xref: /openbmc/linux/arch/riscv/kvm/mmu.c (revision 6db6b729)
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 
22 #ifdef CONFIG_64BIT
23 static unsigned long gstage_mode __ro_after_init = (HGATP_MODE_SV39X4 << HGATP_MODE_SHIFT);
24 static unsigned long gstage_pgd_levels __ro_after_init = 3;
25 #define gstage_index_bits	9
26 #else
27 static unsigned long gstage_mode __ro_after_init = (HGATP_MODE_SV32X4 << HGATP_MODE_SHIFT);
28 static unsigned long gstage_pgd_levels __ro_after_init = 2;
29 #define gstage_index_bits	10
30 #endif
31 
32 #define gstage_pgd_xbits	2
33 #define gstage_pgd_size	(1UL << (HGATP_PAGE_SHIFT + gstage_pgd_xbits))
34 #define gstage_gpa_bits	(HGATP_PAGE_SHIFT + \
35 			 (gstage_pgd_levels * gstage_index_bits) + \
36 			 gstage_pgd_xbits)
37 #define gstage_gpa_size	((gpa_t)(1ULL << gstage_gpa_bits))
38 
39 #define gstage_pte_leaf(__ptep)	\
40 	(pte_val(*(__ptep)) & (_PAGE_READ | _PAGE_WRITE | _PAGE_EXEC))
41 
42 static inline unsigned long gstage_pte_index(gpa_t addr, u32 level)
43 {
44 	unsigned long mask;
45 	unsigned long shift = HGATP_PAGE_SHIFT + (gstage_index_bits * level);
46 
47 	if (level == (gstage_pgd_levels - 1))
48 		mask = (PTRS_PER_PTE * (1UL << gstage_pgd_xbits)) - 1;
49 	else
50 		mask = PTRS_PER_PTE - 1;
51 
52 	return (addr >> shift) & mask;
53 }
54 
55 static inline unsigned long gstage_pte_page_vaddr(pte_t pte)
56 {
57 	return (unsigned long)pfn_to_virt(__page_val_to_pfn(pte_val(pte)));
58 }
59 
60 static int gstage_page_size_to_level(unsigned long page_size, u32 *out_level)
61 {
62 	u32 i;
63 	unsigned long psz = 1UL << 12;
64 
65 	for (i = 0; i < gstage_pgd_levels; i++) {
66 		if (page_size == (psz << (i * gstage_index_bits))) {
67 			*out_level = i;
68 			return 0;
69 		}
70 	}
71 
72 	return -EINVAL;
73 }
74 
75 static int gstage_level_to_page_order(u32 level, unsigned long *out_pgorder)
76 {
77 	if (gstage_pgd_levels < level)
78 		return -EINVAL;
79 
80 	*out_pgorder = 12 + (level * gstage_index_bits);
81 	return 0;
82 }
83 
84 static int gstage_level_to_page_size(u32 level, unsigned long *out_pgsize)
85 {
86 	int rc;
87 	unsigned long page_order = PAGE_SHIFT;
88 
89 	rc = gstage_level_to_page_order(level, &page_order);
90 	if (rc)
91 		return rc;
92 
93 	*out_pgsize = BIT(page_order);
94 	return 0;
95 }
96 
97 static bool gstage_get_leaf_entry(struct kvm *kvm, gpa_t addr,
98 				  pte_t **ptepp, u32 *ptep_level)
99 {
100 	pte_t *ptep;
101 	u32 current_level = gstage_pgd_levels - 1;
102 
103 	*ptep_level = current_level;
104 	ptep = (pte_t *)kvm->arch.pgd;
105 	ptep = &ptep[gstage_pte_index(addr, current_level)];
106 	while (ptep && pte_val(*ptep)) {
107 		if (gstage_pte_leaf(ptep)) {
108 			*ptep_level = current_level;
109 			*ptepp = ptep;
110 			return true;
111 		}
112 
113 		if (current_level) {
114 			current_level--;
115 			*ptep_level = current_level;
116 			ptep = (pte_t *)gstage_pte_page_vaddr(*ptep);
117 			ptep = &ptep[gstage_pte_index(addr, current_level)];
118 		} else {
119 			ptep = NULL;
120 		}
121 	}
122 
123 	return false;
124 }
125 
126 static void gstage_remote_tlb_flush(struct kvm *kvm, u32 level, gpa_t addr)
127 {
128 	unsigned long order = PAGE_SHIFT;
129 
130 	if (gstage_level_to_page_order(level, &order))
131 		return;
132 	addr &= ~(BIT(order) - 1);
133 
134 	kvm_riscv_hfence_gvma_vmid_gpa(kvm, -1UL, 0, addr, BIT(order), order);
135 }
136 
137 static int gstage_set_pte(struct kvm *kvm, u32 level,
138 			   struct kvm_mmu_memory_cache *pcache,
139 			   gpa_t addr, const pte_t *new_pte)
140 {
141 	u32 current_level = gstage_pgd_levels - 1;
142 	pte_t *next_ptep = (pte_t *)kvm->arch.pgd;
143 	pte_t *ptep = &next_ptep[gstage_pte_index(addr, current_level)];
144 
145 	if (current_level < level)
146 		return -EINVAL;
147 
148 	while (current_level != level) {
149 		if (gstage_pte_leaf(ptep))
150 			return -EEXIST;
151 
152 		if (!pte_val(*ptep)) {
153 			if (!pcache)
154 				return -ENOMEM;
155 			next_ptep = kvm_mmu_memory_cache_alloc(pcache);
156 			if (!next_ptep)
157 				return -ENOMEM;
158 			*ptep = pfn_pte(PFN_DOWN(__pa(next_ptep)),
159 					__pgprot(_PAGE_TABLE));
160 		} else {
161 			if (gstage_pte_leaf(ptep))
162 				return -EEXIST;
163 			next_ptep = (pte_t *)gstage_pte_page_vaddr(*ptep);
164 		}
165 
166 		current_level--;
167 		ptep = &next_ptep[gstage_pte_index(addr, current_level)];
168 	}
169 
170 	*ptep = *new_pte;
171 	if (gstage_pte_leaf(ptep))
172 		gstage_remote_tlb_flush(kvm, current_level, addr);
173 
174 	return 0;
175 }
176 
177 static int gstage_map_page(struct kvm *kvm,
178 			   struct kvm_mmu_memory_cache *pcache,
179 			   gpa_t gpa, phys_addr_t hpa,
180 			   unsigned long page_size,
181 			   bool page_rdonly, bool page_exec)
182 {
183 	int ret;
184 	u32 level = 0;
185 	pte_t new_pte;
186 	pgprot_t prot;
187 
188 	ret = gstage_page_size_to_level(page_size, &level);
189 	if (ret)
190 		return ret;
191 
192 	/*
193 	 * A RISC-V implementation can choose to either:
194 	 * 1) Update 'A' and 'D' PTE bits in hardware
195 	 * 2) Generate page fault when 'A' and/or 'D' bits are not set
196 	 *    PTE so that software can update these bits.
197 	 *
198 	 * We support both options mentioned above. To achieve this, we
199 	 * always set 'A' and 'D' PTE bits at time of creating G-stage
200 	 * mapping. To support KVM dirty page logging with both options
201 	 * mentioned above, we will write-protect G-stage PTEs to track
202 	 * dirty pages.
203 	 */
204 
205 	if (page_exec) {
206 		if (page_rdonly)
207 			prot = PAGE_READ_EXEC;
208 		else
209 			prot = PAGE_WRITE_EXEC;
210 	} else {
211 		if (page_rdonly)
212 			prot = PAGE_READ;
213 		else
214 			prot = PAGE_WRITE;
215 	}
216 	new_pte = pfn_pte(PFN_DOWN(hpa), prot);
217 	new_pte = pte_mkdirty(new_pte);
218 
219 	return gstage_set_pte(kvm, level, pcache, gpa, &new_pte);
220 }
221 
222 enum gstage_op {
223 	GSTAGE_OP_NOP = 0,	/* Nothing */
224 	GSTAGE_OP_CLEAR,	/* Clear/Unmap */
225 	GSTAGE_OP_WP,		/* Write-protect */
226 };
227 
228 static void gstage_op_pte(struct kvm *kvm, gpa_t addr,
229 			  pte_t *ptep, u32 ptep_level, enum gstage_op op)
230 {
231 	int i, ret;
232 	pte_t *next_ptep;
233 	u32 next_ptep_level;
234 	unsigned long next_page_size, page_size;
235 
236 	ret = gstage_level_to_page_size(ptep_level, &page_size);
237 	if (ret)
238 		return;
239 
240 	BUG_ON(addr & (page_size - 1));
241 
242 	if (!pte_val(*ptep))
243 		return;
244 
245 	if (ptep_level && !gstage_pte_leaf(ptep)) {
246 		next_ptep = (pte_t *)gstage_pte_page_vaddr(*ptep);
247 		next_ptep_level = ptep_level - 1;
248 		ret = gstage_level_to_page_size(next_ptep_level,
249 						&next_page_size);
250 		if (ret)
251 			return;
252 
253 		if (op == GSTAGE_OP_CLEAR)
254 			set_pte(ptep, __pte(0));
255 		for (i = 0; i < PTRS_PER_PTE; i++)
256 			gstage_op_pte(kvm, addr + i * next_page_size,
257 					&next_ptep[i], next_ptep_level, op);
258 		if (op == GSTAGE_OP_CLEAR)
259 			put_page(virt_to_page(next_ptep));
260 	} else {
261 		if (op == GSTAGE_OP_CLEAR)
262 			set_pte(ptep, __pte(0));
263 		else if (op == GSTAGE_OP_WP)
264 			set_pte(ptep, __pte(pte_val(*ptep) & ~_PAGE_WRITE));
265 		gstage_remote_tlb_flush(kvm, ptep_level, addr);
266 	}
267 }
268 
269 static void gstage_unmap_range(struct kvm *kvm, gpa_t start,
270 			       gpa_t size, bool may_block)
271 {
272 	int ret;
273 	pte_t *ptep;
274 	u32 ptep_level;
275 	bool found_leaf;
276 	unsigned long page_size;
277 	gpa_t addr = start, end = start + size;
278 
279 	while (addr < end) {
280 		found_leaf = gstage_get_leaf_entry(kvm, addr,
281 						   &ptep, &ptep_level);
282 		ret = gstage_level_to_page_size(ptep_level, &page_size);
283 		if (ret)
284 			break;
285 
286 		if (!found_leaf)
287 			goto next;
288 
289 		if (!(addr & (page_size - 1)) && ((end - addr) >= page_size))
290 			gstage_op_pte(kvm, addr, ptep,
291 				      ptep_level, GSTAGE_OP_CLEAR);
292 
293 next:
294 		addr += page_size;
295 
296 		/*
297 		 * If the range is too large, release the kvm->mmu_lock
298 		 * to prevent starvation and lockup detector warnings.
299 		 */
300 		if (may_block && addr < end)
301 			cond_resched_lock(&kvm->mmu_lock);
302 	}
303 }
304 
305 static void gstage_wp_range(struct kvm *kvm, gpa_t start, gpa_t end)
306 {
307 	int ret;
308 	pte_t *ptep;
309 	u32 ptep_level;
310 	bool found_leaf;
311 	gpa_t addr = start;
312 	unsigned long page_size;
313 
314 	while (addr < end) {
315 		found_leaf = gstage_get_leaf_entry(kvm, addr,
316 						   &ptep, &ptep_level);
317 		ret = gstage_level_to_page_size(ptep_level, &page_size);
318 		if (ret)
319 			break;
320 
321 		if (!found_leaf)
322 			goto next;
323 
324 		if (!(addr & (page_size - 1)) && ((end - addr) >= page_size))
325 			gstage_op_pte(kvm, addr, ptep,
326 				      ptep_level, GSTAGE_OP_WP);
327 
328 next:
329 		addr += page_size;
330 	}
331 }
332 
333 static void gstage_wp_memory_region(struct kvm *kvm, int slot)
334 {
335 	struct kvm_memslots *slots = kvm_memslots(kvm);
336 	struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
337 	phys_addr_t start = memslot->base_gfn << PAGE_SHIFT;
338 	phys_addr_t end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
339 
340 	spin_lock(&kvm->mmu_lock);
341 	gstage_wp_range(kvm, start, end);
342 	spin_unlock(&kvm->mmu_lock);
343 	kvm_flush_remote_tlbs(kvm);
344 }
345 
346 int kvm_riscv_gstage_ioremap(struct kvm *kvm, gpa_t gpa,
347 			     phys_addr_t hpa, unsigned long size,
348 			     bool writable, bool in_atomic)
349 {
350 	pte_t pte;
351 	int ret = 0;
352 	unsigned long pfn;
353 	phys_addr_t addr, end;
354 	struct kvm_mmu_memory_cache pcache = {
355 		.gfp_custom = (in_atomic) ? GFP_ATOMIC | __GFP_ACCOUNT : 0,
356 		.gfp_zero = __GFP_ZERO,
357 	};
358 
359 	end = (gpa + size + PAGE_SIZE - 1) & PAGE_MASK;
360 	pfn = __phys_to_pfn(hpa);
361 
362 	for (addr = gpa; addr < end; addr += PAGE_SIZE) {
363 		pte = pfn_pte(pfn, PAGE_KERNEL_IO);
364 
365 		if (!writable)
366 			pte = pte_wrprotect(pte);
367 
368 		ret = kvm_mmu_topup_memory_cache(&pcache, gstage_pgd_levels);
369 		if (ret)
370 			goto out;
371 
372 		spin_lock(&kvm->mmu_lock);
373 		ret = gstage_set_pte(kvm, 0, &pcache, addr, &pte);
374 		spin_unlock(&kvm->mmu_lock);
375 		if (ret)
376 			goto out;
377 
378 		pfn++;
379 	}
380 
381 out:
382 	kvm_mmu_free_memory_cache(&pcache);
383 	return ret;
384 }
385 
386 void kvm_riscv_gstage_iounmap(struct kvm *kvm, gpa_t gpa, unsigned long size)
387 {
388 	spin_lock(&kvm->mmu_lock);
389 	gstage_unmap_range(kvm, gpa, size, false);
390 	spin_unlock(&kvm->mmu_lock);
391 }
392 
393 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
394 					     struct kvm_memory_slot *slot,
395 					     gfn_t gfn_offset,
396 					     unsigned long mask)
397 {
398 	phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
399 	phys_addr_t start = (base_gfn +  __ffs(mask)) << PAGE_SHIFT;
400 	phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
401 
402 	gstage_wp_range(kvm, start, end);
403 }
404 
405 void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot)
406 {
407 }
408 
409 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free)
410 {
411 }
412 
413 void kvm_arch_memslots_updated(struct kvm *kvm, u64 gen)
414 {
415 }
416 
417 void kvm_arch_flush_shadow_all(struct kvm *kvm)
418 {
419 	kvm_riscv_gstage_free_pgd(kvm);
420 }
421 
422 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
423 				   struct kvm_memory_slot *slot)
424 {
425 	gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
426 	phys_addr_t size = slot->npages << PAGE_SHIFT;
427 
428 	spin_lock(&kvm->mmu_lock);
429 	gstage_unmap_range(kvm, gpa, size, false);
430 	spin_unlock(&kvm->mmu_lock);
431 }
432 
433 void kvm_arch_commit_memory_region(struct kvm *kvm,
434 				struct kvm_memory_slot *old,
435 				const struct kvm_memory_slot *new,
436 				enum kvm_mr_change change)
437 {
438 	/*
439 	 * At this point memslot has been committed and there is an
440 	 * allocated dirty_bitmap[], dirty pages will be tracked while
441 	 * the memory slot is write protected.
442 	 */
443 	if (change != KVM_MR_DELETE && new->flags & KVM_MEM_LOG_DIRTY_PAGES)
444 		gstage_wp_memory_region(kvm, new->id);
445 }
446 
447 int kvm_arch_prepare_memory_region(struct kvm *kvm,
448 				const struct kvm_memory_slot *old,
449 				struct kvm_memory_slot *new,
450 				enum kvm_mr_change change)
451 {
452 	hva_t hva, reg_end, size;
453 	gpa_t base_gpa;
454 	bool writable;
455 	int ret = 0;
456 
457 	if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
458 			change != KVM_MR_FLAGS_ONLY)
459 		return 0;
460 
461 	/*
462 	 * Prevent userspace from creating a memory region outside of the GPA
463 	 * space addressable by the KVM guest GPA space.
464 	 */
465 	if ((new->base_gfn + new->npages) >=
466 	    (gstage_gpa_size >> PAGE_SHIFT))
467 		return -EFAULT;
468 
469 	hva = new->userspace_addr;
470 	size = new->npages << PAGE_SHIFT;
471 	reg_end = hva + size;
472 	base_gpa = new->base_gfn << PAGE_SHIFT;
473 	writable = !(new->flags & KVM_MEM_READONLY);
474 
475 	mmap_read_lock(current->mm);
476 
477 	/*
478 	 * A memory region could potentially cover multiple VMAs, and
479 	 * any holes between them, so iterate over all of them to find
480 	 * out if we can map any of them right now.
481 	 *
482 	 *     +--------------------------------------------+
483 	 * +---------------+----------------+   +----------------+
484 	 * |   : VMA 1     |      VMA 2     |   |    VMA 3  :    |
485 	 * +---------------+----------------+   +----------------+
486 	 *     |               memory region                |
487 	 *     +--------------------------------------------+
488 	 */
489 	do {
490 		struct vm_area_struct *vma = find_vma(current->mm, hva);
491 		hva_t vm_start, vm_end;
492 
493 		if (!vma || vma->vm_start >= reg_end)
494 			break;
495 
496 		/*
497 		 * Mapping a read-only VMA is only allowed if the
498 		 * memory region is configured as read-only.
499 		 */
500 		if (writable && !(vma->vm_flags & VM_WRITE)) {
501 			ret = -EPERM;
502 			break;
503 		}
504 
505 		/* Take the intersection of this VMA with the memory region */
506 		vm_start = max(hva, vma->vm_start);
507 		vm_end = min(reg_end, vma->vm_end);
508 
509 		if (vma->vm_flags & VM_PFNMAP) {
510 			gpa_t gpa = base_gpa + (vm_start - hva);
511 			phys_addr_t pa;
512 
513 			pa = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT;
514 			pa += vm_start - vma->vm_start;
515 
516 			/* IO region dirty page logging not allowed */
517 			if (new->flags & KVM_MEM_LOG_DIRTY_PAGES) {
518 				ret = -EINVAL;
519 				goto out;
520 			}
521 
522 			ret = kvm_riscv_gstage_ioremap(kvm, gpa, pa,
523 						       vm_end - vm_start,
524 						       writable, false);
525 			if (ret)
526 				break;
527 		}
528 		hva = vm_end;
529 	} while (hva < reg_end);
530 
531 	if (change == KVM_MR_FLAGS_ONLY)
532 		goto out;
533 
534 	if (ret)
535 		kvm_riscv_gstage_iounmap(kvm, base_gpa, size);
536 
537 out:
538 	mmap_read_unlock(current->mm);
539 	return ret;
540 }
541 
542 bool kvm_unmap_gfn_range(struct kvm *kvm, struct kvm_gfn_range *range)
543 {
544 	if (!kvm->arch.pgd)
545 		return false;
546 
547 	gstage_unmap_range(kvm, range->start << PAGE_SHIFT,
548 			   (range->end - range->start) << PAGE_SHIFT,
549 			   range->may_block);
550 	return false;
551 }
552 
553 bool kvm_set_spte_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
554 {
555 	int ret;
556 	kvm_pfn_t pfn = pte_pfn(range->arg.pte);
557 
558 	if (!kvm->arch.pgd)
559 		return false;
560 
561 	WARN_ON(range->end - range->start != 1);
562 
563 	ret = gstage_map_page(kvm, NULL, range->start << PAGE_SHIFT,
564 			      __pfn_to_phys(pfn), PAGE_SIZE, true, true);
565 	if (ret) {
566 		kvm_debug("Failed to map G-stage page (error %d)\n", ret);
567 		return true;
568 	}
569 
570 	return false;
571 }
572 
573 bool kvm_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
574 {
575 	pte_t *ptep;
576 	u32 ptep_level = 0;
577 	u64 size = (range->end - range->start) << PAGE_SHIFT;
578 
579 	if (!kvm->arch.pgd)
580 		return false;
581 
582 	WARN_ON(size != PAGE_SIZE && size != PMD_SIZE && size != PUD_SIZE);
583 
584 	if (!gstage_get_leaf_entry(kvm, range->start << PAGE_SHIFT,
585 				   &ptep, &ptep_level))
586 		return false;
587 
588 	return ptep_test_and_clear_young(NULL, 0, ptep);
589 }
590 
591 bool kvm_test_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
592 {
593 	pte_t *ptep;
594 	u32 ptep_level = 0;
595 	u64 size = (range->end - range->start) << PAGE_SHIFT;
596 
597 	if (!kvm->arch.pgd)
598 		return false;
599 
600 	WARN_ON(size != PAGE_SIZE && size != PMD_SIZE && size != PUD_SIZE);
601 
602 	if (!gstage_get_leaf_entry(kvm, range->start << PAGE_SHIFT,
603 				   &ptep, &ptep_level))
604 		return false;
605 
606 	return pte_young(*ptep);
607 }
608 
609 int kvm_riscv_gstage_map(struct kvm_vcpu *vcpu,
610 			 struct kvm_memory_slot *memslot,
611 			 gpa_t gpa, unsigned long hva, bool is_write)
612 {
613 	int ret;
614 	kvm_pfn_t hfn;
615 	bool writable;
616 	short vma_pageshift;
617 	gfn_t gfn = gpa >> PAGE_SHIFT;
618 	struct vm_area_struct *vma;
619 	struct kvm *kvm = vcpu->kvm;
620 	struct kvm_mmu_memory_cache *pcache = &vcpu->arch.mmu_page_cache;
621 	bool logging = (memslot->dirty_bitmap &&
622 			!(memslot->flags & KVM_MEM_READONLY)) ? true : false;
623 	unsigned long vma_pagesize, mmu_seq;
624 
625 	/* We need minimum second+third level pages */
626 	ret = kvm_mmu_topup_memory_cache(pcache, gstage_pgd_levels);
627 	if (ret) {
628 		kvm_err("Failed to topup G-stage cache\n");
629 		return ret;
630 	}
631 
632 	mmap_read_lock(current->mm);
633 
634 	vma = vma_lookup(current->mm, hva);
635 	if (unlikely(!vma)) {
636 		kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
637 		mmap_read_unlock(current->mm);
638 		return -EFAULT;
639 	}
640 
641 	if (is_vm_hugetlb_page(vma))
642 		vma_pageshift = huge_page_shift(hstate_vma(vma));
643 	else
644 		vma_pageshift = PAGE_SHIFT;
645 	vma_pagesize = 1ULL << vma_pageshift;
646 	if (logging || (vma->vm_flags & VM_PFNMAP))
647 		vma_pagesize = PAGE_SIZE;
648 
649 	if (vma_pagesize == PMD_SIZE || vma_pagesize == PUD_SIZE)
650 		gfn = (gpa & huge_page_mask(hstate_vma(vma))) >> PAGE_SHIFT;
651 
652 	/*
653 	 * Read mmu_invalidate_seq so that KVM can detect if the results of
654 	 * vma_lookup() or gfn_to_pfn_prot() become stale priort to acquiring
655 	 * kvm->mmu_lock.
656 	 *
657 	 * Rely on mmap_read_unlock() for an implicit smp_rmb(), which pairs
658 	 * with the smp_wmb() in kvm_mmu_invalidate_end().
659 	 */
660 	mmu_seq = kvm->mmu_invalidate_seq;
661 	mmap_read_unlock(current->mm);
662 
663 	if (vma_pagesize != PUD_SIZE &&
664 	    vma_pagesize != PMD_SIZE &&
665 	    vma_pagesize != PAGE_SIZE) {
666 		kvm_err("Invalid VMA page size 0x%lx\n", vma_pagesize);
667 		return -EFAULT;
668 	}
669 
670 	hfn = gfn_to_pfn_prot(kvm, gfn, is_write, &writable);
671 	if (hfn == KVM_PFN_ERR_HWPOISON) {
672 		send_sig_mceerr(BUS_MCEERR_AR, (void __user *)hva,
673 				vma_pageshift, current);
674 		return 0;
675 	}
676 	if (is_error_noslot_pfn(hfn))
677 		return -EFAULT;
678 
679 	/*
680 	 * If logging is active then we allow writable pages only
681 	 * for write faults.
682 	 */
683 	if (logging && !is_write)
684 		writable = false;
685 
686 	spin_lock(&kvm->mmu_lock);
687 
688 	if (mmu_invalidate_retry(kvm, mmu_seq))
689 		goto out_unlock;
690 
691 	if (writable) {
692 		kvm_set_pfn_dirty(hfn);
693 		mark_page_dirty(kvm, gfn);
694 		ret = gstage_map_page(kvm, pcache, gpa, hfn << PAGE_SHIFT,
695 				      vma_pagesize, false, true);
696 	} else {
697 		ret = gstage_map_page(kvm, pcache, gpa, hfn << PAGE_SHIFT,
698 				      vma_pagesize, true, true);
699 	}
700 
701 	if (ret)
702 		kvm_err("Failed to map in G-stage\n");
703 
704 out_unlock:
705 	spin_unlock(&kvm->mmu_lock);
706 	kvm_set_pfn_accessed(hfn);
707 	kvm_release_pfn_clean(hfn);
708 	return ret;
709 }
710 
711 int kvm_riscv_gstage_alloc_pgd(struct kvm *kvm)
712 {
713 	struct page *pgd_page;
714 
715 	if (kvm->arch.pgd != NULL) {
716 		kvm_err("kvm_arch already initialized?\n");
717 		return -EINVAL;
718 	}
719 
720 	pgd_page = alloc_pages(GFP_KERNEL | __GFP_ZERO,
721 				get_order(gstage_pgd_size));
722 	if (!pgd_page)
723 		return -ENOMEM;
724 	kvm->arch.pgd = page_to_virt(pgd_page);
725 	kvm->arch.pgd_phys = page_to_phys(pgd_page);
726 
727 	return 0;
728 }
729 
730 void kvm_riscv_gstage_free_pgd(struct kvm *kvm)
731 {
732 	void *pgd = NULL;
733 
734 	spin_lock(&kvm->mmu_lock);
735 	if (kvm->arch.pgd) {
736 		gstage_unmap_range(kvm, 0UL, gstage_gpa_size, false);
737 		pgd = READ_ONCE(kvm->arch.pgd);
738 		kvm->arch.pgd = NULL;
739 		kvm->arch.pgd_phys = 0;
740 	}
741 	spin_unlock(&kvm->mmu_lock);
742 
743 	if (pgd)
744 		free_pages((unsigned long)pgd, get_order(gstage_pgd_size));
745 }
746 
747 void kvm_riscv_gstage_update_hgatp(struct kvm_vcpu *vcpu)
748 {
749 	unsigned long hgatp = gstage_mode;
750 	struct kvm_arch *k = &vcpu->kvm->arch;
751 
752 	hgatp |= (READ_ONCE(k->vmid.vmid) << HGATP_VMID_SHIFT) & HGATP_VMID;
753 	hgatp |= (k->pgd_phys >> PAGE_SHIFT) & HGATP_PPN;
754 
755 	csr_write(CSR_HGATP, hgatp);
756 
757 	if (!kvm_riscv_gstage_vmid_bits())
758 		kvm_riscv_local_hfence_gvma_all();
759 }
760 
761 void __init kvm_riscv_gstage_mode_detect(void)
762 {
763 #ifdef CONFIG_64BIT
764 	/* Try Sv57x4 G-stage mode */
765 	csr_write(CSR_HGATP, HGATP_MODE_SV57X4 << HGATP_MODE_SHIFT);
766 	if ((csr_read(CSR_HGATP) >> HGATP_MODE_SHIFT) == HGATP_MODE_SV57X4) {
767 		gstage_mode = (HGATP_MODE_SV57X4 << HGATP_MODE_SHIFT);
768 		gstage_pgd_levels = 5;
769 		goto skip_sv48x4_test;
770 	}
771 
772 	/* Try Sv48x4 G-stage mode */
773 	csr_write(CSR_HGATP, HGATP_MODE_SV48X4 << HGATP_MODE_SHIFT);
774 	if ((csr_read(CSR_HGATP) >> HGATP_MODE_SHIFT) == HGATP_MODE_SV48X4) {
775 		gstage_mode = (HGATP_MODE_SV48X4 << HGATP_MODE_SHIFT);
776 		gstage_pgd_levels = 4;
777 	}
778 skip_sv48x4_test:
779 
780 	csr_write(CSR_HGATP, 0);
781 	kvm_riscv_local_hfence_gvma_all();
782 #endif
783 }
784 
785 unsigned long __init kvm_riscv_gstage_mode(void)
786 {
787 	return gstage_mode >> HGATP_MODE_SHIFT;
788 }
789 
790 int kvm_riscv_gstage_gpa_bits(void)
791 {
792 	return gstage_gpa_bits;
793 }
794