               free_pages_exact(hyp_vcpu, PAGE_ALIGN(PKVM_HYP_VCPU_SIZE));
+       }
+
+       hyp_vm_sz = PAGE_ALIGN(size_add(PKVM_HYP_VM_SIZE,
+                                       size_mul(sizeof(void *), nr_vcpus)));
+       pgd_sz = kvm_pgtable_stage2_pgd_size(host_kvm->arch.vtcr);
+
+       free_pages_exact(host_kvm->arch.pkvm.hyp_donations.vm, hyp_vm_sz);
+       free_pages_exact(host_kvm->arch.pkvm.hyp_donations.pgd, pgd_sz);
+}
+
+int pkvm_init_host_vm(struct kvm *host_kvm)
+{
+       mutex_init(&host_kvm->lock);
+       return 0;
+}