개념 설명 전체
arch/arm64/kvm/mmu.c · v6.18.37
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Copyright (C) 2012 - Virtual Open Systems and Columbia University
4 * Author: Christoffer Dall <[email protected]>
5 */
6
7 #include <linux/acpi.h>
8 #include <linux/mman.h>
9 #include <linux/kvm_host.h>
10 #include <linux/io.h>
11 #include <linux/hugetlb.h>
12 #include <linux/sched/signal.h>
13 #include <trace/events/kvm.h>
14 #include <asm/acpi.h>
15 #include <asm/pgalloc.h>
16 #include <asm/cacheflush.h>
17 #include <asm/kvm_arm.h>
18 #include <asm/kvm_mmu.h>
19 #include <asm/kvm_pgtable.h>
20 #include <asm/kvm_pkvm.h>
21 #include <asm/kvm_asm.h>
22 #include <asm/kvm_emulate.h>
23 #include <asm/virt.h>
24
25 #include "trace.h"
26
27 static struct kvm_pgtable *hyp_pgtable;
28 static DEFINE_MUTEX(kvm_hyp_pgd_mutex);
29
30 static unsigned long __ro_after_init hyp_idmap_start;
31 static unsigned long __ro_after_init hyp_idmap_end;
32 static phys_addr_t __ro_after_init hyp_idmap_vector;
33
34 u32 __ro_after_init __hyp_va_bits;
35
36 static unsigned long __ro_after_init io_map_base;
37
38 #define KVM_PGT_FN(fn) (!is_protected_kvm_enabled() ? fn : p ## fn)
39
40 static phys_addr_t __stage2_range_addr_end(phys_addr_t addr, phys_addr_t end,
41 phys_addr_t size)
42 {
43 phys_addr_t boundary = ALIGN_DOWN(addr + size, size);
44
45 return (boundary - 1 < end - 1) ? boundary : end;
46 }
47
48 static phys_addr_t stage2_range_addr_end(phys_addr_t addr, phys_addr_t end)
49 {
50 phys_addr_t size = kvm_granule_size(KVM_PGTABLE_MIN_BLOCK_LEVEL);
51
52 return __stage2_range_addr_end(addr, end, size);
53 }
54
55 /*
56 * Release kvm_mmu_lock periodically if the memory region is large. Otherwise,
57 * we may see kernel panics with CONFIG_DETECT_HUNG_TASK,
58 * CONFIG_LOCKUP_DETECTOR, CONFIG_LOCKDEP. Additionally, holding the lock too
59 * long will also starve other vCPUs. We have to also make sure that the page
60 * tables are not freed while we released the lock.
61 */
62 static int stage2_apply_range(struct kvm_s2_mmu *mmu, phys_addr_t addr,
63 phys_addr_t end,
64 int (*fn)(struct kvm_pgtable *, u64, u64),
65 bool resched)
66 {
67 struct kvm *kvm = kvm_s2_mmu_to_kvm(mmu);
68 int ret;
69 u64 next;
70
71 do {
72 struct kvm_pgtable *pgt = mmu->pgt;
73 if (!pgt)
74 return -EINVAL;
75
76 next = stage2_range_addr_end(addr, end);
77 ret = fn(pgt, addr, next - addr);
78 if (ret)
79 break;
80
81 if (resched && next != end)
82 cond_resched_rwlock_write(&kvm->mmu_lock);
83 } while (addr = next, addr != end);
84
85 return ret;
86 }
87
88 #define stage2_apply_range_resched(mmu, addr, end, fn) \
89 stage2_apply_range(mmu, addr, end, fn, true)
90
91 /*
92 * Get the maximum number of page-tables pages needed to split a range
93 * of blocks into PAGE_SIZE PTEs. It assumes the range is already
94 * mapped at level 2, or at level 1 if allowed.
95 */
96 static int kvm_mmu_split_nr_page_tables(u64 range)
97 {
98 int n = 0;
99
100 if (KVM_PGTABLE_MIN_BLOCK_LEVEL < 2)
101 n += DIV_ROUND_UP(range, PUD_SIZE);
102 n += DIV_ROUND_UP(range, PMD_SIZE);
103 return n;
104 }
105
106 static bool need_split_memcache_topup_or_resched(struct kvm *kvm)
107 {
108 struct kvm_mmu_memory_cache *cache;
109 u64 chunk_size, min;
110
111 if (need_resched() || rwlock_needbreak(&kvm->mmu_lock))
112 return true;
113
114 chunk_size = kvm->arch.mmu.split_page_chunk_size;
115 min = kvm_mmu_split_nr_page_tables(chunk_size);
116 cache = &kvm->arch.mmu.split_page_cache;
117 return kvm_mmu_memory_cache_nr_free_objects(cache) < min;
118 }
119
120 static int kvm_mmu_split_huge_pages(struct kvm *kvm, phys_addr_t addr,
121 phys_addr_t end)
122 {
123 struct kvm_mmu_memory_cache *cache;
124 struct kvm_pgtable *pgt;
125 int ret, cache_capacity;
126 u64 next, chunk_size;
127
128 lockdep_assert_held_write(&kvm->mmu_lock);
129
130 chunk_size = kvm->arch.mmu.split_page_chunk_size;
131 cache_capacity = kvm_mmu_split_nr_page_tables(chunk_size);
132
133 if (chunk_size == 0)
134 return 0;
135
136 cache = &kvm->arch.mmu.split_page_cache;
137
138 do {
139 if (need_split_memcache_topup_or_resched(kvm)) {
140 write_unlock(&kvm->mmu_lock);
141 cond_resched();
142 /* Eager page splitting is best-effort. */
143 ret = __kvm_mmu_topup_memory_cache(cache,
144 cache_capacity,
145 cache_capacity);
146 write_lock(&kvm->mmu_lock);
147 if (ret)
148 break;
149 }
150
151 pgt = kvm->arch.mmu.pgt;
152 if (!pgt)
153 return -EINVAL;
154
155 next = __stage2_range_addr_end(addr, end, chunk_size);
156 ret = KVM_PGT_FN(kvm_pgtable_stage2_split)(pgt, addr, next - addr, cache);
157 if (ret)
158 break;
159 } while (addr = next, addr != end);
160
161 return ret;
162 }
163
164 static bool memslot_is_logging(struct kvm_memory_slot *memslot)
165 {
166 return memslot->dirty_bitmap && !(memslot->flags & KVM_MEM_READONLY);
167 }
168
169 /**
170 * kvm_arch_flush_remote_tlbs() - flush all VM TLB entries for v7/8
171 * @kvm: pointer to kvm structure.
172 *
173 * Interface to HYP function to flush all VM TLB entries
174 */
175 int kvm_arch_flush_remote_tlbs(struct kvm *kvm)
176 {
177 if (is_protected_kvm_enabled())
178 kvm_call_hyp_nvhe(__pkvm_tlb_flush_vmid, kvm->arch.pkvm.handle);
179 else
180 kvm_call_hyp(__kvm_tlb_flush_vmid, &kvm->arch.mmu);
181 return 0;
182 }
183
184 int kvm_arch_flush_remote_tlbs_range(struct kvm *kvm,
185 gfn_t gfn, u64 nr_pages)
186 {
187 u64 size = nr_pages << PAGE_SHIFT;
188 u64 addr = gfn << PAGE_SHIFT;
189
190 if (is_protected_kvm_enabled())
191 kvm_call_hyp_nvhe(__pkvm_tlb_flush_vmid, kvm->arch.pkvm.handle);
192 else
193 kvm_tlb_flush_vmid_range(&kvm->arch.mmu, addr, size);
194 return 0;
195 }
196
197 static void *stage2_memcache_zalloc_page(void *arg)
198 {
199 struct kvm_mmu_memory_cache *mc = arg;
200 void *virt;
201
202 /* Allocated with __GFP_ZERO, so no need to zero */
203 virt = kvm_mmu_memory_cache_alloc(mc);
204 if (virt)
205 kvm_account_pgtable_pages(virt, 1);
206 return virt;
207 }
208
209 static void *kvm_host_zalloc_pages_exact(size_t size)
210 {
211 return alloc_pages_exact(size, GFP_KERNEL_ACCOUNT | __GFP_ZERO);
212 }
213
214 static void *kvm_s2_zalloc_pages_exact(size_t size)
215 {
216 void *virt = kvm_host_zalloc_pages_exact(size);
217
218 if (virt)
219 kvm_account_pgtable_pages(virt, (size >> PAGE_SHIFT));
220 return virt;
221 }
222
223 static void kvm_s2_free_pages_exact(void *virt, size_t size)
224 {
225 kvm_account_pgtable_pages(virt, -(size >> PAGE_SHIFT));
226 free_pages_exact(virt, size);
227 }
228
229 static struct kvm_pgtable_mm_ops kvm_s2_mm_ops;
230
231 static void stage2_free_unlinked_table_rcu_cb(struct rcu_head *head)
232 {
233 struct page *page = container_of(head, struct page, rcu_head);
234 void *pgtable = page_to_virt(page);
235 s8 level = page_private(page);
236
237 KVM_PGT_FN(kvm_pgtable_stage2_free_unlinked)(&kvm_s2_mm_ops, pgtable, level);
238 }
239
240 static void stage2_free_unlinked_table(void *addr, s8 level)
241 {
242 struct page *page = virt_to_page(addr);
243
244 set_page_private(page, (unsigned long)level);
245 call_rcu(&page->rcu_head, stage2_free_unlinked_table_rcu_cb);
246 }
247
248 static void kvm_host_get_page(void *addr)
249 {
250 get_page(virt_to_page(addr));
251 }
252
253 static void kvm_host_put_page(void *addr)
254 {
255 put_page(virt_to_page(addr));
256 }
257
258 static void kvm_s2_put_page(void *addr)
259 {
260 struct page *p = virt_to_page(addr);
261 /* Dropping last refcount, the page will be freed */
262 if (page_count(p) == 1)
263 kvm_account_pgtable_pages(addr, -1);
264 put_page(p);
265 }
266
267 static int kvm_host_page_count(void *addr)
268 {
269 return page_count(virt_to_page(addr));
270 }
271
272 static phys_addr_t kvm_host_pa(void *addr)
273 {
274 return __pa(addr);
275 }
276
277 static void *kvm_host_va(phys_addr_t phys)
278 {
279 return __va(phys);
280 }
281
282 static void clean_dcache_guest_page(void *va, size_t size)
283 {
284 __clean_dcache_guest_page(va, size);
285 }
286
287 static void invalidate_icache_guest_page(void *va, size_t size)
288 {
289 __invalidate_icache_guest_page(va, size);
290 }
291
292 /*
293 * Unmapping vs dcache management:
294 *
295 * If a guest maps certain memory pages as uncached, all writes will
296 * bypass the data cache and go directly to RAM. However, the CPUs
297 * can still speculate reads (not writes) and fill cache lines with
298 * data.
299 *
300 * Those cache lines will be *clean* cache lines though, so a
301 * clean+invalidate operation is equivalent to an invalidate
302 * operation, because no cache lines are marked dirty.
303 *
304 * Those clean cache lines could be filled prior to an uncached write
305 * by the guest, and the cache coherent IO subsystem would therefore
306 * end up writing old data to disk.
307 *
308 * This is why right after unmapping a page/section and invalidating
309 * the corresponding TLBs, we flush to make sure the IO subsystem will
310 * never hit in the cache.
311 *
312 * This is all avoided on systems that have ARM64_HAS_STAGE2_FWB, as
313 * we then fully enforce cacheability of RAM, no matter what the guest
314 * does.
315 */
316 /**
317 * __unmap_stage2_range -- Clear stage2 page table entries to unmap a range
318 * @mmu: The KVM stage-2 MMU pointer
319 * @start: The intermediate physical base address of the range to unmap
320 * @size: The size of the area to unmap
321 * @may_block: Whether or not we are permitted to block
322 *
323 * Clear a range of stage-2 mappings, lowering the various ref-counts. Must
324 * be called while holding mmu_lock (unless for freeing the stage2 pgd before
325 * destroying the VM), otherwise another faulting VCPU may come in and mess
326 * with things behind our backs.
327 */
328 static void __unmap_stage2_range(struct kvm_s2_mmu *mmu, phys_addr_t start, u64 size,
329 bool may_block)
330 {
331 struct kvm *kvm = kvm_s2_mmu_to_kvm(mmu);
332 phys_addr_t end = start + size;
333
334 lockdep_assert_held_write(&kvm->mmu_lock);
335 WARN_ON(size & ~PAGE_MASK);
336 WARN_ON(stage2_apply_range(mmu, start, end, KVM_PGT_FN(kvm_pgtable_stage2_unmap),
337 may_block));
338 }
339
340 void kvm_stage2_unmap_range(struct kvm_s2_mmu *mmu, phys_addr_t start,
341 u64 size, bool may_block)
342 {
343 __unmap_stage2_range(mmu, start, size, may_block);
344 }
345
346 void kvm_stage2_flush_range(struct kvm_s2_mmu *mmu, phys_addr_t addr, phys_addr_t end)
347 {
348 stage2_apply_range_resched(mmu, addr, end, KVM_PGT_FN(kvm_pgtable_stage2_flush));
349 }
350
351 static void stage2_flush_memslot(struct kvm *kvm,
352 struct kvm_memory_slot *memslot)
353 {
354 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
355 phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
356
357 kvm_stage2_flush_range(&kvm->arch.mmu, addr, end);
358 }
359
360 /**
361 * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
362 * @kvm: The struct kvm pointer
363 *
364 * Go through the stage 2 page tables and invalidate any cache lines
365 * backing memory already mapped to the VM.
366 */
367 static void stage2_flush_vm(struct kvm *kvm)
368 {
369 struct kvm_memslots *slots;
370 struct kvm_memory_slot *memslot;
371 int idx, bkt;
372
373 idx = srcu_read_lock(&kvm->srcu);
374 write_lock(&kvm->mmu_lock);
375
376 slots = kvm_memslots(kvm);
377 kvm_for_each_memslot(memslot, bkt, slots)
378 stage2_flush_memslot(kvm, memslot);
379
380 kvm_nested_s2_flush(kvm);
381
382 write_unlock(&kvm->mmu_lock);
383 srcu_read_unlock(&kvm->srcu, idx);
384 }
385
386 /**
387 * free_hyp_pgds - free Hyp-mode page tables
388 */
389 void __init free_hyp_pgds(void)
390 {
391 mutex_lock(&kvm_hyp_pgd_mutex);
392 if (hyp_pgtable) {
393 kvm_pgtable_hyp_destroy(hyp_pgtable);
394 kfree(hyp_pgtable);
395 hyp_pgtable = NULL;
396 }
397 mutex_unlock(&kvm_hyp_pgd_mutex);
398 }
399
400 static bool kvm_host_owns_hyp_mappings(void)
401 {
402 if (is_kernel_in_hyp_mode())
403 return false;
404
405 if (static_branch_likely(&kvm_protected_mode_initialized))
406 return false;
407
408 /*
409 * This can happen at boot time when __create_hyp_mappings() is called
410 * after the hyp protection has been enabled, but the static key has
411 * not been flipped yet.
412 */
413 if (!hyp_pgtable && is_protected_kvm_enabled())
414 return false;
415
416 WARN_ON(!hyp_pgtable);
417
418 return true;
419 }
420
421 int __create_hyp_mappings(unsigned long start, unsigned long size,
422 unsigned long phys, enum kvm_pgtable_prot prot)
423 {
424 int err;
425
426 if (WARN_ON(!kvm_host_owns_hyp_mappings()))
427 return -EINVAL;
428
429 mutex_lock(&kvm_hyp_pgd_mutex);
430 err = kvm_pgtable_hyp_map(hyp_pgtable, start, size, phys, prot);
431 mutex_unlock(&kvm_hyp_pgd_mutex);
432
433 return err;
434 }
435
436 static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
437 {
438 if (!is_vmalloc_addr(kaddr)) {
439 BUG_ON(!virt_addr_valid(kaddr));
440 return __pa(kaddr);
441 } else {
442 return page_to_phys(vmalloc_to_page(kaddr)) +
443 offset_in_page(kaddr);
444 }
445 }
446
447 struct hyp_shared_pfn {
448 u64 pfn;
449 int count;
450 struct rb_node node;
451 };
452
453 static DEFINE_MUTEX(hyp_shared_pfns_lock);
454 static struct rb_root hyp_shared_pfns = RB_ROOT;
455
456 static struct hyp_shared_pfn *find_shared_pfn(u64 pfn, struct rb_node ***node,
457 struct rb_node **parent)
458 {
459 struct hyp_shared_pfn *this;
460
461 *node = &hyp_shared_pfns.rb_node;
462 *parent = NULL;
463 while (**node) {
464 this = container_of(**node, struct hyp_shared_pfn, node);
465 *parent = **node;
466 if (this->pfn < pfn)
467 *node = &((**node)->rb_left);
468 else if (this->pfn > pfn)
469 *node = &((**node)->rb_right);
470 else
471 return this;
472 }
473
474 return NULL;
475 }
476
477 static int share_pfn_hyp(u64 pfn)
478 {
479 struct rb_node **node, *parent;
480 struct hyp_shared_pfn *this;
481 int ret = 0;
482
483 mutex_lock(&hyp_shared_pfns_lock);
484 this = find_shared_pfn(pfn, &node, &parent);
485 if (this) {
486 this->count++;
487 goto unlock;
488 }
489
490 this = kzalloc(sizeof(*this), GFP_KERNEL);
491 if (!this) {
492 ret = -ENOMEM;
493 goto unlock;
494 }
495
496 this->pfn = pfn;
497 this->count = 1;
498 rb_link_node(&this->node, parent, node);
499 rb_insert_color(&this->node, &hyp_shared_pfns);
500 ret = kvm_call_hyp_nvhe(__pkvm_host_share_hyp, pfn, 1);
501 unlock:
502 mutex_unlock(&hyp_shared_pfns_lock);
503
504 return ret;
505 }
506
507 static int unshare_pfn_hyp(u64 pfn)
508 {
509 struct rb_node **node, *parent;
510 struct hyp_shared_pfn *this;
511 int ret = 0;
512
513 mutex_lock(&hyp_shared_pfns_lock);
514 this = find_shared_pfn(pfn, &node, &parent);
515 if (WARN_ON(!this)) {
516 ret = -ENOENT;
517 goto unlock;
518 }
519
520 this->count--;
521 if (this->count)
522 goto unlock;
523
524 rb_erase(&this->node, &hyp_shared_pfns);
525 kfree(this);
526 ret = kvm_call_hyp_nvhe(__pkvm_host_unshare_hyp, pfn, 1);
527 unlock:
528 mutex_unlock(&hyp_shared_pfns_lock);
529
530 return ret;
531 }
532
533 int kvm_share_hyp(void *from, void *to)
534 {
535 phys_addr_t start, end, cur;
536 u64 pfn;
537 int ret;
538
539 if (is_kernel_in_hyp_mode())
540 return 0;
541
542 /*
543 * The share hcall maps things in the 'fixed-offset' region of the hyp
544 * VA space, so we can only share physically contiguous data-structures
545 * for now.
546 */
547 if (is_vmalloc_or_module_addr(from) || is_vmalloc_or_module_addr(to))
548 return -EINVAL;
549
550 if (kvm_host_owns_hyp_mappings())
551 return create_hyp_mappings(from, to, PAGE_HYP);
552
553 start = ALIGN_DOWN(__pa(from), PAGE_SIZE);
554 end = PAGE_ALIGN(__pa(to));
555 for (cur = start; cur < end; cur += PAGE_SIZE) {
556 pfn = __phys_to_pfn(cur);
557 ret = share_pfn_hyp(pfn);
558 if (ret)
559 return ret;
560 }
561
562 return 0;
563 }
564
565 void kvm_unshare_hyp(void *from, void *to)
566 {
567 phys_addr_t start, end, cur;
568 u64 pfn;
569
570 if (is_kernel_in_hyp_mode() || kvm_host_owns_hyp_mappings() || !from)
571 return;
572
573 start = ALIGN_DOWN(__pa(from), PAGE_SIZE);
574 end = PAGE_ALIGN(__pa(to));
575 for (cur = start; cur < end; cur += PAGE_SIZE) {
576 pfn = __phys_to_pfn(cur);
577 WARN_ON(unshare_pfn_hyp(pfn));
578 }
579 }
580
581 /**
582 * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
583 * @from: The virtual kernel start address of the range
584 * @to: The virtual kernel end address of the range (exclusive)
585 * @prot: The protection to be applied to this range
586 *
587 * The same virtual address as the kernel virtual address is also used
588 * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
589 * physical pages.
590 */
591 int create_hyp_mappings(void *from, void *to, enum kvm_pgtable_prot prot)
592 {
593 phys_addr_t phys_addr;
594 unsigned long virt_addr;
595 unsigned long start = kern_hyp_va((unsigned long)from);
596 unsigned long end = kern_hyp_va((unsigned long)to);
597
598 if (is_kernel_in_hyp_mode())
599 return 0;
600
601 if (!kvm_host_owns_hyp_mappings())
602 return -EPERM;
603
604 start = start & PAGE_MASK;
605 end = PAGE_ALIGN(end);
606
607 for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
608 int err;
609
610 phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
611 err = __create_hyp_mappings(virt_addr, PAGE_SIZE, phys_addr,
612 prot);
613 if (err)
614 return err;
615 }
616
617 return 0;
618 }
619
620 static int __hyp_alloc_private_va_range(unsigned long base)
621 {
622 lockdep_assert_held(&kvm_hyp_pgd_mutex);
623
624 if (!PAGE_ALIGNED(base))
625 return -EINVAL;
626
627 /*
628 * Verify that BIT(VA_BITS - 1) hasn't been flipped by
629 * allocating the new area, as it would indicate we've
630 * overflowed the idmap/IO address range.
631 */
632 if ((base ^ io_map_base) & BIT(VA_BITS - 1))
633 return -ENOMEM;
634
635 io_map_base = base;
636
637 return 0;
638 }
639
640 /**
641 * hyp_alloc_private_va_range - Allocates a private VA range.
642 * @size: The size of the VA range to reserve.
643 * @haddr: The hypervisor virtual start address of the allocation.
644 *
645 * The private virtual address (VA) range is allocated below io_map_base
646 * and aligned based on the order of @size.
647 *
648 * Return: 0 on success or negative error code on failure.
649 */
650 int hyp_alloc_private_va_range(size_t size, unsigned long *haddr)
651 {
652 unsigned long base;
653 int ret = 0;
654
655 mutex_lock(&kvm_hyp_pgd_mutex);
656
657 /*
658 * This assumes that we have enough space below the idmap
659 * page to allocate our VAs. If not, the check in
660 * __hyp_alloc_private_va_range() will kick. A potential
661 * alternative would be to detect that overflow and switch
662 * to an allocation above the idmap.
663 *
664 * The allocated size is always a multiple of PAGE_SIZE.
665 */
666 size = PAGE_ALIGN(size);
667 base = io_map_base - size;
668 ret = __hyp_alloc_private_va_range(base);
669
670 mutex_unlock(&kvm_hyp_pgd_mutex);
671
672 if (!ret)
673 *haddr = base;
674
675 return ret;
676 }
677
678 static int __create_hyp_private_mapping(phys_addr_t phys_addr, size_t size,
679 unsigned long *haddr,
680 enum kvm_pgtable_prot prot)
681 {
682 unsigned long addr;
683 int ret = 0;
684
685 if (!kvm_host_owns_hyp_mappings()) {
686 addr = kvm_call_hyp_nvhe(__pkvm_create_private_mapping,
687 phys_addr, size, prot);
688 if (IS_ERR_VALUE(addr))
689 return addr;
690 *haddr = addr;
691
692 return 0;
693 }
694
695 size = PAGE_ALIGN(size + offset_in_page(phys_addr));
696 ret = hyp_alloc_private_va_range(size, &addr);
697 if (ret)
698 return ret;
699
700 ret = __create_hyp_mappings(addr, size, phys_addr, prot);
701 if (ret)
702 return ret;
703
704 *haddr = addr + offset_in_page(phys_addr);
705 return ret;
706 }
707
708 int create_hyp_stack(phys_addr_t phys_addr, unsigned long *haddr)
709 {
710 unsigned long base;
711 size_t size;
712 int ret;
713
714 mutex_lock(&kvm_hyp_pgd_mutex);
715 /*
716 * Efficient stack verification using the NVHE_STACK_SHIFT bit implies
717 * an alignment of our allocation on the order of the size.
718 */
719 size = NVHE_STACK_SIZE * 2;
720 base = ALIGN_DOWN(io_map_base - size, size);
721
722 ret = __hyp_alloc_private_va_range(base);
723
724 mutex_unlock(&kvm_hyp_pgd_mutex);
725
726 if (ret) {
727 kvm_err("Cannot allocate hyp stack guard page\n");
728 return ret;
729 }
730
731 /*
732 * Since the stack grows downwards, map the stack to the page
733 * at the higher address and leave the lower guard page
734 * unbacked.
735 *
736 * Any valid stack address now has the NVHE_STACK_SHIFT bit as 1
737 * and addresses corresponding to the guard page have the
738 * NVHE_STACK_SHIFT bit as 0 - this is used for overflow detection.
739 */
740 ret = __create_hyp_mappings(base + NVHE_STACK_SIZE, NVHE_STACK_SIZE,
741 phys_addr, PAGE_HYP);
742 if (ret)
743 kvm_err("Cannot map hyp stack\n");
744
745 *haddr = base + size;
746
747 return ret;
748 }
749
750 /**
751 * create_hyp_io_mappings - Map IO into both kernel and HYP
752 * @phys_addr: The physical start address which gets mapped
753 * @size: Size of the region being mapped
754 * @kaddr: Kernel VA for this mapping
755 * @haddr: HYP VA for this mapping
756 */
757 int create_hyp_io_mappings(phys_addr_t phys_addr, size_t size,
758 void __iomem **kaddr,
759 void __iomem **haddr)
760 {
761 unsigned long addr;
762 int ret;
763
764 if (is_protected_kvm_enabled())
765 return -EPERM;
766
767 *kaddr = ioremap(phys_addr, size);
768 if (!*kaddr)
769 return -ENOMEM;
770
771 if (is_kernel_in_hyp_mode()) {
772 *haddr = *kaddr;
773 return 0;
774 }
775
776 ret = __create_hyp_private_mapping(phys_addr, size,
777 &addr, PAGE_HYP_DEVICE);
778 if (ret) {
779 iounmap(*kaddr);
780 *kaddr = NULL;
781 *haddr = NULL;
782 return ret;
783 }
784
785 *haddr = (void __iomem *)addr;
786 return 0;
787 }
788
789 /**
790 * create_hyp_exec_mappings - Map an executable range into HYP
791 * @phys_addr: The physical start address which gets mapped
792 * @size: Size of the region being mapped
793 * @haddr: HYP VA for this mapping
794 */
795 int create_hyp_exec_mappings(phys_addr_t phys_addr, size_t size,
796 void **haddr)
797 {
798 unsigned long addr;
799 int ret;
800
801 BUG_ON(is_kernel_in_hyp_mode());
802
803 ret = __create_hyp_private_mapping(phys_addr, size,
804 &addr, PAGE_HYP_EXEC);
805 if (ret) {
806 *haddr = NULL;
807 return ret;
808 }
809
810 *haddr = (void *)addr;
811 return 0;
812 }
813
814 static struct kvm_pgtable_mm_ops kvm_user_mm_ops = {
815 /* We shouldn't need any other callback to walk the PT */
816 .phys_to_virt = kvm_host_va,
817 };
818
819 static int get_user_mapping_size(struct kvm *kvm, u64 addr)
820 {
821 struct kvm_pgtable pgt = {
822 .pgd = (kvm_pteref_t)kvm->mm->pgd,
823 .ia_bits = vabits_actual,
824 .start_level = (KVM_PGTABLE_LAST_LEVEL -
825 ARM64_HW_PGTABLE_LEVELS(pgt.ia_bits) + 1),
826 .mm_ops = &kvm_user_mm_ops,
827 };
828 unsigned long flags;
829 kvm_pte_t pte = 0; /* Keep GCC quiet... */
830 s8 level = S8_MAX;
831 int ret;
832
833 /*
834 * Disable IRQs so that we hazard against a concurrent
835 * teardown of the userspace page tables (which relies on
836 * IPI-ing threads).
837 */
838 local_irq_save(flags);
839 ret = kvm_pgtable_get_leaf(&pgt, addr, &pte, &level);
840 local_irq_restore(flags);
841
842 if (ret)
843 return ret;
844
845 /*
846 * Not seeing an error, but not updating level? Something went
847 * deeply wrong...
848 */
849 if (WARN_ON(level > KVM_PGTABLE_LAST_LEVEL))
850 return -EFAULT;
851 if (WARN_ON(level < KVM_PGTABLE_FIRST_LEVEL))
852 return -EFAULT;
853
854 /* Oops, the userspace PTs are gone... Replay the fault */
855 if (!kvm_pte_valid(pte))
856 return -EAGAIN;
857
858 return BIT(ARM64_HW_PGTABLE_LEVEL_SHIFT(level));
859 }
860
861 static struct kvm_pgtable_mm_ops kvm_s2_mm_ops = {
862 .zalloc_page = stage2_memcache_zalloc_page,
863 .zalloc_pages_exact = kvm_s2_zalloc_pages_exact,
864 .free_pages_exact = kvm_s2_free_pages_exact,
865 .free_unlinked_table = stage2_free_unlinked_table,
866 .get_page = kvm_host_get_page,
867 .put_page = kvm_s2_put_page,
868 .page_count = kvm_host_page_count,
869 .phys_to_virt = kvm_host_va,
870 .virt_to_phys = kvm_host_pa,
871 .dcache_clean_inval_poc = clean_dcache_guest_page,
872 .icache_inval_pou = invalidate_icache_guest_page,
873 };
874
875 static int kvm_init_ipa_range(struct kvm_s2_mmu *mmu, unsigned long type)
876 {
877 u32 kvm_ipa_limit = get_kvm_ipa_limit();
878 u64 mmfr0, mmfr1;
879 u32 phys_shift;
880
881 if (type & ~KVM_VM_TYPE_ARM_IPA_SIZE_MASK)
882 return -EINVAL;
883
884 phys_shift = KVM_VM_TYPE_ARM_IPA_SIZE(type);
885 if (is_protected_kvm_enabled()) {
886 phys_shift = kvm_ipa_limit;
887 } else if (phys_shift) {
888 if (phys_shift > kvm_ipa_limit ||
889 phys_shift < ARM64_MIN_PARANGE_BITS)
890 return -EINVAL;
891 } else {
892 phys_shift = KVM_PHYS_SHIFT;
893 if (phys_shift > kvm_ipa_limit) {
894 pr_warn_once("%s using unsupported default IPA limit, upgrade your VMM\n",
895 current->comm);
896 return -EINVAL;
897 }
898 }
899
900 mmfr0 = read_sanitised_ftr_reg(SYS_ID_AA64MMFR0_EL1);
901 mmfr1 = read_sanitised_ftr_reg(SYS_ID_AA64MMFR1_EL1);
902 mmu->vtcr = kvm_get_vtcr(mmfr0, mmfr1, phys_shift);
903
904 return 0;
905 }
906
907 /**
908 * kvm_init_stage2_mmu - Initialise a S2 MMU structure
909 * @kvm: The pointer to the KVM structure
910 * @mmu: The pointer to the s2 MMU structure
911 * @type: The machine type of the virtual machine
912 *
913 * Allocates only the stage-2 HW PGD level table(s).
914 * Note we don't need locking here as this is only called in two cases:
915 *
916 * - when the VM is created, which can't race against anything
917 *
918 * - when secondary kvm_s2_mmu structures are initialised for NV
919 * guests, and the caller must hold kvm->lock as this is called on a
920 * per-vcpu basis.
921 */
922 int kvm_init_stage2_mmu(struct kvm *kvm, struct kvm_s2_mmu *mmu, unsigned long type)
923 {
924 int cpu, err;
925 struct kvm_pgtable *pgt;
926
927 /*
928 * If we already have our page tables in place, and that the
929 * MMU context is the canonical one, we have a bug somewhere,
930 * as this is only supposed to ever happen once per VM.
931 *
932 * Otherwise, we're building nested page tables, and that's
933 * probably because userspace called KVM_ARM_VCPU_INIT more
934 * than once on the same vcpu. Since that's actually legal,
935 * don't kick a fuss and leave gracefully.
936 */
937 if (mmu->pgt != NULL) {
938 if (kvm_is_nested_s2_mmu(kvm, mmu))
939 return 0;
940
941 kvm_err("kvm_arch already initialized?\n");
942 return -EINVAL;
943 }
944
945 err = kvm_init_ipa_range(mmu, type);
946 if (err)
947 return err;
948
949 pgt = kzalloc(sizeof(*pgt), GFP_KERNEL_ACCOUNT);
950 if (!pgt)
951 return -ENOMEM;
952
953 mmu->arch = &kvm->arch;
954 err = KVM_PGT_FN(kvm_pgtable_stage2_init)(pgt, mmu, &kvm_s2_mm_ops);
955 if (err)
956 goto out_free_pgtable;
957
958 mmu->pgt = pgt;
959 if (is_protected_kvm_enabled())
960 return 0;
961
962 mmu->last_vcpu_ran = alloc_percpu(typeof(*mmu->last_vcpu_ran));
963 if (!mmu->last_vcpu_ran) {
964 err = -ENOMEM;
965 goto out_destroy_pgtable;
966 }
967
968 for_each_possible_cpu(cpu)
969 *per_cpu_ptr(mmu->last_vcpu_ran, cpu) = -1;
970
971 /* The eager page splitting is disabled by default */
972 mmu->split_page_chunk_size = KVM_ARM_EAGER_SPLIT_CHUNK_SIZE_DEFAULT;
973 mmu->split_page_cache.gfp_zero = __GFP_ZERO;
974
975 mmu->pgd_phys = __pa(pgt->pgd);
976
977 if (kvm_is_nested_s2_mmu(kvm, mmu))
978 kvm_init_nested_s2_mmu(mmu);
979
980 return 0;
981
982 out_destroy_pgtable:
983 KVM_PGT_FN(kvm_pgtable_stage2_destroy)(pgt);
984 out_free_pgtable:
985 kfree(pgt);
986 return err;
987 }
988
989 void kvm_uninit_stage2_mmu(struct kvm *kvm)
990 {
991 kvm_free_stage2_pgd(&kvm->arch.mmu);
992 kvm_mmu_free_memory_cache(&kvm->arch.mmu.split_page_cache);
993 }
994
995 static void stage2_unmap_memslot(struct kvm *kvm,
996 struct kvm_memory_slot *memslot)
997 {
998 hva_t hva = memslot->userspace_addr;
999 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
1000 phys_addr_t size = PAGE_SIZE * memslot->npages;
1001 hva_t reg_end = hva + size;
1002
1003 /*
1004 * A memory region could potentially cover multiple VMAs, and any holes
1005 * between them, so iterate over all of them to find out if we should
1006 * unmap any of them.
1007 *
1008 * +--------------------------------------------+
1009 * +---------------+----------------+ +----------------+
1010 * | : VMA 1 | VMA 2 | | VMA 3 : |
1011 * +---------------+----------------+ +----------------+
1012 * | memory region |
1013 * +--------------------------------------------+
1014 */
1015 do {
1016 struct vm_area_struct *vma;
1017 hva_t vm_start, vm_end;
1018
1019 vma = find_vma_intersection(current->mm, hva, reg_end);
1020 if (!vma)
1021 break;
1022
1023 /*
1024 * Take the intersection of this VMA with the memory region
1025 */
1026 vm_start = max(hva, vma->vm_start);
1027 vm_end = min(reg_end, vma->vm_end);
1028
1029 if (!(vma->vm_flags & VM_PFNMAP)) {
1030 gpa_t gpa = addr + (vm_start - memslot->userspace_addr);
1031 kvm_stage2_unmap_range(&kvm->arch.mmu, gpa, vm_end - vm_start, true);
1032 }
1033 hva = vm_end;
1034 } while (hva < reg_end);
1035 }
1036
1037 /**
1038 * stage2_unmap_vm - Unmap Stage-2 RAM mappings
1039 * @kvm: The struct kvm pointer
1040 *
1041 * Go through the memregions and unmap any regular RAM
1042 * backing memory already mapped to the VM.
1043 */
1044 void stage2_unmap_vm(struct kvm *kvm)
1045 {
1046 struct kvm_memslots *slots;
1047 struct kvm_memory_slot *memslot;
1048 int idx, bkt;
1049
1050 idx = srcu_read_lock(&kvm->srcu);
1051 mmap_read_lock(current->mm);
1052 write_lock(&kvm->mmu_lock);
1053
1054 slots = kvm_memslots(kvm);
1055 kvm_for_each_memslot(memslot, bkt, slots)
1056 stage2_unmap_memslot(kvm, memslot);
1057
1058 kvm_nested_s2_unmap(kvm, true);
1059
1060 write_unlock(&kvm->mmu_lock);
1061 mmap_read_unlock(current->mm);
1062 srcu_read_unlock(&kvm->srcu, idx);
1063 }
1064
1065 void kvm_free_stage2_pgd(struct kvm_s2_mmu *mmu)
1066 {
1067 struct kvm *kvm = kvm_s2_mmu_to_kvm(mmu);
1068 struct kvm_pgtable *pgt = NULL;
1069
1070 write_lock(&kvm->mmu_lock);
1071 pgt = mmu->pgt;
1072 if (pgt) {
1073 mmu->pgd_phys = 0;
1074 mmu->pgt = NULL;
1075 free_percpu(mmu->last_vcpu_ran);
1076 }
1077
1078 if (kvm_is_nested_s2_mmu(kvm, mmu))
1079 kvm_init_nested_s2_mmu(mmu);
1080
1081 write_unlock(&kvm->mmu_lock);
1082
1083 if (pgt) {
1084 KVM_PGT_FN(kvm_pgtable_stage2_destroy)(pgt);
1085 kfree(pgt);
1086 }
1087 }
1088
1089 static void hyp_mc_free_fn(void *addr, void *mc)
1090 {
1091 struct kvm_hyp_memcache *memcache = mc;
1092
1093 if (memcache->flags & HYP_MEMCACHE_ACCOUNT_STAGE2)
1094 kvm_account_pgtable_pages(addr, -1);
1095
1096 free_page((unsigned long)addr);
1097 }
1098
1099 static void *hyp_mc_alloc_fn(void *mc)
1100 {
1101 struct kvm_hyp_memcache *memcache = mc;
1102 void *addr;
1103
1104 addr = (void *)__get_free_page(GFP_KERNEL_ACCOUNT);
1105 if (addr && memcache->flags & HYP_MEMCACHE_ACCOUNT_STAGE2)
1106 kvm_account_pgtable_pages(addr, 1);
1107
1108 return addr;
1109 }
1110
1111 void free_hyp_memcache(struct kvm_hyp_memcache *mc)
1112 {
1113 if (!is_protected_kvm_enabled())
1114 return;
1115
1116 kfree(mc->mapping);
1117 __free_hyp_memcache(mc, hyp_mc_free_fn, kvm_host_va, mc);
1118 }
1119
1120 int topup_hyp_memcache(struct kvm_hyp_memcache *mc, unsigned long min_pages)
1121 {
1122 if (!is_protected_kvm_enabled())
1123 return 0;
1124
1125 if (!mc->mapping) {
1126 mc->mapping = kzalloc(sizeof(struct pkvm_mapping), GFP_KERNEL_ACCOUNT);
1127 if (!mc->mapping)
1128 return -ENOMEM;
1129 }
1130
1131 return __topup_hyp_memcache(mc, min_pages, hyp_mc_alloc_fn,
1132 kvm_host_pa, mc);
1133 }
1134
1135 /**
1136 * kvm_phys_addr_ioremap - map a device range to guest IPA
1137 *
1138 * @kvm: The KVM pointer
1139 * @guest_ipa: The IPA at which to insert the mapping
1140 * @pa: The physical address of the device
1141 * @size: The size of the mapping
1142 * @writable: Whether or not to create a writable mapping
1143 */
1144 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
1145 phys_addr_t pa, unsigned long size, bool writable)
1146 {
1147 phys_addr_t addr;
1148 int ret = 0;
1149 struct kvm_mmu_memory_cache cache = { .gfp_zero = __GFP_ZERO };
1150 struct kvm_s2_mmu *mmu = &kvm->arch.mmu;
1151 struct kvm_pgtable *pgt = mmu->pgt;
1152 enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_DEVICE |
1153 KVM_PGTABLE_PROT_R |
1154 (writable ? KVM_PGTABLE_PROT_W : 0);
1155
1156 if (is_protected_kvm_enabled())
1157 return -EPERM;
1158
1159 size += offset_in_page(guest_ipa);
1160 guest_ipa &= PAGE_MASK;
1161
1162 for (addr = guest_ipa; addr < guest_ipa + size; addr += PAGE_SIZE) {
1163 ret = kvm_mmu_topup_memory_cache(&cache,
1164 kvm_mmu_cache_min_pages(mmu));
1165 if (ret)
1166 break;
1167
1168 write_lock(&kvm->mmu_lock);
1169 ret = KVM_PGT_FN(kvm_pgtable_stage2_map)(pgt, addr, PAGE_SIZE,
1170 pa, prot, &cache, 0);
1171 write_unlock(&kvm->mmu_lock);
1172 if (ret)
1173 break;
1174
1175 pa += PAGE_SIZE;
1176 }
1177
1178 kvm_mmu_free_memory_cache(&cache);
1179 return ret;
1180 }
1181
1182 /**
1183 * kvm_stage2_wp_range() - write protect stage2 memory region range
1184 * @mmu: The KVM stage-2 MMU pointer
1185 * @addr: Start address of range
1186 * @end: End address of range
1187 */
1188 void kvm_stage2_wp_range(struct kvm_s2_mmu *mmu, phys_addr_t addr, phys_addr_t end)
1189 {
1190 stage2_apply_range_resched(mmu, addr, end, KVM_PGT_FN(kvm_pgtable_stage2_wrprotect));
1191 }
1192
1193 /**
1194 * kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
1195 * @kvm: The KVM pointer
1196 * @slot: The memory slot to write protect
1197 *
1198 * Called to start logging dirty pages after memory region
1199 * KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
1200 * all present PUD, PMD and PTEs are write protected in the memory region.
1201 * Afterwards read of dirty page log can be called.
1202 *
1203 * Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
1204 * serializing operations for VM memory regions.
1205 */
1206 static void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
1207 {
1208 struct kvm_memslots *slots = kvm_memslots(kvm);
1209 struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
1210 phys_addr_t start, end;
1211
1212 if (WARN_ON_ONCE(!memslot))
1213 return;
1214
1215 start = memslot->base_gfn << PAGE_SHIFT;
1216 end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1217
1218 write_lock(&kvm->mmu_lock);
1219 kvm_stage2_wp_range(&kvm->arch.mmu, start, end);
1220 kvm_nested_s2_wp(kvm);
1221 write_unlock(&kvm->mmu_lock);
1222 kvm_flush_remote_tlbs_memslot(kvm, memslot);
1223 }
1224
1225 /**
1226 * kvm_mmu_split_memory_region() - split the stage 2 blocks into PAGE_SIZE
1227 * pages for memory slot
1228 * @kvm: The KVM pointer
1229 * @slot: The memory slot to split
1230 *
1231 * Acquires kvm->mmu_lock. Called with kvm->slots_lock mutex acquired,
1232 * serializing operations for VM memory regions.
1233 */
1234 static void kvm_mmu_split_memory_region(struct kvm *kvm, int slot)
1235 {
1236 struct kvm_memslots *slots;
1237 struct kvm_memory_slot *memslot;
1238 phys_addr_t start, end;
1239
1240 lockdep_assert_held(&kvm->slots_lock);
1241
1242 slots = kvm_memslots(kvm);
1243 memslot = id_to_memslot(slots, slot);
1244
1245 start = memslot->base_gfn << PAGE_SHIFT;
1246 end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1247
1248 write_lock(&kvm->mmu_lock);
1249 kvm_mmu_split_huge_pages(kvm, start, end);
1250 write_unlock(&kvm->mmu_lock);
1251 }
1252
1253 /*
1254 * kvm_arch_mmu_enable_log_dirty_pt_masked() - enable dirty logging for selected pages.
1255 * @kvm: The KVM pointer
1256 * @slot: The memory slot associated with mask
1257 * @gfn_offset: The gfn offset in memory slot
1258 * @mask: The mask of pages at offset 'gfn_offset' in this memory
1259 * slot to enable dirty logging on
1260 *
1261 * Writes protect selected pages to enable dirty logging, and then
1262 * splits them to PAGE_SIZE. Caller must acquire kvm->mmu_lock.
1263 */
1264 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1265 struct kvm_memory_slot *slot,
1266 gfn_t gfn_offset, unsigned long mask)
1267 {
1268 phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
1269 phys_addr_t start = (base_gfn + __ffs(mask)) << PAGE_SHIFT;
1270 phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
1271
1272 lockdep_assert_held_write(&kvm->mmu_lock);
1273
1274 kvm_stage2_wp_range(&kvm->arch.mmu, start, end);
1275
1276 /*
1277 * Eager-splitting is done when manual-protect is set. We
1278 * also check for initially-all-set because we can avoid
1279 * eager-splitting if initially-all-set is false.
1280 * Initially-all-set equal false implies that huge-pages were
1281 * already split when enabling dirty logging: no need to do it
1282 * again.
1283 */
1284 if (kvm_dirty_log_manual_protect_and_init_set(kvm))
1285 kvm_mmu_split_huge_pages(kvm, start, end);
1286
1287 kvm_nested_s2_wp(kvm);
1288 }
1289
1290 static void kvm_send_hwpoison_signal(unsigned long address, short lsb)
1291 {
1292 send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb, current);
1293 }
1294
1295 static bool fault_supports_stage2_huge_mapping(struct kvm_memory_slot *memslot,
1296 unsigned long hva,
1297 unsigned long map_size)
1298 {
1299 gpa_t gpa_start;
1300 hva_t uaddr_start, uaddr_end;
1301 size_t size;
1302
1303 /* The memslot and the VMA are guaranteed to be aligned to PAGE_SIZE */
1304 if (map_size == PAGE_SIZE)
1305 return true;
1306
1307 /* pKVM only supports PMD_SIZE huge-mappings */
1308 if (is_protected_kvm_enabled() && map_size != PMD_SIZE)
1309 return false;
1310
1311 size = memslot->npages * PAGE_SIZE;
1312
1313 gpa_start = memslot->base_gfn << PAGE_SHIFT;
1314
1315 uaddr_start = memslot->userspace_addr;
1316 uaddr_end = uaddr_start + size;
1317
1318 /*
1319 * Pages belonging to memslots that don't have the same alignment
1320 * within a PMD/PUD for userspace and IPA cannot be mapped with stage-2
1321 * PMD/PUD entries, because we'll end up mapping the wrong pages.
1322 *
1323 * Consider a layout like the following:
1324 *
1325 * memslot->userspace_addr:
1326 * +-----+--------------------+--------------------+---+
1327 * |abcde|fgh Stage-1 block | Stage-1 block tv|xyz|
1328 * +-----+--------------------+--------------------+---+
1329 *
1330 * memslot->base_gfn << PAGE_SHIFT:
1331 * +---+--------------------+--------------------+-----+
1332 * |abc|def Stage-2 block | Stage-2 block |tvxyz|
1333 * +---+--------------------+--------------------+-----+
1334 *
1335 * If we create those stage-2 blocks, we'll end up with this incorrect
1336 * mapping:
1337 * d -> f
1338 * e -> g
1339 * f -> h
1340 */
1341 if ((gpa_start & (map_size - 1)) != (uaddr_start & (map_size - 1)))
1342 return false;
1343
1344 /*
1345 * Next, let's make sure we're not trying to map anything not covered
1346 * by the memslot. This means we have to prohibit block size mappings
1347 * for the beginning and end of a non-block aligned and non-block sized
1348 * memory slot (illustrated by the head and tail parts of the
1349 * userspace view above containing pages 'abcde' and 'xyz',
1350 * respectively).
1351 *
1352 * Note that it doesn't matter if we do the check using the
1353 * userspace_addr or the base_gfn, as both are equally aligned (per
1354 * the check above) and equally sized.
1355 */
1356 return (hva & ~(map_size - 1)) >= uaddr_start &&
1357 (hva & ~(map_size - 1)) + map_size <= uaddr_end;
1358 }
1359
1360 /*
1361 * Check if the given hva is backed by a transparent huge page (THP) and
1362 * whether it can be mapped using block mapping in stage2. If so, adjust
1363 * the stage2 PFN and IPA accordingly. Only PMD_SIZE THPs are currently
1364 * supported. This will need to be updated to support other THP sizes.
1365 *
1366 * Returns the size of the mapping.
1367 */
1368 static long
1369 transparent_hugepage_adjust(struct kvm *kvm, struct kvm_memory_slot *memslot,
1370 unsigned long hva, kvm_pfn_t *pfnp,
1371 phys_addr_t *ipap)
1372 {
1373 kvm_pfn_t pfn = *pfnp;
1374
1375 /*
1376 * Make sure the adjustment is done only for THP pages. Also make
1377 * sure that the HVA and IPA are sufficiently aligned and that the
1378 * block map is contained within the memslot.
1379 */
1380 if (fault_supports_stage2_huge_mapping(memslot, hva, PMD_SIZE)) {
1381 int sz = get_user_mapping_size(kvm, hva);
1382
1383 if (sz < 0)
1384 return sz;
1385
1386 if (sz < PMD_SIZE)
1387 return PAGE_SIZE;
1388
1389 *ipap &= PMD_MASK;
1390 pfn &= ~(PTRS_PER_PMD - 1);
1391 *pfnp = pfn;
1392
1393 return PMD_SIZE;
1394 }
1395
1396 /* Use page mapping if we cannot use block mapping. */
1397 return PAGE_SIZE;
1398 }
1399
1400 static int get_vma_page_shift(struct vm_area_struct *vma, unsigned long hva)
1401 {
1402 unsigned long pa;
1403
1404 if (is_vm_hugetlb_page(vma) && !(vma->vm_flags & VM_PFNMAP))
1405 return huge_page_shift(hstate_vma(vma));
1406
1407 if (!(vma->vm_flags & VM_PFNMAP))
1408 return PAGE_SHIFT;
1409
1410 VM_BUG_ON(is_vm_hugetlb_page(vma));
1411
1412 pa = (vma->vm_pgoff << PAGE_SHIFT) + (hva - vma->vm_start);
1413
1414 #ifndef __PAGETABLE_PMD_FOLDED
1415 if ((hva & (PUD_SIZE - 1)) == (pa & (PUD_SIZE - 1)) &&
1416 ALIGN_DOWN(hva, PUD_SIZE) >= vma->vm_start &&
1417 ALIGN(hva, PUD_SIZE) <= vma->vm_end)
1418 return PUD_SHIFT;
1419 #endif
1420
1421 if ((hva & (PMD_SIZE - 1)) == (pa & (PMD_SIZE - 1)) &&
1422 ALIGN_DOWN(hva, PMD_SIZE) >= vma->vm_start &&
1423 ALIGN(hva, PMD_SIZE) <= vma->vm_end)
1424 return PMD_SHIFT;
1425
1426 return PAGE_SHIFT;
1427 }
1428
1429 /*
1430 * The page will be mapped in stage 2 as Normal Cacheable, so the VM will be
1431 * able to see the page's tags and therefore they must be initialised first. If
1432 * PG_mte_tagged is set, tags have already been initialised.
1433 *
1434 * Must be called with kvm->mmu_lock held to ensure the memory remains mapped
1435 * while the tags are zeroed.
1436 */
1437 static void sanitise_mte_tags(struct kvm *kvm, kvm_pfn_t pfn,
1438 unsigned long size)
1439 {
1440 unsigned long i, nr_pages = size >> PAGE_SHIFT;
1441 struct page *page = pfn_to_page(pfn);
1442 struct folio *folio = page_folio(page);
1443
1444 if (!kvm_has_mte(kvm))
1445 return;
1446
1447 if (folio_test_hugetlb(folio)) {
1448 /* Hugetlb has MTE flags set on head page only */
1449 if (folio_try_hugetlb_mte_tagging(folio)) {
1450 for (i = 0; i < nr_pages; i++, page++)
1451 mte_clear_page_tags(page_address(page));
1452 folio_set_hugetlb_mte_tagged(folio);
1453 }
1454 return;
1455 }
1456
1457 for (i = 0; i < nr_pages; i++, page++) {
1458 if (try_page_mte_tagging(page)) {
1459 mte_clear_page_tags(page_address(page));
1460 set_page_mte_tagged(page);
1461 }
1462 }
1463 }
1464
1465 static bool kvm_vma_mte_allowed(struct vm_area_struct *vma)
1466 {
1467 return vma->vm_flags & VM_MTE_ALLOWED;
1468 }
1469
1470 static bool kvm_vma_is_cacheable(struct vm_area_struct *vma)
1471 {
1472 switch (FIELD_GET(PTE_ATTRINDX_MASK, pgprot_val(vma->vm_page_prot))) {
1473 case MT_NORMAL_NC:
1474 case MT_DEVICE_nGnRnE:
1475 case MT_DEVICE_nGnRE:
1476 return false;
1477 default:
1478 return true;
1479 }
1480 }
1481
1482 static int prepare_mmu_memcache(struct kvm_vcpu *vcpu, bool topup_memcache,
1483 void **memcache)
1484 {
1485 int min_pages;
1486
1487 if (!is_protected_kvm_enabled())
1488 *memcache = &vcpu->arch.mmu_page_cache;
1489 else
1490 *memcache = &vcpu->arch.pkvm_memcache;
1491
1492 if (!topup_memcache)
1493 return 0;
1494
1495 min_pages = kvm_mmu_cache_min_pages(vcpu->arch.hw_mmu);
1496
1497 if (!is_protected_kvm_enabled())
1498 return kvm_mmu_topup_memory_cache(*memcache, min_pages);
1499
1500 return topup_hyp_memcache(*memcache, min_pages);
1501 }
1502
1503 /*
1504 * Potentially reduce shadow S2 permissions to match the guest's own S2. For
1505 * exec faults, we'd only reach this point if the guest actually allowed it (see
1506 * kvm_s2_handle_perm_fault).
1507 *
1508 * Also encode the level of the original translation in the SW bits of the leaf
1509 * entry as a proxy for the span of that translation. This will be retrieved on
1510 * TLB invalidation from the guest and used to limit the invalidation scope if a
1511 * TTL hint or a range isn't provided.
1512 */
1513 static void adjust_nested_fault_perms(struct kvm_s2_trans *nested,
1514 enum kvm_pgtable_prot *prot,
1515 bool *writable)
1516 {
1517 *writable &= kvm_s2_trans_writable(nested);
1518 if (!kvm_s2_trans_readable(nested))
1519 *prot &= ~KVM_PGTABLE_PROT_R;
1520
1521 *prot |= kvm_encode_nested_level(nested);
1522 }
1523
1524 #define KVM_PGTABLE_WALK_MEMABORT_FLAGS (KVM_PGTABLE_WALK_HANDLE_FAULT | KVM_PGTABLE_WALK_SHARED)
1525
1526 static int gmem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1527 struct kvm_s2_trans *nested,
1528 struct kvm_memory_slot *memslot, bool is_perm)
1529 {
1530 bool write_fault, exec_fault, writable;
1531 enum kvm_pgtable_walk_flags flags = KVM_PGTABLE_WALK_MEMABORT_FLAGS;
1532 enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_R;
1533 struct kvm_pgtable *pgt = vcpu->arch.hw_mmu->pgt;
1534 unsigned long mmu_seq;
1535 struct page *page;
1536 struct kvm *kvm = vcpu->kvm;
1537 void *memcache;
1538 kvm_pfn_t pfn;
1539 gfn_t gfn;
1540 int ret;
1541
1542 ret = prepare_mmu_memcache(vcpu, true, &memcache);
1543 if (ret)
1544 return ret;
1545
1546 if (nested)
1547 gfn = kvm_s2_trans_output(nested) >> PAGE_SHIFT;
1548 else
1549 gfn = fault_ipa >> PAGE_SHIFT;
1550
1551 write_fault = kvm_is_write_fault(vcpu);
1552 exec_fault = kvm_vcpu_trap_is_exec_fault(vcpu);
1553
1554 VM_WARN_ON_ONCE(write_fault && exec_fault);
1555
1556 mmu_seq = kvm->mmu_invalidate_seq;
1557 /* Pairs with the smp_wmb() in kvm_mmu_invalidate_end(). */
1558 smp_rmb();
1559
1560 ret = kvm_gmem_get_pfn(kvm, memslot, gfn, &pfn, &page, NULL);
1561 if (ret) {
1562 kvm_prepare_memory_fault_exit(vcpu, fault_ipa, PAGE_SIZE,
1563 write_fault, exec_fault, false);
1564 return ret;
1565 }
1566
1567 writable = !(memslot->flags & KVM_MEM_READONLY);
1568
1569 if (nested)
1570 adjust_nested_fault_perms(nested, &prot, &writable);
1571
1572 if (writable)
1573 prot |= KVM_PGTABLE_PROT_W;
1574
1575 if (exec_fault ||
1576 (cpus_have_final_cap(ARM64_HAS_CACHE_DIC) &&
1577 (!nested || kvm_s2_trans_executable(nested))))
1578 prot |= KVM_PGTABLE_PROT_X;
1579
1580 kvm_fault_lock(kvm);
1581 if (mmu_invalidate_retry(kvm, mmu_seq)) {
1582 ret = -EAGAIN;
1583 goto out_unlock;
1584 }
1585
1586 ret = KVM_PGT_FN(kvm_pgtable_stage2_map)(pgt, fault_ipa, PAGE_SIZE,
1587 __pfn_to_phys(pfn), prot,
1588 memcache, flags);
1589
1590 out_unlock:
1591 kvm_release_faultin_page(kvm, page, !!ret, writable);
1592 kvm_fault_unlock(kvm);
1593
1594 if (writable && !ret)
1595 mark_page_dirty_in_slot(kvm, memslot, gfn);
1596
1597 return ret != -EAGAIN ? ret : 0;
1598 }
1599
1600 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1601 struct kvm_s2_trans *nested,
1602 struct kvm_memory_slot *memslot, unsigned long hva,
1603 bool fault_is_perm)
1604 {
1605 int ret = 0;
1606 bool topup_memcache;
1607 bool write_fault, writable;
1608 bool exec_fault, mte_allowed, is_vma_cacheable;
1609 bool s2_force_noncacheable = false, vfio_allow_any_uc = false;
1610 unsigned long mmu_seq;
1611 phys_addr_t ipa = fault_ipa;
1612 struct kvm *kvm = vcpu->kvm;
1613 struct vm_area_struct *vma;
1614 short vma_shift;
1615 void *memcache;
1616 gfn_t gfn;
1617 kvm_pfn_t pfn;
1618 bool logging_active = memslot_is_logging(memslot);
1619 bool force_pte = logging_active;
1620 long vma_pagesize, fault_granule;
1621 enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_R;
1622 struct kvm_pgtable *pgt;
1623 struct page *page;
1624 vm_flags_t vm_flags;
1625 enum kvm_pgtable_walk_flags flags = KVM_PGTABLE_WALK_MEMABORT_FLAGS;
1626
1627 if (fault_is_perm)
1628 fault_granule = kvm_vcpu_trap_get_perm_fault_granule(vcpu);
1629 write_fault = kvm_is_write_fault(vcpu);
1630 exec_fault = kvm_vcpu_trap_is_exec_fault(vcpu);
1631 VM_WARN_ON_ONCE(write_fault && exec_fault);
1632
1633 /*
1634 * Permission faults just need to update the existing leaf entry,
1635 * and so normally don't require allocations from the memcache. The
1636 * only exception to this is when dirty logging is enabled at runtime
1637 * and a write fault needs to collapse a block entry into a table.
1638 */
1639 topup_memcache = !fault_is_perm || (logging_active && write_fault);
1640 ret = prepare_mmu_memcache(vcpu, topup_memcache, &memcache);
1641 if (ret)
1642 return ret;
1643
1644 /*
1645 * Let's check if we will get back a huge page backed by hugetlbfs, or
1646 * get block mapping for device MMIO region.
1647 */
1648 mmap_read_lock(current->mm);
1649 vma = vma_lookup(current->mm, hva);
1650 if (unlikely(!vma)) {
1651 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
1652 mmap_read_unlock(current->mm);
1653 return -EFAULT;
1654 }
1655
1656 if (force_pte)
1657 vma_shift = PAGE_SHIFT;
1658 else
1659 vma_shift = get_vma_page_shift(vma, hva);
1660
1661 switch (vma_shift) {
1662 #ifndef __PAGETABLE_PMD_FOLDED
1663 case PUD_SHIFT:
1664 if (fault_supports_stage2_huge_mapping(memslot, hva, PUD_SIZE))
1665 break;
1666 fallthrough;
1667 #endif
1668 case CONT_PMD_SHIFT:
1669 vma_shift = PMD_SHIFT;
1670 fallthrough;
1671 case PMD_SHIFT:
1672 if (fault_supports_stage2_huge_mapping(memslot, hva, PMD_SIZE))
1673 break;
1674 fallthrough;
1675 case CONT_PTE_SHIFT:
1676 vma_shift = PAGE_SHIFT;
1677 force_pte = true;
1678 fallthrough;
1679 case PAGE_SHIFT:
1680 break;
1681 default:
1682 WARN_ONCE(1, "Unknown vma_shift %d", vma_shift);
1683 }
1684
1685 vma_pagesize = 1UL << vma_shift;
1686
1687 if (nested) {
1688 unsigned long max_map_size;
1689
1690 max_map_size = force_pte ? PAGE_SIZE : PUD_SIZE;
1691
1692 ipa = kvm_s2_trans_output(nested);
1693
1694 /*
1695 * If we're about to create a shadow stage 2 entry, then we
1696 * can only create a block mapping if the guest stage 2 page
1697 * table uses at least as big a mapping.
1698 */
1699 max_map_size = min(kvm_s2_trans_size(nested), max_map_size);
1700
1701 /*
1702 * Be careful that if the mapping size falls between
1703 * two host sizes, take the smallest of the two.
1704 */
1705 if (max_map_size >= PMD_SIZE && max_map_size < PUD_SIZE)
1706 max_map_size = PMD_SIZE;
1707 else if (max_map_size >= PAGE_SIZE && max_map_size < PMD_SIZE)
1708 max_map_size = PAGE_SIZE;
1709
1710 force_pte = (max_map_size == PAGE_SIZE);
1711 vma_pagesize = min_t(long, vma_pagesize, max_map_size);
1712 }
1713
1714 /*
1715 * Both the canonical IPA and fault IPA must be aligned to the
1716 * mapping size to ensure we find the right PFN and lay down the
1717 * mapping in the right place.
1718 */
1719 fault_ipa = ALIGN_DOWN(fault_ipa, vma_pagesize);
1720 ipa = ALIGN_DOWN(ipa, vma_pagesize);
1721
1722 gfn = ipa >> PAGE_SHIFT;
1723 mte_allowed = kvm_vma_mte_allowed(vma);
1724
1725 vfio_allow_any_uc = vma->vm_flags & VM_ALLOW_ANY_UNCACHED;
1726
1727 vm_flags = vma->vm_flags;
1728
1729 is_vma_cacheable = kvm_vma_is_cacheable(vma);
1730
1731 /* Don't use the VMA after the unlock -- it may have vanished */
1732 vma = NULL;
1733
1734 /*
1735 * Read mmu_invalidate_seq so that KVM can detect if the results of
1736 * vma_lookup() or __kvm_faultin_pfn() become stale prior to
1737 * acquiring kvm->mmu_lock.
1738 *
1739 * Rely on mmap_read_unlock() for an implicit smp_rmb(), which pairs
1740 * with the smp_wmb() in kvm_mmu_invalidate_end().
1741 */
1742 mmu_seq = kvm->mmu_invalidate_seq;
1743 mmap_read_unlock(current->mm);
1744
1745 pfn = __kvm_faultin_pfn(memslot, gfn, write_fault ? FOLL_WRITE : 0,
1746 &writable, &page);
1747 if (pfn == KVM_PFN_ERR_HWPOISON) {
1748 kvm_send_hwpoison_signal(hva, vma_shift);
1749 return 0;
1750 }
1751 if (is_error_noslot_pfn(pfn))
1752 return -EFAULT;
1753
1754 /*
1755 * Check if this is non-struct page memory PFN, and cannot support
1756 * CMOs. It could potentially be unsafe to access as cachable.
1757 */
1758 if (vm_flags & (VM_PFNMAP | VM_MIXEDMAP) && !pfn_is_map_memory(pfn)) {
1759 if (is_vma_cacheable) {
1760 /*
1761 * Whilst the VMA owner expects cacheable mapping to this
1762 * PFN, hardware also has to support the FWB and CACHE DIC
1763 * features.
1764 *
1765 * ARM64 KVM relies on kernel VA mapping to the PFN to
1766 * perform cache maintenance as the CMO instructions work on
1767 * virtual addresses. VM_PFNMAP region are not necessarily
1768 * mapped to a KVA and hence the presence of hardware features
1769 * S2FWB and CACHE DIC are mandatory to avoid the need for
1770 * cache maintenance.
1771 */
1772 if (!kvm_supports_cacheable_pfnmap())
1773 ret = -EFAULT;
1774 } else {
1775 /*
1776 * If the page was identified as device early by looking at
1777 * the VMA flags, vma_pagesize is already representing the
1778 * largest quantity we can map. If instead it was mapped
1779 * via __kvm_faultin_pfn(), vma_pagesize is set to PAGE_SIZE
1780 * and must not be upgraded.
1781 *
1782 * In both cases, we don't let transparent_hugepage_adjust()
1783 * change things at the last minute.
1784 */
1785 s2_force_noncacheable = true;
1786 }
1787 } else if (logging_active && !write_fault) {
1788 /*
1789 * Only actually map the page as writable if this was a write
1790 * fault.
1791 */
1792 writable = false;
1793 }
1794
1795 if (exec_fault && s2_force_noncacheable)
1796 ret = -ENOEXEC;
1797
1798 if (ret) {
1799 kvm_release_page_unused(page);
1800 return ret;
1801 }
1802
1803 if (nested)
1804 adjust_nested_fault_perms(nested, &prot, &writable);
1805
1806 kvm_fault_lock(kvm);
1807 pgt = vcpu->arch.hw_mmu->pgt;
1808 if (mmu_invalidate_retry(kvm, mmu_seq)) {
1809 ret = -EAGAIN;
1810 goto out_unlock;
1811 }
1812
1813 /*
1814 * If we are not forced to use page mapping, check if we are
1815 * backed by a THP and thus use block mapping if possible.
1816 */
1817 if (vma_pagesize == PAGE_SIZE && !(force_pte || s2_force_noncacheable)) {
1818 if (fault_is_perm && fault_granule > PAGE_SIZE)
1819 vma_pagesize = fault_granule;
1820 else
1821 vma_pagesize = transparent_hugepage_adjust(kvm, memslot,
1822 hva, &pfn,
1823 &fault_ipa);
1824
1825 if (vma_pagesize < 0) {
1826 ret = vma_pagesize;
1827 goto out_unlock;
1828 }
1829 }
1830
1831 if (!fault_is_perm && !s2_force_noncacheable && kvm_has_mte(kvm)) {
1832 /* Check the VMM hasn't introduced a new disallowed VMA */
1833 if (mte_allowed) {
1834 sanitise_mte_tags(kvm, pfn, vma_pagesize);
1835 } else {
1836 ret = -EFAULT;
1837 goto out_unlock;
1838 }
1839 }
1840
1841 if (writable)
1842 prot |= KVM_PGTABLE_PROT_W;
1843
1844 if (exec_fault)
1845 prot |= KVM_PGTABLE_PROT_X;
1846
1847 if (s2_force_noncacheable) {
1848 if (vfio_allow_any_uc)
1849 prot |= KVM_PGTABLE_PROT_NORMAL_NC;
1850 else
1851 prot |= KVM_PGTABLE_PROT_DEVICE;
1852 } else if (cpus_have_final_cap(ARM64_HAS_CACHE_DIC) &&
1853 (!nested || kvm_s2_trans_executable(nested))) {
1854 prot |= KVM_PGTABLE_PROT_X;
1855 }
1856
1857 /*
1858 * Under the premise of getting a FSC_PERM fault, we just need to relax
1859 * permissions only if vma_pagesize equals fault_granule. Otherwise,
1860 * kvm_pgtable_stage2_map() should be called to change block size.
1861 */
1862 if (fault_is_perm && vma_pagesize == fault_granule) {
1863 /*
1864 * Drop the SW bits in favour of those stored in the
1865 * PTE, which will be preserved.
1866 */
1867 prot &= ~KVM_NV_GUEST_MAP_SZ;
1868 ret = KVM_PGT_FN(kvm_pgtable_stage2_relax_perms)(pgt, fault_ipa, prot, flags);
1869 } else {
1870 ret = KVM_PGT_FN(kvm_pgtable_stage2_map)(pgt, fault_ipa, vma_pagesize,
1871 __pfn_to_phys(pfn), prot,
1872 memcache, flags);
1873 }
1874
1875 out_unlock:
1876 kvm_release_faultin_page(kvm, page, !!ret, writable);
1877 kvm_fault_unlock(kvm);
1878
1879 /* Mark the page dirty only if the fault is handled successfully */
1880 if (writable && !ret)
1881 mark_page_dirty_in_slot(kvm, memslot, gfn);
1882
1883 return ret != -EAGAIN ? ret : 0;
1884 }
1885
1886 /* Resolve the access fault by making the page young again. */
1887 static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
1888 {
1889 enum kvm_pgtable_walk_flags flags = KVM_PGTABLE_WALK_HANDLE_FAULT | KVM_PGTABLE_WALK_SHARED;
1890 struct kvm_s2_mmu *mmu;
1891
1892 trace_kvm_access_fault(fault_ipa);
1893
1894 read_lock(&vcpu->kvm->mmu_lock);
1895 mmu = vcpu->arch.hw_mmu;
1896 KVM_PGT_FN(kvm_pgtable_stage2_mkyoung)(mmu->pgt, fault_ipa, flags);
1897 read_unlock(&vcpu->kvm->mmu_lock);
1898 }
1899
1900 int kvm_handle_guest_sea(struct kvm_vcpu *vcpu)
1901 {
1902 /*
1903 * Give APEI the opportunity to claim the abort before handling it
1904 * within KVM. apei_claim_sea() expects to be called with IRQs enabled.
1905 */
1906 lockdep_assert_irqs_enabled();
1907 if (apei_claim_sea(NULL) == 0)
1908 return 1;
1909
1910 return kvm_inject_serror(vcpu);
1911 }
1912
1913 /**
1914 * kvm_handle_guest_abort - handles all 2nd stage aborts
1915 * @vcpu: the VCPU pointer
1916 *
1917 * Any abort that gets to the host is almost guaranteed to be caused by a
1918 * missing second stage translation table entry, which can mean that either the
1919 * guest simply needs more memory and we must allocate an appropriate page or it
1920 * can mean that the guest tried to access I/O memory, which is emulated by user
1921 * space. The distinction is based on the IPA causing the fault and whether this
1922 * memory region has been registered as standard RAM by user space.
1923 */
1924 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu)
1925 {
1926 struct kvm_s2_trans nested_trans, *nested = NULL;
1927 unsigned long esr;
1928 phys_addr_t fault_ipa; /* The address we faulted on */
1929 phys_addr_t ipa; /* Always the IPA in the L1 guest phys space */
1930 struct kvm_memory_slot *memslot;
1931 unsigned long hva;
1932 bool is_iabt, write_fault, writable;
1933 gfn_t gfn;
1934 int ret, idx;
1935
1936 if (kvm_vcpu_abt_issea(vcpu))
1937 return kvm_handle_guest_sea(vcpu);
1938
1939 esr = kvm_vcpu_get_esr(vcpu);
1940
1941 /*
1942 * The fault IPA should be reliable at this point as we're not dealing
1943 * with an SEA.
1944 */
1945 ipa = fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1946 if (KVM_BUG_ON(ipa == INVALID_GPA, vcpu->kvm))
1947 return -EFAULT;
1948
1949 is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1950
1951 if (esr_fsc_is_translation_fault(esr)) {
1952 /* Beyond sanitised PARange (which is the IPA limit) */
1953 if (fault_ipa >= BIT_ULL(get_kvm_ipa_limit())) {
1954 kvm_inject_size_fault(vcpu);
1955 return 1;
1956 }
1957
1958 /* Falls between the IPA range and the PARange? */
1959 if (fault_ipa >= BIT_ULL(VTCR_EL2_IPA(vcpu->arch.hw_mmu->vtcr))) {
1960 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & GENMASK(11, 0);
1961
1962 return kvm_inject_sea(vcpu, is_iabt, fault_ipa);
1963 }
1964 }
1965
1966 trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_esr(vcpu),
1967 kvm_vcpu_get_hfar(vcpu), fault_ipa);
1968
1969 /* Check the stage-2 fault is trans. fault or write fault */
1970 if (!esr_fsc_is_translation_fault(esr) &&
1971 !esr_fsc_is_permission_fault(esr) &&
1972 !esr_fsc_is_access_flag_fault(esr)) {
1973 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1974 kvm_vcpu_trap_get_class(vcpu),
1975 (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1976 (unsigned long)kvm_vcpu_get_esr(vcpu));
1977 return -EFAULT;
1978 }
1979
1980 idx = srcu_read_lock(&vcpu->kvm->srcu);
1981
1982 /*
1983 * We may have faulted on a shadow stage 2 page table if we are
1984 * running a nested guest. In this case, we have to resolve the L2
1985 * IPA to the L1 IPA first, before knowing what kind of memory should
1986 * back the L1 IPA.
1987 *
1988 * If the shadow stage 2 page table walk faults, then we simply inject
1989 * this to the guest and carry on.
1990 *
1991 * If there are no shadow S2 PTs because S2 is disabled, there is
1992 * nothing to walk and we treat it as a 1:1 before going through the
1993 * canonical translation.
1994 */
1995 if (kvm_is_nested_s2_mmu(vcpu->kvm,vcpu->arch.hw_mmu) &&
1996 vcpu->arch.hw_mmu->nested_stage2_enabled) {
1997 u32 esr;
1998
1999 ret = kvm_walk_nested_s2(vcpu, fault_ipa, &nested_trans);
2000 if (ret) {
2001 esr = kvm_s2_trans_esr(&nested_trans);
2002 kvm_inject_s2_fault(vcpu, esr);
2003 goto out_unlock;
2004 }
2005
2006 ret = kvm_s2_handle_perm_fault(vcpu, &nested_trans);
2007 if (ret) {
2008 esr = kvm_s2_trans_esr(&nested_trans);
2009 kvm_inject_s2_fault(vcpu, esr);
2010 goto out_unlock;
2011 }
2012
2013 ipa = kvm_s2_trans_output(&nested_trans);
2014 nested = &nested_trans;
2015 }
2016
2017 gfn = ipa >> PAGE_SHIFT;
2018 memslot = gfn_to_memslot(vcpu->kvm, gfn);
2019 hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
2020 write_fault = kvm_is_write_fault(vcpu);
2021 if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
2022 /*
2023 * The guest has put either its instructions or its page-tables
2024 * somewhere it shouldn't have. Userspace won't be able to do
2025 * anything about this (there's no syndrome for a start), so
2026 * re-inject the abort back into the guest.
2027 */
2028 if (is_iabt) {
2029 ret = -ENOEXEC;
2030 goto out;
2031 }
2032
2033 if (kvm_vcpu_abt_iss1tw(vcpu)) {
2034 ret = kvm_inject_sea_dabt(vcpu, kvm_vcpu_get_hfar(vcpu));
2035 goto out_unlock;
2036 }
2037
2038 /*
2039 * Check for a cache maintenance operation. Since we
2040 * ended-up here, we know it is outside of any memory
2041 * slot. But we can't find out if that is for a device,
2042 * or if the guest is just being stupid. The only thing
2043 * we know for sure is that this range cannot be cached.
2044 *
2045 * So let's assume that the guest is just being
2046 * cautious, and skip the instruction.
2047 */
2048 if (kvm_is_error_hva(hva) && kvm_vcpu_dabt_is_cm(vcpu)) {
2049 kvm_incr_pc(vcpu);
2050 ret = 1;
2051 goto out_unlock;
2052 }
2053
2054 /*
2055 * The IPA is reported as [MAX:12], so we need to
2056 * complement it with the bottom 12 bits from the
2057 * faulting VA. This is always 12 bits, irrespective
2058 * of the page size.
2059 */
2060 ipa |= kvm_vcpu_get_hfar(vcpu) & GENMASK(11, 0);
2061 ret = io_mem_abort(vcpu, ipa);
2062 goto out_unlock;
2063 }
2064
2065 /* Userspace should not be able to register out-of-bounds IPAs */
2066 VM_BUG_ON(ipa >= kvm_phys_size(vcpu->arch.hw_mmu));
2067
2068 if (esr_fsc_is_access_flag_fault(esr)) {
2069 handle_access_fault(vcpu, fault_ipa);
2070 ret = 1;
2071 goto out_unlock;
2072 }
2073
2074 VM_WARN_ON_ONCE(kvm_vcpu_trap_is_permission_fault(vcpu) &&
2075 !write_fault && !kvm_vcpu_trap_is_exec_fault(vcpu));
2076
2077 if (kvm_slot_has_gmem(memslot))
2078 ret = gmem_abort(vcpu, fault_ipa, nested, memslot,
2079 esr_fsc_is_permission_fault(esr));
2080 else
2081 ret = user_mem_abort(vcpu, fault_ipa, nested, memslot, hva,
2082 esr_fsc_is_permission_fault(esr));
2083 if (ret == 0)
2084 ret = 1;
2085 out:
2086 if (ret == -ENOEXEC)
2087 ret = kvm_inject_sea_iabt(vcpu, kvm_vcpu_get_hfar(vcpu));
2088 out_unlock:
2089 srcu_read_unlock(&vcpu->kvm->srcu, idx);
2090 return ret;
2091 }
2092
2093 bool kvm_unmap_gfn_range(struct kvm *kvm, struct kvm_gfn_range *range)
2094 {
2095 if (!kvm->arch.mmu.pgt)
2096 return false;
2097
2098 __unmap_stage2_range(&kvm->arch.mmu, range->start << PAGE_SHIFT,
2099 (range->end - range->start) << PAGE_SHIFT,
2100 range->may_block);
2101
2102 kvm_nested_s2_unmap(kvm, range->may_block);
2103 return false;
2104 }
2105
2106 bool kvm_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
2107 {
2108 u64 size = (range->end - range->start) << PAGE_SHIFT;
2109
2110 if (!kvm->arch.mmu.pgt)
2111 return false;
2112
2113 return KVM_PGT_FN(kvm_pgtable_stage2_test_clear_young)(kvm->arch.mmu.pgt,
2114 range->start << PAGE_SHIFT,
2115 size, true);
2116 /*
2117 * TODO: Handle nested_mmu structures here using the reverse mapping in
2118 * a later version of patch series.
2119 */
2120 }
2121
2122 bool kvm_test_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
2123 {
2124 u64 size = (range->end - range->start) << PAGE_SHIFT;
2125
2126 if (!kvm->arch.mmu.pgt)
2127 return false;
2128
2129 return KVM_PGT_FN(kvm_pgtable_stage2_test_clear_young)(kvm->arch.mmu.pgt,
2130 range->start << PAGE_SHIFT,
2131 size, false);
2132 }
2133
2134 phys_addr_t kvm_mmu_get_httbr(void)
2135 {
2136 return __pa(hyp_pgtable->pgd);
2137 }
2138
2139 phys_addr_t kvm_get_idmap_vector(void)
2140 {
2141 return hyp_idmap_vector;
2142 }
2143
2144 static int kvm_map_idmap_text(void)
2145 {
2146 unsigned long size = hyp_idmap_end - hyp_idmap_start;
2147 int err = __create_hyp_mappings(hyp_idmap_start, size, hyp_idmap_start,
2148 PAGE_HYP_EXEC);
2149 if (err)
2150 kvm_err("Failed to idmap %lx-%lx\n",
2151 hyp_idmap_start, hyp_idmap_end);
2152
2153 return err;
2154 }
2155
2156 static void *kvm_hyp_zalloc_page(void *arg)
2157 {
2158 return (void *)get_zeroed_page(GFP_KERNEL);
2159 }
2160
2161 static struct kvm_pgtable_mm_ops kvm_hyp_mm_ops = {
2162 .zalloc_page = kvm_hyp_zalloc_page,
2163 .get_page = kvm_host_get_page,
2164 .put_page = kvm_host_put_page,
2165 .phys_to_virt = kvm_host_va,
2166 .virt_to_phys = kvm_host_pa,
2167 };
2168
2169 int __init kvm_mmu_init(u32 *hyp_va_bits)
2170 {
2171 int err;
2172 u32 idmap_bits;
2173 u32 kernel_bits;
2174
2175 hyp_idmap_start = __pa_symbol(__hyp_idmap_text_start);
2176 hyp_idmap_start = ALIGN_DOWN(hyp_idmap_start, PAGE_SIZE);
2177 hyp_idmap_end = __pa_symbol(__hyp_idmap_text_end);
2178 hyp_idmap_end = ALIGN(hyp_idmap_end, PAGE_SIZE);
2179 hyp_idmap_vector = __pa_symbol(__kvm_hyp_init);
2180
2181 /*
2182 * We rely on the linker script to ensure at build time that the HYP
2183 * init code does not cross a page boundary.
2184 */
2185 BUG_ON((hyp_idmap_start ^ (hyp_idmap_end - 1)) & PAGE_MASK);
2186
2187 /*
2188 * The ID map is always configured for 48 bits of translation, which
2189 * may be fewer than the number of VA bits used by the regular kernel
2190 * stage 1, when VA_BITS=52.
2191 *
2192 * At EL2, there is only one TTBR register, and we can't switch between
2193 * translation tables *and* update TCR_EL2.T0SZ at the same time. Bottom
2194 * line: we need to use the extended range with *both* our translation
2195 * tables.
2196 *
2197 * So use the maximum of the idmap VA bits and the regular kernel stage
2198 * 1 VA bits to assure that the hypervisor can both ID map its code page
2199 * and map any kernel memory.
2200 */
2201 idmap_bits = IDMAP_VA_BITS;
2202 kernel_bits = vabits_actual;
2203 *hyp_va_bits = max(idmap_bits, kernel_bits);
2204
2205 kvm_debug("Using %u-bit virtual addresses at EL2\n", *hyp_va_bits);
2206 kvm_debug("IDMAP page: %lx\n", hyp_idmap_start);
2207 kvm_debug("HYP VA range: %lx:%lx\n",
2208 kern_hyp_va(PAGE_OFFSET),
2209 kern_hyp_va((unsigned long)high_memory - 1));
2210
2211 if (hyp_idmap_start >= kern_hyp_va(PAGE_OFFSET) &&
2212 hyp_idmap_start < kern_hyp_va((unsigned long)high_memory - 1) &&
2213 hyp_idmap_start != (unsigned long)__hyp_idmap_text_start) {
2214 /*
2215 * The idmap page is intersecting with the VA space,
2216 * it is not safe to continue further.
2217 */
2218 kvm_err("IDMAP intersecting with HYP VA, unable to continue\n");
2219 err = -EINVAL;
2220 goto out;
2221 }
2222
2223 hyp_pgtable = kzalloc(sizeof(*hyp_pgtable), GFP_KERNEL);
2224 if (!hyp_pgtable) {
2225 kvm_err("Hyp mode page-table not allocated\n");
2226 err = -ENOMEM;
2227 goto out;
2228 }
2229
2230 err = kvm_pgtable_hyp_init(hyp_pgtable, *hyp_va_bits, &kvm_hyp_mm_ops);
2231 if (err)
2232 goto out_free_pgtable;
2233
2234 err = kvm_map_idmap_text();
2235 if (err)
2236 goto out_destroy_pgtable;
2237
2238 io_map_base = hyp_idmap_start;
2239 __hyp_va_bits = *hyp_va_bits;
2240 return 0;
2241
2242 out_destroy_pgtable:
2243 kvm_pgtable_hyp_destroy(hyp_pgtable);
2244 out_free_pgtable:
2245 kfree(hyp_pgtable);
2246 hyp_pgtable = NULL;
2247 out:
2248 return err;
2249 }
2250
2251 void kvm_arch_commit_memory_region(struct kvm *kvm,
2252 struct kvm_memory_slot *old,
2253 const struct kvm_memory_slot *new,
2254 enum kvm_mr_change change)
2255 {
2256 bool log_dirty_pages = new && new->flags & KVM_MEM_LOG_DIRTY_PAGES;
2257
2258 /*
2259 * At this point memslot has been committed and there is an
2260 * allocated dirty_bitmap[], dirty pages will be tracked while the
2261 * memory slot is write protected.
2262 */
2263 if (log_dirty_pages) {
2264
2265 if (change == KVM_MR_DELETE)
2266 return;
2267
2268 /*
2269 * Huge and normal pages are write-protected and split
2270 * on either of these two cases:
2271 *
2272 * 1. with initial-all-set: gradually with CLEAR ioctls,
2273 */
2274 if (kvm_dirty_log_manual_protect_and_init_set(kvm))
2275 return;
2276 /*
2277 * or
2278 * 2. without initial-all-set: all in one shot when
2279 * enabling dirty logging.
2280 */
2281 kvm_mmu_wp_memory_region(kvm, new->id);
2282 kvm_mmu_split_memory_region(kvm, new->id);
2283 } else {
2284 /*
2285 * Free any leftovers from the eager page splitting cache. Do
2286 * this when deleting, moving, disabling dirty logging, or
2287 * creating the memslot (a nop). Doing it for deletes makes
2288 * sure we don't leak memory, and there's no need to keep the
2289 * cache around for any of the other cases.
2290 */
2291 kvm_mmu_free_memory_cache(&kvm->arch.mmu.split_page_cache);
2292 }
2293 }
2294
2295 int kvm_arch_prepare_memory_region(struct kvm *kvm,
2296 const struct kvm_memory_slot *old,
2297 struct kvm_memory_slot *new,
2298 enum kvm_mr_change change)
2299 {
2300 hva_t hva, reg_end;
2301 int ret = 0;
2302
2303 if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
2304 change != KVM_MR_FLAGS_ONLY)
2305 return 0;
2306
2307 /*
2308 * Prevent userspace from creating a memory region outside of the IPA
2309 * space addressable by the KVM guest IPA space.
2310 */
2311 if ((new->base_gfn + new->npages) > (kvm_phys_size(&kvm->arch.mmu) >> PAGE_SHIFT))
2312 return -EFAULT;
2313
2314 /*
2315 * Only support guest_memfd backed memslots with mappable memory, since
2316 * there aren't any CoCo VMs that support only private memory on arm64.
2317 */
2318 if (kvm_slot_has_gmem(new) && !kvm_memslot_is_gmem_only(new))
2319 return -EINVAL;
2320
2321 hva = new->userspace_addr;
2322 reg_end = hva + (new->npages << PAGE_SHIFT);
2323
2324 mmap_read_lock(current->mm);
2325 /*
2326 * A memory region could potentially cover multiple VMAs, and any holes
2327 * between them, so iterate over all of them.
2328 *
2329 * +--------------------------------------------+
2330 * +---------------+----------------+ +----------------+
2331 * | : VMA 1 | VMA 2 | | VMA 3 : |
2332 * +---------------+----------------+ +----------------+
2333 * | memory region |
2334 * +--------------------------------------------+
2335 */
2336 do {
2337 struct vm_area_struct *vma;
2338
2339 vma = find_vma_intersection(current->mm, hva, reg_end);
2340 if (!vma)
2341 break;
2342
2343 if (kvm_has_mte(kvm) && !kvm_vma_mte_allowed(vma)) {
2344 ret = -EINVAL;
2345 break;
2346 }
2347
2348 if (vma->vm_flags & VM_PFNMAP) {
2349 /* IO region dirty page logging not allowed */
2350 if (new->flags & KVM_MEM_LOG_DIRTY_PAGES) {
2351 ret = -EINVAL;
2352 break;
2353 }
2354
2355 /*
2356 * Cacheable PFNMAP is allowed only if the hardware
2357 * supports it.
2358 */
2359 if (kvm_vma_is_cacheable(vma) && !kvm_supports_cacheable_pfnmap()) {
2360 ret = -EINVAL;
2361 break;
2362 }
2363 }
2364 hva = min(reg_end, vma->vm_end);
2365 } while (hva < reg_end);
2366
2367 mmap_read_unlock(current->mm);
2368 return ret;
2369 }
2370
2371 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *slot)
2372 {
2373 }
2374
2375 void kvm_arch_memslots_updated(struct kvm *kvm, u64 gen)
2376 {
2377 }
2378
2379 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
2380 struct kvm_memory_slot *slot)
2381 {
2382 gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
2383 phys_addr_t size = slot->npages << PAGE_SHIFT;
2384
2385 write_lock(&kvm->mmu_lock);
2386 kvm_stage2_unmap_range(&kvm->arch.mmu, gpa, size, true);
2387 kvm_nested_s2_unmap(kvm, true);
2388 write_unlock(&kvm->mmu_lock);
2389 }
2390
2391 /*
2392 * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
2393 *
2394 * Main problems:
2395 * - S/W ops are local to a CPU (not broadcast)
2396 * - We have line migration behind our back (speculation)
2397 * - System caches don't support S/W at all (damn!)
2398 *
2399 * In the face of the above, the best we can do is to try and convert
2400 * S/W ops to VA ops. Because the guest is not allowed to infer the
2401 * S/W to PA mapping, it can only use S/W to nuke the whole cache,
2402 * which is a rather good thing for us.
2403 *
2404 * Also, it is only used when turning caches on/off ("The expected
2405 * usage of the cache maintenance instructions that operate by set/way
2406 * is associated with the cache maintenance instructions associated
2407 * with the powerdown and powerup of caches, if this is required by
2408 * the implementation.").
2409 *
2410 * We use the following policy:
2411 *
2412 * - If we trap a S/W operation, we enable VM trapping to detect
2413 * caches being turned on/off, and do a full clean.
2414 *
2415 * - We flush the caches on both caches being turned on and off.
2416 *
2417 * - Once the caches are enabled, we stop trapping VM ops.
2418 */
2419 void kvm_set_way_flush(struct kvm_vcpu *vcpu)
2420 {
2421 unsigned long hcr = *vcpu_hcr(vcpu);
2422
2423 /*
2424 * If this is the first time we do a S/W operation
2425 * (i.e. HCR_TVM not set) flush the whole memory, and set the
2426 * VM trapping.
2427 *
2428 * Otherwise, rely on the VM trapping to wait for the MMU +
2429 * Caches to be turned off. At that point, we'll be able to
2430 * clean the caches again.
2431 */
2432 if (!(hcr & HCR_TVM)) {
2433 trace_kvm_set_way_flush(*vcpu_pc(vcpu),
2434 vcpu_has_cache_enabled(vcpu));
2435 stage2_flush_vm(vcpu->kvm);
2436 *vcpu_hcr(vcpu) = hcr | HCR_TVM;
2437 }
2438 }
2439
2440 void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
2441 {
2442 bool now_enabled = vcpu_has_cache_enabled(vcpu);
2443
2444 /*
2445 * If switching the MMU+caches on, need to invalidate the caches.
2446 * If switching it off, need to clean the caches.
2447 * Clean + invalidate does the trick always.
2448 */
2449 if (now_enabled != was_enabled)
2450 stage2_flush_vm(vcpu->kvm);
2451
2452 /* Caches are now on, stop trapping VM ops (until a S/W op) */
2453 if (now_enabled)
2454 *vcpu_hcr(vcpu) &= ~HCR_TVM;
2455
2456 trace_kvm_toggle_cache(*vcpu_pc(vcpu), was_enabled, now_enabled);
2457 }