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/ffa.h>
15 #include <nvhe/gfp.h>
16 #include <nvhe/iommu.h>
17 #include <nvhe/memory.h>
18 #include <nvhe/mem_protect.h>
19 #include <nvhe/mm.h>
20 #include <nvhe/pkvm.h>
21 #include <nvhe/trap_handler.h>
22
23 unsigned long hyp_nr_cpus;
24
25 phys_addr_t pvmfw_base;
26 phys_addr_t pvmfw_size;
27
28 #define hyp_percpu_size ((unsigned long)__per_cpu_end - \
29 (unsigned long)__per_cpu_start)
30
31 static void *vmemmap_base;
32 static void *shadow_table_base;
33 static void *hyp_pgt_base;
34 static void *host_s2_pgt_base;
35 static void *ffa_proxy_pages;
36 static struct kvm_pgtable_mm_ops pkvm_pgtable_mm_ops;
37 static struct hyp_pool hpool;
38
divide_memory_pool(void * virt,unsigned long size)39 static int divide_memory_pool(void *virt, unsigned long size)
40 {
41 unsigned long nr_pages;
42
43 hyp_early_alloc_init(virt, size);
44
45 nr_pages = hyp_vmemmap_pages(sizeof(struct hyp_page));
46 vmemmap_base = hyp_early_alloc_contig(nr_pages);
47 if (!vmemmap_base)
48 return -ENOMEM;
49
50 nr_pages = hyp_shadow_table_pages(sizeof(struct kvm_shadow_vm));
51 shadow_table_base = hyp_early_alloc_contig(nr_pages);
52 if (!shadow_table_base)
53 return -ENOMEM;
54
55 nr_pages = hyp_s1_pgtable_pages();
56 hyp_pgt_base = hyp_early_alloc_contig(nr_pages);
57 if (!hyp_pgt_base)
58 return -ENOMEM;
59
60 nr_pages = host_s2_pgtable_pages();
61 host_s2_pgt_base = hyp_early_alloc_contig(nr_pages);
62 if (!host_s2_pgt_base)
63 return -ENOMEM;
64
65 nr_pages = hyp_ffa_proxy_pages();
66 ffa_proxy_pages = hyp_early_alloc_contig(nr_pages);
67 if (!ffa_proxy_pages)
68 return -ENOMEM;
69
70 return 0;
71 }
72
recreate_hyp_mappings(phys_addr_t phys,unsigned long size,unsigned long * per_cpu_base,u32 hyp_va_bits)73 static int recreate_hyp_mappings(phys_addr_t phys, unsigned long size,
74 unsigned long *per_cpu_base,
75 u32 hyp_va_bits)
76 {
77 void *start, *end, *virt = hyp_phys_to_virt(phys);
78 unsigned long pgt_size = hyp_s1_pgtable_pages() << PAGE_SHIFT;
79 enum kvm_pgtable_prot prot;
80 int ret, i;
81
82 /* Recreate the hyp page-table using the early page allocator */
83 hyp_early_alloc_init(hyp_pgt_base, pgt_size);
84 ret = kvm_pgtable_hyp_init(&pkvm_pgtable, hyp_va_bits,
85 &hyp_early_alloc_mm_ops);
86 if (ret)
87 return ret;
88
89 ret = hyp_create_idmap(hyp_va_bits);
90 if (ret)
91 return ret;
92
93 ret = hyp_map_vectors();
94 if (ret)
95 return ret;
96
97 ret = hyp_back_vmemmap(hyp_virt_to_phys(vmemmap_base));
98 if (ret)
99 return ret;
100
101 ret = pkvm_create_mappings(__hyp_text_start, __hyp_text_end, PAGE_HYP_EXEC);
102 if (ret)
103 return ret;
104
105 ret = pkvm_create_mappings(__hyp_data_start, __hyp_data_end, PAGE_HYP);
106 if (ret)
107 return ret;
108
109 ret = pkvm_create_mappings(__hyp_rodata_start, __hyp_rodata_end, PAGE_HYP_RO);
110 if (ret)
111 return ret;
112
113 ret = pkvm_create_mappings(__hyp_bss_start, __hyp_bss_end, PAGE_HYP);
114 if (ret)
115 return ret;
116
117 ret = pkvm_create_mappings(virt, virt + size, PAGE_HYP);
118 if (ret)
119 return ret;
120
121 for (i = 0; i < hyp_nr_cpus; i++) {
122 start = (void *)kern_hyp_va(per_cpu_base[i]);
123 end = start + PAGE_ALIGN(hyp_percpu_size);
124 ret = pkvm_create_mappings(start, end, PAGE_HYP);
125 if (ret)
126 return ret;
127
128 end = (void *)per_cpu_ptr(&kvm_init_params, i)->stack_hyp_va;
129 start = end - PAGE_SIZE;
130 ret = pkvm_create_mappings(start, end, PAGE_HYP);
131 if (ret)
132 return ret;
133 }
134
135 /*
136 * Map the host sections RO in the hypervisor, but transfer the
137 * ownership from the host to the hypervisor itself to make sure they
138 * can't be donated or shared with another entity.
139 *
140 * The ownership transition requires matching changes in the host
141 * stage-2. This will be done later (see finalize_host_mappings()) once
142 * the hyp_vmemmap is addressable.
143 */
144 prot = pkvm_mkstate(PAGE_HYP_RO, PKVM_PAGE_SHARED_OWNED);
145 ret = pkvm_create_mappings(&kvm_vgic_global_state, &kvm_vgic_global_state + 1, prot);
146 if (ret)
147 return ret;
148
149 start = hyp_phys_to_virt(pvmfw_base);
150 end = start + pvmfw_size;
151 prot = pkvm_mkstate(PAGE_HYP, PKVM_PAGE_OWNED);
152 ret = pkvm_create_mappings(start, end, prot);
153 if (ret)
154 return ret;
155
156 return 0;
157 }
158
update_nvhe_init_params(void)159 static void update_nvhe_init_params(void)
160 {
161 struct kvm_nvhe_init_params *params;
162 unsigned long i;
163
164 for (i = 0; i < hyp_nr_cpus; i++) {
165 params = per_cpu_ptr(&kvm_init_params, i);
166 params->pgd_pa = __hyp_pa(pkvm_pgtable.pgd);
167 dcache_clean_inval_poc((unsigned long)params,
168 (unsigned long)params + sizeof(*params));
169 }
170 }
171
hyp_zalloc_hyp_page(void * arg)172 static void *hyp_zalloc_hyp_page(void *arg)
173 {
174 return hyp_alloc_pages(&hpool, 0);
175 }
176
hpool_get_page(void * addr)177 static void hpool_get_page(void *addr)
178 {
179 hyp_get_page(&hpool, addr);
180 }
181
hpool_put_page(void * addr)182 static void hpool_put_page(void *addr)
183 {
184 hyp_put_page(&hpool, addr);
185 }
186
fix_host_ownership_walker(u64 addr,u64 end,u32 level,kvm_pte_t * ptep,enum kvm_pgtable_walk_flags flag,void * const arg)187 static int fix_host_ownership_walker(u64 addr, u64 end, u32 level,
188 kvm_pte_t *ptep,
189 enum kvm_pgtable_walk_flags flag,
190 void * const arg)
191 {
192 enum kvm_pgtable_prot prot;
193 enum pkvm_page_state state;
194 kvm_pte_t pte = *ptep;
195 phys_addr_t phys;
196
197 if (!kvm_pte_valid(pte))
198 return 0;
199
200 if (level != (KVM_PGTABLE_MAX_LEVELS - 1))
201 return -EINVAL;
202
203 phys = kvm_pte_to_phys(pte);
204 if (!addr_is_memory(phys))
205 return -EINVAL;
206
207 /*
208 * Adjust the host stage-2 mappings to match the ownership attributes
209 * configured in the hypervisor stage-1.
210 */
211 state = pkvm_getstate(kvm_pgtable_hyp_pte_prot(pte));
212 switch (state) {
213 case PKVM_PAGE_OWNED:
214 return host_stage2_set_owner_locked(phys, PAGE_SIZE, pkvm_hyp_id);
215 case PKVM_PAGE_SHARED_OWNED:
216 prot = pkvm_mkstate(PKVM_HOST_MEM_PROT, PKVM_PAGE_SHARED_BORROWED);
217 break;
218 case PKVM_PAGE_SHARED_BORROWED:
219 prot = pkvm_mkstate(PKVM_HOST_MEM_PROT, PKVM_PAGE_SHARED_OWNED);
220 break;
221 default:
222 return -EINVAL;
223 }
224
225 return host_stage2_idmap_locked(phys, PAGE_SIZE, prot, false);
226 }
227
fix_hyp_pgtable_refcnt_walker(u64 addr,u64 end,u32 level,kvm_pte_t * ptep,enum kvm_pgtable_walk_flags flag,void * const arg)228 static int fix_hyp_pgtable_refcnt_walker(u64 addr, u64 end, u32 level,
229 kvm_pte_t *ptep,
230 enum kvm_pgtable_walk_flags flag,
231 void * const arg)
232 {
233 struct kvm_pgtable_mm_ops *mm_ops = arg;
234 kvm_pte_t pte = *ptep;
235
236 /*
237 * Fix-up the refcount for the page-table pages as the early allocator
238 * was unable to access the hyp_vmemmap and so the buddy allocator has
239 * initialised the refcount to '1'.
240 */
241 if (kvm_pte_valid(pte))
242 mm_ops->get_page(ptep);
243
244 return 0;
245 }
246
fix_host_ownership(void)247 static int fix_host_ownership(void)
248 {
249 struct kvm_pgtable_walker walker = {
250 .cb = fix_host_ownership_walker,
251 .flags = KVM_PGTABLE_WALK_LEAF,
252 };
253 int i, ret;
254
255 for (i = 0; i < hyp_memblock_nr; i++) {
256 struct memblock_region *reg = &hyp_memory[i];
257 u64 start = (u64)hyp_phys_to_virt(reg->base);
258
259 ret = kvm_pgtable_walk(&pkvm_pgtable, start, reg->size, &walker);
260 if (ret)
261 return ret;
262 }
263
264 return 0;
265 }
266
fix_hyp_pgtable_refcnt(void)267 static int fix_hyp_pgtable_refcnt(void)
268 {
269 struct kvm_pgtable_walker walker = {
270 .cb = fix_hyp_pgtable_refcnt_walker,
271 .flags = KVM_PGTABLE_WALK_LEAF | KVM_PGTABLE_WALK_TABLE_POST,
272 .arg = pkvm_pgtable.mm_ops,
273 };
274
275 return kvm_pgtable_walk(&pkvm_pgtable, 0, BIT(pkvm_pgtable.ia_bits),
276 &walker);
277 }
278
__pkvm_init_finalise(void)279 void __noreturn __pkvm_init_finalise(void)
280 {
281 struct kvm_host_data *host_data = this_cpu_ptr(&kvm_host_data);
282 struct kvm_cpu_context *host_ctxt = &host_data->host_ctxt;
283 unsigned long nr_pages, reserved_pages, pfn;
284 int ret;
285
286 /* Now that the vmemmap is backed, install the full-fledged allocator */
287 pfn = hyp_virt_to_pfn(hyp_pgt_base);
288 nr_pages = hyp_s1_pgtable_pages();
289 reserved_pages = hyp_early_alloc_nr_used_pages();
290 ret = hyp_pool_init(&hpool, pfn, nr_pages, reserved_pages);
291 if (ret)
292 goto out;
293
294 ret = kvm_host_prepare_stage2(host_s2_pgt_base);
295 if (ret)
296 goto out;
297
298 pkvm_pgtable_mm_ops = (struct kvm_pgtable_mm_ops) {
299 .zalloc_page = hyp_zalloc_hyp_page,
300 .phys_to_virt = hyp_phys_to_virt,
301 .virt_to_phys = hyp_virt_to_phys,
302 .get_page = hpool_get_page,
303 .put_page = hpool_put_page,
304 .page_count = hyp_page_count,
305 };
306 pkvm_pgtable.mm_ops = &pkvm_pgtable_mm_ops;
307
308 ret = fix_host_ownership();
309 if (ret)
310 goto out;
311
312 ret = fix_hyp_pgtable_refcnt();
313 if (ret)
314 goto out;
315
316 ret = hyp_create_pcpu_fixmap();
317 if (ret)
318 goto out;
319
320 ret = hyp_ffa_init(ffa_proxy_pages);
321 if (ret)
322 goto out;
323
324 hyp_shadow_table_init(shadow_table_base);
325 out:
326 /*
327 * We tail-called to here from handle___pkvm_init() and will not return,
328 * so make sure to propagate the return value to the host.
329 */
330 cpu_reg(host_ctxt, 1) = ret;
331
332 __host_enter(host_ctxt);
333 }
334
__pkvm_init(phys_addr_t phys,unsigned long size,unsigned long nr_cpus,unsigned long * per_cpu_base,u32 hyp_va_bits)335 int __pkvm_init(phys_addr_t phys, unsigned long size, unsigned long nr_cpus,
336 unsigned long *per_cpu_base, u32 hyp_va_bits)
337 {
338 struct kvm_nvhe_init_params *params;
339 void *virt = hyp_phys_to_virt(phys);
340 void (*fn)(phys_addr_t params_pa, void *finalize_fn_va);
341 int ret;
342
343 BUG_ON(kvm_check_pvm_sysreg_table());
344
345 if (!PAGE_ALIGNED(phys) || !PAGE_ALIGNED(size))
346 return -EINVAL;
347
348 hyp_spin_lock_init(&pkvm_pgd_lock);
349 hyp_nr_cpus = nr_cpus;
350
351 ret = divide_memory_pool(virt, size);
352 if (ret)
353 return ret;
354
355 ret = recreate_hyp_mappings(phys, size, per_cpu_base, hyp_va_bits);
356 if (ret)
357 return ret;
358
359 update_nvhe_init_params();
360
361 /* Jump in the idmap page to switch to the new page-tables */
362 params = this_cpu_ptr(&kvm_init_params);
363 fn = (typeof(fn))__hyp_pa(__pkvm_init_switch_pgd);
364 fn(__hyp_pa(params), __pkvm_init_finalise);
365
366 unreachable();
367 }
368