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/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(lock: &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(range: chunk_size);
116	cache = &kvm->arch.mmu.split_page_cache;
117	return kvm_mmu_memory_cache_nr_free_objects(mc: 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(range: 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(mc: cache,
144	capacity: cache_capacity,
145	min: 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, size: 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, nr: 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, nr: (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, nr: -(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, private: (unsigned long)level);
245	call_rcu(head: &page->rcu_head, func: 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(page: p) == 1)
263	kvm_account_pgtable_pages(virt: addr, nr: -1);
264	put_page(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(mmu: &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(ssp: &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(ssp: &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(objp: hyp_pgtable);
395	hyp_pgtable = NULL;
396	}
397	mutex_unlock(lock: &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(lock: &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(x: 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: &node, parent: &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(node: &this->node, parent, rb_link: node);
499	rb_insert_color(&this->node, &hyp_shared_pfns);
500	ret = kvm_call_hyp_nvhe(__pkvm_host_share_hyp, pfn);
501	unlock:
502	mutex_unlock(lock: &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: &node, parent: &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(objp: this);
526	ret = kvm_call_hyp_nvhe(__pkvm_host_unshare_hyp, pfn);
527	unlock:
528	mutex_unlock(lock: &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(x: from) || is_vmalloc_or_module_addr(x: 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(kaddr: from + virt_addr - start);
611	err = __create_hyp_mappings(start: virt_addr, PAGE_SIZE, phys: phys_addr,
612	prot: 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(lock: &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, haddr: &addr);
697	if (ret)
698	return ret;
699
700	ret = __create_hyp_mappings(start: addr, size, phys: phys_addr, prot: 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(lock: &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(offset: 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(addr: *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	* Assume that @pgt is valid and unlinked from the KVM MMU to free the
909	* page-table without taking the kvm_mmu_lock and without performing any
910	* TLB invalidations.
911	*
912	* Also, the range of addresses can be large enough to cause need_resched
913	* warnings, for instance on CONFIG_PREEMPT_NONE kernels. Hence, invoke
914	* cond_resched() periodically to prevent hogging the CPU for a long time
915	* and schedule something else, if required.
916	*/
917	static void stage2_destroy_range(struct kvm_pgtable *pgt, phys_addr_t addr,
918	phys_addr_t end)
919	{
920	u64 next;
921
922	do {
923	next = stage2_range_addr_end(addr, end);
924	KVM_PGT_FN(kvm_pgtable_stage2_destroy_range)(pgt, addr,
925	next - addr);
926	if (next != end)
927	cond_resched();
928	} while (addr = next, addr != end);
929	}
930
931	static void kvm_stage2_destroy(struct kvm_pgtable *pgt)
932	{
933	unsigned int ia_bits = VTCR_EL2_IPA(pgt->mmu->vtcr);
934
935	stage2_destroy_range(pgt, addr: 0, BIT(ia_bits));
936	KVM_PGT_FN(kvm_pgtable_stage2_destroy_pgd)(pgt);
937	}
938
939	/**
940	* kvm_init_stage2_mmu - Initialise a S2 MMU structure
941	* @kvm: The pointer to the KVM structure
942	* @mmu: The pointer to the s2 MMU structure
943	* @type: The machine type of the virtual machine
944	*
945	* Allocates only the stage-2 HW PGD level table(s).
946	* Note we don't need locking here as this is only called in two cases:
947	*
948	* - when the VM is created, which can't race against anything
949	*
950	* - when secondary kvm_s2_mmu structures are initialised for NV
951	* guests, and the caller must hold kvm->lock as this is called on a
952	* per-vcpu basis.
953	*/
954	int kvm_init_stage2_mmu(struct kvm *kvm, struct kvm_s2_mmu *mmu, unsigned long type)
955	{
956	int cpu, err;
957	struct kvm_pgtable *pgt;
958
959	/*
960	* If we already have our page tables in place, and that the
961	* MMU context is the canonical one, we have a bug somewhere,
962	* as this is only supposed to ever happen once per VM.
963	*
964	* Otherwise, we're building nested page tables, and that's
965	* probably because userspace called KVM_ARM_VCPU_INIT more
966	* than once on the same vcpu. Since that's actually legal,
967	* don't kick a fuss and leave gracefully.
968	*/
969	if (mmu->pgt != NULL) {
970	if (kvm_is_nested_s2_mmu(kvm, mmu))
971	return 0;
972
973	kvm_err("kvm_arch already initialized?\n");
974	return -EINVAL;
975	}
976
977	err = kvm_init_ipa_range(mmu, type);
978	if (err)
979	return err;
980
981	pgt = kzalloc(sizeof(*pgt), GFP_KERNEL_ACCOUNT);
982	if (!pgt)
983	return -ENOMEM;
984
985	mmu->arch = &kvm->arch;
986	err = KVM_PGT_FN(kvm_pgtable_stage2_init)(pgt, mmu, &kvm_s2_mm_ops);
987	if (err)
988	goto out_free_pgtable;
989
990	mmu->pgt = pgt;
991	if (is_protected_kvm_enabled())
992	return 0;
993
994	mmu->last_vcpu_ran = alloc_percpu(typeof(*mmu->last_vcpu_ran));
995	if (!mmu->last_vcpu_ran) {
996	err = -ENOMEM;
997	goto out_destroy_pgtable;
998	}
999
1000	for_each_possible_cpu(cpu)
1001	*per_cpu_ptr(mmu->last_vcpu_ran, cpu) = -1;
1002
1003	/* The eager page splitting is disabled by default */
1004	mmu->split_page_chunk_size = KVM_ARM_EAGER_SPLIT_CHUNK_SIZE_DEFAULT;
1005	mmu->split_page_cache.gfp_zero = __GFP_ZERO;
1006
1007	mmu->pgd_phys = __pa(pgt->pgd);
1008
1009	if (kvm_is_nested_s2_mmu(kvm, mmu))
1010	kvm_init_nested_s2_mmu(mmu);
1011
1012	return 0;
1013
1014	out_destroy_pgtable:
1015	kvm_stage2_destroy(pgt);
1016	out_free_pgtable:
1017	kfree(objp: pgt);
1018	return err;
1019	}
1020
1021	void kvm_uninit_stage2_mmu(struct kvm *kvm)
1022	{
1023	kvm_free_stage2_pgd(&kvm->arch.mmu);
1024	kvm_mmu_free_memory_cache(mc: &kvm->arch.mmu.split_page_cache);
1025	}
1026
1027	static void stage2_unmap_memslot(struct kvm *kvm,
1028	struct kvm_memory_slot *memslot)
1029	{
1030	hva_t hva = memslot->userspace_addr;
1031	phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
1032	phys_addr_t size = PAGE_SIZE * memslot->npages;
1033	hva_t reg_end = hva + size;
1034
1035	/*
1036	* A memory region could potentially cover multiple VMAs, and any holes
1037	* between them, so iterate over all of them to find out if we should
1038	* unmap any of them.
1039	*
1040	* +--------------------------------------------+
1041	* +---------------+----------------+ +----------------+
1042	* | : VMA 1 | VMA 2 | | VMA 3 : |
1043	* +---------------+----------------+ +----------------+
1044	* | memory region |
1045	* +--------------------------------------------+
1046	*/
1047	do {
1048	struct vm_area_struct *vma;
1049	hva_t vm_start, vm_end;
1050
1051	vma = find_vma_intersection(current->mm, start_addr: hva, end_addr: reg_end);
1052	if (!vma)
1053	break;
1054
1055	/*
1056	* Take the intersection of this VMA with the memory region
1057	*/
1058	vm_start = max(hva, vma->vm_start);
1059	vm_end = min(reg_end, vma->vm_end);
1060
1061	if (!(vma->vm_flags & VM_PFNMAP)) {
1062	gpa_t gpa = addr + (vm_start - memslot->userspace_addr);
1063	kvm_stage2_unmap_range(mmu: &kvm->arch.mmu, start: gpa, size: vm_end - vm_start, may_block: true);
1064	}
1065	hva = vm_end;
1066	} while (hva < reg_end);
1067	}
1068
1069	/**
1070	* stage2_unmap_vm - Unmap Stage-2 RAM mappings
1071	* @kvm: The struct kvm pointer
1072	*
1073	* Go through the memregions and unmap any regular RAM
1074	* backing memory already mapped to the VM.
1075	*/
1076	void stage2_unmap_vm(struct kvm *kvm)
1077	{
1078	struct kvm_memslots *slots;
1079	struct kvm_memory_slot *memslot;
1080	int idx, bkt;
1081
1082	idx = srcu_read_lock(ssp: &kvm->srcu);
1083	mmap_read_lock(current->mm);
1084	write_lock(&kvm->mmu_lock);
1085
1086	slots = kvm_memslots(kvm);
1087	kvm_for_each_memslot(memslot, bkt, slots)
1088	stage2_unmap_memslot(kvm, memslot);
1089
1090	kvm_nested_s2_unmap(kvm, true);
1091
1092	write_unlock(&kvm->mmu_lock);
1093	mmap_read_unlock(current->mm);
1094	srcu_read_unlock(ssp: &kvm->srcu, idx);
1095	}
1096
1097	void kvm_free_stage2_pgd(struct kvm_s2_mmu *mmu)
1098	{
1099	struct kvm *kvm = kvm_s2_mmu_to_kvm(mmu);
1100	struct kvm_pgtable *pgt = NULL;
1101
1102	write_lock(&kvm->mmu_lock);
1103	pgt = mmu->pgt;
1104	if (pgt) {
1105	mmu->pgd_phys = 0;
1106	mmu->pgt = NULL;
1107	free_percpu(pdata: mmu->last_vcpu_ran);
1108	}
1109
1110	if (kvm_is_nested_s2_mmu(kvm, mmu))
1111	kvm_init_nested_s2_mmu(mmu);
1112
1113	write_unlock(&kvm->mmu_lock);
1114
1115	if (pgt) {
1116	kvm_stage2_destroy(pgt);
1117	kfree(objp: pgt);
1118	}
1119	}
1120
1121	static void hyp_mc_free_fn(void *addr, void *mc)
1122	{
1123	struct kvm_hyp_memcache *memcache = mc;
1124
1125	if (memcache->flags & HYP_MEMCACHE_ACCOUNT_STAGE2)
1126	kvm_account_pgtable_pages(virt: addr, nr: -1);
1127
1128	free_page((unsigned long)addr);
1129	}
1130
1131	static void *hyp_mc_alloc_fn(void *mc)
1132	{
1133	struct kvm_hyp_memcache *memcache = mc;
1134	void *addr;
1135
1136	addr = (void *)__get_free_page(GFP_KERNEL_ACCOUNT);
1137	if (addr && memcache->flags & HYP_MEMCACHE_ACCOUNT_STAGE2)
1138	kvm_account_pgtable_pages(virt: addr, nr: 1);
1139
1140	return addr;
1141	}
1142
1143	void free_hyp_memcache(struct kvm_hyp_memcache *mc)
1144	{
1145	if (!is_protected_kvm_enabled())
1146	return;
1147
1148	kfree(objp: mc->mapping);
1149	__free_hyp_memcache(mc, hyp_mc_free_fn, kvm_host_va, mc);
1150	}
1151
1152	int topup_hyp_memcache(struct kvm_hyp_memcache *mc, unsigned long min_pages)
1153	{
1154	if (!is_protected_kvm_enabled())
1155	return 0;
1156
1157	if (!mc->mapping) {
1158	mc->mapping = kzalloc(sizeof(struct pkvm_mapping), GFP_KERNEL_ACCOUNT);
1159	if (!mc->mapping)
1160	return -ENOMEM;
1161	}
1162
1163	return __topup_hyp_memcache(mc, min_pages, hyp_mc_alloc_fn,
1164	kvm_host_pa, mc);
1165	}
1166
1167	/**
1168	* kvm_phys_addr_ioremap - map a device range to guest IPA
1169	*
1170	* @kvm: The KVM pointer
1171	* @guest_ipa: The IPA at which to insert the mapping
1172	* @pa: The physical address of the device
1173	* @size: The size of the mapping
1174	* @writable: Whether or not to create a writable mapping
1175	*/
1176	int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
1177	phys_addr_t pa, unsigned long size, bool writable)
1178	{
1179	phys_addr_t addr;
1180	int ret = 0;
1181	struct kvm_mmu_memory_cache cache = { .gfp_zero = __GFP_ZERO };
1182	struct kvm_s2_mmu *mmu = &kvm->arch.mmu;
1183	struct kvm_pgtable *pgt = mmu->pgt;
1184	enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_DEVICE |
1185	KVM_PGTABLE_PROT_R |
1186	(writable ? KVM_PGTABLE_PROT_W : 0);
1187
1188	if (is_protected_kvm_enabled())
1189	return -EPERM;
1190
1191	size += offset_in_page(guest_ipa);
1192	guest_ipa &= PAGE_MASK;
1193
1194	for (addr = guest_ipa; addr < guest_ipa + size; addr += PAGE_SIZE) {
1195	ret = kvm_mmu_topup_memory_cache(mc: &cache,
1196	min: kvm_mmu_cache_min_pages(mmu));
1197	if (ret)
1198	break;
1199
1200	write_lock(&kvm->mmu_lock);
1201	ret = KVM_PGT_FN(kvm_pgtable_stage2_map)(pgt, addr, PAGE_SIZE,
1202	pa, prot, &cache, 0);
1203	write_unlock(&kvm->mmu_lock);
1204	if (ret)
1205	break;
1206
1207	pa += PAGE_SIZE;
1208	}
1209
1210	kvm_mmu_free_memory_cache(mc: &cache);
1211	return ret;
1212	}
1213
1214	/**
1215	* kvm_stage2_wp_range() - write protect stage2 memory region range
1216	* @mmu: The KVM stage-2 MMU pointer
1217	* @addr: Start address of range
1218	* @end: End address of range
1219	*/
1220	void kvm_stage2_wp_range(struct kvm_s2_mmu *mmu, phys_addr_t addr, phys_addr_t end)
1221	{
1222	stage2_apply_range_resched(mmu, addr, end, KVM_PGT_FN(kvm_pgtable_stage2_wrprotect));
1223	}
1224
1225	/**
1226	* kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
1227	* @kvm: The KVM pointer
1228	* @slot: The memory slot to write protect
1229	*
1230	* Called to start logging dirty pages after memory region
1231	* KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
1232	* all present PUD, PMD and PTEs are write protected in the memory region.
1233	* Afterwards read of dirty page log can be called.
1234	*
1235	* Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
1236	* serializing operations for VM memory regions.
1237	*/
1238	static void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
1239	{
1240	struct kvm_memslots *slots = kvm_memslots(kvm);
1241	struct kvm_memory_slot *memslot = id_to_memslot(slots, id: slot);
1242	phys_addr_t start, end;
1243
1244	if (WARN_ON_ONCE(!memslot))
1245	return;
1246
1247	start = memslot->base_gfn << PAGE_SHIFT;
1248	end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1249
1250	write_lock(&kvm->mmu_lock);
1251	kvm_stage2_wp_range(mmu: &kvm->arch.mmu, addr: start, end);
1252	kvm_nested_s2_wp(kvm);
1253	write_unlock(&kvm->mmu_lock);
1254	kvm_flush_remote_tlbs_memslot(kvm, memslot);
1255	}
1256
1257	/**
1258	* kvm_mmu_split_memory_region() - split the stage 2 blocks into PAGE_SIZE
1259	* pages for memory slot
1260	* @kvm: The KVM pointer
1261	* @slot: The memory slot to split
1262	*
1263	* Acquires kvm->mmu_lock. Called with kvm->slots_lock mutex acquired,
1264	* serializing operations for VM memory regions.
1265	*/
1266	static void kvm_mmu_split_memory_region(struct kvm *kvm, int slot)
1267	{
1268	struct kvm_memslots *slots;
1269	struct kvm_memory_slot *memslot;
1270	phys_addr_t start, end;
1271
1272	lockdep_assert_held(&kvm->slots_lock);
1273
1274	slots = kvm_memslots(kvm);
1275	memslot = id_to_memslot(slots, id: slot);
1276
1277	start = memslot->base_gfn << PAGE_SHIFT;
1278	end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1279
1280	write_lock(&kvm->mmu_lock);
1281	kvm_mmu_split_huge_pages(kvm, addr: start, end);
1282	write_unlock(&kvm->mmu_lock);
1283	}
1284
1285	/*
1286	* kvm_arch_mmu_enable_log_dirty_pt_masked() - enable dirty logging for selected pages.
1287	* @kvm: The KVM pointer
1288	* @slot: The memory slot associated with mask
1289	* @gfn_offset: The gfn offset in memory slot
1290	* @mask: The mask of pages at offset 'gfn_offset' in this memory
1291	* slot to enable dirty logging on
1292	*
1293	* Writes protect selected pages to enable dirty logging, and then
1294	* splits them to PAGE_SIZE. Caller must acquire kvm->mmu_lock.
1295	*/
1296	void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1297	struct kvm_memory_slot *slot,
1298	gfn_t gfn_offset, unsigned long mask)
1299	{
1300	phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
1301	phys_addr_t start = (base_gfn + __ffs(mask)) << PAGE_SHIFT;
1302	phys_addr_t end = (base_gfn + __fls(word: mask) + 1) << PAGE_SHIFT;
1303
1304	lockdep_assert_held_write(&kvm->mmu_lock);
1305
1306	kvm_stage2_wp_range(mmu: &kvm->arch.mmu, addr: start, end);
1307
1308	/*
1309	* Eager-splitting is done when manual-protect is set. We
1310	* also check for initially-all-set because we can avoid
1311	* eager-splitting if initially-all-set is false.
1312	* Initially-all-set equal false implies that huge-pages were
1313	* already split when enabling dirty logging: no need to do it
1314	* again.
1315	*/
1316	if (kvm_dirty_log_manual_protect_and_init_set(kvm))
1317	kvm_mmu_split_huge_pages(kvm, addr: start, end);
1318
1319	kvm_nested_s2_wp(kvm);
1320	}
1321
1322	static void kvm_send_hwpoison_signal(unsigned long address, short lsb)
1323	{
1324	send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb, current);
1325	}
1326
1327	static bool fault_supports_stage2_huge_mapping(struct kvm_memory_slot *memslot,
1328	unsigned long hva,
1329	unsigned long map_size)
1330	{
1331	gpa_t gpa_start;
1332	hva_t uaddr_start, uaddr_end;
1333	size_t size;
1334
1335	/* The memslot and the VMA are guaranteed to be aligned to PAGE_SIZE */
1336	if (map_size == PAGE_SIZE)
1337	return true;
1338
1339	/* pKVM only supports PMD_SIZE huge-mappings */
1340	if (is_protected_kvm_enabled() && map_size != PMD_SIZE)
1341	return false;
1342
1343	size = memslot->npages * PAGE_SIZE;
1344
1345	gpa_start = memslot->base_gfn << PAGE_SHIFT;
1346
1347	uaddr_start = memslot->userspace_addr;
1348	uaddr_end = uaddr_start + size;
1349
1350	/*
1351	* Pages belonging to memslots that don't have the same alignment
1352	* within a PMD/PUD for userspace and IPA cannot be mapped with stage-2
1353	* PMD/PUD entries, because we'll end up mapping the wrong pages.
1354	*
1355	* Consider a layout like the following:
1356	*
1357	* memslot->userspace_addr:
1358	* +-----+--------------------+--------------------+---+
1359	* |abcde|fgh Stage-1 block | Stage-1 block tv|xyz|
1360	* +-----+--------------------+--------------------+---+
1361	*
1362	* memslot->base_gfn << PAGE_SHIFT:
1363	* +---+--------------------+--------------------+-----+
1364	* |abc|def Stage-2 block | Stage-2 block |tvxyz|
1365	* +---+--------------------+--------------------+-----+
1366	*
1367	* If we create those stage-2 blocks, we'll end up with this incorrect
1368	* mapping:
1369	* d -> f
1370	* e -> g
1371	* f -> h
1372	*/
1373	if ((gpa_start & (map_size - 1)) != (uaddr_start & (map_size - 1)))
1374	return false;
1375
1376	/*
1377	* Next, let's make sure we're not trying to map anything not covered
1378	* by the memslot. This means we have to prohibit block size mappings
1379	* for the beginning and end of a non-block aligned and non-block sized
1380	* memory slot (illustrated by the head and tail parts of the
1381	* userspace view above containing pages 'abcde' and 'xyz',
1382	* respectively).
1383	*
1384	* Note that it doesn't matter if we do the check using the
1385	* userspace_addr or the base_gfn, as both are equally aligned (per
1386	* the check above) and equally sized.
1387	*/
1388	return (hva & ~(map_size - 1)) >= uaddr_start &&
1389	(hva & ~(map_size - 1)) + map_size <= uaddr_end;
1390	}
1391
1392	/*
1393	* Check if the given hva is backed by a transparent huge page (THP) and
1394	* whether it can be mapped using block mapping in stage2. If so, adjust
1395	* the stage2 PFN and IPA accordingly. Only PMD_SIZE THPs are currently
1396	* supported. This will need to be updated to support other THP sizes.
1397	*
1398	* Returns the size of the mapping.
1399	*/
1400	static long
1401	transparent_hugepage_adjust(struct kvm *kvm, struct kvm_memory_slot *memslot,
1402	unsigned long hva, kvm_pfn_t *pfnp,
1403	phys_addr_t *ipap)
1404	{
1405	kvm_pfn_t pfn = *pfnp;
1406
1407	/*
1408	* Make sure the adjustment is done only for THP pages. Also make
1409	* sure that the HVA and IPA are sufficiently aligned and that the
1410	* block map is contained within the memslot.
1411	*/
1412	if (fault_supports_stage2_huge_mapping(memslot, hva, PMD_SIZE)) {
1413	int sz = get_user_mapping_size(kvm, addr: hva);
1414
1415	if (sz < 0)
1416	return sz;
1417
1418	if (sz < PMD_SIZE)
1419	return PAGE_SIZE;
1420
1421	*ipap &= PMD_MASK;
1422	pfn &= ~(PTRS_PER_PMD - 1);
1423	*pfnp = pfn;
1424
1425	return PMD_SIZE;
1426	}
1427
1428	/* Use page mapping if we cannot use block mapping. */
1429	return PAGE_SIZE;
1430	}
1431
1432	static int get_vma_page_shift(struct vm_area_struct *vma, unsigned long hva)
1433	{
1434	unsigned long pa;
1435
1436	if (is_vm_hugetlb_page(vma) && !(vma->vm_flags & VM_PFNMAP))
1437	return huge_page_shift(h: hstate_vma(vma));
1438
1439	if (!(vma->vm_flags & VM_PFNMAP))
1440	return PAGE_SHIFT;
1441
1442	VM_BUG_ON(is_vm_hugetlb_page(vma));
1443
1444	pa = (vma->vm_pgoff << PAGE_SHIFT) + (hva - vma->vm_start);
1445
1446	#ifndef __PAGETABLE_PMD_FOLDED
1447	if ((hva & (PUD_SIZE - 1)) == (pa & (PUD_SIZE - 1)) &&
1448	ALIGN_DOWN(hva, PUD_SIZE) >= vma->vm_start &&
1449	ALIGN(hva, PUD_SIZE) <= vma->vm_end)
1450	return PUD_SHIFT;
1451	#endif
1452
1453	if ((hva & (PMD_SIZE - 1)) == (pa & (PMD_SIZE - 1)) &&
1454	ALIGN_DOWN(hva, PMD_SIZE) >= vma->vm_start &&
1455	ALIGN(hva, PMD_SIZE) <= vma->vm_end)
1456	return PMD_SHIFT;
1457
1458	return PAGE_SHIFT;
1459	}
1460
1461	/*
1462	* The page will be mapped in stage 2 as Normal Cacheable, so the VM will be
1463	* able to see the page's tags and therefore they must be initialised first. If
1464	* PG_mte_tagged is set, tags have already been initialised.
1465	*
1466	* Must be called with kvm->mmu_lock held to ensure the memory remains mapped
1467	* while the tags are zeroed.
1468	*/
1469	static void sanitise_mte_tags(struct kvm *kvm, kvm_pfn_t pfn,
1470	unsigned long size)
1471	{
1472	unsigned long i, nr_pages = size >> PAGE_SHIFT;
1473	struct page *page = pfn_to_page(pfn);
1474	struct folio *folio = page_folio(page);
1475
1476	if (!kvm_has_mte(kvm))
1477	return;
1478
1479	if (folio_test_hugetlb(folio)) {
1480	/* Hugetlb has MTE flags set on head page only */
1481	if (folio_try_hugetlb_mte_tagging(folio)) {
1482	for (i = 0; i < nr_pages; i++, page++)
1483	mte_clear_page_tags(page_address(page));
1484	folio_set_hugetlb_mte_tagged(folio);
1485	}
1486	return;
1487	}
1488
1489	for (i = 0; i < nr_pages; i++, page++) {
1490	if (try_page_mte_tagging(page)) {
1491	mte_clear_page_tags(page_address(page));
1492	set_page_mte_tagged(page);
1493	}
1494	}
1495	}
1496
1497	static bool kvm_vma_mte_allowed(struct vm_area_struct *vma)
1498	{
1499	return vma->vm_flags & VM_MTE_ALLOWED;
1500	}
1501
1502	static bool kvm_vma_is_cacheable(struct vm_area_struct *vma)
1503	{
1504	switch (FIELD_GET(PTE_ATTRINDX_MASK, pgprot_val(vma->vm_page_prot))) {
1505	case MT_NORMAL_NC:
1506	case MT_DEVICE_nGnRnE:
1507	case MT_DEVICE_nGnRE:
1508	return false;
1509	default:
1510	return true;
1511	}
1512	}
1513
1514	static int prepare_mmu_memcache(struct kvm_vcpu *vcpu, bool topup_memcache,
1515	void **memcache)
1516	{
1517	int min_pages;
1518
1519	if (!is_protected_kvm_enabled())
1520	*memcache = &vcpu->arch.mmu_page_cache;
1521	else
1522	*memcache = &vcpu->arch.pkvm_memcache;
1523
1524	if (!topup_memcache)
1525	return 0;
1526
1527	min_pages = kvm_mmu_cache_min_pages(vcpu->arch.hw_mmu);
1528
1529	if (!is_protected_kvm_enabled())
1530	return kvm_mmu_topup_memory_cache(mc: *memcache, min: min_pages);
1531
1532	return topup_hyp_memcache(mc: *memcache, min_pages);
1533	}
1534
1535	/*
1536	* Potentially reduce shadow S2 permissions to match the guest's own S2. For
1537	* exec faults, we'd only reach this point if the guest actually allowed it (see
1538	* kvm_s2_handle_perm_fault).
1539	*
1540	* Also encode the level of the original translation in the SW bits of the leaf
1541	* entry as a proxy for the span of that translation. This will be retrieved on
1542	* TLB invalidation from the guest and used to limit the invalidation scope if a
1543	* TTL hint or a range isn't provided.
1544	*/
1545	static void adjust_nested_fault_perms(struct kvm_s2_trans *nested,
1546	enum kvm_pgtable_prot *prot,
1547	bool *writable)
1548	{
1549	*writable &= kvm_s2_trans_writable(nested);
1550	if (!kvm_s2_trans_readable(nested))
1551	*prot &= ~KVM_PGTABLE_PROT_R;
1552
1553	*prot |= kvm_encode_nested_level(nested);
1554	}
1555
1556	static void adjust_nested_exec_perms(struct kvm *kvm,
1557	struct kvm_s2_trans *nested,
1558	enum kvm_pgtable_prot *prot)
1559	{
1560	if (!kvm_s2_trans_exec_el0(kvm, nested))
1561	*prot &= ~KVM_PGTABLE_PROT_UX;
1562	if (!kvm_s2_trans_exec_el1(kvm, nested))
1563	*prot &= ~KVM_PGTABLE_PROT_PX;
1564	}
1565
1566	static int gmem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1567	struct kvm_s2_trans *nested,
1568	struct kvm_memory_slot *memslot, bool is_perm)
1569	{
1570	bool write_fault, exec_fault, writable;
1571	enum kvm_pgtable_walk_flags flags = KVM_PGTABLE_WALK_SHARED;
1572	enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_R;
1573	struct kvm_pgtable *pgt = vcpu->arch.hw_mmu->pgt;
1574	unsigned long mmu_seq;
1575	struct page *page;
1576	struct kvm *kvm = vcpu->kvm;
1577	void *memcache;
1578	kvm_pfn_t pfn;
1579	gfn_t gfn;
1580	int ret;
1581
1582	ret = prepare_mmu_memcache(vcpu, topup_memcache: true, memcache: &memcache);
1583	if (ret)
1584	return ret;
1585
1586	if (nested)
1587	gfn = kvm_s2_trans_output(nested) >> PAGE_SHIFT;
1588	else
1589	gfn = fault_ipa >> PAGE_SHIFT;
1590
1591	write_fault = kvm_is_write_fault(vcpu);
1592	exec_fault = kvm_vcpu_trap_is_exec_fault(vcpu);
1593
1594	VM_WARN_ON_ONCE(write_fault && exec_fault);
1595
1596	mmu_seq = kvm->mmu_invalidate_seq;
1597	/* Pairs with the smp_wmb() in kvm_mmu_invalidate_end(). */
1598	smp_rmb();
1599
1600	ret = kvm_gmem_get_pfn(kvm, slot: memslot, gfn, pfn: &pfn, page: &page, NULL);
1601	if (ret) {
1602	kvm_prepare_memory_fault_exit(vcpu, gpa: fault_ipa, PAGE_SIZE,
1603	is_write: write_fault, is_exec: exec_fault, is_private: false);
1604	return ret;
1605	}
1606
1607	writable = !(memslot->flags & KVM_MEM_READONLY);
1608
1609	if (nested)
1610	adjust_nested_fault_perms(nested, prot: &prot, writable: &writable);
1611
1612	if (writable)
1613	prot |= KVM_PGTABLE_PROT_W;
1614
1615	if (exec_fault || cpus_have_final_cap(ARM64_HAS_CACHE_DIC))
1616	prot |= KVM_PGTABLE_PROT_X;
1617
1618	if (nested)
1619	adjust_nested_exec_perms(kvm, nested, prot: &prot);
1620
1621	kvm_fault_lock(kvm);
1622	if (mmu_invalidate_retry(kvm, mmu_seq)) {
1623	ret = -EAGAIN;
1624	goto out_unlock;
1625	}
1626
1627	ret = KVM_PGT_FN(kvm_pgtable_stage2_map)(pgt, fault_ipa, PAGE_SIZE,
1628	__pfn_to_phys(pfn), prot,
1629	memcache, flags);
1630
1631	out_unlock:
1632	kvm_release_faultin_page(kvm, page, unused: !!ret, dirty: writable);
1633	kvm_fault_unlock(kvm);
1634
1635	if (writable && !ret)
1636	mark_page_dirty_in_slot(kvm, memslot, gfn);
1637
1638	return ret != -EAGAIN ? ret : 0;
1639	}
1640
1641	static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1642	struct kvm_s2_trans *nested,
1643	struct kvm_memory_slot *memslot, unsigned long hva,
1644	bool fault_is_perm)
1645	{
1646	int ret = 0;
1647	bool topup_memcache;
1648	bool write_fault, writable;
1649	bool exec_fault, mte_allowed, is_vma_cacheable;
1650	bool s2_force_noncacheable = false, vfio_allow_any_uc = false;
1651	unsigned long mmu_seq;
1652	phys_addr_t ipa = fault_ipa;
1653	struct kvm *kvm = vcpu->kvm;
1654	struct vm_area_struct *vma;
1655	short vma_shift;
1656	void *memcache;
1657	gfn_t gfn;
1658	kvm_pfn_t pfn;
1659	bool logging_active = memslot_is_logging(memslot);
1660	bool force_pte = logging_active;
1661	long vma_pagesize, fault_granule;
1662	enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_R;
1663	struct kvm_pgtable *pgt;
1664	struct page *page;
1665	vm_flags_t vm_flags;
1666	enum kvm_pgtable_walk_flags flags = KVM_PGTABLE_WALK_SHARED;
1667
1668	if (fault_is_perm)
1669	fault_granule = kvm_vcpu_trap_get_perm_fault_granule(vcpu);
1670	write_fault = kvm_is_write_fault(vcpu);
1671	exec_fault = kvm_vcpu_trap_is_exec_fault(vcpu);
1672	VM_WARN_ON_ONCE(write_fault && exec_fault);
1673
1674	/*
1675	* Permission faults just need to update the existing leaf entry,
1676	* and so normally don't require allocations from the memcache. The
1677	* only exception to this is when dirty logging is enabled at runtime
1678	* and a write fault needs to collapse a block entry into a table.
1679	*/
1680	topup_memcache = !fault_is_perm || (logging_active && write_fault);
1681	ret = prepare_mmu_memcache(vcpu, topup_memcache, memcache: &memcache);
1682	if (ret)
1683	return ret;
1684
1685	/*
1686	* Let's check if we will get back a huge page backed by hugetlbfs, or
1687	* get block mapping for device MMIO region.
1688	*/
1689	mmap_read_lock(current->mm);
1690	vma = vma_lookup(current->mm, addr: hva);
1691	if (unlikely(!vma)) {
1692	kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
1693	mmap_read_unlock(current->mm);
1694	return -EFAULT;
1695	}
1696
1697	if (force_pte)
1698	vma_shift = PAGE_SHIFT;
1699	else
1700	vma_shift = get_vma_page_shift(vma, hva);
1701
1702	switch (vma_shift) {
1703	#ifndef __PAGETABLE_PMD_FOLDED
1704	case PUD_SHIFT:
1705	if (fault_supports_stage2_huge_mapping(memslot, hva, PUD_SIZE))
1706	break;
1707	fallthrough;
1708	#endif
1709	case CONT_PMD_SHIFT:
1710	vma_shift = PMD_SHIFT;
1711	fallthrough;
1712	case PMD_SHIFT:
1713	if (fault_supports_stage2_huge_mapping(memslot, hva, PMD_SIZE))
1714	break;
1715	fallthrough;
1716	case CONT_PTE_SHIFT:
1717	vma_shift = PAGE_SHIFT;
1718	force_pte = true;
1719	fallthrough;
1720	case PAGE_SHIFT:
1721	break;
1722	default:
1723	WARN_ONCE(1, "Unknown vma_shift %d", vma_shift);
1724	}
1725
1726	vma_pagesize = 1UL << vma_shift;
1727
1728	if (nested) {
1729	unsigned long max_map_size;
1730
1731	max_map_size = force_pte ? PAGE_SIZE : PUD_SIZE;
1732
1733	ipa = kvm_s2_trans_output(nested);
1734
1735	/*
1736	* If we're about to create a shadow stage 2 entry, then we
1737	* can only create a block mapping if the guest stage 2 page
1738	* table uses at least as big a mapping.
1739	*/
1740	max_map_size = min(kvm_s2_trans_size(nested), max_map_size);
1741
1742	/*
1743	* Be careful that if the mapping size falls between
1744	* two host sizes, take the smallest of the two.
1745	*/
1746	if (max_map_size >= PMD_SIZE && max_map_size < PUD_SIZE)
1747	max_map_size = PMD_SIZE;
1748	else if (max_map_size >= PAGE_SIZE && max_map_size < PMD_SIZE)
1749	max_map_size = PAGE_SIZE;
1750
1751	force_pte = (max_map_size == PAGE_SIZE);
1752	vma_pagesize = min_t(long, vma_pagesize, max_map_size);
1753	}
1754
1755	/*
1756	* Both the canonical IPA and fault IPA must be hugepage-aligned to
1757	* ensure we find the right PFN and lay down the mapping in the right
1758	* place.
1759	*/
1760	if (vma_pagesize == PMD_SIZE || vma_pagesize == PUD_SIZE) {
1761	fault_ipa &= ~(vma_pagesize - 1);
1762	ipa &= ~(vma_pagesize - 1);
1763	}
1764
1765	gfn = ipa >> PAGE_SHIFT;
1766	mte_allowed = kvm_vma_mte_allowed(vma);
1767
1768	vfio_allow_any_uc = vma->vm_flags & VM_ALLOW_ANY_UNCACHED;
1769
1770	vm_flags = vma->vm_flags;
1771
1772	is_vma_cacheable = kvm_vma_is_cacheable(vma);
1773
1774	/* Don't use the VMA after the unlock -- it may have vanished */
1775	vma = NULL;
1776
1777	/*
1778	* Read mmu_invalidate_seq so that KVM can detect if the results of
1779	* vma_lookup() or __kvm_faultin_pfn() become stale prior to
1780	* acquiring kvm->mmu_lock.
1781	*
1782	* Rely on mmap_read_unlock() for an implicit smp_rmb(), which pairs
1783	* with the smp_wmb() in kvm_mmu_invalidate_end().
1784	*/
1785	mmu_seq = kvm->mmu_invalidate_seq;
1786	mmap_read_unlock(current->mm);
1787
1788	pfn = __kvm_faultin_pfn(slot: memslot, gfn, foll: write_fault ? FOLL_WRITE : 0,
1789	writable: &writable, refcounted_page: &page);
1790	if (pfn == KVM_PFN_ERR_HWPOISON) {
1791	kvm_send_hwpoison_signal(address: hva, lsb: vma_shift);
1792	return 0;
1793	}
1794	if (is_error_noslot_pfn(pfn))
1795	return -EFAULT;
1796
1797	/*
1798	* Check if this is non-struct page memory PFN, and cannot support
1799	* CMOs. It could potentially be unsafe to access as cacheable.
1800	*/
1801	if (vm_flags & (VM_PFNMAP | VM_MIXEDMAP) && !pfn_is_map_memory(pfn)) {
1802	if (is_vma_cacheable) {
1803	/*
1804	* Whilst the VMA owner expects cacheable mapping to this
1805	* PFN, hardware also has to support the FWB and CACHE DIC
1806	* features.
1807	*
1808	* ARM64 KVM relies on kernel VA mapping to the PFN to
1809	* perform cache maintenance as the CMO instructions work on
1810	* virtual addresses. VM_PFNMAP region are not necessarily
1811	* mapped to a KVA and hence the presence of hardware features
1812	* S2FWB and CACHE DIC are mandatory to avoid the need for
1813	* cache maintenance.
1814	*/
1815	if (!kvm_supports_cacheable_pfnmap())
1816	ret = -EFAULT;
1817	} else {
1818	/*
1819	* If the page was identified as device early by looking at
1820	* the VMA flags, vma_pagesize is already representing the
1821	* largest quantity we can map. If instead it was mapped
1822	* via __kvm_faultin_pfn(), vma_pagesize is set to PAGE_SIZE
1823	* and must not be upgraded.
1824	*
1825	* In both cases, we don't let transparent_hugepage_adjust()
1826	* change things at the last minute.
1827	*/
1828	s2_force_noncacheable = true;
1829	}
1830	} else if (logging_active && !write_fault) {
1831	/*
1832	* Only actually map the page as writable if this was a write
1833	* fault.
1834	*/
1835	writable = false;
1836	}
1837
1838	if (exec_fault && s2_force_noncacheable)
1839	ret = -ENOEXEC;
1840
1841	if (ret) {
1842	kvm_release_page_unused(page);
1843	return ret;
1844	}
1845
1846	if (nested)
1847	adjust_nested_fault_perms(nested, prot: &prot, writable: &writable);
1848
1849	kvm_fault_lock(kvm);
1850	pgt = vcpu->arch.hw_mmu->pgt;
1851	if (mmu_invalidate_retry(kvm, mmu_seq)) {
1852	ret = -EAGAIN;
1853	goto out_unlock;
1854	}
1855
1856	/*
1857	* If we are not forced to use page mapping, check if we are
1858	* backed by a THP and thus use block mapping if possible.
1859	*/
1860	if (vma_pagesize == PAGE_SIZE && !(force_pte || s2_force_noncacheable)) {
1861	if (fault_is_perm && fault_granule > PAGE_SIZE)
1862	vma_pagesize = fault_granule;
1863	else
1864	vma_pagesize = transparent_hugepage_adjust(kvm, memslot,
1865	hva, pfnp: &pfn,
1866	ipap: &fault_ipa);
1867
1868	if (vma_pagesize < 0) {
1869	ret = vma_pagesize;
1870	goto out_unlock;
1871	}
1872	}
1873
1874	if (!fault_is_perm && !s2_force_noncacheable && kvm_has_mte(kvm)) {
1875	/* Check the VMM hasn't introduced a new disallowed VMA */
1876	if (mte_allowed) {
1877	sanitise_mte_tags(kvm, pfn, size: vma_pagesize);
1878	} else {
1879	ret = -EFAULT;
1880	goto out_unlock;
1881	}
1882	}
1883
1884	if (writable)
1885	prot |= KVM_PGTABLE_PROT_W;
1886
1887	if (exec_fault)
1888	prot |= KVM_PGTABLE_PROT_X;
1889
1890	if (s2_force_noncacheable) {
1891	if (vfio_allow_any_uc)
1892	prot |= KVM_PGTABLE_PROT_NORMAL_NC;
1893	else
1894	prot |= KVM_PGTABLE_PROT_DEVICE;
1895	} else if (cpus_have_final_cap(ARM64_HAS_CACHE_DIC)) {
1896	prot |= KVM_PGTABLE_PROT_X;
1897	}
1898
1899	if (nested)
1900	adjust_nested_exec_perms(kvm, nested, prot: &prot);
1901
1902	/*
1903	* Under the premise of getting a FSC_PERM fault, we just need to relax
1904	* permissions only if vma_pagesize equals fault_granule. Otherwise,
1905	* kvm_pgtable_stage2_map() should be called to change block size.
1906	*/
1907	if (fault_is_perm && vma_pagesize == fault_granule) {
1908	/*
1909	* Drop the SW bits in favour of those stored in the
1910	* PTE, which will be preserved.
1911	*/
1912	prot &= ~KVM_NV_GUEST_MAP_SZ;
1913	ret = KVM_PGT_FN(kvm_pgtable_stage2_relax_perms)(pgt, fault_ipa, prot, flags);
1914	} else {
1915	ret = KVM_PGT_FN(kvm_pgtable_stage2_map)(pgt, fault_ipa, vma_pagesize,
1916	__pfn_to_phys(pfn), prot,
1917	memcache, flags);
1918	}
1919
1920	out_unlock:
1921	kvm_release_faultin_page(kvm, page, unused: !!ret, dirty: writable);
1922	kvm_fault_unlock(kvm);
1923
1924	/* Mark the page dirty only if the fault is handled successfully */
1925	if (writable && !ret)
1926	mark_page_dirty_in_slot(kvm, memslot, gfn);
1927
1928	return ret != -EAGAIN ? ret : 0;
1929	}
1930
1931	/* Resolve the access fault by making the page young again. */
1932	static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
1933	{
1934	enum kvm_pgtable_walk_flags flags = KVM_PGTABLE_WALK_SHARED;
1935	struct kvm_s2_mmu *mmu;
1936
1937	trace_kvm_access_fault(ipa: fault_ipa);
1938
1939	read_lock(&vcpu->kvm->mmu_lock);
1940	mmu = vcpu->arch.hw_mmu;
1941	KVM_PGT_FN(kvm_pgtable_stage2_mkyoung)(mmu->pgt, fault_ipa, flags);
1942	read_unlock(&vcpu->kvm->mmu_lock);
1943	}
1944
1945	/*
1946	* Returns true if the SEA should be handled locally within KVM if the abort
1947	* is caused by a kernel memory allocation (e.g. stage-2 table memory).
1948	*/
1949	static bool host_owns_sea(struct kvm_vcpu *vcpu, u64 esr)
1950	{
1951	/*
1952	* Without FEAT_RAS HCR_EL2.TEA is RES0, meaning any external abort
1953	* taken from a guest EL to EL2 is due to a host-imposed access (e.g.
1954	* stage-2 PTW).
1955	*/
1956	if (!cpus_have_final_cap(ARM64_HAS_RAS_EXTN))
1957	return true;
1958
1959	/* KVM owns the VNCR when the vCPU isn't in a nested context. */
1960	if (is_hyp_ctxt(vcpu) && !kvm_vcpu_trap_is_iabt(vcpu) && (esr & ESR_ELx_VNCR))
1961	return true;
1962
1963	/*
1964	* Determining if an external abort during a table walk happened at
1965	* stage-2 is only possible with S1PTW is set. Otherwise, since KVM
1966	* sets HCR_EL2.TEA, SEAs due to a stage-1 walk (i.e. accessing the
1967	* PA of the stage-1 descriptor) can reach here and are reported
1968	* with a TTW ESR value.
1969	*/
1970	return (esr_fsc_is_sea_ttw(esr) && (esr & ESR_ELx_S1PTW));
1971	}
1972
1973	int kvm_handle_guest_sea(struct kvm_vcpu *vcpu)
1974	{
1975	struct kvm *kvm = vcpu->kvm;
1976	struct kvm_run *run = vcpu->run;
1977	u64 esr = kvm_vcpu_get_esr(vcpu);
1978	u64 esr_mask = ESR_ELx_EC_MASK |
1979	ESR_ELx_IL |
1980	ESR_ELx_FnV |
1981	ESR_ELx_EA |
1982	ESR_ELx_CM |
1983	ESR_ELx_WNR |
1984	ESR_ELx_FSC;
1985	u64 ipa;
1986
1987	/*
1988	* Give APEI the opportunity to claim the abort before handling it
1989	* within KVM. apei_claim_sea() expects to be called with IRQs enabled.
1990	*/
1991	lockdep_assert_irqs_enabled();
1992	if (apei_claim_sea(NULL) == 0)
1993	return 1;
1994
1995	if (host_owns_sea(vcpu, esr) ||
1996	!test_bit(KVM_ARCH_FLAG_EXIT_SEA, &vcpu->kvm->arch.flags))
1997	return kvm_inject_serror(vcpu);
1998
1999	/* ESR_ELx.SET is RES0 when FEAT_RAS isn't implemented. */
2000	if (kvm_has_ras(kvm))
2001	esr_mask |= ESR_ELx_SET_MASK;
2002
2003	/*
2004	* Exit to userspace, and provide faulting guest virtual and physical
2005	* addresses in case userspace wants to emulate SEA to guest by
2006	* writing to FAR_ELx and HPFAR_ELx registers.
2007	*/
2008	memset(&run->arm_sea, 0, sizeof(run->arm_sea));
2009	run->exit_reason = KVM_EXIT_ARM_SEA;
2010	run->arm_sea.esr = esr & esr_mask;
2011
2012	if (!(esr & ESR_ELx_FnV))
2013	run->arm_sea.gva = kvm_vcpu_get_hfar(vcpu);
2014
2015	ipa = kvm_vcpu_get_fault_ipa(vcpu);
2016	if (ipa != INVALID_GPA) {
2017	run->arm_sea.flags |= KVM_EXIT_ARM_SEA_FLAG_GPA_VALID;
2018	run->arm_sea.gpa = ipa;
2019	}
2020
2021	return 0;
2022	}
2023
2024	/**
2025	* kvm_handle_guest_abort - handles all 2nd stage aborts
2026	* @vcpu: the VCPU pointer
2027	*
2028	* Any abort that gets to the host is almost guaranteed to be caused by a
2029	* missing second stage translation table entry, which can mean that either the
2030	* guest simply needs more memory and we must allocate an appropriate page or it
2031	* can mean that the guest tried to access I/O memory, which is emulated by user
2032	* space. The distinction is based on the IPA causing the fault and whether this
2033	* memory region has been registered as standard RAM by user space.
2034	*/
2035	int kvm_handle_guest_abort(struct kvm_vcpu *vcpu)
2036	{
2037	struct kvm_s2_trans nested_trans, *nested = NULL;
2038	unsigned long esr;
2039	phys_addr_t fault_ipa; /* The address we faulted on */
2040	phys_addr_t ipa; /* Always the IPA in the L1 guest phys space */
2041	struct kvm_memory_slot *memslot;
2042	unsigned long hva;
2043	bool is_iabt, write_fault, writable;
2044	gfn_t gfn;
2045	int ret, idx;
2046
2047	if (kvm_vcpu_abt_issea(vcpu))
2048	return kvm_handle_guest_sea(vcpu);
2049
2050	esr = kvm_vcpu_get_esr(vcpu);
2051
2052	/*
2053	* The fault IPA should be reliable at this point as we're not dealing
2054	* with an SEA.
2055	*/
2056	ipa = fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
2057	if (KVM_BUG_ON(ipa == INVALID_GPA, vcpu->kvm))
2058	return -EFAULT;
2059
2060	is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
2061
2062	if (esr_fsc_is_translation_fault(esr)) {
2063	/* Beyond sanitised PARange (which is the IPA limit) */
2064	if (fault_ipa >= BIT_ULL(get_kvm_ipa_limit())) {
2065	kvm_inject_size_fault(vcpu);
2066	return 1;
2067	}
2068
2069	/* Falls between the IPA range and the PARange? */
2070	if (fault_ipa >= BIT_ULL(VTCR_EL2_IPA(vcpu->arch.hw_mmu->vtcr))) {
2071	fault_ipa |= kvm_vcpu_get_hfar(vcpu) & GENMASK(11, 0);
2072
2073	return kvm_inject_sea(vcpu, is_iabt, fault_ipa);
2074	}
2075	}
2076
2077	trace_kvm_guest_fault(vcpu_pc: *vcpu_pc(vcpu), hsr: kvm_vcpu_get_esr(vcpu),
2078	hxfar: kvm_vcpu_get_hfar(vcpu), ipa: fault_ipa);
2079
2080	/* Check the stage-2 fault is trans. fault or write fault */
2081	if (!esr_fsc_is_translation_fault(esr) &&
2082	!esr_fsc_is_permission_fault(esr) &&
2083	!esr_fsc_is_access_flag_fault(esr)) {
2084	kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
2085	kvm_vcpu_trap_get_class(vcpu),
2086	(unsigned long)kvm_vcpu_trap_get_fault(vcpu),
2087	(unsigned long)kvm_vcpu_get_esr(vcpu));
2088	return -EFAULT;
2089	}
2090
2091	idx = srcu_read_lock(ssp: &vcpu->kvm->srcu);
2092
2093	/*
2094	* We may have faulted on a shadow stage 2 page table if we are
2095	* running a nested guest. In this case, we have to resolve the L2
2096	* IPA to the L1 IPA first, before knowing what kind of memory should
2097	* back the L1 IPA.
2098	*
2099	* If the shadow stage 2 page table walk faults, then we simply inject
2100	* this to the guest and carry on.
2101	*
2102	* If there are no shadow S2 PTs because S2 is disabled, there is
2103	* nothing to walk and we treat it as a 1:1 before going through the
2104	* canonical translation.
2105	*/
2106	if (kvm_is_nested_s2_mmu(vcpu->kvm,vcpu->arch.hw_mmu) &&
2107	vcpu->arch.hw_mmu->nested_stage2_enabled) {
2108	u32 esr;
2109
2110	ret = kvm_walk_nested_s2(vcpu, fault_ipa, &nested_trans);
2111	if (ret == -EAGAIN) {
2112	ret = 1;
2113	goto out_unlock;
2114	}
2115
2116	if (ret) {
2117	esr = kvm_s2_trans_esr(&nested_trans);
2118	kvm_inject_s2_fault(vcpu, esr);
2119	goto out_unlock;
2120	}
2121
2122	ret = kvm_s2_handle_perm_fault(vcpu, &nested_trans);
2123	if (ret) {
2124	esr = kvm_s2_trans_esr(&nested_trans);
2125	kvm_inject_s2_fault(vcpu, esr);
2126	goto out_unlock;
2127	}
2128
2129	ipa = kvm_s2_trans_output(&nested_trans);
2130	nested = &nested_trans;
2131	}
2132
2133	gfn = ipa >> PAGE_SHIFT;
2134	memslot = gfn_to_memslot(kvm: vcpu->kvm, gfn);
2135	hva = gfn_to_hva_memslot_prot(slot: memslot, gfn, writable: &writable);
2136	write_fault = kvm_is_write_fault(vcpu);
2137	if (kvm_is_error_hva(addr: hva) || (write_fault && !writable)) {
2138	/*
2139	* The guest has put either its instructions or its page-tables
2140	* somewhere it shouldn't have. Userspace won't be able to do
2141	* anything about this (there's no syndrome for a start), so
2142	* re-inject the abort back into the guest.
2143	*/
2144	if (is_iabt) {
2145	ret = -ENOEXEC;
2146	goto out;
2147	}
2148
2149	if (kvm_vcpu_abt_iss1tw(vcpu)) {
2150	ret = kvm_inject_sea_dabt(vcpu, kvm_vcpu_get_hfar(vcpu));
2151	goto out_unlock;
2152	}
2153
2154	/*
2155	* Check for a cache maintenance operation. Since we
2156	* ended-up here, we know it is outside of any memory
2157	* slot. But we can't find out if that is for a device,
2158	* or if the guest is just being stupid. The only thing
2159	* we know for sure is that this range cannot be cached.
2160	*
2161	* So let's assume that the guest is just being
2162	* cautious, and skip the instruction.
2163	*/
2164	if (kvm_is_error_hva(addr: hva) && kvm_vcpu_dabt_is_cm(vcpu)) {
2165	kvm_incr_pc(vcpu);
2166	ret = 1;
2167	goto out_unlock;
2168	}
2169
2170	/*
2171	* The IPA is reported as [MAX:12], so we need to
2172	* complement it with the bottom 12 bits from the
2173	* faulting VA. This is always 12 bits, irrespective
2174	* of the page size.
2175	*/
2176	ipa |= kvm_vcpu_get_hfar(vcpu) & GENMASK(11, 0);
2177	ret = io_mem_abort(vcpu, ipa);
2178	goto out_unlock;
2179	}
2180
2181	/* Userspace should not be able to register out-of-bounds IPAs */
2182	VM_BUG_ON(ipa >= kvm_phys_size(vcpu->arch.hw_mmu));
2183
2184	if (esr_fsc_is_access_flag_fault(esr)) {
2185	handle_access_fault(vcpu, fault_ipa);
2186	ret = 1;
2187	goto out_unlock;
2188	}
2189
2190	VM_WARN_ON_ONCE(kvm_vcpu_trap_is_permission_fault(vcpu) &&
2191	!write_fault && !kvm_vcpu_trap_is_exec_fault(vcpu));
2192
2193	if (kvm_slot_has_gmem(slot: memslot))
2194	ret = gmem_abort(vcpu, fault_ipa, nested, memslot,
2195	is_perm: esr_fsc_is_permission_fault(esr));
2196	else
2197	ret = user_mem_abort(vcpu, fault_ipa, nested, memslot, hva,
2198	fault_is_perm: esr_fsc_is_permission_fault(esr));
2199	if (ret == 0)
2200	ret = 1;
2201	out:
2202	if (ret == -ENOEXEC)
2203	ret = kvm_inject_sea_iabt(vcpu, kvm_vcpu_get_hfar(vcpu));
2204	out_unlock:
2205	srcu_read_unlock(ssp: &vcpu->kvm->srcu, idx);
2206	return ret;
2207	}
2208
2209	bool kvm_unmap_gfn_range(struct kvm *kvm, struct kvm_gfn_range *range)
2210	{
2211	if (!kvm->arch.mmu.pgt)
2212	return false;
2213
2214	__unmap_stage2_range(mmu: &kvm->arch.mmu, start: range->start << PAGE_SHIFT,
2215	size: (range->end - range->start) << PAGE_SHIFT,
2216	may_block: range->may_block);
2217
2218	kvm_nested_s2_unmap(kvm, range->may_block);
2219	return false;
2220	}
2221
2222	bool kvm_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
2223	{
2224	u64 size = (range->end - range->start) << PAGE_SHIFT;
2225
2226	if (!kvm->arch.mmu.pgt)
2227	return false;
2228
2229	return KVM_PGT_FN(kvm_pgtable_stage2_test_clear_young)(kvm->arch.mmu.pgt,
2230	range->start << PAGE_SHIFT,
2231	size, true);
2232	/*
2233	* TODO: Handle nested_mmu structures here using the reverse mapping in
2234	* a later version of patch series.
2235	*/
2236	}
2237
2238	bool kvm_test_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
2239	{
2240	u64 size = (range->end - range->start) << PAGE_SHIFT;
2241
2242	if (!kvm->arch.mmu.pgt)
2243	return false;
2244
2245	return KVM_PGT_FN(kvm_pgtable_stage2_test_clear_young)(kvm->arch.mmu.pgt,
2246	range->start << PAGE_SHIFT,
2247	size, false);
2248	}
2249
2250	phys_addr_t kvm_mmu_get_httbr(void)
2251	{
2252	return __pa(hyp_pgtable->pgd);
2253	}
2254
2255	phys_addr_t kvm_get_idmap_vector(void)
2256	{
2257	return hyp_idmap_vector;
2258	}
2259
2260	static int kvm_map_idmap_text(void)
2261	{
2262	unsigned long size = hyp_idmap_end - hyp_idmap_start;
2263	int err = __create_hyp_mappings(hyp_idmap_start, size, hyp_idmap_start,
2264	PAGE_HYP_EXEC);
2265	if (err)
2266	kvm_err("Failed to idmap %lx-%lx\n",
2267	hyp_idmap_start, hyp_idmap_end);
2268
2269	return err;
2270	}
2271
2272	static void *kvm_hyp_zalloc_page(void *arg)
2273	{
2274	return (void *)get_zeroed_page(GFP_KERNEL);
2275	}
2276
2277	static struct kvm_pgtable_mm_ops kvm_hyp_mm_ops = {
2278	.zalloc_page = kvm_hyp_zalloc_page,
2279	.get_page = kvm_host_get_page,
2280	.put_page = kvm_host_put_page,
2281	.phys_to_virt = kvm_host_va,
2282	.virt_to_phys = kvm_host_pa,
2283	};
2284
2285	int __init kvm_mmu_init(u32 *hyp_va_bits)
2286	{
2287	int err;
2288	u32 idmap_bits;
2289	u32 kernel_bits;
2290
2291	hyp_idmap_start = __pa_symbol(__hyp_idmap_text_start);
2292	hyp_idmap_start = ALIGN_DOWN(hyp_idmap_start, PAGE_SIZE);
2293	hyp_idmap_end = __pa_symbol(__hyp_idmap_text_end);
2294	hyp_idmap_end = ALIGN(hyp_idmap_end, PAGE_SIZE);
2295	hyp_idmap_vector = __pa_symbol(__kvm_hyp_init);
2296
2297	/*
2298	* We rely on the linker script to ensure at build time that the HYP
2299	* init code does not cross a page boundary.
2300	*/
2301	BUG_ON((hyp_idmap_start ^ (hyp_idmap_end - 1)) & PAGE_MASK);
2302
2303	/*
2304	* The ID map is always configured for 48 bits of translation, which
2305	* may be fewer than the number of VA bits used by the regular kernel
2306	* stage 1, when VA_BITS=52.
2307	*
2308	* At EL2, there is only one TTBR register, and we can't switch between
2309	* translation tables *and* update TCR_EL2.T0SZ at the same time. Bottom
2310	* line: we need to use the extended range with *both* our translation
2311	* tables.
2312	*
2313	* So use the maximum of the idmap VA bits and the regular kernel stage
2314	* 1 VA bits to assure that the hypervisor can both ID map its code page
2315	* and map any kernel memory.
2316	*/
2317	idmap_bits = IDMAP_VA_BITS;
2318	kernel_bits = vabits_actual;
2319	*hyp_va_bits = max(idmap_bits, kernel_bits);
2320
2321	kvm_debug("Using %u-bit virtual addresses at EL2\n", *hyp_va_bits);
2322	kvm_debug("IDMAP page: %lx\n", hyp_idmap_start);
2323	kvm_debug("HYP VA range: %lx:%lx\n",
2324	kern_hyp_va(PAGE_OFFSET),
2325	kern_hyp_va((unsigned long)high_memory - 1));
2326
2327	if (hyp_idmap_start >= kern_hyp_va(PAGE_OFFSET) &&
2328	hyp_idmap_start < kern_hyp_va((unsigned long)high_memory - 1) &&
2329	hyp_idmap_start != (unsigned long)__hyp_idmap_text_start) {
2330	/*
2331	* The idmap page is intersecting with the VA space,
2332	* it is not safe to continue further.
2333	*/
2334	kvm_err("IDMAP intersecting with HYP VA, unable to continue\n");
2335	err = -EINVAL;
2336	goto out;
2337	}
2338
2339	hyp_pgtable = kzalloc(sizeof(*hyp_pgtable), GFP_KERNEL);
2340	if (!hyp_pgtable) {
2341	kvm_err("Hyp mode page-table not allocated\n");
2342	err = -ENOMEM;
2343	goto out;
2344	}
2345
2346	err = kvm_pgtable_hyp_init(hyp_pgtable, *hyp_va_bits, &kvm_hyp_mm_ops);
2347	if (err)
2348	goto out_free_pgtable;
2349
2350	err = kvm_map_idmap_text();
2351	if (err)
2352	goto out_destroy_pgtable;
2353
2354	io_map_base = hyp_idmap_start;
2355	__hyp_va_bits = *hyp_va_bits;
2356	return 0;
2357
2358	out_destroy_pgtable:
2359	kvm_pgtable_hyp_destroy(hyp_pgtable);
2360	out_free_pgtable:
2361	kfree(objp: hyp_pgtable);
2362	hyp_pgtable = NULL;
2363	out:
2364	return err;
2365	}
2366
2367	void kvm_arch_commit_memory_region(struct kvm *kvm,
2368	struct kvm_memory_slot *old,
2369	const struct kvm_memory_slot *new,
2370	enum kvm_mr_change change)
2371	{
2372	bool log_dirty_pages = new && new->flags & KVM_MEM_LOG_DIRTY_PAGES;
2373
2374	/*
2375	* At this point memslot has been committed and there is an
2376	* allocated dirty_bitmap[], dirty pages will be tracked while the
2377	* memory slot is write protected.
2378	*/
2379	if (log_dirty_pages) {
2380
2381	if (change == KVM_MR_DELETE)
2382	return;
2383
2384	/*
2385	* Huge and normal pages are write-protected and split
2386	* on either of these two cases:
2387	*
2388	* 1. with initial-all-set: gradually with CLEAR ioctls,
2389	*/
2390	if (kvm_dirty_log_manual_protect_and_init_set(kvm))
2391	return;
2392	/*
2393	* or
2394	* 2. without initial-all-set: all in one shot when
2395	* enabling dirty logging.
2396	*/
2397	kvm_mmu_wp_memory_region(kvm, slot: new->id);
2398	kvm_mmu_split_memory_region(kvm, slot: new->id);
2399	} else {
2400	/*
2401	* Free any leftovers from the eager page splitting cache. Do
2402	* this when deleting, moving, disabling dirty logging, or
2403	* creating the memslot (a nop). Doing it for deletes makes
2404	* sure we don't leak memory, and there's no need to keep the
2405	* cache around for any of the other cases.
2406	*/
2407	kvm_mmu_free_memory_cache(mc: &kvm->arch.mmu.split_page_cache);
2408	}
2409	}
2410
2411	int kvm_arch_prepare_memory_region(struct kvm *kvm,
2412	const struct kvm_memory_slot *old,
2413	struct kvm_memory_slot *new,
2414	enum kvm_mr_change change)
2415	{
2416	hva_t hva, reg_end;
2417	int ret = 0;
2418
2419	if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
2420	change != KVM_MR_FLAGS_ONLY)
2421	return 0;
2422
2423	/*
2424	* Prevent userspace from creating a memory region outside of the IPA
2425	* space addressable by the KVM guest IPA space.
2426	*/
2427	if ((new->base_gfn + new->npages) > (kvm_phys_size(&kvm->arch.mmu) >> PAGE_SHIFT))
2428	return -EFAULT;
2429
2430	/*
2431	* Only support guest_memfd backed memslots with mappable memory, since
2432	* there aren't any CoCo VMs that support only private memory on arm64.
2433	*/
2434	if (kvm_slot_has_gmem(slot: new) && !kvm_memslot_is_gmem_only(slot: new))
2435	return -EINVAL;
2436
2437	hva = new->userspace_addr;
2438	reg_end = hva + (new->npages << PAGE_SHIFT);
2439
2440	mmap_read_lock(current->mm);
2441	/*
2442	* A memory region could potentially cover multiple VMAs, and any holes
2443	* between them, so iterate over all of them.
2444	*
2445	* +--------------------------------------------+
2446	* +---------------+----------------+ +----------------+
2447	* | : VMA 1 | VMA 2 | | VMA 3 : |
2448	* +---------------+----------------+ +----------------+
2449	* | memory region |
2450	* +--------------------------------------------+
2451	*/
2452	do {
2453	struct vm_area_struct *vma;
2454
2455	vma = find_vma_intersection(current->mm, start_addr: hva, end_addr: reg_end);
2456	if (!vma)
2457	break;
2458
2459	if (kvm_has_mte(kvm) && !kvm_vma_mte_allowed(vma)) {
2460	ret = -EINVAL;
2461	break;
2462	}
2463
2464	if (vma->vm_flags & VM_PFNMAP) {
2465	/* IO region dirty page logging not allowed */
2466	if (new->flags & KVM_MEM_LOG_DIRTY_PAGES) {
2467	ret = -EINVAL;
2468	break;
2469	}
2470
2471	/*
2472	* Cacheable PFNMAP is allowed only if the hardware
2473	* supports it.
2474	*/
2475	if (kvm_vma_is_cacheable(vma) && !kvm_supports_cacheable_pfnmap()) {
2476	ret = -EINVAL;
2477	break;
2478	}
2479	}
2480	hva = min(reg_end, vma->vm_end);
2481	} while (hva < reg_end);
2482
2483	mmap_read_unlock(current->mm);
2484	return ret;
2485	}
2486
2487	void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *slot)
2488	{
2489	}
2490
2491	void kvm_arch_memslots_updated(struct kvm *kvm, u64 gen)
2492	{
2493	}
2494
2495	void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
2496	struct kvm_memory_slot *slot)
2497	{
2498	gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
2499	phys_addr_t size = slot->npages << PAGE_SHIFT;
2500
2501	write_lock(&kvm->mmu_lock);
2502	kvm_stage2_unmap_range(mmu: &kvm->arch.mmu, start: gpa, size, may_block: true);
2503	kvm_nested_s2_unmap(kvm, true);
2504	write_unlock(&kvm->mmu_lock);
2505	}
2506
2507	/*
2508	* See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
2509	*
2510	* Main problems:
2511	* - S/W ops are local to a CPU (not broadcast)
2512	* - We have line migration behind our back (speculation)
2513	* - System caches don't support S/W at all (damn!)
2514	*
2515	* In the face of the above, the best we can do is to try and convert
2516	* S/W ops to VA ops. Because the guest is not allowed to infer the
2517	* S/W to PA mapping, it can only use S/W to nuke the whole cache,
2518	* which is a rather good thing for us.
2519	*
2520	* Also, it is only used when turning caches on/off ("The expected
2521	* usage of the cache maintenance instructions that operate by set/way
2522	* is associated with the cache maintenance instructions associated
2523	* with the powerdown and powerup of caches, if this is required by
2524	* the implementation.").
2525	*
2526	* We use the following policy:
2527	*
2528	* - If we trap a S/W operation, we enable VM trapping to detect
2529	* caches being turned on/off, and do a full clean.
2530	*
2531	* - We flush the caches on both caches being turned on and off.
2532	*
2533	* - Once the caches are enabled, we stop trapping VM ops.
2534	*/
2535	void kvm_set_way_flush(struct kvm_vcpu *vcpu)
2536	{
2537	unsigned long hcr = *vcpu_hcr(vcpu);
2538
2539	/*
2540	* If this is the first time we do a S/W operation
2541	* (i.e. HCR_TVM not set) flush the whole memory, and set the
2542	* VM trapping.
2543	*
2544	* Otherwise, rely on the VM trapping to wait for the MMU +
2545	* Caches to be turned off. At that point, we'll be able to
2546	* clean the caches again.
2547	*/
2548	if (!(hcr & HCR_TVM)) {
2549	trace_kvm_set_way_flush(vcpu_pc: *vcpu_pc(vcpu),
2550	cache: vcpu_has_cache_enabled(vcpu));
2551	stage2_flush_vm(kvm: vcpu->kvm);
2552	*vcpu_hcr(vcpu) = hcr | HCR_TVM;
2553	}
2554	}
2555
2556	void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
2557	{
2558	bool now_enabled = vcpu_has_cache_enabled(vcpu);
2559
2560	/*
2561	* If switching the MMU+caches on, need to invalidate the caches.
2562	* If switching it off, need to clean the caches.
2563	* Clean + invalidate does the trick always.
2564	*/
2565	if (now_enabled != was_enabled)
2566	stage2_flush_vm(kvm: vcpu->kvm);
2567
2568	/* Caches are now on, stop trapping VM ops (until a S/W op) */
2569	if (now_enabled)
2570	*vcpu_hcr(vcpu) &= ~HCR_TVM;
2571
2572	trace_kvm_toggle_cache(vcpu_pc: *vcpu_pc(vcpu), was: was_enabled, now: now_enabled);
2573	}
2574