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/alloc.h>
14 #include <nvhe/early_alloc.h>
15 #include <nvhe/ffa.h>
16 #include <nvhe/gfp.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/serial.h>
22 #include <nvhe/trace.h>
23 #include <nvhe/trap_handler.h>
24
25 unsigned long hyp_nr_cpus;
26
27 phys_addr_t pvmfw_base;
28 phys_addr_t pvmfw_size;
29
30 #define hyp_percpu_size ((unsigned long)__per_cpu_end - \
31 (unsigned long)__per_cpu_start)
32
33 u64 hyp_lm_size_mb;
34
35 static void *vmemmap_base;
36 static void *vm_table_base;
37 static void *hyp_pgt_base;
38 static void *host_s2_pgt_base;
39 static void *selftest_base;
40 static void *ffa_proxy_pages;
41 static struct kvm_pgtable_mm_ops pkvm_pgtable_mm_ops;
42 static struct hyp_pool hpool;
43
divide_memory_pool(void * virt,unsigned long size)44 static int divide_memory_pool(void *virt, unsigned long size)
45 {
46 unsigned long nr_pages;
47
48 hyp_early_alloc_init(virt, size);
49
50 nr_pages = pkvm_selftest_pages();
51 selftest_base = hyp_early_alloc_contig(nr_pages);
52 if (nr_pages && !selftest_base)
53 return -ENOMEM;
54
55 nr_pages = hyp_vmemmap_pages(sizeof(struct hyp_page));
56 vmemmap_base = hyp_early_alloc_contig(nr_pages);
57 if (!vmemmap_base)
58 return -ENOMEM;
59
60 nr_pages = hyp_vm_table_pages();
61 vm_table_base = hyp_early_alloc_contig(nr_pages);
62 if (!vm_table_base)
63 return -ENOMEM;
64
65 nr_pages = hyp_s1_pgtable_pages();
66 hyp_pgt_base = hyp_early_alloc_contig(nr_pages);
67 if (!hyp_pgt_base)
68 return -ENOMEM;
69
70 nr_pages = host_s2_pgtable_pages();
71 host_s2_pgt_base = hyp_early_alloc_contig(nr_pages);
72 if (!host_s2_pgt_base)
73 return -ENOMEM;
74
75 nr_pages = hyp_ffa_proxy_pages();
76 ffa_proxy_pages = hyp_early_alloc_contig(nr_pages);
77 if (!ffa_proxy_pages)
78 return -ENOMEM;
79
80 hyp_ppages = hyp_early_alloc_contig(1);
81 if (!hyp_ppages)
82 return -ENOMEM;
83
84 return 0;
85 }
86
pkvm_create_host_sve_mappings(void)87 static int pkvm_create_host_sve_mappings(void)
88 {
89 void *start, *end;
90 int ret, i;
91
92 if (!system_supports_sve())
93 return 0;
94
95 for (i = 0; i < hyp_nr_cpus; i++) {
96 struct kvm_host_data *host_data = per_cpu_ptr(&kvm_host_data, i);
97 struct cpu_sve_state *sve_state = host_data->sve_state;
98
99 start = kern_hyp_va(sve_state);
100 end = start + PAGE_ALIGN(pkvm_host_sve_state_size());
101 ret = pkvm_create_mappings(start, end, PAGE_HYP);
102 if (ret)
103 return ret;
104 }
105
106 return 0;
107 }
108
recreate_hyp_mappings(phys_addr_t phys,unsigned long size,unsigned long * per_cpu_base,u32 hyp_va_bits)109 static int recreate_hyp_mappings(phys_addr_t phys, unsigned long size,
110 unsigned long *per_cpu_base,
111 u32 hyp_va_bits)
112 {
113 void *start, *end, *virt = hyp_phys_to_virt(phys);
114 unsigned long pgt_size = hyp_s1_pgtable_pages() << PAGE_SHIFT;
115 enum kvm_pgtable_prot prot;
116 int ret, i;
117
118 /* Recreate the hyp page-table using the early page allocator */
119 hyp_early_alloc_init(hyp_pgt_base, pgt_size);
120 ret = kvm_pgtable_hyp_init(&pkvm_pgtable, hyp_va_bits,
121 &hyp_early_alloc_mm_ops);
122 if (ret)
123 return ret;
124
125 ret = hyp_create_idmap(hyp_va_bits);
126 if (ret)
127 return ret;
128
129 ret = hyp_map_vectors();
130 if (ret)
131 return ret;
132
133 ret = hyp_back_vmemmap(hyp_virt_to_phys(vmemmap_base));
134 if (ret)
135 return ret;
136
137 ret = pkvm_create_mappings(__hyp_text_start, __hyp_text_end, PAGE_HYP_EXEC);
138 if (ret)
139 return ret;
140
141 ret = pkvm_create_mappings(__hyp_data_start, __hyp_data_end, PAGE_HYP);
142 if (ret)
143 return ret;
144
145 ret = pkvm_create_mappings(__hyp_rodata_start, __hyp_rodata_end, PAGE_HYP_RO);
146 if (ret)
147 return ret;
148
149 ret = pkvm_create_mappings(__hyp_bss_start, __hyp_bss_end, PAGE_HYP);
150 if (ret)
151 return ret;
152
153 ret = pkvm_create_mappings(virt, virt + size, PAGE_HYP);
154 if (ret)
155 return ret;
156
157 for (i = 0; i < hyp_nr_cpus; i++) {
158 struct kvm_nvhe_init_params *params = per_cpu_ptr(&kvm_init_params, i);
159
160 start = (void *)kern_hyp_va(per_cpu_base[i]);
161 end = start + PAGE_ALIGN(hyp_percpu_size);
162 ret = pkvm_create_mappings(start, end, PAGE_HYP);
163 if (ret)
164 return ret;
165
166 ret = pkvm_create_stack(params->stack_pa, ¶ms->stack_hyp_va);
167 if (ret)
168 return ret;
169 }
170
171 ret = pkvm_create_host_sve_mappings();
172 if (ret)
173 return ret;
174
175 /*
176 * Map the pvmfw section RO in the hypervisor, but transfer the
177 * ownership from the host to the hypervisor itself to make sure that it
178 * can't be donated or shared with another entity.
179 *
180 * The ownership transition requires matching changes in the host
181 * stage-2. This will be done later (see finalize_host_mappings()) once
182 * the hyp_vmemmap is addressable.
183 */
184 start = hyp_phys_to_virt(pvmfw_base);
185 end = start + pvmfw_size;
186 prot = pkvm_mkstate(PAGE_HYP_RO, PKVM_PAGE_OWNED);
187 ret = pkvm_create_mappings(start, end, prot);
188 if (ret)
189 return ret;
190
191 return 0;
192 }
193
update_nvhe_init_params(void)194 static void update_nvhe_init_params(void)
195 {
196 struct kvm_nvhe_init_params *params;
197 unsigned long i;
198
199 for (i = 0; i < hyp_nr_cpus; i++) {
200 params = per_cpu_ptr(&kvm_init_params, i);
201 params->pgd_pa = __hyp_pa(pkvm_pgtable.pgd);
202 dcache_clean_inval_poc((unsigned long)params,
203 (unsigned long)params + sizeof(*params));
204 }
205 }
206
hyp_zalloc_hyp_page(void * arg)207 static void *hyp_zalloc_hyp_page(void *arg)
208 {
209 return hyp_alloc_pages(&hpool, 0);
210 }
211
hpool_get_page(void * addr)212 static void hpool_get_page(void *addr)
213 {
214 hyp_get_page(&hpool, addr);
215 }
216
hpool_put_page(void * addr)217 static void hpool_put_page(void *addr)
218 {
219 hyp_put_page(&hpool, addr);
220 }
221
fix_host_ownership_walker(const struct kvm_pgtable_visit_ctx * ctx,enum kvm_pgtable_walk_flags visit)222 static int fix_host_ownership_walker(const struct kvm_pgtable_visit_ctx *ctx,
223 enum kvm_pgtable_walk_flags visit)
224 {
225 enum pkvm_page_state state;
226 phys_addr_t phys;
227
228 if (!kvm_pte_valid(ctx->old))
229 return 0;
230
231 if (ctx->level != KVM_PGTABLE_LAST_LEVEL)
232 return -EINVAL;
233
234 phys = kvm_pte_to_phys(ctx->old);
235 if (!addr_is_memory(phys))
236 return -EINVAL;
237
238 /*
239 * Adjust the host stage-2 mappings to match the ownership attributes
240 * configured in the hypervisor stage-1.
241 */
242 state = pkvm_getstate(kvm_pgtable_hyp_pte_prot(ctx->old));
243 switch (state) {
244 case PKVM_PAGE_OWNED:
245 return host_stage2_set_owner_locked(phys, PAGE_SIZE, PKVM_ID_HYP);
246 case PKVM_PAGE_SHARED_OWNED:
247 hyp_phys_to_page(phys)->host_state = PKVM_PAGE_SHARED_BORROWED;
248 break;
249 case PKVM_PAGE_SHARED_BORROWED:
250 hyp_phys_to_page(phys)->host_state = PKVM_PAGE_SHARED_OWNED;
251 break;
252 default:
253 return -EINVAL;
254 }
255
256 return 0;
257 }
258
fix_hyp_pgtable_refcnt_walker(const struct kvm_pgtable_visit_ctx * ctx,enum kvm_pgtable_walk_flags visit)259 static int fix_hyp_pgtable_refcnt_walker(const struct kvm_pgtable_visit_ctx *ctx,
260 enum kvm_pgtable_walk_flags visit)
261 {
262 /*
263 * Fix-up the refcount for the page-table pages as the early allocator
264 * was unable to access the hyp_vmemmap and so the buddy allocator has
265 * initialised the refcount to '1'.
266 */
267 if (kvm_pte_valid(ctx->old))
268 ctx->mm_ops->get_page(ctx->ptep);
269
270 return 0;
271 }
272
pin_table_walker(const struct kvm_pgtable_visit_ctx * ctx,enum kvm_pgtable_walk_flags visit)273 static int pin_table_walker(const struct kvm_pgtable_visit_ctx *ctx,
274 enum kvm_pgtable_walk_flags visit)
275 {
276 struct kvm_pgtable_mm_ops *mm_ops = ctx->mm_ops;
277 kvm_pte_t pte = *(ctx->ptep);
278
279 if (kvm_pte_valid(pte))
280 mm_ops->get_page(kvm_pte_follow(pte, mm_ops));
281
282 return 0;
283 }
284
pin_host_tables(void)285 static int pin_host_tables(void)
286 {
287 struct kvm_pgtable_walker walker = {
288 .cb = pin_table_walker,
289 .flags = KVM_PGTABLE_WALK_TABLE_POST,
290 .arg = &host_mmu.mm_ops,
291 };
292
293 return kvm_pgtable_walk(&host_mmu.pgt, 0, BIT(host_mmu.pgt.ia_bits), &walker);
294 }
295
fix_host_ownership(void)296 static int fix_host_ownership(void)
297 {
298 struct kvm_pgtable_walker walker = {
299 .cb = fix_host_ownership_walker,
300 .flags = KVM_PGTABLE_WALK_LEAF,
301 };
302 int i, ret;
303
304 for (i = 0; i < hyp_memblock_nr; i++) {
305 struct memblock_region *reg = &hyp_memory[i];
306 u64 start = (u64)hyp_phys_to_virt(reg->base);
307
308 ret = kvm_pgtable_walk(&pkvm_pgtable, start, reg->size, &walker);
309 if (ret)
310 return ret;
311 }
312
313 return 0;
314 }
315
fix_hyp_pgtable_refcnt(void)316 static int fix_hyp_pgtable_refcnt(void)
317 {
318 struct kvm_pgtable_walker walker = {
319 .cb = fix_hyp_pgtable_refcnt_walker,
320 .flags = KVM_PGTABLE_WALK_LEAF | KVM_PGTABLE_WALK_TABLE_POST,
321 .arg = pkvm_pgtable.mm_ops,
322 };
323
324 return kvm_pgtable_walk(&pkvm_pgtable, 0, BIT(pkvm_pgtable.ia_bits),
325 &walker);
326 }
327
unmap_protected_regions(void)328 static int unmap_protected_regions(void)
329 {
330 struct pkvm_moveable_reg *reg;
331 int i, ret;
332
333 for (i = 0; i < pkvm_moveable_regs_nr; i++) {
334 reg = &pkvm_moveable_regs[i];
335 if (reg->type != PKVM_MREG_PROTECTED_RANGE)
336 continue;
337
338 ret = host_stage2_set_owner_locked(reg->start, reg->size,
339 PKVM_ID_PROTECTED);
340 if (ret)
341 return ret;
342 }
343
344 return 0;
345 }
346
__pkvm_init_finalise(void)347 void __noreturn __pkvm_init_finalise(void)
348 {
349 struct kvm_cpu_context *host_ctxt = host_data_ptr(host_ctxt);
350 unsigned long nr_pages, reserved_pages, pfn;
351 int ret;
352
353 /* Now that the vmemmap is backed, install the full-fledged allocator */
354 pfn = hyp_virt_to_pfn(hyp_pgt_base);
355 nr_pages = hyp_s1_pgtable_pages();
356 reserved_pages = hyp_early_alloc_nr_used_pages();
357 ret = hyp_pool_init(&hpool, pfn, nr_pages, reserved_pages);
358 if (ret)
359 goto out;
360
361 ret = kvm_host_prepare_stage2(host_s2_pgt_base);
362 if (ret)
363 goto out;
364
365 pkvm_pgtable_mm_ops = (struct kvm_pgtable_mm_ops) {
366 .zalloc_page = hyp_zalloc_hyp_page,
367 .phys_to_virt = hyp_phys_to_virt,
368 .virt_to_phys = hyp_virt_to_phys,
369 .get_page = hpool_get_page,
370 .put_page = hpool_put_page,
371 .page_count = hyp_page_count,
372 };
373 pkvm_pgtable.mm_ops = &pkvm_pgtable_mm_ops;
374
375 ret = fix_hyp_pgtable_refcnt();
376 if (ret)
377 goto out;
378
379 ret = hyp_create_fixmap();
380 if (ret)
381 goto out;
382
383 ret = pkvm_timer_init();
384 if (ret)
385 goto out;
386
387 hyp_ftrace_setup_core();
388
389 ret = fix_host_ownership();
390 if (ret)
391 goto out;
392
393 ret = unmap_protected_regions();
394 if (ret)
395 goto out;
396
397 ret = pin_host_tables();
398 if (ret)
399 goto out;
400
401 ret = hyp_ffa_init(ffa_proxy_pages);
402 if (ret)
403 goto out;
404
405 pkvm_hyp_vm_table_init(vm_table_base);
406
407 pkvm_ownership_selftest(selftest_base);
408 out:
409 /*
410 * We tail-called to here from handle___pkvm_init() and will not return,
411 * so make sure to propagate the return value to the host.
412 */
413 cpu_reg(host_ctxt, 1) = ret;
414
415 __host_enter(host_ctxt);
416 }
417
__pkvm_init(phys_addr_t phys,unsigned long size,unsigned long nr_cpus,unsigned long * per_cpu_base,u32 hyp_va_bits)418 int __pkvm_init(phys_addr_t phys, unsigned long size, unsigned long nr_cpus,
419 unsigned long *per_cpu_base, u32 hyp_va_bits)
420 {
421 struct kvm_nvhe_init_params *params;
422 void *virt = hyp_phys_to_virt(phys);
423 typeof(__pkvm_init_switch_pgd) *fn;
424 int ret;
425
426 BUG_ON(kvm_check_pvm_sysreg_table());
427
428 if (!PAGE_ALIGNED(phys) || !PAGE_ALIGNED(size))
429 return -EINVAL;
430
431 hyp_spin_lock_init(&pkvm_pgd_lock);
432 hyp_nr_cpus = nr_cpus;
433
434 ret = divide_memory_pool(virt, size);
435 if (ret)
436 return ret;
437
438 ret = recreate_hyp_mappings(phys, size, per_cpu_base, hyp_va_bits);
439 if (ret)
440 return ret;
441
442 ret = hyp_alloc_init(SZ_128M);
443 if (ret)
444 return ret;
445
446 update_nvhe_init_params();
447
448 /* Jump in the idmap page to switch to the new page-tables */
449 params = this_cpu_ptr(&kvm_init_params);
450 fn = (typeof(fn))__hyp_pa(__pkvm_init_switch_pgd);
451 fn(params->pgd_pa, params->stack_hyp_va, __pkvm_init_finalise);
452
453 unreachable();
454 }
455