xref: /openbmc/linux/arch/arm64/kvm/pkvm.c (revision ffcdf473)
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 
82 	/*
83 	 * Try to allocate a PMD-aligned region to reduce TLB pressure once
84 	 * this is unmapped from the host stage-2, and fallback to PAGE_SIZE.
85 	 */
86 	hyp_mem_size = hyp_mem_pages << PAGE_SHIFT;
87 	hyp_mem_base = memblock_phys_alloc(ALIGN(hyp_mem_size, PMD_SIZE),
88 					   PMD_SIZE);
89 	if (!hyp_mem_base)
90 		hyp_mem_base = memblock_phys_alloc(hyp_mem_size, PAGE_SIZE);
91 	else
92 		hyp_mem_size = ALIGN(hyp_mem_size, PMD_SIZE);
93 
94 	if (!hyp_mem_base) {
95 		kvm_err("Failed to reserve hyp memory\n");
96 		return;
97 	}
98 
99 	kvm_info("Reserved %lld MiB at 0x%llx\n", hyp_mem_size >> 20,
100 		 hyp_mem_base);
101 }
102 
103 /*
104  * Allocates and donates memory for hypervisor VM structs at EL2.
105  *
106  * Allocates space for the VM state, which includes the hyp vm as well as
107  * the hyp vcpus.
108  *
109  * Stores an opaque handler in the kvm struct for future reference.
110  *
111  * Return 0 on success, negative error code on failure.
112  */
113 static int __pkvm_create_hyp_vm(struct kvm *host_kvm)
114 {
115 	size_t pgd_sz, hyp_vm_sz, hyp_vcpu_sz;
116 	struct kvm_vcpu *host_vcpu;
117 	pkvm_handle_t handle;
118 	void *pgd, *hyp_vm;
119 	unsigned long idx;
120 	int ret;
121 
122 	if (host_kvm->created_vcpus < 1)
123 		return -EINVAL;
124 
125 	pgd_sz = kvm_pgtable_stage2_pgd_size(host_kvm->arch.vtcr);
126 
127 	/*
128 	 * The PGD pages will be reclaimed using a hyp_memcache which implies
129 	 * page granularity. So, use alloc_pages_exact() to get individual
130 	 * refcounts.
131 	 */
132 	pgd = alloc_pages_exact(pgd_sz, GFP_KERNEL_ACCOUNT);
133 	if (!pgd)
134 		return -ENOMEM;
135 
136 	/* Allocate memory to donate to hyp for vm and vcpu pointers. */
137 	hyp_vm_sz = PAGE_ALIGN(size_add(PKVM_HYP_VM_SIZE,
138 					size_mul(sizeof(void *),
139 						 host_kvm->created_vcpus)));
140 	hyp_vm = alloc_pages_exact(hyp_vm_sz, GFP_KERNEL_ACCOUNT);
141 	if (!hyp_vm) {
142 		ret = -ENOMEM;
143 		goto free_pgd;
144 	}
145 
146 	/* Donate the VM memory to hyp and let hyp initialize it. */
147 	ret = kvm_call_hyp_nvhe(__pkvm_init_vm, host_kvm, hyp_vm, pgd);
148 	if (ret < 0)
149 		goto free_vm;
150 
151 	handle = ret;
152 
153 	host_kvm->arch.pkvm.handle = handle;
154 
155 	/* Donate memory for the vcpus at hyp and initialize it. */
156 	hyp_vcpu_sz = PAGE_ALIGN(PKVM_HYP_VCPU_SIZE);
157 	kvm_for_each_vcpu(idx, host_vcpu, host_kvm) {
158 		void *hyp_vcpu;
159 
160 		/* Indexing of the vcpus to be sequential starting at 0. */
161 		if (WARN_ON(host_vcpu->vcpu_idx != idx)) {
162 			ret = -EINVAL;
163 			goto destroy_vm;
164 		}
165 
166 		hyp_vcpu = alloc_pages_exact(hyp_vcpu_sz, GFP_KERNEL_ACCOUNT);
167 		if (!hyp_vcpu) {
168 			ret = -ENOMEM;
169 			goto destroy_vm;
170 		}
171 
172 		ret = kvm_call_hyp_nvhe(__pkvm_init_vcpu, handle, host_vcpu,
173 					hyp_vcpu);
174 		if (ret) {
175 			free_pages_exact(hyp_vcpu, hyp_vcpu_sz);
176 			goto destroy_vm;
177 		}
178 	}
179 
180 	return 0;
181 
182 destroy_vm:
183 	pkvm_destroy_hyp_vm(host_kvm);
184 	return ret;
185 free_vm:
186 	free_pages_exact(hyp_vm, hyp_vm_sz);
187 free_pgd:
188 	free_pages_exact(pgd, pgd_sz);
189 	return ret;
190 }
191 
192 int pkvm_create_hyp_vm(struct kvm *host_kvm)
193 {
194 	int ret = 0;
195 
196 	mutex_lock(&host_kvm->lock);
197 	if (!host_kvm->arch.pkvm.handle)
198 		ret = __pkvm_create_hyp_vm(host_kvm);
199 	mutex_unlock(&host_kvm->lock);
200 
201 	return ret;
202 }
203 
204 void pkvm_destroy_hyp_vm(struct kvm *host_kvm)
205 {
206 	if (host_kvm->arch.pkvm.handle) {
207 		WARN_ON(kvm_call_hyp_nvhe(__pkvm_teardown_vm,
208 					  host_kvm->arch.pkvm.handle));
209 	}
210 
211 	host_kvm->arch.pkvm.handle = 0;
212 	free_hyp_memcache(&host_kvm->arch.pkvm.teardown_mc);
213 }
214 
215 int pkvm_init_host_vm(struct kvm *host_kvm)
216 {
217 	mutex_init(&host_kvm->lock);
218 	return 0;
219 }
220 
221 static void __init _kvm_host_prot_finalize(void *arg)
222 {
223 	int *err = arg;
224 
225 	if (WARN_ON(kvm_call_hyp_nvhe(__pkvm_prot_finalize)))
226 		WRITE_ONCE(*err, -EINVAL);
227 }
228 
229 static int __init pkvm_drop_host_privileges(void)
230 {
231 	int ret = 0;
232 
233 	/*
234 	 * Flip the static key upfront as that may no longer be possible
235 	 * once the host stage 2 is installed.
236 	 */
237 	static_branch_enable(&kvm_protected_mode_initialized);
238 	on_each_cpu(_kvm_host_prot_finalize, &ret, 1);
239 	return ret;
240 }
241 
242 static int __init finalize_pkvm(void)
243 {
244 	int ret;
245 
246 	if (!is_protected_kvm_enabled())
247 		return 0;
248 
249 	/*
250 	 * Exclude HYP sections from kmemleak so that they don't get peeked
251 	 * at, which would end badly once inaccessible.
252 	 */
253 	kmemleak_free_part(__hyp_bss_start, __hyp_bss_end - __hyp_bss_start);
254 	kmemleak_free_part_phys(hyp_mem_base, hyp_mem_size);
255 
256 	ret = pkvm_drop_host_privileges();
257 	if (ret)
258 		pr_err("Failed to finalize Hyp protection: %d\n", ret);
259 
260 	return ret;
261 }
262 device_initcall_sync(finalize_pkvm);
263