1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Copyright (C) 2012 - Virtual Open Systems and Columbia University
4 * Author: Christoffer Dall <c.dall@virtualopensystems.com>
5 */
6
7 #include <linux/mman.h>
8 #include <linux/kvm_host.h>
9 #include <linux/io.h>
10 #include <linux/hugetlb.h>
11 #include <linux/sched/signal.h>
12 #include <trace/events/kvm.h>
13 #include <asm/pgalloc.h>
14 #include <asm/cacheflush.h>
15 #include <asm/kvm_arm.h>
16 #include <asm/kvm_mmu.h>
17 #include <asm/kvm_pgtable.h>
18 #include <asm/kvm_ras.h>
19 #include <asm/kvm_asm.h>
20 #include <asm/kvm_emulate.h>
21 #include <asm/virt.h>
22
23 #include "trace.h"
24
25 static struct kvm_pgtable *hyp_pgtable;
26 static DEFINE_MUTEX(kvm_hyp_pgd_mutex);
27
28 static unsigned long __ro_after_init hyp_idmap_start;
29 static unsigned long __ro_after_init hyp_idmap_end;
30 static phys_addr_t __ro_after_init hyp_idmap_vector;
31
32 static unsigned long __ro_after_init io_map_base;
33
__stage2_range_addr_end(phys_addr_t addr,phys_addr_t end,phys_addr_t size)34 static phys_addr_t __stage2_range_addr_end(phys_addr_t addr, phys_addr_t end,
35 phys_addr_t size)
36 {
37 phys_addr_t boundary = ALIGN_DOWN(addr + size, size);
38
39 return (boundary - 1 < end - 1) ? boundary : end;
40 }
41
stage2_range_addr_end(phys_addr_t addr,phys_addr_t end)42 static phys_addr_t stage2_range_addr_end(phys_addr_t addr, phys_addr_t end)
43 {
44 phys_addr_t size = kvm_granule_size(KVM_PGTABLE_MIN_BLOCK_LEVEL);
45
46 return __stage2_range_addr_end(addr, end, size);
47 }
48
49 /*
50 * Release kvm_mmu_lock periodically if the memory region is large. Otherwise,
51 * we may see kernel panics with CONFIG_DETECT_HUNG_TASK,
52 * CONFIG_LOCKUP_DETECTOR, CONFIG_LOCKDEP. Additionally, holding the lock too
53 * long will also starve other vCPUs. We have to also make sure that the page
54 * tables are not freed while we released the lock.
55 */
stage2_apply_range(struct kvm_s2_mmu * mmu,phys_addr_t addr,phys_addr_t end,int (* fn)(struct kvm_pgtable *,u64,u64),bool resched)56 static int stage2_apply_range(struct kvm_s2_mmu *mmu, phys_addr_t addr,
57 phys_addr_t end,
58 int (*fn)(struct kvm_pgtable *, u64, u64),
59 bool resched)
60 {
61 struct kvm *kvm = kvm_s2_mmu_to_kvm(mmu);
62 int ret;
63 u64 next;
64
65 do {
66 struct kvm_pgtable *pgt = mmu->pgt;
67 if (!pgt)
68 return -EINVAL;
69
70 next = stage2_range_addr_end(addr, end);
71 ret = fn(pgt, addr, next - addr);
72 if (ret)
73 break;
74
75 if (resched && next != end)
76 cond_resched_rwlock_write(&kvm->mmu_lock);
77 } while (addr = next, addr != end);
78
79 return ret;
80 }
81
82 #define stage2_apply_range_resched(mmu, addr, end, fn) \
83 stage2_apply_range(mmu, addr, end, fn, true)
84
85 /*
86 * Get the maximum number of page-tables pages needed to split a range
87 * of blocks into PAGE_SIZE PTEs. It assumes the range is already
88 * mapped at level 2, or at level 1 if allowed.
89 */
kvm_mmu_split_nr_page_tables(u64 range)90 static int kvm_mmu_split_nr_page_tables(u64 range)
91 {
92 int n = 0;
93
94 if (KVM_PGTABLE_MIN_BLOCK_LEVEL < 2)
95 n += DIV_ROUND_UP(range, PUD_SIZE);
96 n += DIV_ROUND_UP(range, PMD_SIZE);
97 return n;
98 }
99
need_split_memcache_topup_or_resched(struct kvm * kvm)100 static bool need_split_memcache_topup_or_resched(struct kvm *kvm)
101 {
102 struct kvm_mmu_memory_cache *cache;
103 u64 chunk_size, min;
104
105 if (need_resched() || rwlock_needbreak(&kvm->mmu_lock))
106 return true;
107
108 chunk_size = kvm->arch.mmu.split_page_chunk_size;
109 min = kvm_mmu_split_nr_page_tables(chunk_size);
110 cache = &kvm->arch.mmu.split_page_cache;
111 return kvm_mmu_memory_cache_nr_free_objects(cache) < min;
112 }
113
kvm_mmu_split_huge_pages(struct kvm * kvm,phys_addr_t addr,phys_addr_t end)114 static int kvm_mmu_split_huge_pages(struct kvm *kvm, phys_addr_t addr,
115 phys_addr_t end)
116 {
117 struct kvm_mmu_memory_cache *cache;
118 struct kvm_pgtable *pgt;
119 int ret, cache_capacity;
120 u64 next, chunk_size;
121
122 lockdep_assert_held_write(&kvm->mmu_lock);
123
124 chunk_size = kvm->arch.mmu.split_page_chunk_size;
125 cache_capacity = kvm_mmu_split_nr_page_tables(chunk_size);
126
127 if (chunk_size == 0)
128 return 0;
129
130 cache = &kvm->arch.mmu.split_page_cache;
131
132 do {
133 if (need_split_memcache_topup_or_resched(kvm)) {
134 write_unlock(&kvm->mmu_lock);
135 cond_resched();
136 /* Eager page splitting is best-effort. */
137 ret = __kvm_mmu_topup_memory_cache(cache,
138 cache_capacity,
139 cache_capacity);
140 write_lock(&kvm->mmu_lock);
141 if (ret)
142 break;
143 }
144
145 pgt = kvm->arch.mmu.pgt;
146 if (!pgt)
147 return -EINVAL;
148
149 next = __stage2_range_addr_end(addr, end, chunk_size);
150 ret = kvm_pgtable_stage2_split(pgt, addr, next - addr, cache);
151 if (ret)
152 break;
153 } while (addr = next, addr != end);
154
155 return ret;
156 }
157
memslot_is_logging(struct kvm_memory_slot * memslot)158 static bool memslot_is_logging(struct kvm_memory_slot *memslot)
159 {
160 return memslot->dirty_bitmap && !(memslot->flags & KVM_MEM_READONLY);
161 }
162
163 /**
164 * kvm_arch_flush_remote_tlbs() - flush all VM TLB entries for v7/8
165 * @kvm: pointer to kvm structure.
166 *
167 * Interface to HYP function to flush all VM TLB entries
168 */
kvm_arch_flush_remote_tlbs(struct kvm * kvm)169 int kvm_arch_flush_remote_tlbs(struct kvm *kvm)
170 {
171 kvm_call_hyp(__kvm_tlb_flush_vmid, &kvm->arch.mmu);
172 return 0;
173 }
174
kvm_arch_flush_remote_tlbs_range(struct kvm * kvm,gfn_t gfn,u64 nr_pages)175 int kvm_arch_flush_remote_tlbs_range(struct kvm *kvm,
176 gfn_t gfn, u64 nr_pages)
177 {
178 kvm_tlb_flush_vmid_range(&kvm->arch.mmu,
179 gfn << PAGE_SHIFT, nr_pages << PAGE_SHIFT);
180 return 0;
181 }
182
kvm_is_device_pfn(unsigned long pfn)183 static bool kvm_is_device_pfn(unsigned long pfn)
184 {
185 return !pfn_is_map_memory(pfn);
186 }
187
stage2_memcache_zalloc_page(void * arg)188 static void *stage2_memcache_zalloc_page(void *arg)
189 {
190 struct kvm_mmu_memory_cache *mc = arg;
191 void *virt;
192
193 /* Allocated with __GFP_ZERO, so no need to zero */
194 virt = kvm_mmu_memory_cache_alloc(mc);
195 if (virt)
196 kvm_account_pgtable_pages(virt, 1);
197 return virt;
198 }
199
kvm_host_zalloc_pages_exact(size_t size)200 static void *kvm_host_zalloc_pages_exact(size_t size)
201 {
202 return alloc_pages_exact(size, GFP_KERNEL_ACCOUNT | __GFP_ZERO);
203 }
204
kvm_s2_zalloc_pages_exact(size_t size)205 static void *kvm_s2_zalloc_pages_exact(size_t size)
206 {
207 void *virt = kvm_host_zalloc_pages_exact(size);
208
209 if (virt)
210 kvm_account_pgtable_pages(virt, (size >> PAGE_SHIFT));
211 return virt;
212 }
213
kvm_s2_free_pages_exact(void * virt,size_t size)214 static void kvm_s2_free_pages_exact(void *virt, size_t size)
215 {
216 kvm_account_pgtable_pages(virt, -(size >> PAGE_SHIFT));
217 free_pages_exact(virt, size);
218 }
219
220 static struct kvm_pgtable_mm_ops kvm_s2_mm_ops;
221
stage2_free_unlinked_table_rcu_cb(struct rcu_head * head)222 static void stage2_free_unlinked_table_rcu_cb(struct rcu_head *head)
223 {
224 struct page *page = container_of(head, struct page, rcu_head);
225 void *pgtable = page_to_virt(page);
226 s8 level = page_private(page);
227
228 kvm_pgtable_stage2_free_unlinked(&kvm_s2_mm_ops, pgtable, level);
229 }
230
stage2_free_unlinked_table(void * addr,s8 level)231 static void stage2_free_unlinked_table(void *addr, s8 level)
232 {
233 struct page *page = virt_to_page(addr);
234
235 set_page_private(page, (unsigned long)level);
236 call_rcu(&page->rcu_head, stage2_free_unlinked_table_rcu_cb);
237 }
238
kvm_host_get_page(void * addr)239 static void kvm_host_get_page(void *addr)
240 {
241 get_page(virt_to_page(addr));
242 }
243
kvm_host_put_page(void * addr)244 static void kvm_host_put_page(void *addr)
245 {
246 put_page(virt_to_page(addr));
247 }
248
kvm_s2_put_page(void * addr)249 static void kvm_s2_put_page(void *addr)
250 {
251 struct page *p = virt_to_page(addr);
252 /* Dropping last refcount, the page will be freed */
253 if (page_count(p) == 1)
254 kvm_account_pgtable_pages(addr, -1);
255 put_page(p);
256 }
257
kvm_host_page_count(void * addr)258 static int kvm_host_page_count(void *addr)
259 {
260 return page_count(virt_to_page(addr));
261 }
262
kvm_host_pa(void * addr)263 static phys_addr_t kvm_host_pa(void *addr)
264 {
265 return __pa(addr);
266 }
267
kvm_host_va(phys_addr_t phys)268 static void *kvm_host_va(phys_addr_t phys)
269 {
270 return __va(phys);
271 }
272
clean_dcache_guest_page(void * va,size_t size)273 static void clean_dcache_guest_page(void *va, size_t size)
274 {
275 __clean_dcache_guest_page(va, size);
276 }
277
invalidate_icache_guest_page(void * va,size_t size)278 static void invalidate_icache_guest_page(void *va, size_t size)
279 {
280 __invalidate_icache_guest_page(va, size);
281 }
282
283 /*
284 * Unmapping vs dcache management:
285 *
286 * If a guest maps certain memory pages as uncached, all writes will
287 * bypass the data cache and go directly to RAM. However, the CPUs
288 * can still speculate reads (not writes) and fill cache lines with
289 * data.
290 *
291 * Those cache lines will be *clean* cache lines though, so a
292 * clean+invalidate operation is equivalent to an invalidate
293 * operation, because no cache lines are marked dirty.
294 *
295 * Those clean cache lines could be filled prior to an uncached write
296 * by the guest, and the cache coherent IO subsystem would therefore
297 * end up writing old data to disk.
298 *
299 * This is why right after unmapping a page/section and invalidating
300 * the corresponding TLBs, we flush to make sure the IO subsystem will
301 * never hit in the cache.
302 *
303 * This is all avoided on systems that have ARM64_HAS_STAGE2_FWB, as
304 * we then fully enforce cacheability of RAM, no matter what the guest
305 * does.
306 */
307 /**
308 * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
309 * @mmu: The KVM stage-2 MMU pointer
310 * @start: The intermediate physical base address of the range to unmap
311 * @size: The size of the area to unmap
312 * @may_block: Whether or not we are permitted to block
313 *
314 * Clear a range of stage-2 mappings, lowering the various ref-counts. Must
315 * be called while holding mmu_lock (unless for freeing the stage2 pgd before
316 * destroying the VM), otherwise another faulting VCPU may come in and mess
317 * with things behind our backs.
318 */
__unmap_stage2_range(struct kvm_s2_mmu * mmu,phys_addr_t start,u64 size,bool may_block)319 static void __unmap_stage2_range(struct kvm_s2_mmu *mmu, phys_addr_t start, u64 size,
320 bool may_block)
321 {
322 struct kvm *kvm = kvm_s2_mmu_to_kvm(mmu);
323 phys_addr_t end = start + size;
324
325 lockdep_assert_held_write(&kvm->mmu_lock);
326 WARN_ON(size & ~PAGE_MASK);
327 WARN_ON(stage2_apply_range(mmu, start, end, kvm_pgtable_stage2_unmap,
328 may_block));
329 }
330
unmap_stage2_range(struct kvm_s2_mmu * mmu,phys_addr_t start,u64 size)331 static void unmap_stage2_range(struct kvm_s2_mmu *mmu, phys_addr_t start, u64 size)
332 {
333 __unmap_stage2_range(mmu, start, size, true);
334 }
335
stage2_flush_memslot(struct kvm * kvm,struct kvm_memory_slot * memslot)336 static void stage2_flush_memslot(struct kvm *kvm,
337 struct kvm_memory_slot *memslot)
338 {
339 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
340 phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
341
342 stage2_apply_range_resched(&kvm->arch.mmu, addr, end, kvm_pgtable_stage2_flush);
343 }
344
345 /**
346 * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
347 * @kvm: The struct kvm pointer
348 *
349 * Go through the stage 2 page tables and invalidate any cache lines
350 * backing memory already mapped to the VM.
351 */
stage2_flush_vm(struct kvm * kvm)352 static void stage2_flush_vm(struct kvm *kvm)
353 {
354 struct kvm_memslots *slots;
355 struct kvm_memory_slot *memslot;
356 int idx, bkt;
357
358 idx = srcu_read_lock(&kvm->srcu);
359 write_lock(&kvm->mmu_lock);
360
361 slots = kvm_memslots(kvm);
362 kvm_for_each_memslot(memslot, bkt, slots)
363 stage2_flush_memslot(kvm, memslot);
364
365 write_unlock(&kvm->mmu_lock);
366 srcu_read_unlock(&kvm->srcu, idx);
367 }
368
369 /**
370 * free_hyp_pgds - free Hyp-mode page tables
371 */
free_hyp_pgds(void)372 void __init free_hyp_pgds(void)
373 {
374 mutex_lock(&kvm_hyp_pgd_mutex);
375 if (hyp_pgtable) {
376 kvm_pgtable_hyp_destroy(hyp_pgtable);
377 kfree(hyp_pgtable);
378 hyp_pgtable = NULL;
379 }
380 mutex_unlock(&kvm_hyp_pgd_mutex);
381 }
382
kvm_host_owns_hyp_mappings(void)383 static bool kvm_host_owns_hyp_mappings(void)
384 {
385 if (is_kernel_in_hyp_mode())
386 return false;
387
388 if (static_branch_likely(&kvm_protected_mode_initialized))
389 return false;
390
391 /*
392 * This can happen at boot time when __create_hyp_mappings() is called
393 * after the hyp protection has been enabled, but the static key has
394 * not been flipped yet.
395 */
396 if (!hyp_pgtable && is_protected_kvm_enabled())
397 return false;
398
399 WARN_ON(!hyp_pgtable);
400
401 return true;
402 }
403
__create_hyp_mappings(unsigned long start,unsigned long size,unsigned long phys,enum kvm_pgtable_prot prot)404 int __create_hyp_mappings(unsigned long start, unsigned long size,
405 unsigned long phys, enum kvm_pgtable_prot prot)
406 {
407 int err;
408
409 if (WARN_ON(!kvm_host_owns_hyp_mappings()))
410 return -EINVAL;
411
412 mutex_lock(&kvm_hyp_pgd_mutex);
413 err = kvm_pgtable_hyp_map(hyp_pgtable, start, size, phys, prot);
414 mutex_unlock(&kvm_hyp_pgd_mutex);
415
416 return err;
417 }
418
kvm_kaddr_to_phys(void * kaddr)419 static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
420 {
421 if (!is_vmalloc_addr(kaddr)) {
422 BUG_ON(!virt_addr_valid(kaddr));
423 return __pa(kaddr);
424 } else {
425 return page_to_phys(vmalloc_to_page(kaddr)) +
426 offset_in_page(kaddr);
427 }
428 }
429
430 struct hyp_shared_pfn {
431 u64 pfn;
432 int count;
433 struct rb_node node;
434 };
435
436 static DEFINE_MUTEX(hyp_shared_pfns_lock);
437 static struct rb_root hyp_shared_pfns = RB_ROOT;
438
find_shared_pfn(u64 pfn,struct rb_node *** node,struct rb_node ** parent)439 static struct hyp_shared_pfn *find_shared_pfn(u64 pfn, struct rb_node ***node,
440 struct rb_node **parent)
441 {
442 struct hyp_shared_pfn *this;
443
444 *node = &hyp_shared_pfns.rb_node;
445 *parent = NULL;
446 while (**node) {
447 this = container_of(**node, struct hyp_shared_pfn, node);
448 *parent = **node;
449 if (this->pfn < pfn)
450 *node = &((**node)->rb_left);
451 else if (this->pfn > pfn)
452 *node = &((**node)->rb_right);
453 else
454 return this;
455 }
456
457 return NULL;
458 }
459
share_pfn_hyp(u64 pfn)460 static int share_pfn_hyp(u64 pfn)
461 {
462 struct rb_node **node, *parent;
463 struct hyp_shared_pfn *this;
464 int ret = 0;
465
466 mutex_lock(&hyp_shared_pfns_lock);
467 this = find_shared_pfn(pfn, &node, &parent);
468 if (this) {
469 this->count++;
470 goto unlock;
471 }
472
473 this = kzalloc(sizeof(*this), GFP_KERNEL);
474 if (!this) {
475 ret = -ENOMEM;
476 goto unlock;
477 }
478
479 this->pfn = pfn;
480 this->count = 1;
481 rb_link_node(&this->node, parent, node);
482 rb_insert_color(&this->node, &hyp_shared_pfns);
483 ret = kvm_call_hyp_nvhe(__pkvm_host_share_hyp, pfn, 1);
484 unlock:
485 mutex_unlock(&hyp_shared_pfns_lock);
486
487 return ret;
488 }
489
unshare_pfn_hyp(u64 pfn)490 static int unshare_pfn_hyp(u64 pfn)
491 {
492 struct rb_node **node, *parent;
493 struct hyp_shared_pfn *this;
494 int ret = 0;
495
496 mutex_lock(&hyp_shared_pfns_lock);
497 this = find_shared_pfn(pfn, &node, &parent);
498 if (WARN_ON(!this)) {
499 ret = -ENOENT;
500 goto unlock;
501 }
502
503 this->count--;
504 if (this->count)
505 goto unlock;
506
507 rb_erase(&this->node, &hyp_shared_pfns);
508 kfree(this);
509 ret = kvm_call_hyp_nvhe(__pkvm_host_unshare_hyp, pfn, 1);
510 unlock:
511 mutex_unlock(&hyp_shared_pfns_lock);
512
513 return ret;
514 }
515
kvm_share_hyp(void * from,void * to)516 int kvm_share_hyp(void *from, void *to)
517 {
518 phys_addr_t start, end, cur;
519 u64 pfn;
520 int ret;
521
522 if (is_kernel_in_hyp_mode())
523 return 0;
524
525 /*
526 * The share hcall maps things in the 'fixed-offset' region of the hyp
527 * VA space, so we can only share physically contiguous data-structures
528 * for now.
529 */
530 if (is_vmalloc_or_module_addr(from) || is_vmalloc_or_module_addr(to))
531 return -EINVAL;
532
533 if (kvm_host_owns_hyp_mappings())
534 return create_hyp_mappings(from, to, PAGE_HYP);
535
536 start = ALIGN_DOWN(__pa(from), PAGE_SIZE);
537 end = PAGE_ALIGN(__pa(to));
538 for (cur = start; cur < end; cur += PAGE_SIZE) {
539 pfn = __phys_to_pfn(cur);
540 ret = share_pfn_hyp(pfn);
541 if (ret)
542 return ret;
543 }
544
545 return 0;
546 }
547
kvm_unshare_hyp(void * from,void * to)548 void kvm_unshare_hyp(void *from, void *to)
549 {
550 phys_addr_t start, end, cur;
551 u64 pfn;
552
553 if (is_kernel_in_hyp_mode() || kvm_host_owns_hyp_mappings() || !from)
554 return;
555
556 start = ALIGN_DOWN(__pa(from), PAGE_SIZE);
557 end = PAGE_ALIGN(__pa(to));
558 for (cur = start; cur < end; cur += PAGE_SIZE) {
559 pfn = __phys_to_pfn(cur);
560 WARN_ON(unshare_pfn_hyp(pfn));
561 }
562 }
563
564 /**
565 * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
566 * @from: The virtual kernel start address of the range
567 * @to: The virtual kernel end address of the range (exclusive)
568 * @prot: The protection to be applied to this range
569 *
570 * The same virtual address as the kernel virtual address is also used
571 * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
572 * physical pages.
573 */
create_hyp_mappings(void * from,void * to,enum kvm_pgtable_prot prot)574 int create_hyp_mappings(void *from, void *to, enum kvm_pgtable_prot prot)
575 {
576 phys_addr_t phys_addr;
577 unsigned long virt_addr;
578 unsigned long start = kern_hyp_va((unsigned long)from);
579 unsigned long end = kern_hyp_va((unsigned long)to);
580
581 if (is_kernel_in_hyp_mode())
582 return 0;
583
584 if (!kvm_host_owns_hyp_mappings())
585 return -EPERM;
586
587 start = start & PAGE_MASK;
588 end = PAGE_ALIGN(end);
589
590 for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
591 int err;
592
593 phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
594 err = __create_hyp_mappings(virt_addr, PAGE_SIZE, phys_addr,
595 prot);
596 if (err)
597 return err;
598 }
599
600 return 0;
601 }
602
__hyp_alloc_private_va_range(unsigned long base)603 static int __hyp_alloc_private_va_range(unsigned long base)
604 {
605 lockdep_assert_held(&kvm_hyp_pgd_mutex);
606
607 if (!PAGE_ALIGNED(base))
608 return -EINVAL;
609
610 /*
611 * Verify that BIT(VA_BITS - 1) hasn't been flipped by
612 * allocating the new area, as it would indicate we've
613 * overflowed the idmap/IO address range.
614 */
615 if ((base ^ io_map_base) & BIT(VA_BITS - 1))
616 return -ENOMEM;
617
618 io_map_base = base;
619
620 return 0;
621 }
622
623 /**
624 * hyp_alloc_private_va_range - Allocates a private VA range.
625 * @size: The size of the VA range to reserve.
626 * @haddr: The hypervisor virtual start address of the allocation.
627 *
628 * The private virtual address (VA) range is allocated below io_map_base
629 * and aligned based on the order of @size.
630 *
631 * Return: 0 on success or negative error code on failure.
632 */
hyp_alloc_private_va_range(size_t size,unsigned long * haddr)633 int hyp_alloc_private_va_range(size_t size, unsigned long *haddr)
634 {
635 unsigned long base;
636 int ret = 0;
637
638 mutex_lock(&kvm_hyp_pgd_mutex);
639
640 /*
641 * This assumes that we have enough space below the idmap
642 * page to allocate our VAs. If not, the check in
643 * __hyp_alloc_private_va_range() will kick. A potential
644 * alternative would be to detect that overflow and switch
645 * to an allocation above the idmap.
646 *
647 * The allocated size is always a multiple of PAGE_SIZE.
648 */
649 size = PAGE_ALIGN(size);
650 base = io_map_base - size;
651 ret = __hyp_alloc_private_va_range(base);
652
653 mutex_unlock(&kvm_hyp_pgd_mutex);
654
655 if (!ret)
656 *haddr = base;
657
658 return ret;
659 }
660
__create_hyp_private_mapping(phys_addr_t phys_addr,size_t size,unsigned long * haddr,enum kvm_pgtable_prot prot)661 static int __create_hyp_private_mapping(phys_addr_t phys_addr, size_t size,
662 unsigned long *haddr,
663 enum kvm_pgtable_prot prot)
664 {
665 unsigned long addr;
666 int ret = 0;
667
668 if (!kvm_host_owns_hyp_mappings()) {
669 addr = kvm_call_hyp_nvhe(__pkvm_create_private_mapping,
670 phys_addr, size, prot);
671 if (IS_ERR_VALUE(addr))
672 return addr;
673 *haddr = addr;
674
675 return 0;
676 }
677
678 size = PAGE_ALIGN(size + offset_in_page(phys_addr));
679 ret = hyp_alloc_private_va_range(size, &addr);
680 if (ret)
681 return ret;
682
683 ret = __create_hyp_mappings(addr, size, phys_addr, prot);
684 if (ret)
685 return ret;
686
687 *haddr = addr + offset_in_page(phys_addr);
688 return ret;
689 }
690
create_hyp_stack(phys_addr_t phys_addr,unsigned long * haddr)691 int create_hyp_stack(phys_addr_t phys_addr, unsigned long *haddr)
692 {
693 unsigned long base;
694 size_t size;
695 int ret;
696
697 mutex_lock(&kvm_hyp_pgd_mutex);
698 /*
699 * Efficient stack verification using the PAGE_SHIFT bit implies
700 * an alignment of our allocation on the order of the size.
701 */
702 size = PAGE_SIZE * 2;
703 base = ALIGN_DOWN(io_map_base - size, size);
704
705 ret = __hyp_alloc_private_va_range(base);
706
707 mutex_unlock(&kvm_hyp_pgd_mutex);
708
709 if (ret) {
710 kvm_err("Cannot allocate hyp stack guard page\n");
711 return ret;
712 }
713
714 /*
715 * Since the stack grows downwards, map the stack to the page
716 * at the higher address and leave the lower guard page
717 * unbacked.
718 *
719 * Any valid stack address now has the PAGE_SHIFT bit as 1
720 * and addresses corresponding to the guard page have the
721 * PAGE_SHIFT bit as 0 - this is used for overflow detection.
722 */
723 ret = __create_hyp_mappings(base + PAGE_SIZE, PAGE_SIZE, phys_addr,
724 PAGE_HYP);
725 if (ret)
726 kvm_err("Cannot map hyp stack\n");
727
728 *haddr = base + size;
729
730 return ret;
731 }
732
733 /**
734 * create_hyp_io_mappings - Map IO into both kernel and HYP
735 * @phys_addr: The physical start address which gets mapped
736 * @size: Size of the region being mapped
737 * @kaddr: Kernel VA for this mapping
738 * @haddr: HYP VA for this mapping
739 */
create_hyp_io_mappings(phys_addr_t phys_addr,size_t size,void __iomem ** kaddr,void __iomem ** haddr)740 int create_hyp_io_mappings(phys_addr_t phys_addr, size_t size,
741 void __iomem **kaddr,
742 void __iomem **haddr)
743 {
744 unsigned long addr;
745 int ret;
746
747 if (is_protected_kvm_enabled())
748 return -EPERM;
749
750 *kaddr = ioremap(phys_addr, size);
751 if (!*kaddr)
752 return -ENOMEM;
753
754 if (is_kernel_in_hyp_mode()) {
755 *haddr = *kaddr;
756 return 0;
757 }
758
759 ret = __create_hyp_private_mapping(phys_addr, size,
760 &addr, PAGE_HYP_DEVICE);
761 if (ret) {
762 iounmap(*kaddr);
763 *kaddr = NULL;
764 *haddr = NULL;
765 return ret;
766 }
767
768 *haddr = (void __iomem *)addr;
769 return 0;
770 }
771
772 /**
773 * create_hyp_exec_mappings - Map an executable range into HYP
774 * @phys_addr: The physical start address which gets mapped
775 * @size: Size of the region being mapped
776 * @haddr: HYP VA for this mapping
777 */
create_hyp_exec_mappings(phys_addr_t phys_addr,size_t size,void ** haddr)778 int create_hyp_exec_mappings(phys_addr_t phys_addr, size_t size,
779 void **haddr)
780 {
781 unsigned long addr;
782 int ret;
783
784 BUG_ON(is_kernel_in_hyp_mode());
785
786 ret = __create_hyp_private_mapping(phys_addr, size,
787 &addr, PAGE_HYP_EXEC);
788 if (ret) {
789 *haddr = NULL;
790 return ret;
791 }
792
793 *haddr = (void *)addr;
794 return 0;
795 }
796
797 static struct kvm_pgtable_mm_ops kvm_user_mm_ops = {
798 /* We shouldn't need any other callback to walk the PT */
799 .phys_to_virt = kvm_host_va,
800 };
801
get_user_mapping_size(struct kvm * kvm,u64 addr)802 static int get_user_mapping_size(struct kvm *kvm, u64 addr)
803 {
804 struct kvm_pgtable pgt = {
805 .pgd = (kvm_pteref_t)kvm->mm->pgd,
806 .ia_bits = vabits_actual,
807 .start_level = (KVM_PGTABLE_LAST_LEVEL -
808 CONFIG_PGTABLE_LEVELS + 1),
809 .mm_ops = &kvm_user_mm_ops,
810 };
811 unsigned long flags;
812 kvm_pte_t pte = 0; /* Keep GCC quiet... */
813 s8 level = S8_MAX;
814 int ret;
815
816 /*
817 * Disable IRQs so that we hazard against a concurrent
818 * teardown of the userspace page tables (which relies on
819 * IPI-ing threads).
820 */
821 local_irq_save(flags);
822 ret = kvm_pgtable_get_leaf(&pgt, addr, &pte, &level);
823 local_irq_restore(flags);
824
825 if (ret)
826 return ret;
827
828 /*
829 * Not seeing an error, but not updating level? Something went
830 * deeply wrong...
831 */
832 if (WARN_ON(level > KVM_PGTABLE_LAST_LEVEL))
833 return -EFAULT;
834 if (WARN_ON(level < KVM_PGTABLE_FIRST_LEVEL))
835 return -EFAULT;
836
837 /* Oops, the userspace PTs are gone... Replay the fault */
838 if (!kvm_pte_valid(pte))
839 return -EAGAIN;
840
841 return BIT(ARM64_HW_PGTABLE_LEVEL_SHIFT(level));
842 }
843
844 static struct kvm_pgtable_mm_ops kvm_s2_mm_ops = {
845 .zalloc_page = stage2_memcache_zalloc_page,
846 .zalloc_pages_exact = kvm_s2_zalloc_pages_exact,
847 .free_pages_exact = kvm_s2_free_pages_exact,
848 .free_unlinked_table = stage2_free_unlinked_table,
849 .get_page = kvm_host_get_page,
850 .put_page = kvm_s2_put_page,
851 .page_count = kvm_host_page_count,
852 .phys_to_virt = kvm_host_va,
853 .virt_to_phys = kvm_host_pa,
854 .dcache_clean_inval_poc = clean_dcache_guest_page,
855 .icache_inval_pou = invalidate_icache_guest_page,
856 };
857
858 /**
859 * kvm_init_stage2_mmu - Initialise a S2 MMU structure
860 * @kvm: The pointer to the KVM structure
861 * @mmu: The pointer to the s2 MMU structure
862 * @type: The machine type of the virtual machine
863 *
864 * Allocates only the stage-2 HW PGD level table(s).
865 * Note we don't need locking here as this is only called when the VM is
866 * created, which can only be done once.
867 */
kvm_init_stage2_mmu(struct kvm * kvm,struct kvm_s2_mmu * mmu,unsigned long type)868 int kvm_init_stage2_mmu(struct kvm *kvm, struct kvm_s2_mmu *mmu, unsigned long type)
869 {
870 u32 kvm_ipa_limit = get_kvm_ipa_limit();
871 int cpu, err;
872 struct kvm_pgtable *pgt;
873 u64 mmfr0, mmfr1;
874 u32 phys_shift;
875
876 if (type & ~KVM_VM_TYPE_ARM_IPA_SIZE_MASK)
877 return -EINVAL;
878
879 phys_shift = KVM_VM_TYPE_ARM_IPA_SIZE(type);
880 if (is_protected_kvm_enabled()) {
881 phys_shift = kvm_ipa_limit;
882 } else if (phys_shift) {
883 if (phys_shift > kvm_ipa_limit ||
884 phys_shift < ARM64_MIN_PARANGE_BITS)
885 return -EINVAL;
886 } else {
887 phys_shift = KVM_PHYS_SHIFT;
888 if (phys_shift > kvm_ipa_limit) {
889 pr_warn_once("%s using unsupported default IPA limit, upgrade your VMM\n",
890 current->comm);
891 return -EINVAL;
892 }
893 }
894
895 mmfr0 = read_sanitised_ftr_reg(SYS_ID_AA64MMFR0_EL1);
896 mmfr1 = read_sanitised_ftr_reg(SYS_ID_AA64MMFR1_EL1);
897 mmu->vtcr = kvm_get_vtcr(mmfr0, mmfr1, phys_shift);
898
899 if (mmu->pgt != NULL) {
900 kvm_err("kvm_arch already initialized?\n");
901 return -EINVAL;
902 }
903
904 pgt = kzalloc(sizeof(*pgt), GFP_KERNEL_ACCOUNT);
905 if (!pgt)
906 return -ENOMEM;
907
908 mmu->arch = &kvm->arch;
909 err = kvm_pgtable_stage2_init(pgt, mmu, &kvm_s2_mm_ops);
910 if (err)
911 goto out_free_pgtable;
912
913 mmu->last_vcpu_ran = alloc_percpu(typeof(*mmu->last_vcpu_ran));
914 if (!mmu->last_vcpu_ran) {
915 err = -ENOMEM;
916 goto out_destroy_pgtable;
917 }
918
919 for_each_possible_cpu(cpu)
920 *per_cpu_ptr(mmu->last_vcpu_ran, cpu) = -1;
921
922 /* The eager page splitting is disabled by default */
923 mmu->split_page_chunk_size = KVM_ARM_EAGER_SPLIT_CHUNK_SIZE_DEFAULT;
924 mmu->split_page_cache.gfp_zero = __GFP_ZERO;
925
926 mmu->pgt = pgt;
927 mmu->pgd_phys = __pa(pgt->pgd);
928 return 0;
929
930 out_destroy_pgtable:
931 kvm_pgtable_stage2_destroy(pgt);
932 out_free_pgtable:
933 kfree(pgt);
934 return err;
935 }
936
kvm_uninit_stage2_mmu(struct kvm * kvm)937 void kvm_uninit_stage2_mmu(struct kvm *kvm)
938 {
939 kvm_free_stage2_pgd(&kvm->arch.mmu);
940 kvm_mmu_free_memory_cache(&kvm->arch.mmu.split_page_cache);
941 }
942
stage2_unmap_memslot(struct kvm * kvm,struct kvm_memory_slot * memslot)943 static void stage2_unmap_memslot(struct kvm *kvm,
944 struct kvm_memory_slot *memslot)
945 {
946 hva_t hva = memslot->userspace_addr;
947 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
948 phys_addr_t size = PAGE_SIZE * memslot->npages;
949 hva_t reg_end = hva + size;
950
951 /*
952 * A memory region could potentially cover multiple VMAs, and any holes
953 * between them, so iterate over all of them to find out if we should
954 * unmap any of them.
955 *
956 * +--------------------------------------------+
957 * +---------------+----------------+ +----------------+
958 * | : VMA 1 | VMA 2 | | VMA 3 : |
959 * +---------------+----------------+ +----------------+
960 * | memory region |
961 * +--------------------------------------------+
962 */
963 do {
964 struct vm_area_struct *vma;
965 hva_t vm_start, vm_end;
966
967 vma = find_vma_intersection(current->mm, hva, reg_end);
968 if (!vma)
969 break;
970
971 /*
972 * Take the intersection of this VMA with the memory region
973 */
974 vm_start = max(hva, vma->vm_start);
975 vm_end = min(reg_end, vma->vm_end);
976
977 if (!(vma->vm_flags & VM_PFNMAP)) {
978 gpa_t gpa = addr + (vm_start - memslot->userspace_addr);
979 unmap_stage2_range(&kvm->arch.mmu, gpa, vm_end - vm_start);
980 }
981 hva = vm_end;
982 } while (hva < reg_end);
983 }
984
985 /**
986 * stage2_unmap_vm - Unmap Stage-2 RAM mappings
987 * @kvm: The struct kvm pointer
988 *
989 * Go through the memregions and unmap any regular RAM
990 * backing memory already mapped to the VM.
991 */
stage2_unmap_vm(struct kvm * kvm)992 void stage2_unmap_vm(struct kvm *kvm)
993 {
994 struct kvm_memslots *slots;
995 struct kvm_memory_slot *memslot;
996 int idx, bkt;
997
998 idx = srcu_read_lock(&kvm->srcu);
999 mmap_read_lock(current->mm);
1000 write_lock(&kvm->mmu_lock);
1001
1002 slots = kvm_memslots(kvm);
1003 kvm_for_each_memslot(memslot, bkt, slots)
1004 stage2_unmap_memslot(kvm, memslot);
1005
1006 write_unlock(&kvm->mmu_lock);
1007 mmap_read_unlock(current->mm);
1008 srcu_read_unlock(&kvm->srcu, idx);
1009 }
1010
kvm_free_stage2_pgd(struct kvm_s2_mmu * mmu)1011 void kvm_free_stage2_pgd(struct kvm_s2_mmu *mmu)
1012 {
1013 struct kvm *kvm = kvm_s2_mmu_to_kvm(mmu);
1014 struct kvm_pgtable *pgt = NULL;
1015
1016 write_lock(&kvm->mmu_lock);
1017 pgt = mmu->pgt;
1018 if (pgt) {
1019 mmu->pgd_phys = 0;
1020 mmu->pgt = NULL;
1021 free_percpu(mmu->last_vcpu_ran);
1022 }
1023 write_unlock(&kvm->mmu_lock);
1024
1025 if (pgt) {
1026 kvm_pgtable_stage2_destroy(pgt);
1027 kfree(pgt);
1028 }
1029 }
1030
hyp_mc_free_fn(void * addr,void * unused)1031 static void hyp_mc_free_fn(void *addr, void *unused)
1032 {
1033 free_page((unsigned long)addr);
1034 }
1035
hyp_mc_alloc_fn(void * unused)1036 static void *hyp_mc_alloc_fn(void *unused)
1037 {
1038 return (void *)__get_free_page(GFP_KERNEL_ACCOUNT);
1039 }
1040
free_hyp_memcache(struct kvm_hyp_memcache * mc)1041 void free_hyp_memcache(struct kvm_hyp_memcache *mc)
1042 {
1043 if (is_protected_kvm_enabled())
1044 __free_hyp_memcache(mc, hyp_mc_free_fn,
1045 kvm_host_va, NULL);
1046 }
1047
topup_hyp_memcache(struct kvm_hyp_memcache * mc,unsigned long min_pages)1048 int topup_hyp_memcache(struct kvm_hyp_memcache *mc, unsigned long min_pages)
1049 {
1050 if (!is_protected_kvm_enabled())
1051 return 0;
1052
1053 return __topup_hyp_memcache(mc, min_pages, hyp_mc_alloc_fn,
1054 kvm_host_pa, NULL);
1055 }
1056
1057 /**
1058 * kvm_phys_addr_ioremap - map a device range to guest IPA
1059 *
1060 * @kvm: The KVM pointer
1061 * @guest_ipa: The IPA at which to insert the mapping
1062 * @pa: The physical address of the device
1063 * @size: The size of the mapping
1064 * @writable: Whether or not to create a writable mapping
1065 */
kvm_phys_addr_ioremap(struct kvm * kvm,phys_addr_t guest_ipa,phys_addr_t pa,unsigned long size,bool writable)1066 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
1067 phys_addr_t pa, unsigned long size, bool writable)
1068 {
1069 phys_addr_t addr;
1070 int ret = 0;
1071 struct kvm_mmu_memory_cache cache = { .gfp_zero = __GFP_ZERO };
1072 struct kvm_s2_mmu *mmu = &kvm->arch.mmu;
1073 struct kvm_pgtable *pgt = mmu->pgt;
1074 enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_DEVICE |
1075 KVM_PGTABLE_PROT_R |
1076 (writable ? KVM_PGTABLE_PROT_W : 0);
1077
1078 if (is_protected_kvm_enabled())
1079 return -EPERM;
1080
1081 size += offset_in_page(guest_ipa);
1082 guest_ipa &= PAGE_MASK;
1083
1084 for (addr = guest_ipa; addr < guest_ipa + size; addr += PAGE_SIZE) {
1085 ret = kvm_mmu_topup_memory_cache(&cache,
1086 kvm_mmu_cache_min_pages(mmu));
1087 if (ret)
1088 break;
1089
1090 write_lock(&kvm->mmu_lock);
1091 ret = kvm_pgtable_stage2_map(pgt, addr, PAGE_SIZE, pa, prot,
1092 &cache, 0);
1093 write_unlock(&kvm->mmu_lock);
1094 if (ret)
1095 break;
1096
1097 pa += PAGE_SIZE;
1098 }
1099
1100 kvm_mmu_free_memory_cache(&cache);
1101 return ret;
1102 }
1103
1104 /**
1105 * stage2_wp_range() - write protect stage2 memory region range
1106 * @mmu: The KVM stage-2 MMU pointer
1107 * @addr: Start address of range
1108 * @end: End address of range
1109 */
stage2_wp_range(struct kvm_s2_mmu * mmu,phys_addr_t addr,phys_addr_t end)1110 static void stage2_wp_range(struct kvm_s2_mmu *mmu, phys_addr_t addr, phys_addr_t end)
1111 {
1112 stage2_apply_range_resched(mmu, addr, end, kvm_pgtable_stage2_wrprotect);
1113 }
1114
1115 /**
1116 * kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
1117 * @kvm: The KVM pointer
1118 * @slot: The memory slot to write protect
1119 *
1120 * Called to start logging dirty pages after memory region
1121 * KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
1122 * all present PUD, PMD and PTEs are write protected in the memory region.
1123 * Afterwards read of dirty page log can be called.
1124 *
1125 * Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
1126 * serializing operations for VM memory regions.
1127 */
kvm_mmu_wp_memory_region(struct kvm * kvm,int slot)1128 static void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
1129 {
1130 struct kvm_memslots *slots = kvm_memslots(kvm);
1131 struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
1132 phys_addr_t start, end;
1133
1134 if (WARN_ON_ONCE(!memslot))
1135 return;
1136
1137 start = memslot->base_gfn << PAGE_SHIFT;
1138 end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1139
1140 write_lock(&kvm->mmu_lock);
1141 stage2_wp_range(&kvm->arch.mmu, start, end);
1142 write_unlock(&kvm->mmu_lock);
1143 kvm_flush_remote_tlbs_memslot(kvm, memslot);
1144 }
1145
1146 /**
1147 * kvm_mmu_split_memory_region() - split the stage 2 blocks into PAGE_SIZE
1148 * pages for memory slot
1149 * @kvm: The KVM pointer
1150 * @slot: The memory slot to split
1151 *
1152 * Acquires kvm->mmu_lock. Called with kvm->slots_lock mutex acquired,
1153 * serializing operations for VM memory regions.
1154 */
kvm_mmu_split_memory_region(struct kvm * kvm,int slot)1155 static void kvm_mmu_split_memory_region(struct kvm *kvm, int slot)
1156 {
1157 struct kvm_memslots *slots;
1158 struct kvm_memory_slot *memslot;
1159 phys_addr_t start, end;
1160
1161 lockdep_assert_held(&kvm->slots_lock);
1162
1163 slots = kvm_memslots(kvm);
1164 memslot = id_to_memslot(slots, slot);
1165
1166 start = memslot->base_gfn << PAGE_SHIFT;
1167 end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1168
1169 write_lock(&kvm->mmu_lock);
1170 kvm_mmu_split_huge_pages(kvm, start, end);
1171 write_unlock(&kvm->mmu_lock);
1172 }
1173
1174 /*
1175 * kvm_arch_mmu_enable_log_dirty_pt_masked() - enable dirty logging for selected pages.
1176 * @kvm: The KVM pointer
1177 * @slot: The memory slot associated with mask
1178 * @gfn_offset: The gfn offset in memory slot
1179 * @mask: The mask of pages at offset 'gfn_offset' in this memory
1180 * slot to enable dirty logging on
1181 *
1182 * Writes protect selected pages to enable dirty logging, and then
1183 * splits them to PAGE_SIZE. Caller must acquire kvm->mmu_lock.
1184 */
kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm * kvm,struct kvm_memory_slot * slot,gfn_t gfn_offset,unsigned long mask)1185 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1186 struct kvm_memory_slot *slot,
1187 gfn_t gfn_offset, unsigned long mask)
1188 {
1189 phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
1190 phys_addr_t start = (base_gfn + __ffs(mask)) << PAGE_SHIFT;
1191 phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
1192
1193 lockdep_assert_held_write(&kvm->mmu_lock);
1194
1195 stage2_wp_range(&kvm->arch.mmu, start, end);
1196
1197 /*
1198 * Eager-splitting is done when manual-protect is set. We
1199 * also check for initially-all-set because we can avoid
1200 * eager-splitting if initially-all-set is false.
1201 * Initially-all-set equal false implies that huge-pages were
1202 * already split when enabling dirty logging: no need to do it
1203 * again.
1204 */
1205 if (kvm_dirty_log_manual_protect_and_init_set(kvm))
1206 kvm_mmu_split_huge_pages(kvm, start, end);
1207 }
1208
kvm_send_hwpoison_signal(unsigned long address,short lsb)1209 static void kvm_send_hwpoison_signal(unsigned long address, short lsb)
1210 {
1211 send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb, current);
1212 }
1213
fault_supports_stage2_huge_mapping(struct kvm_memory_slot * memslot,unsigned long hva,unsigned long map_size)1214 static bool fault_supports_stage2_huge_mapping(struct kvm_memory_slot *memslot,
1215 unsigned long hva,
1216 unsigned long map_size)
1217 {
1218 gpa_t gpa_start;
1219 hva_t uaddr_start, uaddr_end;
1220 size_t size;
1221
1222 /* The memslot and the VMA are guaranteed to be aligned to PAGE_SIZE */
1223 if (map_size == PAGE_SIZE)
1224 return true;
1225
1226 size = memslot->npages * PAGE_SIZE;
1227
1228 gpa_start = memslot->base_gfn << PAGE_SHIFT;
1229
1230 uaddr_start = memslot->userspace_addr;
1231 uaddr_end = uaddr_start + size;
1232
1233 /*
1234 * Pages belonging to memslots that don't have the same alignment
1235 * within a PMD/PUD for userspace and IPA cannot be mapped with stage-2
1236 * PMD/PUD entries, because we'll end up mapping the wrong pages.
1237 *
1238 * Consider a layout like the following:
1239 *
1240 * memslot->userspace_addr:
1241 * +-----+--------------------+--------------------+---+
1242 * |abcde|fgh Stage-1 block | Stage-1 block tv|xyz|
1243 * +-----+--------------------+--------------------+---+
1244 *
1245 * memslot->base_gfn << PAGE_SHIFT:
1246 * +---+--------------------+--------------------+-----+
1247 * |abc|def Stage-2 block | Stage-2 block |tvxyz|
1248 * +---+--------------------+--------------------+-----+
1249 *
1250 * If we create those stage-2 blocks, we'll end up with this incorrect
1251 * mapping:
1252 * d -> f
1253 * e -> g
1254 * f -> h
1255 */
1256 if ((gpa_start & (map_size - 1)) != (uaddr_start & (map_size - 1)))
1257 return false;
1258
1259 /*
1260 * Next, let's make sure we're not trying to map anything not covered
1261 * by the memslot. This means we have to prohibit block size mappings
1262 * for the beginning and end of a non-block aligned and non-block sized
1263 * memory slot (illustrated by the head and tail parts of the
1264 * userspace view above containing pages 'abcde' and 'xyz',
1265 * respectively).
1266 *
1267 * Note that it doesn't matter if we do the check using the
1268 * userspace_addr or the base_gfn, as both are equally aligned (per
1269 * the check above) and equally sized.
1270 */
1271 return (hva & ~(map_size - 1)) >= uaddr_start &&
1272 (hva & ~(map_size - 1)) + map_size <= uaddr_end;
1273 }
1274
1275 /*
1276 * Check if the given hva is backed by a transparent huge page (THP) and
1277 * whether it can be mapped using block mapping in stage2. If so, adjust
1278 * the stage2 PFN and IPA accordingly. Only PMD_SIZE THPs are currently
1279 * supported. This will need to be updated to support other THP sizes.
1280 *
1281 * Returns the size of the mapping.
1282 */
1283 static long
transparent_hugepage_adjust(struct kvm * kvm,struct kvm_memory_slot * memslot,unsigned long hva,kvm_pfn_t * pfnp,phys_addr_t * ipap)1284 transparent_hugepage_adjust(struct kvm *kvm, struct kvm_memory_slot *memslot,
1285 unsigned long hva, kvm_pfn_t *pfnp,
1286 phys_addr_t *ipap)
1287 {
1288 kvm_pfn_t pfn = *pfnp;
1289
1290 /*
1291 * Make sure the adjustment is done only for THP pages. Also make
1292 * sure that the HVA and IPA are sufficiently aligned and that the
1293 * block map is contained within the memslot.
1294 */
1295 if (fault_supports_stage2_huge_mapping(memslot, hva, PMD_SIZE)) {
1296 int sz = get_user_mapping_size(kvm, hva);
1297
1298 if (sz < 0)
1299 return sz;
1300
1301 if (sz < PMD_SIZE)
1302 return PAGE_SIZE;
1303
1304 *ipap &= PMD_MASK;
1305 pfn &= ~(PTRS_PER_PMD - 1);
1306 *pfnp = pfn;
1307
1308 return PMD_SIZE;
1309 }
1310
1311 /* Use page mapping if we cannot use block mapping. */
1312 return PAGE_SIZE;
1313 }
1314
get_vma_page_shift(struct vm_area_struct * vma,unsigned long hva)1315 static int get_vma_page_shift(struct vm_area_struct *vma, unsigned long hva)
1316 {
1317 unsigned long pa;
1318
1319 if (is_vm_hugetlb_page(vma) && !(vma->vm_flags & VM_PFNMAP))
1320 return huge_page_shift(hstate_vma(vma));
1321
1322 if (!(vma->vm_flags & VM_PFNMAP))
1323 return PAGE_SHIFT;
1324
1325 VM_BUG_ON(is_vm_hugetlb_page(vma));
1326
1327 pa = (vma->vm_pgoff << PAGE_SHIFT) + (hva - vma->vm_start);
1328
1329 #ifndef __PAGETABLE_PMD_FOLDED
1330 if ((hva & (PUD_SIZE - 1)) == (pa & (PUD_SIZE - 1)) &&
1331 ALIGN_DOWN(hva, PUD_SIZE) >= vma->vm_start &&
1332 ALIGN(hva, PUD_SIZE) <= vma->vm_end)
1333 return PUD_SHIFT;
1334 #endif
1335
1336 if ((hva & (PMD_SIZE - 1)) == (pa & (PMD_SIZE - 1)) &&
1337 ALIGN_DOWN(hva, PMD_SIZE) >= vma->vm_start &&
1338 ALIGN(hva, PMD_SIZE) <= vma->vm_end)
1339 return PMD_SHIFT;
1340
1341 return PAGE_SHIFT;
1342 }
1343
1344 /*
1345 * The page will be mapped in stage 2 as Normal Cacheable, so the VM will be
1346 * able to see the page's tags and therefore they must be initialised first. If
1347 * PG_mte_tagged is set, tags have already been initialised.
1348 *
1349 * The race in the test/set of the PG_mte_tagged flag is handled by:
1350 * - preventing VM_SHARED mappings in a memslot with MTE preventing two VMs
1351 * racing to santise the same page
1352 * - mmap_lock protects between a VM faulting a page in and the VMM performing
1353 * an mprotect() to add VM_MTE
1354 */
sanitise_mte_tags(struct kvm * kvm,kvm_pfn_t pfn,unsigned long size)1355 static void sanitise_mte_tags(struct kvm *kvm, kvm_pfn_t pfn,
1356 unsigned long size)
1357 {
1358 unsigned long i, nr_pages = size >> PAGE_SHIFT;
1359 struct page *page = pfn_to_page(pfn);
1360
1361 if (!kvm_has_mte(kvm))
1362 return;
1363
1364 for (i = 0; i < nr_pages; i++, page++) {
1365 if (try_page_mte_tagging(page)) {
1366 mte_clear_page_tags(page_address(page));
1367 set_page_mte_tagged(page);
1368 }
1369 }
1370 }
1371
kvm_vma_mte_allowed(struct vm_area_struct * vma)1372 static bool kvm_vma_mte_allowed(struct vm_area_struct *vma)
1373 {
1374 return vma->vm_flags & VM_MTE_ALLOWED;
1375 }
1376
user_mem_abort(struct kvm_vcpu * vcpu,phys_addr_t fault_ipa,struct kvm_memory_slot * memslot,unsigned long hva,bool fault_is_perm)1377 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1378 struct kvm_memory_slot *memslot, unsigned long hva,
1379 bool fault_is_perm)
1380 {
1381 int ret = 0;
1382 bool write_fault, writable, force_pte = false;
1383 bool exec_fault, mte_allowed;
1384 bool device = false;
1385 unsigned long mmu_seq;
1386 struct kvm *kvm = vcpu->kvm;
1387 struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
1388 struct vm_area_struct *vma;
1389 short vma_shift;
1390 gfn_t gfn;
1391 kvm_pfn_t pfn;
1392 bool logging_active = memslot_is_logging(memslot);
1393 long vma_pagesize, fault_granule;
1394 enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_R;
1395 struct kvm_pgtable *pgt;
1396
1397 if (fault_is_perm)
1398 fault_granule = kvm_vcpu_trap_get_perm_fault_granule(vcpu);
1399 write_fault = kvm_is_write_fault(vcpu);
1400 exec_fault = kvm_vcpu_trap_is_exec_fault(vcpu);
1401 VM_BUG_ON(write_fault && exec_fault);
1402
1403 if (fault_is_perm && !write_fault && !exec_fault) {
1404 kvm_err("Unexpected L2 read permission error\n");
1405 return -EFAULT;
1406 }
1407
1408 /*
1409 * Permission faults just need to update the existing leaf entry,
1410 * and so normally don't require allocations from the memcache. The
1411 * only exception to this is when dirty logging is enabled at runtime
1412 * and a write fault needs to collapse a block entry into a table.
1413 */
1414 if (!fault_is_perm || (logging_active && write_fault)) {
1415 ret = kvm_mmu_topup_memory_cache(memcache,
1416 kvm_mmu_cache_min_pages(vcpu->arch.hw_mmu));
1417 if (ret)
1418 return ret;
1419 }
1420
1421 /*
1422 * Let's check if we will get back a huge page backed by hugetlbfs, or
1423 * get block mapping for device MMIO region.
1424 */
1425 mmap_read_lock(current->mm);
1426 vma = vma_lookup(current->mm, hva);
1427 if (unlikely(!vma)) {
1428 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
1429 mmap_read_unlock(current->mm);
1430 return -EFAULT;
1431 }
1432
1433 /*
1434 * logging_active is guaranteed to never be true for VM_PFNMAP
1435 * memslots.
1436 */
1437 if (logging_active) {
1438 force_pte = true;
1439 vma_shift = PAGE_SHIFT;
1440 } else {
1441 vma_shift = get_vma_page_shift(vma, hva);
1442 }
1443
1444 switch (vma_shift) {
1445 #ifndef __PAGETABLE_PMD_FOLDED
1446 case PUD_SHIFT:
1447 if (fault_supports_stage2_huge_mapping(memslot, hva, PUD_SIZE))
1448 break;
1449 fallthrough;
1450 #endif
1451 case CONT_PMD_SHIFT:
1452 vma_shift = PMD_SHIFT;
1453 fallthrough;
1454 case PMD_SHIFT:
1455 if (fault_supports_stage2_huge_mapping(memslot, hva, PMD_SIZE))
1456 break;
1457 fallthrough;
1458 case CONT_PTE_SHIFT:
1459 vma_shift = PAGE_SHIFT;
1460 force_pte = true;
1461 fallthrough;
1462 case PAGE_SHIFT:
1463 break;
1464 default:
1465 WARN_ONCE(1, "Unknown vma_shift %d", vma_shift);
1466 }
1467
1468 vma_pagesize = 1UL << vma_shift;
1469 if (vma_pagesize == PMD_SIZE || vma_pagesize == PUD_SIZE)
1470 fault_ipa &= ~(vma_pagesize - 1);
1471
1472 gfn = fault_ipa >> PAGE_SHIFT;
1473 mte_allowed = kvm_vma_mte_allowed(vma);
1474
1475 /* Don't use the VMA after the unlock -- it may have vanished */
1476 vma = NULL;
1477
1478 /*
1479 * Read mmu_invalidate_seq so that KVM can detect if the results of
1480 * vma_lookup() or __gfn_to_pfn_memslot() become stale prior to
1481 * acquiring kvm->mmu_lock.
1482 *
1483 * Rely on mmap_read_unlock() for an implicit smp_rmb(), which pairs
1484 * with the smp_wmb() in kvm_mmu_invalidate_end().
1485 */
1486 mmu_seq = vcpu->kvm->mmu_invalidate_seq;
1487 mmap_read_unlock(current->mm);
1488
1489 pfn = __gfn_to_pfn_memslot(memslot, gfn, false, false, NULL,
1490 write_fault, &writable, NULL);
1491 if (pfn == KVM_PFN_ERR_HWPOISON) {
1492 kvm_send_hwpoison_signal(hva, vma_shift);
1493 return 0;
1494 }
1495 if (is_error_noslot_pfn(pfn))
1496 return -EFAULT;
1497
1498 if (kvm_is_device_pfn(pfn)) {
1499 /*
1500 * If the page was identified as device early by looking at
1501 * the VMA flags, vma_pagesize is already representing the
1502 * largest quantity we can map. If instead it was mapped
1503 * via gfn_to_pfn_prot(), vma_pagesize is set to PAGE_SIZE
1504 * and must not be upgraded.
1505 *
1506 * In both cases, we don't let transparent_hugepage_adjust()
1507 * change things at the last minute.
1508 */
1509 device = true;
1510 } else if (logging_active && !write_fault) {
1511 /*
1512 * Only actually map the page as writable if this was a write
1513 * fault.
1514 */
1515 writable = false;
1516 }
1517
1518 if (exec_fault && device)
1519 return -ENOEXEC;
1520
1521 read_lock(&kvm->mmu_lock);
1522 pgt = vcpu->arch.hw_mmu->pgt;
1523 if (mmu_invalidate_retry(kvm, mmu_seq))
1524 goto out_unlock;
1525
1526 /*
1527 * If we are not forced to use page mapping, check if we are
1528 * backed by a THP and thus use block mapping if possible.
1529 */
1530 if (vma_pagesize == PAGE_SIZE && !(force_pte || device)) {
1531 if (fault_is_perm && fault_granule > PAGE_SIZE)
1532 vma_pagesize = fault_granule;
1533 else
1534 vma_pagesize = transparent_hugepage_adjust(kvm, memslot,
1535 hva, &pfn,
1536 &fault_ipa);
1537
1538 if (vma_pagesize < 0) {
1539 ret = vma_pagesize;
1540 goto out_unlock;
1541 }
1542 }
1543
1544 if (!fault_is_perm && !device && kvm_has_mte(kvm)) {
1545 /* Check the VMM hasn't introduced a new disallowed VMA */
1546 if (mte_allowed) {
1547 sanitise_mte_tags(kvm, pfn, vma_pagesize);
1548 } else {
1549 ret = -EFAULT;
1550 goto out_unlock;
1551 }
1552 }
1553
1554 if (writable)
1555 prot |= KVM_PGTABLE_PROT_W;
1556
1557 if (exec_fault)
1558 prot |= KVM_PGTABLE_PROT_X;
1559
1560 if (device)
1561 prot |= KVM_PGTABLE_PROT_DEVICE;
1562 else if (cpus_have_final_cap(ARM64_HAS_CACHE_DIC))
1563 prot |= KVM_PGTABLE_PROT_X;
1564
1565 /*
1566 * Under the premise of getting a FSC_PERM fault, we just need to relax
1567 * permissions only if vma_pagesize equals fault_granule. Otherwise,
1568 * kvm_pgtable_stage2_map() should be called to change block size.
1569 */
1570 if (fault_is_perm && vma_pagesize == fault_granule)
1571 ret = kvm_pgtable_stage2_relax_perms(pgt, fault_ipa, prot);
1572 else
1573 ret = kvm_pgtable_stage2_map(pgt, fault_ipa, vma_pagesize,
1574 __pfn_to_phys(pfn), prot,
1575 memcache,
1576 KVM_PGTABLE_WALK_HANDLE_FAULT |
1577 KVM_PGTABLE_WALK_SHARED);
1578
1579 /* Mark the page dirty only if the fault is handled successfully */
1580 if (writable && !ret) {
1581 kvm_set_pfn_dirty(pfn);
1582 mark_page_dirty_in_slot(kvm, memslot, gfn);
1583 }
1584
1585 out_unlock:
1586 read_unlock(&kvm->mmu_lock);
1587 kvm_release_pfn_clean(pfn);
1588 return ret != -EAGAIN ? ret : 0;
1589 }
1590
1591 /* Resolve the access fault by making the page young again. */
handle_access_fault(struct kvm_vcpu * vcpu,phys_addr_t fault_ipa)1592 static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
1593 {
1594 kvm_pte_t pte;
1595 struct kvm_s2_mmu *mmu;
1596
1597 trace_kvm_access_fault(fault_ipa);
1598
1599 read_lock(&vcpu->kvm->mmu_lock);
1600 mmu = vcpu->arch.hw_mmu;
1601 pte = kvm_pgtable_stage2_mkyoung(mmu->pgt, fault_ipa);
1602 read_unlock(&vcpu->kvm->mmu_lock);
1603
1604 if (kvm_pte_valid(pte))
1605 kvm_set_pfn_accessed(kvm_pte_to_pfn(pte));
1606 }
1607
1608 /**
1609 * kvm_handle_guest_abort - handles all 2nd stage aborts
1610 * @vcpu: the VCPU pointer
1611 *
1612 * Any abort that gets to the host is almost guaranteed to be caused by a
1613 * missing second stage translation table entry, which can mean that either the
1614 * guest simply needs more memory and we must allocate an appropriate page or it
1615 * can mean that the guest tried to access I/O memory, which is emulated by user
1616 * space. The distinction is based on the IPA causing the fault and whether this
1617 * memory region has been registered as standard RAM by user space.
1618 */
kvm_handle_guest_abort(struct kvm_vcpu * vcpu)1619 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu)
1620 {
1621 unsigned long esr;
1622 phys_addr_t fault_ipa;
1623 struct kvm_memory_slot *memslot;
1624 unsigned long hva;
1625 bool is_iabt, write_fault, writable;
1626 gfn_t gfn;
1627 int ret, idx;
1628
1629 esr = kvm_vcpu_get_esr(vcpu);
1630
1631 fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1632 is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1633
1634 if (esr_fsc_is_permission_fault(esr)) {
1635 /* Beyond sanitised PARange (which is the IPA limit) */
1636 if (fault_ipa >= BIT_ULL(get_kvm_ipa_limit())) {
1637 kvm_inject_size_fault(vcpu);
1638 return 1;
1639 }
1640
1641 /* Falls between the IPA range and the PARange? */
1642 if (fault_ipa >= BIT_ULL(vcpu->arch.hw_mmu->pgt->ia_bits)) {
1643 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & GENMASK(11, 0);
1644
1645 if (is_iabt)
1646 kvm_inject_pabt(vcpu, fault_ipa);
1647 else
1648 kvm_inject_dabt(vcpu, fault_ipa);
1649 return 1;
1650 }
1651 }
1652
1653 /* Synchronous External Abort? */
1654 if (kvm_vcpu_abt_issea(vcpu)) {
1655 /*
1656 * For RAS the host kernel may handle this abort.
1657 * There is no need to pass the error into the guest.
1658 */
1659 if (kvm_handle_guest_sea(fault_ipa, kvm_vcpu_get_esr(vcpu)))
1660 kvm_inject_vabt(vcpu);
1661
1662 return 1;
1663 }
1664
1665 trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_esr(vcpu),
1666 kvm_vcpu_get_hfar(vcpu), fault_ipa);
1667
1668 /* Check the stage-2 fault is trans. fault or write fault */
1669 if (!esr_fsc_is_translation_fault(esr) &&
1670 !esr_fsc_is_permission_fault(esr) &&
1671 !esr_fsc_is_access_flag_fault(esr)) {
1672 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1673 kvm_vcpu_trap_get_class(vcpu),
1674 (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1675 (unsigned long)kvm_vcpu_get_esr(vcpu));
1676 return -EFAULT;
1677 }
1678
1679 idx = srcu_read_lock(&vcpu->kvm->srcu);
1680
1681 gfn = fault_ipa >> PAGE_SHIFT;
1682 memslot = gfn_to_memslot(vcpu->kvm, gfn);
1683 hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
1684 write_fault = kvm_is_write_fault(vcpu);
1685 if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
1686 /*
1687 * The guest has put either its instructions or its page-tables
1688 * somewhere it shouldn't have. Userspace won't be able to do
1689 * anything about this (there's no syndrome for a start), so
1690 * re-inject the abort back into the guest.
1691 */
1692 if (is_iabt) {
1693 ret = -ENOEXEC;
1694 goto out;
1695 }
1696
1697 if (kvm_vcpu_abt_iss1tw(vcpu)) {
1698 kvm_inject_dabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1699 ret = 1;
1700 goto out_unlock;
1701 }
1702
1703 /*
1704 * Check for a cache maintenance operation. Since we
1705 * ended-up here, we know it is outside of any memory
1706 * slot. But we can't find out if that is for a device,
1707 * or if the guest is just being stupid. The only thing
1708 * we know for sure is that this range cannot be cached.
1709 *
1710 * So let's assume that the guest is just being
1711 * cautious, and skip the instruction.
1712 */
1713 if (kvm_is_error_hva(hva) && kvm_vcpu_dabt_is_cm(vcpu)) {
1714 kvm_incr_pc(vcpu);
1715 ret = 1;
1716 goto out_unlock;
1717 }
1718
1719 /*
1720 * The IPA is reported as [MAX:12], so we need to
1721 * complement it with the bottom 12 bits from the
1722 * faulting VA. This is always 12 bits, irrespective
1723 * of the page size.
1724 */
1725 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1726 ret = io_mem_abort(vcpu, fault_ipa);
1727 goto out_unlock;
1728 }
1729
1730 /* Userspace should not be able to register out-of-bounds IPAs */
1731 VM_BUG_ON(fault_ipa >= kvm_phys_size(vcpu->arch.hw_mmu));
1732
1733 if (esr_fsc_is_access_flag_fault(esr)) {
1734 handle_access_fault(vcpu, fault_ipa);
1735 ret = 1;
1736 goto out_unlock;
1737 }
1738
1739 ret = user_mem_abort(vcpu, fault_ipa, memslot, hva,
1740 esr_fsc_is_permission_fault(esr));
1741 if (ret == 0)
1742 ret = 1;
1743 out:
1744 if (ret == -ENOEXEC) {
1745 kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1746 ret = 1;
1747 }
1748 out_unlock:
1749 srcu_read_unlock(&vcpu->kvm->srcu, idx);
1750 return ret;
1751 }
1752
kvm_unmap_gfn_range(struct kvm * kvm,struct kvm_gfn_range * range)1753 bool kvm_unmap_gfn_range(struct kvm *kvm, struct kvm_gfn_range *range)
1754 {
1755 if (!kvm->arch.mmu.pgt)
1756 return false;
1757
1758 __unmap_stage2_range(&kvm->arch.mmu, range->start << PAGE_SHIFT,
1759 (range->end - range->start) << PAGE_SHIFT,
1760 range->may_block);
1761
1762 return false;
1763 }
1764
kvm_set_spte_gfn(struct kvm * kvm,struct kvm_gfn_range * range)1765 bool kvm_set_spte_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
1766 {
1767 kvm_pfn_t pfn = pte_pfn(range->arg.pte);
1768
1769 if (!kvm->arch.mmu.pgt)
1770 return false;
1771
1772 WARN_ON(range->end - range->start != 1);
1773
1774 /*
1775 * If the page isn't tagged, defer to user_mem_abort() for sanitising
1776 * the MTE tags. The S2 pte should have been unmapped by
1777 * mmu_notifier_invalidate_range_end().
1778 */
1779 if (kvm_has_mte(kvm) && !page_mte_tagged(pfn_to_page(pfn)))
1780 return false;
1781
1782 /*
1783 * We've moved a page around, probably through CoW, so let's treat
1784 * it just like a translation fault and the map handler will clean
1785 * the cache to the PoC.
1786 *
1787 * The MMU notifiers will have unmapped a huge PMD before calling
1788 * ->change_pte() (which in turn calls kvm_set_spte_gfn()) and
1789 * therefore we never need to clear out a huge PMD through this
1790 * calling path and a memcache is not required.
1791 */
1792 kvm_pgtable_stage2_map(kvm->arch.mmu.pgt, range->start << PAGE_SHIFT,
1793 PAGE_SIZE, __pfn_to_phys(pfn),
1794 KVM_PGTABLE_PROT_R, NULL, 0);
1795
1796 return false;
1797 }
1798
kvm_age_gfn(struct kvm * kvm,struct kvm_gfn_range * range)1799 bool kvm_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
1800 {
1801 u64 size = (range->end - range->start) << PAGE_SHIFT;
1802
1803 if (!kvm->arch.mmu.pgt)
1804 return false;
1805
1806 return kvm_pgtable_stage2_test_clear_young(kvm->arch.mmu.pgt,
1807 range->start << PAGE_SHIFT,
1808 size, true);
1809 }
1810
kvm_test_age_gfn(struct kvm * kvm,struct kvm_gfn_range * range)1811 bool kvm_test_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
1812 {
1813 u64 size = (range->end - range->start) << PAGE_SHIFT;
1814
1815 if (!kvm->arch.mmu.pgt)
1816 return false;
1817
1818 return kvm_pgtable_stage2_test_clear_young(kvm->arch.mmu.pgt,
1819 range->start << PAGE_SHIFT,
1820 size, false);
1821 }
1822
kvm_mmu_get_httbr(void)1823 phys_addr_t kvm_mmu_get_httbr(void)
1824 {
1825 return __pa(hyp_pgtable->pgd);
1826 }
1827
kvm_get_idmap_vector(void)1828 phys_addr_t kvm_get_idmap_vector(void)
1829 {
1830 return hyp_idmap_vector;
1831 }
1832
kvm_map_idmap_text(void)1833 static int kvm_map_idmap_text(void)
1834 {
1835 unsigned long size = hyp_idmap_end - hyp_idmap_start;
1836 int err = __create_hyp_mappings(hyp_idmap_start, size, hyp_idmap_start,
1837 PAGE_HYP_EXEC);
1838 if (err)
1839 kvm_err("Failed to idmap %lx-%lx\n",
1840 hyp_idmap_start, hyp_idmap_end);
1841
1842 return err;
1843 }
1844
kvm_hyp_zalloc_page(void * arg)1845 static void *kvm_hyp_zalloc_page(void *arg)
1846 {
1847 return (void *)get_zeroed_page(GFP_KERNEL);
1848 }
1849
1850 static struct kvm_pgtable_mm_ops kvm_hyp_mm_ops = {
1851 .zalloc_page = kvm_hyp_zalloc_page,
1852 .get_page = kvm_host_get_page,
1853 .put_page = kvm_host_put_page,
1854 .phys_to_virt = kvm_host_va,
1855 .virt_to_phys = kvm_host_pa,
1856 };
1857
kvm_mmu_init(u32 * hyp_va_bits)1858 int __init kvm_mmu_init(u32 *hyp_va_bits)
1859 {
1860 int err;
1861 u32 idmap_bits;
1862 u32 kernel_bits;
1863
1864 hyp_idmap_start = __pa_symbol(__hyp_idmap_text_start);
1865 hyp_idmap_start = ALIGN_DOWN(hyp_idmap_start, PAGE_SIZE);
1866 hyp_idmap_end = __pa_symbol(__hyp_idmap_text_end);
1867 hyp_idmap_end = ALIGN(hyp_idmap_end, PAGE_SIZE);
1868 hyp_idmap_vector = __pa_symbol(__kvm_hyp_init);
1869
1870 /*
1871 * We rely on the linker script to ensure at build time that the HYP
1872 * init code does not cross a page boundary.
1873 */
1874 BUG_ON((hyp_idmap_start ^ (hyp_idmap_end - 1)) & PAGE_MASK);
1875
1876 /*
1877 * The ID map may be configured to use an extended virtual address
1878 * range. This is only the case if system RAM is out of range for the
1879 * currently configured page size and VA_BITS_MIN, in which case we will
1880 * also need the extended virtual range for the HYP ID map, or we won't
1881 * be able to enable the EL2 MMU.
1882 *
1883 * However, in some cases the ID map may be configured for fewer than
1884 * the number of VA bits used by the regular kernel stage 1. This
1885 * happens when VA_BITS=52 and the kernel image is placed in PA space
1886 * below 48 bits.
1887 *
1888 * At EL2, there is only one TTBR register, and we can't switch between
1889 * translation tables *and* update TCR_EL2.T0SZ at the same time. Bottom
1890 * line: we need to use the extended range with *both* our translation
1891 * tables.
1892 *
1893 * So use the maximum of the idmap VA bits and the regular kernel stage
1894 * 1 VA bits to assure that the hypervisor can both ID map its code page
1895 * and map any kernel memory.
1896 */
1897 idmap_bits = 64 - ((idmap_t0sz & TCR_T0SZ_MASK) >> TCR_T0SZ_OFFSET);
1898 kernel_bits = vabits_actual;
1899 *hyp_va_bits = max(idmap_bits, kernel_bits);
1900
1901 kvm_debug("Using %u-bit virtual addresses at EL2\n", *hyp_va_bits);
1902 kvm_debug("IDMAP page: %lx\n", hyp_idmap_start);
1903 kvm_debug("HYP VA range: %lx:%lx\n",
1904 kern_hyp_va(PAGE_OFFSET),
1905 kern_hyp_va((unsigned long)high_memory - 1));
1906
1907 if (hyp_idmap_start >= kern_hyp_va(PAGE_OFFSET) &&
1908 hyp_idmap_start < kern_hyp_va((unsigned long)high_memory - 1) &&
1909 hyp_idmap_start != (unsigned long)__hyp_idmap_text_start) {
1910 /*
1911 * The idmap page is intersecting with the VA space,
1912 * it is not safe to continue further.
1913 */
1914 kvm_err("IDMAP intersecting with HYP VA, unable to continue\n");
1915 err = -EINVAL;
1916 goto out;
1917 }
1918
1919 hyp_pgtable = kzalloc(sizeof(*hyp_pgtable), GFP_KERNEL);
1920 if (!hyp_pgtable) {
1921 kvm_err("Hyp mode page-table not allocated\n");
1922 err = -ENOMEM;
1923 goto out;
1924 }
1925
1926 err = kvm_pgtable_hyp_init(hyp_pgtable, *hyp_va_bits, &kvm_hyp_mm_ops);
1927 if (err)
1928 goto out_free_pgtable;
1929
1930 err = kvm_map_idmap_text();
1931 if (err)
1932 goto out_destroy_pgtable;
1933
1934 io_map_base = hyp_idmap_start;
1935 return 0;
1936
1937 out_destroy_pgtable:
1938 kvm_pgtable_hyp_destroy(hyp_pgtable);
1939 out_free_pgtable:
1940 kfree(hyp_pgtable);
1941 hyp_pgtable = NULL;
1942 out:
1943 return err;
1944 }
1945
kvm_arch_commit_memory_region(struct kvm * kvm,struct kvm_memory_slot * old,const struct kvm_memory_slot * new,enum kvm_mr_change change)1946 void kvm_arch_commit_memory_region(struct kvm *kvm,
1947 struct kvm_memory_slot *old,
1948 const struct kvm_memory_slot *new,
1949 enum kvm_mr_change change)
1950 {
1951 bool log_dirty_pages = new && new->flags & KVM_MEM_LOG_DIRTY_PAGES;
1952
1953 /*
1954 * At this point memslot has been committed and there is an
1955 * allocated dirty_bitmap[], dirty pages will be tracked while the
1956 * memory slot is write protected.
1957 */
1958 if (log_dirty_pages) {
1959
1960 if (change == KVM_MR_DELETE)
1961 return;
1962
1963 /*
1964 * Huge and normal pages are write-protected and split
1965 * on either of these two cases:
1966 *
1967 * 1. with initial-all-set: gradually with CLEAR ioctls,
1968 */
1969 if (kvm_dirty_log_manual_protect_and_init_set(kvm))
1970 return;
1971 /*
1972 * or
1973 * 2. without initial-all-set: all in one shot when
1974 * enabling dirty logging.
1975 */
1976 kvm_mmu_wp_memory_region(kvm, new->id);
1977 kvm_mmu_split_memory_region(kvm, new->id);
1978 } else {
1979 /*
1980 * Free any leftovers from the eager page splitting cache. Do
1981 * this when deleting, moving, disabling dirty logging, or
1982 * creating the memslot (a nop). Doing it for deletes makes
1983 * sure we don't leak memory, and there's no need to keep the
1984 * cache around for any of the other cases.
1985 */
1986 kvm_mmu_free_memory_cache(&kvm->arch.mmu.split_page_cache);
1987 }
1988 }
1989
kvm_arch_prepare_memory_region(struct kvm * kvm,const struct kvm_memory_slot * old,struct kvm_memory_slot * new,enum kvm_mr_change change)1990 int kvm_arch_prepare_memory_region(struct kvm *kvm,
1991 const struct kvm_memory_slot *old,
1992 struct kvm_memory_slot *new,
1993 enum kvm_mr_change change)
1994 {
1995 hva_t hva, reg_end;
1996 int ret = 0;
1997
1998 if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
1999 change != KVM_MR_FLAGS_ONLY)
2000 return 0;
2001
2002 /*
2003 * Prevent userspace from creating a memory region outside of the IPA
2004 * space addressable by the KVM guest IPA space.
2005 */
2006 if ((new->base_gfn + new->npages) > (kvm_phys_size(&kvm->arch.mmu) >> PAGE_SHIFT))
2007 return -EFAULT;
2008
2009 hva = new->userspace_addr;
2010 reg_end = hva + (new->npages << PAGE_SHIFT);
2011
2012 mmap_read_lock(current->mm);
2013 /*
2014 * A memory region could potentially cover multiple VMAs, and any holes
2015 * between them, so iterate over all of them.
2016 *
2017 * +--------------------------------------------+
2018 * +---------------+----------------+ +----------------+
2019 * | : VMA 1 | VMA 2 | | VMA 3 : |
2020 * +---------------+----------------+ +----------------+
2021 * | memory region |
2022 * +--------------------------------------------+
2023 */
2024 do {
2025 struct vm_area_struct *vma;
2026
2027 vma = find_vma_intersection(current->mm, hva, reg_end);
2028 if (!vma)
2029 break;
2030
2031 if (kvm_has_mte(kvm) && !kvm_vma_mte_allowed(vma)) {
2032 ret = -EINVAL;
2033 break;
2034 }
2035
2036 if (vma->vm_flags & VM_PFNMAP) {
2037 /* IO region dirty page logging not allowed */
2038 if (new->flags & KVM_MEM_LOG_DIRTY_PAGES) {
2039 ret = -EINVAL;
2040 break;
2041 }
2042 }
2043 hva = min(reg_end, vma->vm_end);
2044 } while (hva < reg_end);
2045
2046 mmap_read_unlock(current->mm);
2047 return ret;
2048 }
2049
kvm_arch_free_memslot(struct kvm * kvm,struct kvm_memory_slot * slot)2050 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *slot)
2051 {
2052 }
2053
kvm_arch_memslots_updated(struct kvm * kvm,u64 gen)2054 void kvm_arch_memslots_updated(struct kvm *kvm, u64 gen)
2055 {
2056 }
2057
kvm_arch_flush_shadow_all(struct kvm * kvm)2058 void kvm_arch_flush_shadow_all(struct kvm *kvm)
2059 {
2060 kvm_uninit_stage2_mmu(kvm);
2061 }
2062
kvm_arch_flush_shadow_memslot(struct kvm * kvm,struct kvm_memory_slot * slot)2063 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
2064 struct kvm_memory_slot *slot)
2065 {
2066 gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
2067 phys_addr_t size = slot->npages << PAGE_SHIFT;
2068
2069 write_lock(&kvm->mmu_lock);
2070 unmap_stage2_range(&kvm->arch.mmu, gpa, size);
2071 write_unlock(&kvm->mmu_lock);
2072 }
2073
2074 /*
2075 * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
2076 *
2077 * Main problems:
2078 * - S/W ops are local to a CPU (not broadcast)
2079 * - We have line migration behind our back (speculation)
2080 * - System caches don't support S/W at all (damn!)
2081 *
2082 * In the face of the above, the best we can do is to try and convert
2083 * S/W ops to VA ops. Because the guest is not allowed to infer the
2084 * S/W to PA mapping, it can only use S/W to nuke the whole cache,
2085 * which is a rather good thing for us.
2086 *
2087 * Also, it is only used when turning caches on/off ("The expected
2088 * usage of the cache maintenance instructions that operate by set/way
2089 * is associated with the cache maintenance instructions associated
2090 * with the powerdown and powerup of caches, if this is required by
2091 * the implementation.").
2092 *
2093 * We use the following policy:
2094 *
2095 * - If we trap a S/W operation, we enable VM trapping to detect
2096 * caches being turned on/off, and do a full clean.
2097 *
2098 * - We flush the caches on both caches being turned on and off.
2099 *
2100 * - Once the caches are enabled, we stop trapping VM ops.
2101 */
kvm_set_way_flush(struct kvm_vcpu * vcpu)2102 void kvm_set_way_flush(struct kvm_vcpu *vcpu)
2103 {
2104 unsigned long hcr = *vcpu_hcr(vcpu);
2105
2106 /*
2107 * If this is the first time we do a S/W operation
2108 * (i.e. HCR_TVM not set) flush the whole memory, and set the
2109 * VM trapping.
2110 *
2111 * Otherwise, rely on the VM trapping to wait for the MMU +
2112 * Caches to be turned off. At that point, we'll be able to
2113 * clean the caches again.
2114 */
2115 if (!(hcr & HCR_TVM)) {
2116 trace_kvm_set_way_flush(*vcpu_pc(vcpu),
2117 vcpu_has_cache_enabled(vcpu));
2118 stage2_flush_vm(vcpu->kvm);
2119 *vcpu_hcr(vcpu) = hcr | HCR_TVM;
2120 }
2121 }
2122
kvm_toggle_cache(struct kvm_vcpu * vcpu,bool was_enabled)2123 void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
2124 {
2125 bool now_enabled = vcpu_has_cache_enabled(vcpu);
2126
2127 /*
2128 * If switching the MMU+caches on, need to invalidate the caches.
2129 * If switching it off, need to clean the caches.
2130 * Clean + invalidate does the trick always.
2131 */
2132 if (now_enabled != was_enabled)
2133 stage2_flush_vm(vcpu->kvm);
2134
2135 /* Caches are now on, stop trapping VM ops (until a S/W op) */
2136 if (now_enabled)
2137 *vcpu_hcr(vcpu) &= ~HCR_TVM;
2138
2139 trace_kvm_toggle_cache(*vcpu_pc(vcpu), was_enabled, now_enabled);
2140 }
2141