xref: /linux/arch/riscv/kvm/mmu.c (revision cbac924200b838cfb8d8b1415113d788089dc50b)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2019 Western Digital Corporation or its affiliates.
4  *
5  * Authors:
6  *     Anup Patel <anup.patel@wdc.com>
7  */
8 
9 #include <linux/bitops.h>
10 #include <linux/errno.h>
11 #include <linux/err.h>
12 #include <linux/hugetlb.h>
13 #include <linux/module.h>
14 #include <linux/uaccess.h>
15 #include <linux/vmalloc.h>
16 #include <linux/kvm_host.h>
17 #include <linux/sched/signal.h>
18 #include <asm/csr.h>
19 #include <asm/page.h>
20 #include <asm/pgtable.h>
21 #include <asm/sbi.h>
22 
23 #ifdef CONFIG_64BIT
24 static unsigned long stage2_mode = (HGATP_MODE_SV39X4 << HGATP_MODE_SHIFT);
25 static unsigned long stage2_pgd_levels = 3;
26 #define stage2_index_bits	9
27 #else
28 static unsigned long stage2_mode = (HGATP_MODE_SV32X4 << HGATP_MODE_SHIFT);
29 static unsigned long stage2_pgd_levels = 2;
30 #define stage2_index_bits	10
31 #endif
32 
33 #define stage2_pgd_xbits	2
34 #define stage2_pgd_size	(1UL << (HGATP_PAGE_SHIFT + stage2_pgd_xbits))
35 #define stage2_gpa_bits	(HGATP_PAGE_SHIFT + \
36 			 (stage2_pgd_levels * stage2_index_bits) + \
37 			 stage2_pgd_xbits)
38 #define stage2_gpa_size	((gpa_t)(1ULL << stage2_gpa_bits))
39 
40 #define stage2_pte_leaf(__ptep)	\
41 	(pte_val(*(__ptep)) & (_PAGE_READ | _PAGE_WRITE | _PAGE_EXEC))
42 
43 static inline unsigned long stage2_pte_index(gpa_t addr, u32 level)
44 {
45 	unsigned long mask;
46 	unsigned long shift = HGATP_PAGE_SHIFT + (stage2_index_bits * level);
47 
48 	if (level == (stage2_pgd_levels - 1))
49 		mask = (PTRS_PER_PTE * (1UL << stage2_pgd_xbits)) - 1;
50 	else
51 		mask = PTRS_PER_PTE - 1;
52 
53 	return (addr >> shift) & mask;
54 }
55 
56 static inline unsigned long stage2_pte_page_vaddr(pte_t pte)
57 {
58 	return (unsigned long)pfn_to_virt(pte_val(pte) >> _PAGE_PFN_SHIFT);
59 }
60 
61 static int stage2_page_size_to_level(unsigned long page_size, u32 *out_level)
62 {
63 	u32 i;
64 	unsigned long psz = 1UL << 12;
65 
66 	for (i = 0; i < stage2_pgd_levels; i++) {
67 		if (page_size == (psz << (i * stage2_index_bits))) {
68 			*out_level = i;
69 			return 0;
70 		}
71 	}
72 
73 	return -EINVAL;
74 }
75 
76 static int stage2_level_to_page_size(u32 level, unsigned long *out_pgsize)
77 {
78 	if (stage2_pgd_levels < level)
79 		return -EINVAL;
80 
81 	*out_pgsize = 1UL << (12 + (level * stage2_index_bits));
82 
83 	return 0;
84 }
85 
86 static bool stage2_get_leaf_entry(struct kvm *kvm, gpa_t addr,
87 				  pte_t **ptepp, u32 *ptep_level)
88 {
89 	pte_t *ptep;
90 	u32 current_level = stage2_pgd_levels - 1;
91 
92 	*ptep_level = current_level;
93 	ptep = (pte_t *)kvm->arch.pgd;
94 	ptep = &ptep[stage2_pte_index(addr, current_level)];
95 	while (ptep && pte_val(*ptep)) {
96 		if (stage2_pte_leaf(ptep)) {
97 			*ptep_level = current_level;
98 			*ptepp = ptep;
99 			return true;
100 		}
101 
102 		if (current_level) {
103 			current_level--;
104 			*ptep_level = current_level;
105 			ptep = (pte_t *)stage2_pte_page_vaddr(*ptep);
106 			ptep = &ptep[stage2_pte_index(addr, current_level)];
107 		} else {
108 			ptep = NULL;
109 		}
110 	}
111 
112 	return false;
113 }
114 
115 static void stage2_remote_tlb_flush(struct kvm *kvm, u32 level, gpa_t addr)
116 {
117 	unsigned long size = PAGE_SIZE;
118 	struct kvm_vmid *vmid = &kvm->arch.vmid;
119 
120 	if (stage2_level_to_page_size(level, &size))
121 		return;
122 	addr &= ~(size - 1);
123 
124 	/*
125 	 * TODO: Instead of cpu_online_mask, we should only target CPUs
126 	 * where the Guest/VM is running.
127 	 */
128 	preempt_disable();
129 	sbi_remote_hfence_gvma_vmid(cpu_online_mask, addr, size,
130 				    READ_ONCE(vmid->vmid));
131 	preempt_enable();
132 }
133 
134 static int stage2_set_pte(struct kvm *kvm, u32 level,
135 			   struct kvm_mmu_memory_cache *pcache,
136 			   gpa_t addr, const pte_t *new_pte)
137 {
138 	u32 current_level = stage2_pgd_levels - 1;
139 	pte_t *next_ptep = (pte_t *)kvm->arch.pgd;
140 	pte_t *ptep = &next_ptep[stage2_pte_index(addr, current_level)];
141 
142 	if (current_level < level)
143 		return -EINVAL;
144 
145 	while (current_level != level) {
146 		if (stage2_pte_leaf(ptep))
147 			return -EEXIST;
148 
149 		if (!pte_val(*ptep)) {
150 			if (!pcache)
151 				return -ENOMEM;
152 			next_ptep = kvm_mmu_memory_cache_alloc(pcache);
153 			if (!next_ptep)
154 				return -ENOMEM;
155 			*ptep = pfn_pte(PFN_DOWN(__pa(next_ptep)),
156 					__pgprot(_PAGE_TABLE));
157 		} else {
158 			if (stage2_pte_leaf(ptep))
159 				return -EEXIST;
160 			next_ptep = (pte_t *)stage2_pte_page_vaddr(*ptep);
161 		}
162 
163 		current_level--;
164 		ptep = &next_ptep[stage2_pte_index(addr, current_level)];
165 	}
166 
167 	*ptep = *new_pte;
168 	if (stage2_pte_leaf(ptep))
169 		stage2_remote_tlb_flush(kvm, current_level, addr);
170 
171 	return 0;
172 }
173 
174 static int stage2_map_page(struct kvm *kvm,
175 			   struct kvm_mmu_memory_cache *pcache,
176 			   gpa_t gpa, phys_addr_t hpa,
177 			   unsigned long page_size,
178 			   bool page_rdonly, bool page_exec)
179 {
180 	int ret;
181 	u32 level = 0;
182 	pte_t new_pte;
183 	pgprot_t prot;
184 
185 	ret = stage2_page_size_to_level(page_size, &level);
186 	if (ret)
187 		return ret;
188 
189 	/*
190 	 * A RISC-V implementation can choose to either:
191 	 * 1) Update 'A' and 'D' PTE bits in hardware
192 	 * 2) Generate page fault when 'A' and/or 'D' bits are not set
193 	 *    PTE so that software can update these bits.
194 	 *
195 	 * We support both options mentioned above. To achieve this, we
196 	 * always set 'A' and 'D' PTE bits at time of creating stage2
197 	 * mapping. To support KVM dirty page logging with both options
198 	 * mentioned above, we will write-protect stage2 PTEs to track
199 	 * dirty pages.
200 	 */
201 
202 	if (page_exec) {
203 		if (page_rdonly)
204 			prot = PAGE_READ_EXEC;
205 		else
206 			prot = PAGE_WRITE_EXEC;
207 	} else {
208 		if (page_rdonly)
209 			prot = PAGE_READ;
210 		else
211 			prot = PAGE_WRITE;
212 	}
213 	new_pte = pfn_pte(PFN_DOWN(hpa), prot);
214 	new_pte = pte_mkdirty(new_pte);
215 
216 	return stage2_set_pte(kvm, level, pcache, gpa, &new_pte);
217 }
218 
219 enum stage2_op {
220 	STAGE2_OP_NOP = 0,	/* Nothing */
221 	STAGE2_OP_CLEAR,	/* Clear/Unmap */
222 	STAGE2_OP_WP,		/* Write-protect */
223 };
224 
225 static void stage2_op_pte(struct kvm *kvm, gpa_t addr,
226 			  pte_t *ptep, u32 ptep_level, enum stage2_op op)
227 {
228 	int i, ret;
229 	pte_t *next_ptep;
230 	u32 next_ptep_level;
231 	unsigned long next_page_size, page_size;
232 
233 	ret = stage2_level_to_page_size(ptep_level, &page_size);
234 	if (ret)
235 		return;
236 
237 	BUG_ON(addr & (page_size - 1));
238 
239 	if (!pte_val(*ptep))
240 		return;
241 
242 	if (ptep_level && !stage2_pte_leaf(ptep)) {
243 		next_ptep = (pte_t *)stage2_pte_page_vaddr(*ptep);
244 		next_ptep_level = ptep_level - 1;
245 		ret = stage2_level_to_page_size(next_ptep_level,
246 						&next_page_size);
247 		if (ret)
248 			return;
249 
250 		if (op == STAGE2_OP_CLEAR)
251 			set_pte(ptep, __pte(0));
252 		for (i = 0; i < PTRS_PER_PTE; i++)
253 			stage2_op_pte(kvm, addr + i * next_page_size,
254 					&next_ptep[i], next_ptep_level, op);
255 		if (op == STAGE2_OP_CLEAR)
256 			put_page(virt_to_page(next_ptep));
257 	} else {
258 		if (op == STAGE2_OP_CLEAR)
259 			set_pte(ptep, __pte(0));
260 		else if (op == STAGE2_OP_WP)
261 			set_pte(ptep, __pte(pte_val(*ptep) & ~_PAGE_WRITE));
262 		stage2_remote_tlb_flush(kvm, ptep_level, addr);
263 	}
264 }
265 
266 static void stage2_unmap_range(struct kvm *kvm, gpa_t start,
267 			       gpa_t size, bool may_block)
268 {
269 	int ret;
270 	pte_t *ptep;
271 	u32 ptep_level;
272 	bool found_leaf;
273 	unsigned long page_size;
274 	gpa_t addr = start, end = start + size;
275 
276 	while (addr < end) {
277 		found_leaf = stage2_get_leaf_entry(kvm, addr,
278 						   &ptep, &ptep_level);
279 		ret = stage2_level_to_page_size(ptep_level, &page_size);
280 		if (ret)
281 			break;
282 
283 		if (!found_leaf)
284 			goto next;
285 
286 		if (!(addr & (page_size - 1)) && ((end - addr) >= page_size))
287 			stage2_op_pte(kvm, addr, ptep,
288 				      ptep_level, STAGE2_OP_CLEAR);
289 
290 next:
291 		addr += page_size;
292 
293 		/*
294 		 * If the range is too large, release the kvm->mmu_lock
295 		 * to prevent starvation and lockup detector warnings.
296 		 */
297 		if (may_block && addr < end)
298 			cond_resched_lock(&kvm->mmu_lock);
299 	}
300 }
301 
302 static void stage2_wp_range(struct kvm *kvm, gpa_t start, gpa_t end)
303 {
304 	int ret;
305 	pte_t *ptep;
306 	u32 ptep_level;
307 	bool found_leaf;
308 	gpa_t addr = start;
309 	unsigned long page_size;
310 
311 	while (addr < end) {
312 		found_leaf = stage2_get_leaf_entry(kvm, addr,
313 						   &ptep, &ptep_level);
314 		ret = stage2_level_to_page_size(ptep_level, &page_size);
315 		if (ret)
316 			break;
317 
318 		if (!found_leaf)
319 			goto next;
320 
321 		if (!(addr & (page_size - 1)) && ((end - addr) >= page_size))
322 			stage2_op_pte(kvm, addr, ptep,
323 				      ptep_level, STAGE2_OP_WP);
324 
325 next:
326 		addr += page_size;
327 	}
328 }
329 
330 static void stage2_wp_memory_region(struct kvm *kvm, int slot)
331 {
332 	struct kvm_memslots *slots = kvm_memslots(kvm);
333 	struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
334 	phys_addr_t start = memslot->base_gfn << PAGE_SHIFT;
335 	phys_addr_t end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
336 
337 	spin_lock(&kvm->mmu_lock);
338 	stage2_wp_range(kvm, start, end);
339 	spin_unlock(&kvm->mmu_lock);
340 	kvm_flush_remote_tlbs(kvm);
341 }
342 
343 static int stage2_ioremap(struct kvm *kvm, gpa_t gpa, phys_addr_t hpa,
344 			  unsigned long size, bool writable)
345 {
346 	pte_t pte;
347 	int ret = 0;
348 	unsigned long pfn;
349 	phys_addr_t addr, end;
350 	struct kvm_mmu_memory_cache pcache;
351 
352 	memset(&pcache, 0, sizeof(pcache));
353 	pcache.gfp_zero = __GFP_ZERO;
354 
355 	end = (gpa + size + PAGE_SIZE - 1) & PAGE_MASK;
356 	pfn = __phys_to_pfn(hpa);
357 
358 	for (addr = gpa; addr < end; addr += PAGE_SIZE) {
359 		pte = pfn_pte(pfn, PAGE_KERNEL);
360 
361 		if (!writable)
362 			pte = pte_wrprotect(pte);
363 
364 		ret = kvm_mmu_topup_memory_cache(&pcache, stage2_pgd_levels);
365 		if (ret)
366 			goto out;
367 
368 		spin_lock(&kvm->mmu_lock);
369 		ret = stage2_set_pte(kvm, 0, &pcache, addr, &pte);
370 		spin_unlock(&kvm->mmu_lock);
371 		if (ret)
372 			goto out;
373 
374 		pfn++;
375 	}
376 
377 out:
378 	kvm_mmu_free_memory_cache(&pcache);
379 	return ret;
380 }
381 
382 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
383 					     struct kvm_memory_slot *slot,
384 					     gfn_t gfn_offset,
385 					     unsigned long mask)
386 {
387 	phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
388 	phys_addr_t start = (base_gfn +  __ffs(mask)) << PAGE_SHIFT;
389 	phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
390 
391 	stage2_wp_range(kvm, start, end);
392 }
393 
394 void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot)
395 {
396 }
397 
398 void kvm_arch_flush_remote_tlbs_memslot(struct kvm *kvm,
399 					const struct kvm_memory_slot *memslot)
400 {
401 	kvm_flush_remote_tlbs(kvm);
402 }
403 
404 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free)
405 {
406 }
407 
408 void kvm_arch_memslots_updated(struct kvm *kvm, u64 gen)
409 {
410 }
411 
412 void kvm_arch_flush_shadow_all(struct kvm *kvm)
413 {
414 	kvm_riscv_stage2_free_pgd(kvm);
415 }
416 
417 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
418 				   struct kvm_memory_slot *slot)
419 {
420 	gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
421 	phys_addr_t size = slot->npages << PAGE_SHIFT;
422 
423 	spin_lock(&kvm->mmu_lock);
424 	stage2_unmap_range(kvm, gpa, size, false);
425 	spin_unlock(&kvm->mmu_lock);
426 }
427 
428 void kvm_arch_commit_memory_region(struct kvm *kvm,
429 				struct kvm_memory_slot *old,
430 				const struct kvm_memory_slot *new,
431 				enum kvm_mr_change change)
432 {
433 	/*
434 	 * At this point memslot has been committed and there is an
435 	 * allocated dirty_bitmap[], dirty pages will be tracked while
436 	 * the memory slot is write protected.
437 	 */
438 	if (change != KVM_MR_DELETE && new->flags & KVM_MEM_LOG_DIRTY_PAGES)
439 		stage2_wp_memory_region(kvm, new->id);
440 }
441 
442 int kvm_arch_prepare_memory_region(struct kvm *kvm,
443 				const struct kvm_memory_slot *old,
444 				struct kvm_memory_slot *new,
445 				enum kvm_mr_change change)
446 {
447 	hva_t hva, reg_end, size;
448 	gpa_t base_gpa;
449 	bool writable;
450 	int ret = 0;
451 
452 	if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
453 			change != KVM_MR_FLAGS_ONLY)
454 		return 0;
455 
456 	/*
457 	 * Prevent userspace from creating a memory region outside of the GPA
458 	 * space addressable by the KVM guest GPA space.
459 	 */
460 	if ((new->base_gfn + new->npages) >=
461 	    (stage2_gpa_size >> PAGE_SHIFT))
462 		return -EFAULT;
463 
464 	hva = new->userspace_addr;
465 	size = new->npages << PAGE_SHIFT;
466 	reg_end = hva + size;
467 	base_gpa = new->base_gfn << PAGE_SHIFT;
468 	writable = !(new->flags & KVM_MEM_READONLY);
469 
470 	mmap_read_lock(current->mm);
471 
472 	/*
473 	 * A memory region could potentially cover multiple VMAs, and
474 	 * any holes between them, so iterate over all of them to find
475 	 * out if we can map any of them right now.
476 	 *
477 	 *     +--------------------------------------------+
478 	 * +---------------+----------------+   +----------------+
479 	 * |   : VMA 1     |      VMA 2     |   |    VMA 3  :    |
480 	 * +---------------+----------------+   +----------------+
481 	 *     |               memory region                |
482 	 *     +--------------------------------------------+
483 	 */
484 	do {
485 		struct vm_area_struct *vma = find_vma(current->mm, hva);
486 		hva_t vm_start, vm_end;
487 
488 		if (!vma || vma->vm_start >= reg_end)
489 			break;
490 
491 		/*
492 		 * Mapping a read-only VMA is only allowed if the
493 		 * memory region is configured as read-only.
494 		 */
495 		if (writable && !(vma->vm_flags & VM_WRITE)) {
496 			ret = -EPERM;
497 			break;
498 		}
499 
500 		/* Take the intersection of this VMA with the memory region */
501 		vm_start = max(hva, vma->vm_start);
502 		vm_end = min(reg_end, vma->vm_end);
503 
504 		if (vma->vm_flags & VM_PFNMAP) {
505 			gpa_t gpa = base_gpa + (vm_start - hva);
506 			phys_addr_t pa;
507 
508 			pa = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT;
509 			pa += vm_start - vma->vm_start;
510 
511 			/* IO region dirty page logging not allowed */
512 			if (new->flags & KVM_MEM_LOG_DIRTY_PAGES) {
513 				ret = -EINVAL;
514 				goto out;
515 			}
516 
517 			ret = stage2_ioremap(kvm, gpa, pa,
518 					     vm_end - vm_start, writable);
519 			if (ret)
520 				break;
521 		}
522 		hva = vm_end;
523 	} while (hva < reg_end);
524 
525 	if (change == KVM_MR_FLAGS_ONLY)
526 		goto out;
527 
528 	spin_lock(&kvm->mmu_lock);
529 	if (ret)
530 		stage2_unmap_range(kvm, base_gpa, size, false);
531 	spin_unlock(&kvm->mmu_lock);
532 
533 out:
534 	mmap_read_unlock(current->mm);
535 	return ret;
536 }
537 
538 bool kvm_unmap_gfn_range(struct kvm *kvm, struct kvm_gfn_range *range)
539 {
540 	if (!kvm->arch.pgd)
541 		return false;
542 
543 	stage2_unmap_range(kvm, range->start << PAGE_SHIFT,
544 			   (range->end - range->start) << PAGE_SHIFT,
545 			   range->may_block);
546 	return false;
547 }
548 
549 bool kvm_set_spte_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
550 {
551 	int ret;
552 	kvm_pfn_t pfn = pte_pfn(range->pte);
553 
554 	if (!kvm->arch.pgd)
555 		return false;
556 
557 	WARN_ON(range->end - range->start != 1);
558 
559 	ret = stage2_map_page(kvm, NULL, range->start << PAGE_SHIFT,
560 			      __pfn_to_phys(pfn), PAGE_SIZE, true, true);
561 	if (ret) {
562 		kvm_debug("Failed to map stage2 page (error %d)\n", ret);
563 		return true;
564 	}
565 
566 	return false;
567 }
568 
569 bool kvm_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
570 {
571 	pte_t *ptep;
572 	u32 ptep_level = 0;
573 	u64 size = (range->end - range->start) << PAGE_SHIFT;
574 
575 	if (!kvm->arch.pgd)
576 		return false;
577 
578 	WARN_ON(size != PAGE_SIZE && size != PMD_SIZE && size != PGDIR_SIZE);
579 
580 	if (!stage2_get_leaf_entry(kvm, range->start << PAGE_SHIFT,
581 				   &ptep, &ptep_level))
582 		return false;
583 
584 	return ptep_test_and_clear_young(NULL, 0, ptep);
585 }
586 
587 bool kvm_test_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
588 {
589 	pte_t *ptep;
590 	u32 ptep_level = 0;
591 	u64 size = (range->end - range->start) << PAGE_SHIFT;
592 
593 	if (!kvm->arch.pgd)
594 		return false;
595 
596 	WARN_ON(size != PAGE_SIZE && size != PMD_SIZE && size != PGDIR_SIZE);
597 
598 	if (!stage2_get_leaf_entry(kvm, range->start << PAGE_SHIFT,
599 				   &ptep, &ptep_level))
600 		return false;
601 
602 	return pte_young(*ptep);
603 }
604 
605 int kvm_riscv_stage2_map(struct kvm_vcpu *vcpu,
606 			 struct kvm_memory_slot *memslot,
607 			 gpa_t gpa, unsigned long hva, bool is_write)
608 {
609 	int ret;
610 	kvm_pfn_t hfn;
611 	bool writeable;
612 	short vma_pageshift;
613 	gfn_t gfn = gpa >> PAGE_SHIFT;
614 	struct vm_area_struct *vma;
615 	struct kvm *kvm = vcpu->kvm;
616 	struct kvm_mmu_memory_cache *pcache = &vcpu->arch.mmu_page_cache;
617 	bool logging = (memslot->dirty_bitmap &&
618 			!(memslot->flags & KVM_MEM_READONLY)) ? true : false;
619 	unsigned long vma_pagesize, mmu_seq;
620 
621 	mmap_read_lock(current->mm);
622 
623 	vma = find_vma_intersection(current->mm, hva, hva + 1);
624 	if (unlikely(!vma)) {
625 		kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
626 		mmap_read_unlock(current->mm);
627 		return -EFAULT;
628 	}
629 
630 	if (is_vm_hugetlb_page(vma))
631 		vma_pageshift = huge_page_shift(hstate_vma(vma));
632 	else
633 		vma_pageshift = PAGE_SHIFT;
634 	vma_pagesize = 1ULL << vma_pageshift;
635 	if (logging || (vma->vm_flags & VM_PFNMAP))
636 		vma_pagesize = PAGE_SIZE;
637 
638 	if (vma_pagesize == PMD_SIZE || vma_pagesize == PGDIR_SIZE)
639 		gfn = (gpa & huge_page_mask(hstate_vma(vma))) >> PAGE_SHIFT;
640 
641 	mmap_read_unlock(current->mm);
642 
643 	if (vma_pagesize != PGDIR_SIZE &&
644 	    vma_pagesize != PMD_SIZE &&
645 	    vma_pagesize != PAGE_SIZE) {
646 		kvm_err("Invalid VMA page size 0x%lx\n", vma_pagesize);
647 		return -EFAULT;
648 	}
649 
650 	/* We need minimum second+third level pages */
651 	ret = kvm_mmu_topup_memory_cache(pcache, stage2_pgd_levels);
652 	if (ret) {
653 		kvm_err("Failed to topup stage2 cache\n");
654 		return ret;
655 	}
656 
657 	mmu_seq = kvm->mmu_notifier_seq;
658 
659 	hfn = gfn_to_pfn_prot(kvm, gfn, is_write, &writeable);
660 	if (hfn == KVM_PFN_ERR_HWPOISON) {
661 		send_sig_mceerr(BUS_MCEERR_AR, (void __user *)hva,
662 				vma_pageshift, current);
663 		return 0;
664 	}
665 	if (is_error_noslot_pfn(hfn))
666 		return -EFAULT;
667 
668 	/*
669 	 * If logging is active then we allow writable pages only
670 	 * for write faults.
671 	 */
672 	if (logging && !is_write)
673 		writeable = false;
674 
675 	spin_lock(&kvm->mmu_lock);
676 
677 	if (mmu_notifier_retry(kvm, mmu_seq))
678 		goto out_unlock;
679 
680 	if (writeable) {
681 		kvm_set_pfn_dirty(hfn);
682 		mark_page_dirty(kvm, gfn);
683 		ret = stage2_map_page(kvm, pcache, gpa, hfn << PAGE_SHIFT,
684 				      vma_pagesize, false, true);
685 	} else {
686 		ret = stage2_map_page(kvm, pcache, gpa, hfn << PAGE_SHIFT,
687 				      vma_pagesize, true, true);
688 	}
689 
690 	if (ret)
691 		kvm_err("Failed to map in stage2\n");
692 
693 out_unlock:
694 	spin_unlock(&kvm->mmu_lock);
695 	kvm_set_pfn_accessed(hfn);
696 	kvm_release_pfn_clean(hfn);
697 	return ret;
698 }
699 
700 int kvm_riscv_stage2_alloc_pgd(struct kvm *kvm)
701 {
702 	struct page *pgd_page;
703 
704 	if (kvm->arch.pgd != NULL) {
705 		kvm_err("kvm_arch already initialized?\n");
706 		return -EINVAL;
707 	}
708 
709 	pgd_page = alloc_pages(GFP_KERNEL | __GFP_ZERO,
710 				get_order(stage2_pgd_size));
711 	if (!pgd_page)
712 		return -ENOMEM;
713 	kvm->arch.pgd = page_to_virt(pgd_page);
714 	kvm->arch.pgd_phys = page_to_phys(pgd_page);
715 
716 	return 0;
717 }
718 
719 void kvm_riscv_stage2_free_pgd(struct kvm *kvm)
720 {
721 	void *pgd = NULL;
722 
723 	spin_lock(&kvm->mmu_lock);
724 	if (kvm->arch.pgd) {
725 		stage2_unmap_range(kvm, 0UL, stage2_gpa_size, false);
726 		pgd = READ_ONCE(kvm->arch.pgd);
727 		kvm->arch.pgd = NULL;
728 		kvm->arch.pgd_phys = 0;
729 	}
730 	spin_unlock(&kvm->mmu_lock);
731 
732 	if (pgd)
733 		free_pages((unsigned long)pgd, get_order(stage2_pgd_size));
734 }
735 
736 void kvm_riscv_stage2_update_hgatp(struct kvm_vcpu *vcpu)
737 {
738 	unsigned long hgatp = stage2_mode;
739 	struct kvm_arch *k = &vcpu->kvm->arch;
740 
741 	hgatp |= (READ_ONCE(k->vmid.vmid) << HGATP_VMID_SHIFT) &
742 		 HGATP_VMID_MASK;
743 	hgatp |= (k->pgd_phys >> PAGE_SHIFT) & HGATP_PPN;
744 
745 	csr_write(CSR_HGATP, hgatp);
746 
747 	if (!kvm_riscv_stage2_vmid_bits())
748 		__kvm_riscv_hfence_gvma_all();
749 }
750 
751 void kvm_riscv_stage2_mode_detect(void)
752 {
753 #ifdef CONFIG_64BIT
754 	/* Try Sv48x4 stage2 mode */
755 	csr_write(CSR_HGATP, HGATP_MODE_SV48X4 << HGATP_MODE_SHIFT);
756 	if ((csr_read(CSR_HGATP) >> HGATP_MODE_SHIFT) == HGATP_MODE_SV48X4) {
757 		stage2_mode = (HGATP_MODE_SV48X4 << HGATP_MODE_SHIFT);
758 		stage2_pgd_levels = 4;
759 	}
760 	csr_write(CSR_HGATP, 0);
761 
762 	__kvm_riscv_hfence_gvma_all();
763 #endif
764 }
765 
766 unsigned long kvm_riscv_stage2_mode(void)
767 {
768 	return stage2_mode >> HGATP_MODE_SHIFT;
769 }
770 
771 int kvm_riscv_stage2_gpa_bits(void)
772 {
773 	return stage2_gpa_bits;
774 }
775