KVM: arm64: Return guest memory from EL2 via dedicated teardown memcache

Rather than relying on the host to free the previously-donated pKVM
hypervisor VM pages explicitly on teardown, introduce a dedicated
teardown memcache which allows the host to reclaim guest memory
resources without having to keep track of all of the allocations made by
the pKVM hypervisor at EL2.

Tested-by: Vincent Donnefort <[e-mail]>
Co-developed-by: Fuad Tabba <[e-mail]>
Signed-off-by: Fuad Tabba <[e-mail]>
Signed-off-by: Quentin Perret <[e-mail]>
Signed-off-by: Will Deacon <[e-mail]>
[maz: dropped __maybe_unused from unmap_donated_memory_noclear()]
Signed-off-by: Marc Zyngier <[e-mail]>
Link: https://lore.kernel.org/r/[e-mail]

arch/arm64/include/asm/kvm_host.h

···
 
 struct kvm_protected_vm {
 	pkvm_handle_t handle;
-
-	struct {
-		void *pgd;
-		void *vm;
-		void *vcpus[KVM_MAX_VCPUS];
-	} hyp_donations;
+	struct kvm_hyp_memcache teardown_mc;
 };
 
 struct kvm_arch {

arch/arm64/kvm/hyp/include/nvhe/mem_protect.h

···
 
 int hyp_pin_shared_mem(void *from, void *to);
 void hyp_unpin_shared_mem(void *from, void *to);
-void reclaim_guest_pages(struct pkvm_hyp_vm *vm);
+void reclaim_guest_pages(struct pkvm_hyp_vm *vm, struct kvm_hyp_memcache *mc);
 int refill_memcache(struct kvm_hyp_memcache *mc, unsigned long min_pages,
 		    struct kvm_hyp_memcache *host_mc);
 

arch/arm64/kvm/hyp/nvhe/mem_protect.c

···
 	return 0;
 }
 
-void reclaim_guest_pages(struct pkvm_hyp_vm *vm)
+void reclaim_guest_pages(struct pkvm_hyp_vm *vm, struct kvm_hyp_memcache *mc)
 {
-	void *pgd = vm->pgt.pgd;
-	unsigned long nr_pages;
+	void *addr;
 
-	nr_pages = kvm_pgtable_stage2_pgd_size(vm->kvm.arch.vtcr) >> PAGE_SHIFT;
-
+	/* Dump all pgtable pages in the hyp_pool */
 	guest_lock_component(vm);
 	kvm_pgtable_stage2_destroy(&vm->pgt);
 	vm->kvm.arch.mmu.pgd_phys = 0ULL;
 	guest_unlock_component(vm);
 
-	WARN_ON(__pkvm_hyp_donate_host(hyp_virt_to_pfn(pgd), nr_pages));
+	/* Drain the hyp_pool into the memcache */
+	addr = hyp_alloc_pages(&vm->pool, 0);
+	while (addr) {
+		memset(hyp_virt_to_page(addr), 0, sizeof(struct hyp_page));
+		push_hyp_memcache(mc, addr, hyp_virt_to_phys);
+		WARN_ON(__pkvm_hyp_donate_host(hyp_virt_to_pfn(addr), 1));
+		addr = hyp_alloc_pages(&vm->pool, 0);
+	}
 }
 
 int __pkvm_prot_finalize(void)

arch/arm64/kvm/hyp/nvhe/pkvm.c

···
 	__unmap_donated_memory(va, size);
 }
 
-static void __maybe_unused unmap_donated_memory_noclear(void *va, size_t size)
+static void unmap_donated_memory_noclear(void *va, size_t size)
 {
 	if (!va)
 		return;
···
 	return ret;
 }
 
+static void
+teardown_donated_memory(struct kvm_hyp_memcache *mc, void *addr, size_t size)
+{
+	size = PAGE_ALIGN(size);
+	memset(addr, 0, size);
+
+	for (void *start = addr; start < addr + size; start += PAGE_SIZE)
+		push_hyp_memcache(mc, start, hyp_virt_to_phys);
+
+	unmap_donated_memory_noclear(addr, size);
+}
+
 int __pkvm_teardown_vm(pkvm_handle_t handle)
 {
+	struct kvm_hyp_memcache *mc;
 	struct pkvm_hyp_vm *hyp_vm;
 	struct kvm *host_kvm;
 	unsigned int idx;
···
 		goto err_unlock;
 	}
 
+	host_kvm = hyp_vm->host_kvm;
+
 	/* Ensure the VMID is clean before it can be reallocated */
 	__kvm_tlb_flush_vmid(&hyp_vm->kvm.arch.mmu);
 	remove_vm_table_entry(handle);
 	hyp_spin_unlock(&vm_table_lock);
 
 	/* Reclaim guest pages (including page-table pages) */
-	reclaim_guest_pages(hyp_vm);
+	mc = &host_kvm->arch.pkvm.teardown_mc;
+	reclaim_guest_pages(hyp_vm, mc);
 	unpin_host_vcpus(hyp_vm->vcpus, hyp_vm->nr_vcpus);
 
-	/* Return the metadata pages to the host */
+	/* Push the metadata pages to the teardown memcache */
 	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));
+		teardown_donated_memory(mc, 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);
+	teardown_donated_memory(mc, hyp_vm, vm_size);
 	hyp_unpin_shared_mem(host_kvm, host_kvm + 1);
 	return 0;
 

arch/arm64/kvm/pkvm.c

···
 	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);
···
 			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)
+		if (ret) {
+			free_pages_exact(hyp_vcpu, hyp_vcpu_sz);
 			goto destroy_vm;
+		}
 	}
 
 	return 0;
···
 
 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)
+	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);
+	host_kvm->arch.pkvm.handle = 0;
+	free_hyp_memcache(&host_kvm->arch.pkvm.teardown_mc);
 }
 
 int pkvm_init_host_vm(struct kvm *host_kvm)