1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Copyright (C) 2020 Google LLC
4 * Author: Quentin Perret <qperret@google.com>
5 */
6
7 #include <linux/kvm_host.h>
8 #include <asm/kvm_hyp.h>
9 #include <asm/kvm_mmu.h>
10 #include <asm/kvm_pgtable.h>
11 #include <asm/kvm_pkvm.h>
12
13 #include <nvhe/early_alloc.h>
14 #include <nvhe/fixed_config.h>
15 #include <nvhe/gfp.h>
16 #include <nvhe/memory.h>
17 #include <nvhe/mem_protect.h>
18 #include <nvhe/mm.h>
19 #include <nvhe/trap_handler.h>
20
21 unsigned long hyp_nr_cpus;
22
23 #define hyp_percpu_size ((unsigned long)__per_cpu_end - \
24 (unsigned long)__per_cpu_start)
25
26 static void *vmemmap_base;
27 static void *hyp_pgt_base;
28 static void *host_s2_pgt_base;
29 static struct kvm_pgtable_mm_ops pkvm_pgtable_mm_ops;
30 static struct hyp_pool hpool;
31
divide_memory_pool(void * virt,unsigned long size)32 static int divide_memory_pool(void *virt, unsigned long size)
33 {
34 unsigned long vstart, vend, nr_pages;
35
36 hyp_early_alloc_init(virt, size);
37
38 hyp_vmemmap_range(__hyp_pa(virt), size, &vstart, &vend);
39 nr_pages = (vend - vstart) >> PAGE_SHIFT;
40 vmemmap_base = hyp_early_alloc_contig(nr_pages);
41 if (!vmemmap_base)
42 return -ENOMEM;
43
44 nr_pages = hyp_s1_pgtable_pages();
45 hyp_pgt_base = hyp_early_alloc_contig(nr_pages);
46 if (!hyp_pgt_base)
47 return -ENOMEM;
48
49 nr_pages = host_s2_pgtable_pages();
50 host_s2_pgt_base = hyp_early_alloc_contig(nr_pages);
51 if (!host_s2_pgt_base)
52 return -ENOMEM;
53
54 return 0;
55 }
56
recreate_hyp_mappings(phys_addr_t phys,unsigned long size,unsigned long * per_cpu_base,u32 hyp_va_bits)57 static int recreate_hyp_mappings(phys_addr_t phys, unsigned long size,
58 unsigned long *per_cpu_base,
59 u32 hyp_va_bits)
60 {
61 void *start, *end, *virt = hyp_phys_to_virt(phys);
62 unsigned long pgt_size = hyp_s1_pgtable_pages() << PAGE_SHIFT;
63 enum kvm_pgtable_prot prot;
64 int ret, i;
65
66 /* Recreate the hyp page-table using the early page allocator */
67 hyp_early_alloc_init(hyp_pgt_base, pgt_size);
68 ret = kvm_pgtable_hyp_init(&pkvm_pgtable, hyp_va_bits,
69 &hyp_early_alloc_mm_ops);
70 if (ret)
71 return ret;
72
73 ret = hyp_create_idmap(hyp_va_bits);
74 if (ret)
75 return ret;
76
77 ret = hyp_map_vectors();
78 if (ret)
79 return ret;
80
81 ret = hyp_back_vmemmap(phys, size, hyp_virt_to_phys(vmemmap_base));
82 if (ret)
83 return ret;
84
85 ret = pkvm_create_mappings(__hyp_text_start, __hyp_text_end, PAGE_HYP_EXEC);
86 if (ret)
87 return ret;
88
89 ret = pkvm_create_mappings(__hyp_rodata_start, __hyp_rodata_end, PAGE_HYP_RO);
90 if (ret)
91 return ret;
92
93 ret = pkvm_create_mappings(__hyp_bss_start, __hyp_bss_end, PAGE_HYP);
94 if (ret)
95 return ret;
96
97 ret = pkvm_create_mappings(virt, virt + size, PAGE_HYP);
98 if (ret)
99 return ret;
100
101 for (i = 0; i < hyp_nr_cpus; i++) {
102 struct kvm_nvhe_init_params *params = per_cpu_ptr(&kvm_init_params, i);
103 unsigned long hyp_addr;
104
105 start = (void *)kern_hyp_va(per_cpu_base[i]);
106 end = start + PAGE_ALIGN(hyp_percpu_size);
107 ret = pkvm_create_mappings(start, end, PAGE_HYP);
108 if (ret)
109 return ret;
110
111 /*
112 * Allocate a contiguous HYP private VA range for the stack
113 * and guard page. The allocation is also aligned based on
114 * the order of its size.
115 */
116 ret = pkvm_alloc_private_va_range(PAGE_SIZE * 2, &hyp_addr);
117 if (ret)
118 return ret;
119
120 /*
121 * Since the stack grows downwards, map the stack to the page
122 * at the higher address and leave the lower guard page
123 * unbacked.
124 *
125 * Any valid stack address now has the PAGE_SHIFT bit as 1
126 * and addresses corresponding to the guard page have the
127 * PAGE_SHIFT bit as 0 - this is used for overflow detection.
128 */
129 hyp_spin_lock(&pkvm_pgd_lock);
130 ret = kvm_pgtable_hyp_map(&pkvm_pgtable, hyp_addr + PAGE_SIZE,
131 PAGE_SIZE, params->stack_pa, PAGE_HYP);
132 hyp_spin_unlock(&pkvm_pgd_lock);
133 if (ret)
134 return ret;
135
136 /* Update stack_hyp_va to end of the stack's private VA range */
137 params->stack_hyp_va = hyp_addr + (2 * PAGE_SIZE);
138 }
139
140 /*
141 * Map the host's .bss and .rodata sections RO in the hypervisor, but
142 * transfer the ownership from the host to the hypervisor itself to
143 * make sure it can't be donated or shared with another entity.
144 *
145 * The ownership transition requires matching changes in the host
146 * stage-2. This will be done later (see finalize_host_mappings()) once
147 * the hyp_vmemmap is addressable.
148 */
149 prot = pkvm_mkstate(PAGE_HYP_RO, PKVM_PAGE_SHARED_OWNED);
150 ret = pkvm_create_mappings(__start_rodata, __end_rodata, prot);
151 if (ret)
152 return ret;
153
154 ret = pkvm_create_mappings(__hyp_bss_end, __bss_stop, prot);
155 if (ret)
156 return ret;
157
158 return 0;
159 }
160
update_nvhe_init_params(void)161 static void update_nvhe_init_params(void)
162 {
163 struct kvm_nvhe_init_params *params;
164 unsigned long i;
165
166 for (i = 0; i < hyp_nr_cpus; i++) {
167 params = per_cpu_ptr(&kvm_init_params, i);
168 params->pgd_pa = __hyp_pa(pkvm_pgtable.pgd);
169 dcache_clean_inval_poc((unsigned long)params,
170 (unsigned long)params + sizeof(*params));
171 }
172 }
173
hyp_zalloc_hyp_page(void * arg)174 static void *hyp_zalloc_hyp_page(void *arg)
175 {
176 return hyp_alloc_pages(&hpool, 0);
177 }
178
hpool_get_page(void * addr)179 static void hpool_get_page(void *addr)
180 {
181 hyp_get_page(&hpool, addr);
182 }
183
hpool_put_page(void * addr)184 static void hpool_put_page(void *addr)
185 {
186 hyp_put_page(&hpool, addr);
187 }
188
finalize_host_mappings_walker(u64 addr,u64 end,u32 level,kvm_pte_t * ptep,enum kvm_pgtable_walk_flags flag,void * const arg)189 static int finalize_host_mappings_walker(u64 addr, u64 end, u32 level,
190 kvm_pte_t *ptep,
191 enum kvm_pgtable_walk_flags flag,
192 void * const arg)
193 {
194 struct kvm_pgtable_mm_ops *mm_ops = arg;
195 enum kvm_pgtable_prot prot;
196 enum pkvm_page_state state;
197 kvm_pte_t pte = *ptep;
198 phys_addr_t phys;
199
200 if (!kvm_pte_valid(pte))
201 return 0;
202
203 /*
204 * Fix-up the refcount for the page-table pages as the early allocator
205 * was unable to access the hyp_vmemmap and so the buddy allocator has
206 * initialised the refcount to '1'.
207 */
208 mm_ops->get_page(ptep);
209 if (flag != KVM_PGTABLE_WALK_LEAF)
210 return 0;
211
212 if (level != (KVM_PGTABLE_MAX_LEVELS - 1))
213 return -EINVAL;
214
215 phys = kvm_pte_to_phys(pte);
216 if (!addr_is_memory(phys))
217 return -EINVAL;
218
219 /*
220 * Adjust the host stage-2 mappings to match the ownership attributes
221 * configured in the hypervisor stage-1.
222 */
223 state = pkvm_getstate(kvm_pgtable_hyp_pte_prot(pte));
224 switch (state) {
225 case PKVM_PAGE_OWNED:
226 return host_stage2_set_owner_locked(phys, PAGE_SIZE, pkvm_hyp_id);
227 case PKVM_PAGE_SHARED_OWNED:
228 prot = pkvm_mkstate(PKVM_HOST_MEM_PROT, PKVM_PAGE_SHARED_BORROWED);
229 break;
230 case PKVM_PAGE_SHARED_BORROWED:
231 prot = pkvm_mkstate(PKVM_HOST_MEM_PROT, PKVM_PAGE_SHARED_OWNED);
232 break;
233 default:
234 return -EINVAL;
235 }
236
237 return host_stage2_idmap_locked(phys, PAGE_SIZE, prot);
238 }
239
finalize_host_mappings(void)240 static int finalize_host_mappings(void)
241 {
242 struct kvm_pgtable_walker walker = {
243 .cb = finalize_host_mappings_walker,
244 .flags = KVM_PGTABLE_WALK_LEAF | KVM_PGTABLE_WALK_TABLE_POST,
245 .arg = pkvm_pgtable.mm_ops,
246 };
247 int i, ret;
248
249 for (i = 0; i < hyp_memblock_nr; i++) {
250 struct memblock_region *reg = &hyp_memory[i];
251 u64 start = (u64)hyp_phys_to_virt(reg->base);
252
253 ret = kvm_pgtable_walk(&pkvm_pgtable, start, reg->size, &walker);
254 if (ret)
255 return ret;
256 }
257
258 return 0;
259 }
260
__pkvm_init_finalise(void)261 void __noreturn __pkvm_init_finalise(void)
262 {
263 struct kvm_host_data *host_data = this_cpu_ptr(&kvm_host_data);
264 struct kvm_cpu_context *host_ctxt = &host_data->host_ctxt;
265 unsigned long nr_pages, reserved_pages, pfn;
266 int ret;
267
268 /* Now that the vmemmap is backed, install the full-fledged allocator */
269 pfn = hyp_virt_to_pfn(hyp_pgt_base);
270 nr_pages = hyp_s1_pgtable_pages();
271 reserved_pages = hyp_early_alloc_nr_used_pages();
272 ret = hyp_pool_init(&hpool, pfn, nr_pages, reserved_pages);
273 if (ret)
274 goto out;
275
276 ret = kvm_host_prepare_stage2(host_s2_pgt_base);
277 if (ret)
278 goto out;
279
280 pkvm_pgtable_mm_ops = (struct kvm_pgtable_mm_ops) {
281 .zalloc_page = hyp_zalloc_hyp_page,
282 .phys_to_virt = hyp_phys_to_virt,
283 .virt_to_phys = hyp_virt_to_phys,
284 .get_page = hpool_get_page,
285 .put_page = hpool_put_page,
286 .page_count = hyp_page_count,
287 };
288 pkvm_pgtable.mm_ops = &pkvm_pgtable_mm_ops;
289
290 ret = finalize_host_mappings();
291 if (ret)
292 goto out;
293
294 out:
295 /*
296 * We tail-called to here from handle___pkvm_init() and will not return,
297 * so make sure to propagate the return value to the host.
298 */
299 cpu_reg(host_ctxt, 1) = ret;
300
301 __host_enter(host_ctxt);
302 }
303
__pkvm_init(phys_addr_t phys,unsigned long size,unsigned long nr_cpus,unsigned long * per_cpu_base,u32 hyp_va_bits)304 int __pkvm_init(phys_addr_t phys, unsigned long size, unsigned long nr_cpus,
305 unsigned long *per_cpu_base, u32 hyp_va_bits)
306 {
307 struct kvm_nvhe_init_params *params;
308 void *virt = hyp_phys_to_virt(phys);
309 void (*fn)(phys_addr_t params_pa, void *finalize_fn_va);
310 int ret;
311
312 BUG_ON(kvm_check_pvm_sysreg_table());
313
314 if (!PAGE_ALIGNED(phys) || !PAGE_ALIGNED(size))
315 return -EINVAL;
316
317 hyp_spin_lock_init(&pkvm_pgd_lock);
318 hyp_nr_cpus = nr_cpus;
319
320 ret = divide_memory_pool(virt, size);
321 if (ret)
322 return ret;
323
324 ret = recreate_hyp_mappings(phys, size, per_cpu_base, hyp_va_bits);
325 if (ret)
326 return ret;
327
328 update_nvhe_init_params();
329
330 /* Jump in the idmap page to switch to the new page-tables */
331 params = this_cpu_ptr(&kvm_init_params);
332 fn = (typeof(fn))__hyp_pa(__pkvm_init_switch_pgd);
333 fn(__hyp_pa(params), __pkvm_init_finalise);
334
335 unreachable();
336 }
337