xref: /openbmc/linux/arch/arm64/kvm/pkvm.c (revision 078b39c9)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2020 - Google LLC
4  * Author: Quentin Perret <qperret@google.com>
5  */
6 
7 #include <linux/init.h>
8 #include <linux/kmemleak.h>
9 #include <linux/kvm_host.h>
10 #include <linux/memblock.h>
11 #include <linux/mutex.h>
12 #include <linux/sort.h>
13 
14 #include <asm/kvm_pkvm.h>
15 
16 #include "hyp_constants.h"
17 
18 DEFINE_STATIC_KEY_FALSE(kvm_protected_mode_initialized);
19 
20 static struct memblock_region *hyp_memory = kvm_nvhe_sym(hyp_memory);
21 static unsigned int *hyp_memblock_nr_ptr = &kvm_nvhe_sym(hyp_memblock_nr);
22 
23 phys_addr_t hyp_mem_base;
24 phys_addr_t hyp_mem_size;
25 
26 static int cmp_hyp_memblock(const void *p1, const void *p2)
27 {
28 	const struct memblock_region *r1 = p1;
29 	const struct memblock_region *r2 = p2;
30 
31 	return r1->base < r2->base ? -1 : (r1->base > r2->base);
32 }
33 
34 static void __init sort_memblock_regions(void)
35 {
36 	sort(hyp_memory,
37 	     *hyp_memblock_nr_ptr,
38 	     sizeof(struct memblock_region),
39 	     cmp_hyp_memblock,
40 	     NULL);
41 }
42 
43 static int __init register_memblock_regions(void)
44 {
45 	struct memblock_region *reg;
46 
47 	for_each_mem_region(reg) {
48 		if (*hyp_memblock_nr_ptr >= HYP_MEMBLOCK_REGIONS)
49 			return -ENOMEM;
50 
51 		hyp_memory[*hyp_memblock_nr_ptr] = *reg;
52 		(*hyp_memblock_nr_ptr)++;
53 	}
54 	sort_memblock_regions();
55 
56 	return 0;
57 }
58 
59 void __init kvm_hyp_reserve(void)
60 {
61 	u64 hyp_mem_pages = 0;
62 	int ret;
63 
64 	if (!is_hyp_mode_available() || is_kernel_in_hyp_mode())
65 		return;
66 
67 	if (kvm_get_mode() != KVM_MODE_PROTECTED)
68 		return;
69 
70 	ret = register_memblock_regions();
71 	if (ret) {
72 		*hyp_memblock_nr_ptr = 0;
73 		kvm_err("Failed to register hyp memblocks: %d\n", ret);
74 		return;
75 	}
76 
77 	hyp_mem_pages += hyp_s1_pgtable_pages();
78 	hyp_mem_pages += host_s2_pgtable_pages();
79 	hyp_mem_pages += hyp_vm_table_pages();
80 	hyp_mem_pages += hyp_vmemmap_pages(STRUCT_HYP_PAGE_SIZE);
81 	hyp_mem_pages += hyp_ffa_proxy_pages();
82 
83 	/*
84 	 * Try to allocate a PMD-aligned region to reduce TLB pressure once
85 	 * this is unmapped from the host stage-2, and fallback to PAGE_SIZE.
86 	 */
87 	hyp_mem_size = hyp_mem_pages << PAGE_SHIFT;
88 	hyp_mem_base = memblock_phys_alloc(ALIGN(hyp_mem_size, PMD_SIZE),
89 					   PMD_SIZE);
90 	if (!hyp_mem_base)
91 		hyp_mem_base = memblock_phys_alloc(hyp_mem_size, PAGE_SIZE);
92 	else
93 		hyp_mem_size = ALIGN(hyp_mem_size, PMD_SIZE);
94 
95 	if (!hyp_mem_base) {
96 		kvm_err("Failed to reserve hyp memory\n");
97 		return;
98 	}
99 
100 	kvm_info("Reserved %lld MiB at 0x%llx\n", hyp_mem_size >> 20,
101 		 hyp_mem_base);
102 }
103 
104 /*
105  * Allocates and donates memory for hypervisor VM structs at EL2.
106  *
107  * Allocates space for the VM state, which includes the hyp vm as well as
108  * the hyp vcpus.
109  *
110  * Stores an opaque handler in the kvm struct for future reference.
111  *
112  * Return 0 on success, negative error code on failure.
113  */
114 static int __pkvm_create_hyp_vm(struct kvm *host_kvm)
115 {
116 	size_t pgd_sz, hyp_vm_sz, hyp_vcpu_sz;
117 	struct kvm_vcpu *host_vcpu;
118 	pkvm_handle_t handle;
119 	void *pgd, *hyp_vm;
120 	unsigned long idx;
121 	int ret;
122 
123 	if (host_kvm->created_vcpus < 1)
124 		return -EINVAL;
125 
126 	pgd_sz = kvm_pgtable_stage2_pgd_size(host_kvm->arch.vtcr);
127 
128 	/*
129 	 * The PGD pages will be reclaimed using a hyp_memcache which implies
130 	 * page granularity. So, use alloc_pages_exact() to get individual
131 	 * refcounts.
132 	 */
133 	pgd = alloc_pages_exact(pgd_sz, GFP_KERNEL_ACCOUNT);
134 	if (!pgd)
135 		return -ENOMEM;
136 
137 	/* Allocate memory to donate to hyp for vm and vcpu pointers. */
138 	hyp_vm_sz = PAGE_ALIGN(size_add(PKVM_HYP_VM_SIZE,
139 					size_mul(sizeof(void *),
140 						 host_kvm->created_vcpus)));
141 	hyp_vm = alloc_pages_exact(hyp_vm_sz, GFP_KERNEL_ACCOUNT);
142 	if (!hyp_vm) {
143 		ret = -ENOMEM;
144 		goto free_pgd;
145 	}
146 
147 	/* Donate the VM memory to hyp and let hyp initialize it. */
148 	ret = kvm_call_hyp_nvhe(__pkvm_init_vm, host_kvm, hyp_vm, pgd);
149 	if (ret < 0)
150 		goto free_vm;
151 
152 	handle = ret;
153 
154 	host_kvm->arch.pkvm.handle = handle;
155 
156 	/* Donate memory for the vcpus at hyp and initialize it. */
157 	hyp_vcpu_sz = PAGE_ALIGN(PKVM_HYP_VCPU_SIZE);
158 	kvm_for_each_vcpu(idx, host_vcpu, host_kvm) {
159 		void *hyp_vcpu;
160 
161 		/* Indexing of the vcpus to be sequential starting at 0. */
162 		if (WARN_ON(host_vcpu->vcpu_idx != idx)) {
163 			ret = -EINVAL;
164 			goto destroy_vm;
165 		}
166 
167 		hyp_vcpu = alloc_pages_exact(hyp_vcpu_sz, GFP_KERNEL_ACCOUNT);
168 		if (!hyp_vcpu) {
169 			ret = -ENOMEM;
170 			goto destroy_vm;
171 		}
172 
173 		ret = kvm_call_hyp_nvhe(__pkvm_init_vcpu, handle, host_vcpu,
174 					hyp_vcpu);
175 		if (ret) {
176 			free_pages_exact(hyp_vcpu, hyp_vcpu_sz);
177 			goto destroy_vm;
178 		}
179 	}
180 
181 	return 0;
182 
183 destroy_vm:
184 	pkvm_destroy_hyp_vm(host_kvm);
185 	return ret;
186 free_vm:
187 	free_pages_exact(hyp_vm, hyp_vm_sz);
188 free_pgd:
189 	free_pages_exact(pgd, pgd_sz);
190 	return ret;
191 }
192 
193 int pkvm_create_hyp_vm(struct kvm *host_kvm)
194 {
195 	int ret = 0;
196 
197 	mutex_lock(&host_kvm->lock);
198 	if (!host_kvm->arch.pkvm.handle)
199 		ret = __pkvm_create_hyp_vm(host_kvm);
200 	mutex_unlock(&host_kvm->lock);
201 
202 	return ret;
203 }
204 
205 void pkvm_destroy_hyp_vm(struct kvm *host_kvm)
206 {
207 	if (host_kvm->arch.pkvm.handle) {
208 		WARN_ON(kvm_call_hyp_nvhe(__pkvm_teardown_vm,
209 					  host_kvm->arch.pkvm.handle));
210 	}
211 
212 	host_kvm->arch.pkvm.handle = 0;
213 	free_hyp_memcache(&host_kvm->arch.pkvm.teardown_mc);
214 }
215 
216 int pkvm_init_host_vm(struct kvm *host_kvm)
217 {
218 	mutex_init(&host_kvm->lock);
219 	return 0;
220 }
221 
222 static void __init _kvm_host_prot_finalize(void *arg)
223 {
224 	int *err = arg;
225 
226 	if (WARN_ON(kvm_call_hyp_nvhe(__pkvm_prot_finalize)))
227 		WRITE_ONCE(*err, -EINVAL);
228 }
229 
230 static int __init pkvm_drop_host_privileges(void)
231 {
232 	int ret = 0;
233 
234 	/*
235 	 * Flip the static key upfront as that may no longer be possible
236 	 * once the host stage 2 is installed.
237 	 */
238 	static_branch_enable(&kvm_protected_mode_initialized);
239 	on_each_cpu(_kvm_host_prot_finalize, &ret, 1);
240 	return ret;
241 }
242 
243 static int __init finalize_pkvm(void)
244 {
245 	int ret;
246 
247 	if (!is_protected_kvm_enabled())
248 		return 0;
249 
250 	/*
251 	 * Exclude HYP sections from kmemleak so that they don't get peeked
252 	 * at, which would end badly once inaccessible.
253 	 */
254 	kmemleak_free_part(__hyp_bss_start, __hyp_bss_end - __hyp_bss_start);
255 	kmemleak_free_part_phys(hyp_mem_base, hyp_mem_size);
256 
257 	ret = pkvm_drop_host_privileges();
258 	if (ret)
259 		pr_err("Failed to finalize Hyp protection: %d\n", ret);
260 
261 	return ret;
262 }
263 device_initcall_sync(finalize_pkvm);
264